diff --git a/.gitattributes b/.gitattributes
index a6344aac8c09253b3b630fb776ae94478aa0275b..9bd8f458bcd6e73e3471aa00a704d01d79ff517b 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -33,3 +33,12 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
*.zip filter=lfs diff=lfs merge=lfs -text
*.zst filter=lfs diff=lfs merge=lfs -text
*tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/LBM.jpg filter=lfs diff=lfs merge=lfs -text
+assets/depth_normal.jpg filter=lfs diff=lfs merge=lfs -text
+assets/object_removal.jpg filter=lfs diff=lfs merge=lfs -text
+assets/relight.gif filter=lfs diff=lfs merge=lfs -text
+assets/relight.jpg filter=lfs diff=lfs merge=lfs -text
+assets/relight_2.gif filter=lfs diff=lfs merge=lfs -text
+assets/shadow_control.gif filter=lfs diff=lfs merge=lfs -text
+assets/upscaler.jpg filter=lfs diff=lfs merge=lfs -text
+frpc_linux_amd64_v0.3 filter=lfs diff=lfs merge=lfs -text
diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..2266da3fb51505227fbf25a71748023d29e85174
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,10 @@
+*.ipynb
+*.png
+*.pyc
+*.gradio
+*.safetensors
+examples/inference/ckpts/*
+examples/inference/examples/*
+envs/*
+checkpoints/*
+*.sh
\ No newline at end of file
diff --git a/LICENCE.txt b/LICENCE.txt
new file mode 100644
index 0000000000000000000000000000000000000000..12d4ca74d98cfb36269464dced75e9441b7bf645
--- /dev/null
+++ b/LICENCE.txt
@@ -0,0 +1,407 @@
+Attribution-NonCommercial 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+ Considerations for licensors: Our public licenses are
+ intended for use by those authorized to give the public
+ permission to use material in ways otherwise restricted by
+ copyright and certain other rights. Our licenses are
+ irrevocable. Licensors should read and understand the terms
+ and conditions of the license they choose before applying it.
+ Licensors should also secure all rights necessary before
+ applying our licenses so that the public can reuse the
+ material as expected. Licensors should clearly mark any
+ material not subject to the license. This includes other CC-
+ licensed material, or material used under an exception or
+ limitation to copyright. More considerations for licensors:
+ wiki.creativecommons.org/Considerations_for_licensors
+
+ Considerations for the public: By using one of our public
+ licenses, a licensor grants the public permission to use the
+ licensed material under specified terms and conditions. If
+ the licensor's permission is not necessary for any reason--for
+ example, because of any applicable exception or limitation to
+ copyright--then that use is not regulated by the license. Our
+ licenses grant only permissions under copyright and certain
+ other rights that a licensor has authority to grant. Use of
+ the licensed material may still be restricted for other
+ reasons, including because others have copyright or other
+ rights in the material. A licensor may make special requests,
+ such as asking that all changes be marked or described.
+ Although not required by our licenses, you are encouraged to
+ respect those requests where reasonable. More considerations
+ for the public:
+ wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and
+conditions.
+
+
+Section 1 -- Definitions.
+
+ a. Adapted Material means material subject to Copyright and Similar
+ Rights that is derived from or based upon the Licensed Material
+ and in which the Licensed Material is translated, altered,
+ arranged, transformed, or otherwise modified in a manner requiring
+ permission under the Copyright and Similar Rights held by the
+ Licensor. For purposes of this Public License, where the Licensed
+ Material is a musical work, performance, or sound recording,
+ Adapted Material is always produced where the Licensed Material is
+ synched in timed relation with a moving image.
+
+ b. Adapter's License means the license You apply to Your Copyright
+ and Similar Rights in Your contributions to Adapted Material in
+ accordance with the terms and conditions of this Public License.
+
+ c. Copyright and Similar Rights means copyright and/or similar rights
+ closely related to copyright including, without limitation,
+ performance, broadcast, sound recording, and Sui Generis Database
+ Rights, without regard to how the rights are labeled or
+ categorized. For purposes of this Public License, the rights
+ specified in Section 2(b)(1)-(2) are not Copyright and Similar
+ Rights.
+ d. Effective Technological Measures means those measures that, in the
+ absence of proper authority, may not be circumvented under laws
+ fulfilling obligations under Article 11 of the WIPO Copyright
+ Treaty adopted on December 20, 1996, and/or similar international
+ agreements.
+
+ e. Exceptions and Limitations means fair use, fair dealing, and/or
+ any other exception or limitation to Copyright and Similar Rights
+ that applies to Your use of the Licensed Material.
+
+ f. Licensed Material means the artistic or literary work, database,
+ or other material to which the Licensor applied this Public
+ License.
+
+ g. Licensed Rights means the rights granted to You subject to the
+ terms and conditions of this Public License, which are limited to
+ all Copyright and Similar Rights that apply to Your use of the
+ Licensed Material and that the Licensor has authority to license.
+
+ h. Licensor means the individual(s) or entity(ies) granting rights
+ under this Public License.
+
+ i. NonCommercial means not primarily intended for or directed towards
+ commercial advantage or monetary compensation. For purposes of
+ this Public License, the exchange of the Licensed Material for
+ other material subject to Copyright and Similar Rights by digital
+ file-sharing or similar means is NonCommercial provided there is
+ no payment of monetary compensation in connection with the
+ exchange.
+
+ j. Share means to provide material to the public by any means or
+ process that requires permission under the Licensed Rights, such
+ as reproduction, public display, public performance, distribution,
+ dissemination, communication, or importation, and to make material
+ available to the public including in ways that members of the
+ public may access the material from a place and at a time
+ individually chosen by them.
+
+ k. Sui Generis Database Rights means rights other than copyright
+ resulting from Directive 96/9/EC of the European Parliament and of
+ the Council of 11 March 1996 on the legal protection of databases,
+ as amended and/or succeeded, as well as other essentially
+ equivalent rights anywhere in the world.
+
+ l. You means the individual or entity exercising the Licensed Rights
+ under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+ a. License grant.
+
+ 1. Subject to the terms and conditions of this Public License,
+ the Licensor hereby grants You a worldwide, royalty-free,
+ non-sublicensable, non-exclusive, irrevocable license to
+ exercise the Licensed Rights in the Licensed Material to:
+
+ a. reproduce and Share the Licensed Material, in whole or
+ in part, for NonCommercial purposes only; and
+
+ b. produce, reproduce, and Share Adapted Material for
+ NonCommercial purposes only.
+
+ 2. Exceptions and Limitations. For the avoidance of doubt, where
+ Exceptions and Limitations apply to Your use, this Public
+ License does not apply, and You do not need to comply with
+ its terms and conditions.
+
+ 3. Term. The term of this Public License is specified in Section
+ 6(a).
+
+ 4. Media and formats; technical modifications allowed. The
+ Licensor authorizes You to exercise the Licensed Rights in
+ all media and formats whether now known or hereafter created,
+ and to make technical modifications necessary to do so. The
+ Licensor waives and/or agrees not to assert any right or
+ authority to forbid You from making technical modifications
+ necessary to exercise the Licensed Rights, including
+ technical modifications necessary to circumvent Effective
+ Technological Measures. For purposes of this Public License,
+ simply making modifications authorized by this Section 2(a)
+ (4) never produces Adapted Material.
+
+ 5. Downstream recipients.
+
+ a. Offer from the Licensor -- Licensed Material. Every
+ recipient of the Licensed Material automatically
+ receives an offer from the Licensor to exercise the
+ Licensed Rights under the terms and conditions of this
+ Public License.
+
+ b. No downstream restrictions. You may not offer or impose
+ any additional or different terms or conditions on, or
+ apply any Effective Technological Measures to, the
+ Licensed Material if doing so restricts exercise of the
+ Licensed Rights by any recipient of the Licensed
+ Material.
+
+ 6. No endorsement. Nothing in this Public License constitutes or
+ may be construed as permission to assert or imply that You
+ are, or that Your use of the Licensed Material is, connected
+ with, or sponsored, endorsed, or granted official status by,
+ the Licensor or others designated to receive attribution as
+ provided in Section 3(a)(1)(A)(i).
+
+ b. Other rights.
+
+ 1. Moral rights, such as the right of integrity, are not
+ licensed under this Public License, nor are publicity,
+ privacy, and/or other similar personality rights; however, to
+ the extent possible, the Licensor waives and/or agrees not to
+ assert any such rights held by the Licensor to the limited
+ extent necessary to allow You to exercise the Licensed
+ Rights, but not otherwise.
+
+ 2. Patent and trademark rights are not licensed under this
+ Public License.
+
+ 3. To the extent possible, the Licensor waives any right to
+ collect royalties from You for the exercise of the Licensed
+ Rights, whether directly or through a collecting society
+ under any voluntary or waivable statutory or compulsory
+ licensing scheme. In all other cases the Licensor expressly
+ reserves any right to collect such royalties, including when
+ the Licensed Material is used other than for NonCommercial
+ purposes.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+ a. Attribution.
+
+ 1. If You Share the Licensed Material (including in modified
+ form), You must:
+
+ a. retain the following if it is supplied by the Licensor
+ with the Licensed Material:
+
+ i. identification of the creator(s) of the Licensed
+ Material and any others designated to receive
+ attribution, in any reasonable manner requested by
+ the Licensor (including by pseudonym if
+ designated);
+
+ ii. a copyright notice;
+
+ iii. a notice that refers to this Public License;
+
+ iv. a notice that refers to the disclaimer of
+ warranties;
+
+ v. a URI or hyperlink to the Licensed Material to the
+ extent reasonably practicable;
+
+ b. indicate if You modified the Licensed Material and
+ retain an indication of any previous modifications; and
+
+ c. indicate the Licensed Material is licensed under this
+ Public License, and include the text of, or the URI or
+ hyperlink to, this Public License.
+
+ 2. You may satisfy the conditions in Section 3(a)(1) in any
+ reasonable manner based on the medium, means, and context in
+ which You Share the Licensed Material. For example, it may be
+ reasonable to satisfy the conditions by providing a URI or
+ hyperlink to a resource that includes the required
+ information.
+
+ 3. If requested by the Licensor, You must remove any of the
+ information required by Section 3(a)(1)(A) to the extent
+ reasonably practicable.
+
+ 4. If You Share Adapted Material You produce, the Adapter's
+ License You apply must not prevent recipients of the Adapted
+ Material from complying with this Public License.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+ a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+ to extract, reuse, reproduce, and Share all or a substantial
+ portion of the contents of the database for NonCommercial purposes
+ only;
+
+ b. if You include all or a substantial portion of the database
+ contents in a database in which You have Sui Generis Database
+ Rights, then the database in which You have Sui Generis Database
+ Rights (but not its individual contents) is Adapted Material; and
+
+ c. You must comply with the conditions in Section 3(a) if You Share
+ all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+ a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+ EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+ AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+ ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+ IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+ WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+ PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+ ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+ KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+ ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+ b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+ TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+ NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+ INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+ COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+ USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+ ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+ DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+ IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+ c. The disclaimer of warranties and limitation of liability provided
+ above shall be interpreted in a manner that, to the extent
+ possible, most closely approximates an absolute disclaimer and
+ waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+ a. This Public License applies for the term of the Copyright and
+ Similar Rights licensed here. However, if You fail to comply with
+ this Public License, then Your rights under this Public License
+ terminate automatically.
+
+ b. Where Your right to use the Licensed Material has terminated under
+ Section 6(a), it reinstates:
+
+ 1. automatically as of the date the violation is cured, provided
+ it is cured within 30 days of Your discovery of the
+ violation; or
+
+ 2. upon express reinstatement by the Licensor.
+
+ For the avoidance of doubt, this Section 6(b) does not affect any
+ right the Licensor may have to seek remedies for Your violations
+ of this Public License.
+
+ c. For the avoidance of doubt, the Licensor may also offer the
+ Licensed Material under separate terms or conditions or stop
+ distributing the Licensed Material at any time; however, doing so
+ will not terminate this Public License.
+
+ d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+ License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+ a. The Licensor shall not be bound by any additional or different
+ terms or conditions communicated by You unless expressly agreed.
+
+ b. Any arrangements, understandings, or agreements regarding the
+ Licensed Material not stated herein are separate from and
+ independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+ a. For the avoidance of doubt, this Public License does not, and
+ shall not be interpreted to, reduce, limit, restrict, or impose
+ conditions on any use of the Licensed Material that could lawfully
+ be made without permission under this Public License.
+
+ b. To the extent possible, if any provision of this Public License is
+ deemed unenforceable, it shall be automatically reformed to the
+ minimum extent necessary to make it enforceable. If the provision
+ cannot be reformed, it shall be severed from this Public License
+ without affecting the enforceability of the remaining terms and
+ conditions.
+
+ c. No term or condition of this Public License will be waived and no
+ failure to comply consented to unless expressly agreed to by the
+ Licensor.
+
+ d. Nothing in this Public License constitutes or may be interpreted
+ as a limitation upon, or waiver of, any privileges and immunities
+ that apply to the Licensor or You, including from the legal
+ processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.
\ No newline at end of file
diff --git a/README.md b/README.md
index 388f552a9072263bb9b4b00006b858cac58d378f..f3d984eb13ddb5ec8d1fce91f317bb3dbd6d03c8 100644
--- a/README.md
+++ b/README.md
@@ -1,12 +1,183 @@
---
-title: Homeoppoer.cachehuggingfacegradiofrpc
-emoji: 📉
-colorFrom: indigo
-colorTo: yellow
+title: homeoppoer.cachehuggingfacegradiofrpc
+app_file: frpc_linux_amd64_v0.3
sdk: gradio
-sdk_version: 5.35.0
-app_file: app.py
-pinned: false
+sdk_version: 5.29.0
---
+# Latent Bridge Matching (LBM)
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+This repository is the official implementation of the paper [LBM: Latent Bridge Matching for Fast Image-to-Image Translation](http://arxiv.org/abs/2503.07535).
+
+
+
+ 
+
+
+ 
+
+
+ 
+
+
+ 
+
+
+
+
+ 
+
+
+ 
+
+
+ 
+
+
+ 
+
+
+
+
+
+
+ DEMO space
+
+
+
+
+## Abstract
+In this paper, we introduce Latent Bridge Matching (LBM), a new, versatile and scalable method that relies on Bridge Matching in a latent space to achieve fast image-to-image translation. We show that the method can reach state-of-the-art results for various image-to-image tasks using only a single inference step. In addition to its efficiency, we also demonstrate the versatility of the method across different image translation tasks such as object removal, normal and depth estimation, and object relighting. We also derive a conditional framework of LBM and demonstrate its effectiveness by tackling the tasks of controllable image relighting and shadow generation.
+
+
+ 
+
+
+## License
+This code is released under the **Creative Commons BY-NC 4.0 license**.
+
+## Considered Use-cases
+We validate the method on various use-cases such as object relighting, image restoration, object removal, depth and normal maps estimation as well as controllable object relighting and shadow generation.
+
+ Image Relighting 🔦
+
+For object relighting, the method should translate the encoded source images created by pasting the foreground onto the target background image to the desired target relighted image.
+
+
+
+
+
+
+ Image Restoration 🧹
+
+In the context of image restoration, the method shall transport the distribution of the degraded images to the distribution of the clean images.
+
+
+ 
+
+
+
+ Object Removal ✂️
+ For object removal, the model is trained to find a transport map from the masked images to the images without the objects
+
+ 
+
+
+
+ Controllable Image Relighting and Shadow Generation🕹️
+
+ We also derive a conditional framework of LBM and demonstrate its effectiveness by tackling the tasks of controllable image relighting and shadow generation
+
+
+ 
+
+
+
+ Normals and Depth Maps Estimation 🗺️
+
+ Finally, we also consider common tasks such as normal and depth estimation where the model should translate an input image into a normal or depth map
+
+
+ 
+
+
+
+
+
+## Setup
+To be up and running, you need first to create a virtual env with at least python3.10 installed and activate it
+
+### With venv
+```bash
+python3.10 -m venv envs/lbm
+source envs/lbm/bin/activate
+```
+
+### With conda
+```bash
+conda create -n lbm python=3.10
+conda activate lbm
+```
+
+Then install the required dependencies and the repo in editable mode
+
+```bash
+pip install --upgrade pip
+pip install -e .
+```
+
+## Inference
+
+We provide in `examples` a simple script to perform depth and normal estimation using the proposed method.
+
+```bash
+python examples/inference/inference.py \
+--model_name [depth|normals|relighting] \
+--source_image path_to_your_image.jpg \
+--output_path output_images
+```
+
+See the trained models on the HF Hub 🤗
+- [Surface normals Checkpoint](https://huggingface.co/jasperai/LBM_normals)
+- [Depth Checkpoint](https://huggingface.co/jasperai/LBM_depth)
+- [Relighting Checkpoint](https://huggingface.co/jasperai/LBM_relighting)
+
+## Local Gradio Demo
+To run the local gradio demo, just run the following command:
+```bash
+python examples/inference/gradio_demo.py
+```
+It will download the pretrained model from the HF Hub as well as example images.
+
+## Training
+We provide in `examples\training` an example of a script to train a LBM for surface normal predictions on [`hypersim`](https://github.com/apple/ml-hypersim) see [this](https://github.com/prs-eth/Marigold/blob/main/script/dataset_preprocess/hypersim/README.md) for data processing.
+
+In `examples\trainig\configs`, you will find the configuration `yaml` associated to the training script. The only thing you need to do is to amend the `SHARDS_PATH_OR_URLS` section of the `yaml` so the model is trained on your own data.
+
+Please note that this package uses [`webdataset`](https://github.com/webdataset/webdataset) to handle the datastream and so the urls you use should be fomatted according to the [`webdataset format`](https://github.com/webdataset/webdataset?tab=readme-ov-file#the-webdataset-format). In particular, for this example, each sample in your `.tar` files needs to be composed of a `jpg` file containing the image, a `normal.png` file containing the target normals as well as a `mask.png` containing a mask indicating the valid pixels
+
+```
+sample = {
+ "jpg": source_image,
+ "normal.png": normals # target_image
+ "mask.png": mask # mask of valid pixels
+}
+```
+
+To train the model, you can use the following command:
+
+```bash
+python examples/training/train_lbm_surface.py examples/training/config/surface.yaml
+```
+
+*Note*: Make sure to update the relevant section of the `yaml` file to use your own data and log the results on your own [WandB](https://wandb.ai/site).
+
+## Citation
+If you find this work useful or use it in your research, please consider citing us
+```bibtex
+@article{chadebec2025lbm,
+ title={LBM: Latent Bridge Matching for Fast Image-to-Image Translation},
+ author={Clément Chadebec and Onur Tasar and Sanjeev Sreetharan and Benjamin Aubin},
+ year={2025},
+ journal = {arXiv preprint arXiv:2503.07535},
+}
+```
diff --git a/assets/LBM.jpg b/assets/LBM.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..06f3d0a53d8d41ef4baf47731dc1de64c6aa11e5
--- /dev/null
+++ b/assets/LBM.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f167343baabeab0fbded65d459f49e482fa3d2c3bbdc1d3fb7c957121183e04
+size 456600
diff --git a/assets/depth_normal.jpg b/assets/depth_normal.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..41290bc8f8040e1d15794b485e0f8ec9bd4f11b3
--- /dev/null
+++ b/assets/depth_normal.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92acb3a213336df332c10c67ab585eb8f6fe9dcdeabee36cf9b54324d07e00f4
+size 371683
diff --git a/assets/object_removal.jpg b/assets/object_removal.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2f94d5551e7ac6b2145fc560e4e6dbbf72fa0aad
--- /dev/null
+++ b/assets/object_removal.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a975d260a52741814260801fe1fedbc2f10d83e65d0b162eecba4cb0337c9a4c
+size 427230
diff --git a/assets/relight.gif b/assets/relight.gif
new file mode 100644
index 0000000000000000000000000000000000000000..71f3139b41bde08c3f3f0a65be80ce377d9c84c8
--- /dev/null
+++ b/assets/relight.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d3e522c726333f27616ef2c5f8d8bdc8cfc96fe266505ccefc09d2350cce4a4
+size 6268001
diff --git a/assets/relight.jpg b/assets/relight.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3e4850429d26ab970dbf0afcba4d9058e1002b29
--- /dev/null
+++ b/assets/relight.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4e1469e4c10e91929fdb0a4c99d696fe5dcaf18751cb1e4822e7acd1230bcfe
+size 1485748
diff --git a/assets/relight_2.gif b/assets/relight_2.gif
new file mode 100644
index 0000000000000000000000000000000000000000..10845ab70c36c0f426e56a7c5bc1579159614568
--- /dev/null
+++ b/assets/relight_2.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86802558bd7406c2f2289dff88782ba829f6b64f06648bbded25232789f3d7e4
+size 697917
diff --git a/assets/shadow_control.gif b/assets/shadow_control.gif
new file mode 100644
index 0000000000000000000000000000000000000000..5dcb7b2bbef9b68fc210cf070856d48e817c0dbc
--- /dev/null
+++ b/assets/shadow_control.gif
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3940de426dd87b7f11f6f75b666ddf3f49701bd92c76b654c4c8d039c3b87bc
+size 11311720
diff --git a/assets/upscaler.jpg b/assets/upscaler.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..a703ee389d45c36ce409c56f5b14dca786cc0f6e
--- /dev/null
+++ b/assets/upscaler.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d09ebb0b32567a2b751b637cb49708a9931ff22335db30c2d35528f6f0fa3f03
+size 2605156
diff --git a/examples/inference/gradio_demo.py b/examples/inference/gradio_demo.py
new file mode 100644
index 0000000000000000000000000000000000000000..905568d09970b408de841ba4e32f5a6620ea7ce5
--- /dev/null
+++ b/examples/inference/gradio_demo.py
@@ -0,0 +1,234 @@
+import glob
+import logging
+import os
+from copy import deepcopy
+
+import gradio as gr
+import numpy as np
+import PIL
+import torch
+from huggingface_hub import snapshot_download
+from PIL import Image
+from torchvision.transforms import ToPILImage, ToTensor
+from transformers import AutoModelForImageSegmentation
+from utils import extract_object, resize_and_center_crop
+
+from lbm.inference import get_model
+
+PATH = os.path.dirname(os.path.abspath(__file__))
+os.environ["GRADIO_TEMP_DIR"] = ".gradio"
+
+
+if not os.path.exists(os.path.join(PATH, "ckpts", "relighting")):
+ logging.info(f"Downloading relighting LBM model from HF hub...")
+ model = get_model(
+ f"jasperai/LBM_relighting",
+ save_dir=os.path.join(PATH, "ckpts", "relighting"),
+ torch_dtype=torch.bfloat16,
+ device="cuda",
+ )
+else:
+ model_dir = os.path.join(PATH, "ckpts", "relighting")
+ logging.info(f"Loading relighting LBM model from local...")
+ model = get_model(
+ os.path.join(PATH, "ckpts", "relighting"),
+ torch_dtype=torch.bfloat16,
+ device="cuda",
+ )
+
+ASPECT_RATIOS = {
+ str(512 / 2048): (512, 2048),
+ str(1024 / 1024): (1024, 1024),
+ str(2048 / 512): (2048, 512),
+ str(896 / 1152): (896, 1152),
+ str(1152 / 896): (1152, 896),
+ str(512 / 1920): (512, 1920),
+ str(640 / 1536): (640, 1536),
+ str(768 / 1280): (768, 1280),
+ str(1280 / 768): (1280, 768),
+ str(1536 / 640): (1536, 640),
+ str(1920 / 512): (1920, 512),
+}
+
+birefnet = AutoModelForImageSegmentation.from_pretrained(
+ "ZhengPeng7/BiRefNet", trust_remote_code=True
+).cuda()
+image_size = (1024, 1024)
+
+if not os.path.exists(os.path.join(PATH, "examples")):
+ logging.info(f"Downloading backgrounds from HF hub...")
+ _ = snapshot_download(
+ "jasperai/LBM_relighting",
+ repo_type="space",
+ allow_patterns="*.jpg",
+ local_dir=PATH,
+ )
+
+
+def evaluate(
+ fg_image: PIL.Image.Image,
+ bg_image: PIL.Image.Image,
+ num_sampling_steps: int = 1,
+):
+
+ ori_h_bg, ori_w_bg = fg_image.size
+ ar_bg = ori_h_bg / ori_w_bg
+ closest_ar_bg = min(ASPECT_RATIOS, key=lambda x: abs(float(x) - ar_bg))
+ dimensions_bg = ASPECT_RATIOS[closest_ar_bg]
+
+ _, fg_mask = extract_object(birefnet, deepcopy(fg_image))
+
+ fg_image = resize_and_center_crop(fg_image, dimensions_bg[0], dimensions_bg[1])
+ fg_mask = resize_and_center_crop(fg_mask, dimensions_bg[0], dimensions_bg[1])
+ bg_image = resize_and_center_crop(bg_image, dimensions_bg[0], dimensions_bg[1])
+
+ img_pasted = Image.composite(fg_image, bg_image, fg_mask)
+
+ img_pasted_tensor = ToTensor()(img_pasted).unsqueeze(0) * 2 - 1
+ batch = {
+ "source_image": img_pasted_tensor.cuda().to(torch.bfloat16),
+ }
+
+ z_source = model.vae.encode(batch[model.source_key])
+
+ output_image = model.sample(
+ z=z_source,
+ num_steps=num_sampling_steps,
+ conditioner_inputs=batch,
+ max_samples=1,
+ ).clamp(-1, 1)
+
+ output_image = (output_image[0].float().cpu() + 1) / 2
+ output_image = ToPILImage()(output_image)
+
+ # paste the output image on the background image
+ output_image = Image.composite(output_image, bg_image, fg_mask)
+
+ output_image.resize((ori_h_bg, ori_w_bg))
+
+ return (np.array(img_pasted), np.array(output_image))
+
+
+with gr.Blocks(title="LBM Object Relighting") as demo:
+ gr.Markdown(
+ f"""
+ # Object Relighting with Latent Bridge Matching
+ This is an interactive demo of [LBM: Latent Bridge Matching for Fast Image-to-Image Translation](https://arxiv.org/abs/2503.07535) *by Jasper Research*. This demo is based on the [LBM relighting checkpoint](https://huggingface.co/jasperai/LBM_relighting).
+ """
+ )
+ gr.Markdown(
+ """
+ If you enjoy the space, please also promote *open-source* by giving a ⭐ to the Github Repo.
+ """
+ )
+
+ with gr.Row():
+ with gr.Column():
+ with gr.Row():
+ fg_image = gr.Image(
+ type="pil",
+ label="Input Image",
+ image_mode="RGB",
+ height=360,
+ # width=360,
+ )
+ bg_image = gr.Image(
+ type="pil",
+ label="Target Background",
+ image_mode="RGB",
+ height=360,
+ # width=360,
+ )
+
+ with gr.Row():
+ submit_button = gr.Button("Relight", variant="primary")
+ with gr.Row():
+ num_inference_steps = gr.Slider(
+ minimum=1,
+ maximum=4,
+ value=1,
+ step=1,
+ label="Number of Inference Steps",
+ )
+
+ bg_gallery = gr.Gallery(
+ # height=450,
+ object_fit="contain",
+ label="Background List",
+ value=[
+ path
+ for path in glob.glob(
+ os.path.join(PATH, "examples/backgrounds/*.jpg")
+ )
+ ],
+ columns=5,
+ allow_preview=False,
+ )
+
+ with gr.Column():
+ output_slider = gr.ImageSlider(label="Composite vs LBM", type="numpy")
+ output_slider.upload(
+ fn=evaluate,
+ inputs=[fg_image, bg_image, num_inference_steps],
+ outputs=[output_slider],
+ )
+
+ submit_button.click(
+ evaluate,
+ inputs=[fg_image, bg_image, num_inference_steps],
+ outputs=[output_slider],
+ )
+
+ with gr.Row():
+ gr.Examples(
+ fn=evaluate,
+ examples=[
+ [
+ os.path.join(PATH, "examples/foregrounds/2.jpg"),
+ os.path.join(PATH, "examples/backgrounds/14.jpg"),
+ 1,
+ ],
+ [
+ os.path.join(PATH, "examples/foregrounds/10.jpg"),
+ os.path.join(PATH, "examples/backgrounds/4.jpg"),
+ 1,
+ ],
+ [
+ os.path.join(PATH, "examples/foregrounds/11.jpg"),
+ os.path.join(PATH, "examples/backgrounds/24.jpg"),
+ 1,
+ ],
+ [
+ os.path.join(PATH, "examples/foregrounds/19.jpg"),
+ os.path.join(PATH, "examples/backgrounds/3.jpg"),
+ 1,
+ ],
+ [
+ os.path.join(PATH, "examples/foregrounds/4.jpg"),
+ os.path.join(PATH, "examples/backgrounds/6.jpg"),
+ 1,
+ ],
+ [
+ os.path.join(PATH, "examples/foregrounds/14.jpg"),
+ os.path.join(PATH, "examples/backgrounds/22.jpg"),
+ 1,
+ ],
+ [
+ os.path.join(PATH, "examples/foregrounds/12.jpg"),
+ os.path.join(PATH, "examples/backgrounds/1.jpg"),
+ 1,
+ ],
+ ],
+ inputs=[fg_image, bg_image, num_inference_steps],
+ outputs=[output_slider],
+ run_on_click=True,
+ )
+
+ def bg_gallery_selected(gal, evt: gr.SelectData):
+ return gal[evt.index][0]
+
+ bg_gallery.select(bg_gallery_selected, inputs=bg_gallery, outputs=bg_image)
+
+if __name__ == "__main__":
+
+ demo.launch(share=True)
diff --git a/examples/inference/inference.py b/examples/inference/inference.py
new file mode 100644
index 0000000000000000000000000000000000000000..fa8db1c8a795a101abacbca564b8ccc209adf603
--- /dev/null
+++ b/examples/inference/inference.py
@@ -0,0 +1,59 @@
+import argparse
+import logging
+import os
+
+import torch
+from PIL import Image
+
+from lbm.inference import evaluate, get_model
+
+PATH = os.path.dirname(os.path.abspath(__file__))
+
+logging.basicConfig(level=logging.INFO)
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--source_image", type=str, required=True)
+parser.add_argument("--output_path", type=str, required=True)
+parser.add_argument("--num_inference_steps", type=int, default=1)
+parser.add_argument(
+ "--model_name",
+ type=str,
+ default="normals",
+ choices=["normals", "depth", "relighting"],
+)
+
+
+args = parser.parse_args()
+
+
+def main():
+ # download the weights from HF hub
+ if not os.path.exists(os.path.join(PATH, "ckpts", f"{args.model_name}")):
+ logging.info(f"Downloading {args.model_name} LBM model from HF hub...")
+ model = get_model(
+ f"jasperai/LBM_{args.model_name}",
+ save_dir=os.path.join(PATH, "ckpts", f"{args.model_name}"),
+ torch_dtype=torch.bfloat16,
+ device="cuda",
+ )
+
+ else:
+ model_dir = os.path.join(PATH, "ckpts", f"{args.model_name}")
+ logging.info(f"Loading {args.model_name} LBM model from local...")
+ model = get_model(model_dir, torch_dtype=torch.bfloat16, device="cuda")
+
+ source_image = Image.open(args.source_image).convert("RGB")
+
+ output_image = evaluate(model, source_image, args.num_inference_steps)
+
+ os.makedirs(args.output_path, exist_ok=True)
+
+ source_image.save(os.path.join(args.output_path, "source_image.jpg"))
+ output_image.save(os.path.join(args.output_path, "output_image.jpg"))
+
+ del model
+ torch.cuda.empty_cache()
+
+
+if __name__ == "__main__":
+ main()
diff --git a/examples/inference/utils.py b/examples/inference/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..4109eeb831678dcd5eeb075162b64a1ba25d874a
--- /dev/null
+++ b/examples/inference/utils.py
@@ -0,0 +1,41 @@
+import torch
+from PIL import Image
+from torchvision import transforms
+
+
+def extract_object(birefnet, img):
+ # Data settings
+ image_size = (1024, 1024)
+ transform_image = transforms.Compose(
+ [
+ transforms.Resize(image_size),
+ transforms.ToTensor(),
+ transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+ ]
+ )
+
+ image = img
+ input_images = transform_image(image).unsqueeze(0).cuda()
+
+ # Prediction
+ with torch.no_grad():
+ preds = birefnet(input_images)[-1].sigmoid().cpu()
+ pred = preds[0].squeeze()
+ pred_pil = transforms.ToPILImage()(pred)
+ mask = pred_pil.resize(image.size)
+ image = Image.composite(image, Image.new("RGB", image.size, (127, 127, 127)), mask)
+ return image, mask
+
+
+def resize_and_center_crop(image, target_width, target_height):
+ original_width, original_height = image.size
+ scale_factor = max(target_width / original_width, target_height / original_height)
+ resized_width = int(round(original_width * scale_factor))
+ resized_height = int(round(original_height * scale_factor))
+ resized_image = image.resize((resized_width, resized_height), Image.LANCZOS)
+ left = (resized_width - target_width) / 2
+ top = (resized_height - target_height) / 2
+ right = (resized_width + target_width) / 2
+ bottom = (resized_height + target_height) / 2
+ cropped_image = resized_image.crop((left, top, right, bottom))
+ return cropped_image
diff --git a/examples/training/config/surface.yaml b/examples/training/config/surface.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c7eaa8f75ef7e1e08bb2c3b7d198db336467c202
--- /dev/null
+++ b/examples/training/config/surface.yaml
@@ -0,0 +1,31 @@
+# wandb
+wandb_project: lbm-surface-flows
+timestep_sampling: custom_timesteps
+unet_input_channels: 4
+vae_num_channels: 4
+selected_timesteps: [250, 500, 750, 1000]
+prob: [0.25, 0.25, 0.25, 0.25]
+pixel_loss_type: lpips # l1 l2
+pixel_loss_weight: 10.0
+latent_loss_type: l2 # l1 l2
+latent_loss_weight: 1.0
+bridge_noise_sigma: 0.005
+conditioning_images_keys: []
+conditioning_masks_keys: []
+
+# SHARDS_PATH_OR_URLS
+train_shards:
+ - pipe:cat PATH_TO_TRAIN_TARS
+
+validation_shards:
+ - pipe:cat PATH_TO_VAL_TARS
+
+batch_size: 4
+learning_rate: 4e-5
+optimizer: AdamW
+num_steps: [1, 4]
+log_interval: 500
+resume_from_checkpoint: true
+max_epochs: 50
+save_interval: 5000
+save_ckpt_path: ./checkpoints
diff --git a/examples/training/train_lbm_surface.py b/examples/training/train_lbm_surface.py
new file mode 100644
index 0000000000000000000000000000000000000000..e4e90d4dd7ff1396aa155d291ba141d414cc1dbc
--- /dev/null
+++ b/examples/training/train_lbm_surface.py
@@ -0,0 +1,546 @@
+import datetime
+import logging
+import os
+import random
+import re
+import shutil
+from typing import List, Optional
+
+import braceexpand
+import fire
+import torch
+import yaml
+from diffusers import FlowMatchEulerDiscreteScheduler, StableDiffusionXLPipeline
+from diffusers.models import UNet2DConditionModel
+from diffusers.models.attention import BasicTransformerBlock
+from diffusers.models.resnet import ResnetBlock2D
+from pytorch_lightning import Trainer, loggers
+from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint
+from pytorch_lightning.strategies import FSDPStrategy
+from torch.distributed.fsdp.wrap import ModuleWrapPolicy
+from torchvision.transforms import InterpolationMode
+
+from lbm.data.datasets import DataModule, DataModuleConfig
+from lbm.data.filters import KeyFilter, KeyFilterConfig
+from lbm.data.mappers import (
+ KeyRenameMapper,
+ KeyRenameMapperConfig,
+ MapperWrapper,
+ RescaleMapper,
+ RescaleMapperConfig,
+ TorchvisionMapper,
+ TorchvisionMapperConfig,
+)
+from lbm.models.embedders import (
+ ConditionerWrapper,
+ LatentsConcatEmbedder,
+ LatentsConcatEmbedderConfig,
+)
+from lbm.models.lbm import LBMConfig, LBMModel
+from lbm.models.unets import DiffusersUNet2DCondWrapper
+from lbm.models.vae import AutoencoderKLDiffusers, AutoencoderKLDiffusersConfig
+from lbm.trainer import TrainingConfig, TrainingPipeline
+from lbm.trainer.loggers import WandbSampleLogger
+from lbm.trainer.utils import StateDictAdapter
+
+
+def get_model(
+ backbone_signature: str = "stabilityai/stable-diffusion-xl-base-1.0",
+ vae_num_channels: int = 4,
+ unet_input_channels: int = 4,
+ timestep_sampling: str = "log_normal",
+ selected_timesteps: Optional[List[float]] = None,
+ prob: Optional[List[float]] = None,
+ conditioning_images_keys: Optional[List[str]] = [],
+ conditioning_masks_keys: Optional[List[str]] = [],
+ source_key: str = "source_image",
+ target_key: str = "source_image_paste",
+ mask_key: str = "mask",
+ bridge_noise_sigma: float = 0.0,
+ logit_mean: float = 0.0,
+ logit_std: float = 1.0,
+ pixel_loss_type: str = "lpips",
+ latent_loss_type: str = "l2",
+ latent_loss_weight: float = 1.0,
+ pixel_loss_weight: float = 0.0,
+):
+
+ conditioners = []
+
+ # Load pretrained model as base
+ pipe = StableDiffusionXLPipeline.from_pretrained(
+ backbone_signature,
+ torch_dtype=torch.bfloat16,
+ )
+
+ ### MMMDiT ###
+ # Get Architecture
+ denoiser = DiffusersUNet2DCondWrapper(
+ in_channels=unet_input_channels, # Add downsampled_image
+ out_channels=vae_num_channels,
+ center_input_sample=False,
+ flip_sin_to_cos=True,
+ freq_shift=0,
+ down_block_types=[
+ "DownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ ],
+ mid_block_type="UNetMidBlock2DCrossAttn",
+ up_block_types=["CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "UpBlock2D"],
+ only_cross_attention=False,
+ block_out_channels=[320, 640, 1280],
+ layers_per_block=2,
+ downsample_padding=1,
+ mid_block_scale_factor=1,
+ dropout=0.0,
+ act_fn="silu",
+ norm_num_groups=32,
+ norm_eps=1e-05,
+ cross_attention_dim=[320, 640, 1280],
+ transformer_layers_per_block=[1, 2, 10],
+ reverse_transformer_layers_per_block=None,
+ encoder_hid_dim=None,
+ encoder_hid_dim_type=None,
+ attention_head_dim=[5, 10, 20],
+ num_attention_heads=None,
+ dual_cross_attention=False,
+ use_linear_projection=True,
+ class_embed_type=None,
+ addition_embed_type=None,
+ addition_time_embed_dim=None,
+ num_class_embeds=None,
+ upcast_attention=None,
+ resnet_time_scale_shift="default",
+ resnet_skip_time_act=False,
+ resnet_out_scale_factor=1.0,
+ time_embedding_type="positional",
+ time_embedding_dim=None,
+ time_embedding_act_fn=None,
+ timestep_post_act=None,
+ time_cond_proj_dim=None,
+ conv_in_kernel=3,
+ conv_out_kernel=3,
+ projection_class_embeddings_input_dim=None,
+ attention_type="default",
+ class_embeddings_concat=False,
+ mid_block_only_cross_attention=None,
+ cross_attention_norm=None,
+ addition_embed_type_num_heads=64,
+ ).to(torch.bfloat16)
+
+ state_dict = pipe.unet.state_dict()
+
+ del state_dict["add_embedding.linear_1.weight"]
+ del state_dict["add_embedding.linear_1.bias"]
+ del state_dict["add_embedding.linear_2.weight"]
+ del state_dict["add_embedding.linear_2.bias"]
+
+ # Adapt the shapes
+ state_dict_adapter = StateDictAdapter()
+ state_dict = state_dict_adapter(
+ model_state_dict=denoiser.state_dict(),
+ checkpoint_state_dict=state_dict,
+ regex_keys=[
+ r"class_embedding.linear_\d+.(weight|bias)",
+ r"conv_in.weight",
+ r"(down_blocks|up_blocks)\.\d+\.attentions\.\d+\.transformer_blocks\.\d+\.attn\d+\.(to_k|to_v)\.weight",
+ r"mid_block\.attentions\.\d+\.transformer_blocks\.\d+\.attn\d+\.(to_k|to_v)\.weight",
+ ],
+ strategy="zeros",
+ )
+
+ denoiser.load_state_dict(state_dict, strict=True)
+
+ del pipe
+
+ if conditioning_images_keys != [] or conditioning_masks_keys != []:
+
+ latents_concat_embedder_config = LatentsConcatEmbedderConfig(
+ image_keys=conditioning_images_keys,
+ mask_keys=conditioning_masks_keys,
+ )
+ latent_concat_embedder = LatentsConcatEmbedder(latents_concat_embedder_config)
+ latent_concat_embedder.freeze()
+ conditioners.append(latent_concat_embedder)
+
+ # Wrap conditioners and set to device
+ conditioner = ConditionerWrapper(
+ conditioners=conditioners,
+ )
+
+ ## VAE ##
+ # Get VAE model
+ vae_config = AutoencoderKLDiffusersConfig(
+ version=backbone_signature,
+ subfolder="vae",
+ tiling_size=(128, 128),
+ )
+ vae = AutoencoderKLDiffusers(vae_config)
+ vae.freeze()
+ vae.to(torch.bfloat16)
+
+ # LBM Config
+ config = LBMConfig(
+ ucg_keys=None,
+ source_key=source_key,
+ target_key=target_key,
+ mask_key=mask_key,
+ latent_loss_weight=latent_loss_weight,
+ latent_loss_type=latent_loss_type,
+ pixel_loss_type=pixel_loss_type,
+ pixel_loss_weight=pixel_loss_weight,
+ timestep_sampling=timestep_sampling,
+ logit_mean=logit_mean,
+ logit_std=logit_std,
+ selected_timesteps=selected_timesteps,
+ prob=prob,
+ bridge_noise_sigma=bridge_noise_sigma,
+ )
+
+ training_noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
+ backbone_signature,
+ subfolder="scheduler",
+ )
+ sampling_noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
+ backbone_signature,
+ subfolder="scheduler",
+ )
+
+ # LBM Model
+ model = LBMModel(
+ config,
+ denoiser=denoiser,
+ training_noise_scheduler=training_noise_scheduler,
+ sampling_noise_scheduler=sampling_noise_scheduler,
+ vae=vae,
+ conditioner=conditioner,
+ ).to(torch.bfloat16)
+
+ return model
+
+
+def get_filter_mappers():
+ filters_mappers = [
+ KeyFilter(KeyFilterConfig(keys=["jpg", "normal_aligned.png", "mask.png"])),
+ MapperWrapper(
+ [
+ KeyRenameMapper(
+ KeyRenameMapperConfig(
+ key_map={
+ "jpg": "image",
+ "normal_aligned.png": "normal",
+ "mask.png": "mask",
+ }
+ )
+ ),
+ TorchvisionMapper(
+ TorchvisionMapperConfig(
+ key="image",
+ transforms=["ToTensor", "Resize"],
+ transforms_kwargs=[
+ {},
+ {
+ "size": (480, 640),
+ "interpolation": InterpolationMode.NEAREST_EXACT,
+ },
+ ],
+ )
+ ),
+ TorchvisionMapper(
+ TorchvisionMapperConfig(
+ key="normal",
+ transforms=["ToTensor", "Resize"],
+ transforms_kwargs=[
+ {},
+ {
+ "size": (480, 640),
+ "interpolation": InterpolationMode.NEAREST_EXACT,
+ },
+ ],
+ )
+ ),
+ TorchvisionMapper(
+ TorchvisionMapperConfig(
+ key="mask",
+ transforms=["ToTensor", "Resize", "Normalize"],
+ transforms_kwargs=[
+ {},
+ {
+ "size": (480, 640),
+ "interpolation": InterpolationMode.NEAREST_EXACT,
+ },
+ {"mean": 0.0, "std": 1.0},
+ ],
+ )
+ ),
+ RescaleMapper(RescaleMapperConfig(key="image")),
+ RescaleMapper(RescaleMapperConfig(key="normal")),
+ ],
+ ),
+ ]
+
+ return filters_mappers
+
+
+def get_data_module(
+ train_shards: List[str],
+ validation_shards: List[str],
+ batch_size: int,
+):
+
+ # TRAIN
+ train_filters_mappers = get_filter_mappers()
+
+ # unbrace urls
+ train_shards_path_or_urls_unbraced = []
+ for train_shards_path_or_url in train_shards:
+ train_shards_path_or_urls_unbraced.extend(
+ braceexpand.braceexpand(train_shards_path_or_url)
+ )
+
+ # shuffle shards
+ random.shuffle(train_shards_path_or_urls_unbraced)
+
+ # data config
+ data_config = DataModuleConfig(
+ shards_path_or_urls=train_shards_path_or_urls_unbraced,
+ decoder="pil",
+ shuffle_before_split_by_node_buffer_size=20,
+ shuffle_before_split_by_workers_buffer_size=20,
+ shuffle_before_filter_mappers_buffer_size=20,
+ shuffle_after_filter_mappers_buffer_size=20,
+ per_worker_batch_size=batch_size,
+ num_workers=min(10, len(train_shards_path_or_urls_unbraced)),
+ )
+
+ train_data_config = data_config
+
+ # VALIDATION
+ validation_filters_mappers = get_filter_mappers()
+
+ # unbrace urls
+ validation_shards_path_or_urls_unbraced = []
+ for validation_shards_path_or_url in validation_shards:
+ validation_shards_path_or_urls_unbraced.extend(
+ braceexpand.braceexpand(validation_shards_path_or_url)
+ )
+
+ data_config = DataModuleConfig(
+ shards_path_or_urls=validation_shards_path_or_urls_unbraced,
+ decoder="pil",
+ shuffle_before_split_by_node_buffer_size=10,
+ shuffle_before_split_by_workers_buffer_size=10,
+ shuffle_before_filter_mappers_buffer_size=10,
+ shuffle_after_filter_mappers_buffer_size=10,
+ per_worker_batch_size=batch_size,
+ num_workers=min(10, len(train_shards_path_or_urls_unbraced)),
+ )
+
+ validation_data_config = data_config
+
+ # data module
+ data_module = DataModule(
+ train_config=train_data_config,
+ train_filters_mappers=train_filters_mappers,
+ eval_config=validation_data_config,
+ eval_filters_mappers=validation_filters_mappers,
+ )
+
+ return data_module
+
+
+def main(
+ train_shards: List[str] = ["pipe:cat path/to/train/shards"],
+ validation_shards: List[str] = ["pipe:cat path/to/validation/shards"],
+ backbone_signature: str = "stabilityai/stable-diffusion-xl-base-1.0",
+ vae_num_channels: int = 4,
+ unet_input_channels: int = 4,
+ source_key: str = "image",
+ target_key: str = "normal",
+ mask_key: str = "mask",
+ wandb_project: str = "lbm-surface",
+ batch_size: int = 8,
+ num_steps: List[int] = [1, 2, 4],
+ learning_rate: float = 5e-5,
+ learning_rate_scheduler: str = None,
+ learning_rate_scheduler_kwargs: dict = {},
+ optimizer: str = "AdamW",
+ optimizer_kwargs: dict = {},
+ timestep_sampling: str = "uniform",
+ logit_mean: float = 0.0,
+ logit_std: float = 1.0,
+ pixel_loss_type: str = "lpips",
+ latent_loss_type: str = "l2",
+ latent_loss_weight: float = 1.0,
+ pixel_loss_weight: float = 0.0,
+ selected_timesteps: List[float] = None,
+ prob: List[float] = None,
+ conditioning_images_keys: Optional[List[str]] = [],
+ conditioning_masks_keys: Optional[List[str]] = [],
+ config_yaml: dict = None,
+ save_ckpt_path: str = "./checkpoints",
+ log_interval: int = 100,
+ resume_from_checkpoint: bool = True,
+ max_epochs: int = 100,
+ bridge_noise_sigma: float = 0.005,
+ save_interval: int = 1000,
+ path_config: str = None,
+):
+ model = get_model(
+ backbone_signature=backbone_signature,
+ vae_num_channels=vae_num_channels,
+ unet_input_channels=unet_input_channels,
+ source_key=source_key,
+ target_key=target_key,
+ mask_key=mask_key,
+ timestep_sampling=timestep_sampling,
+ logit_mean=logit_mean,
+ logit_std=logit_std,
+ pixel_loss_type=pixel_loss_type,
+ latent_loss_type=latent_loss_type,
+ latent_loss_weight=latent_loss_weight,
+ pixel_loss_weight=pixel_loss_weight,
+ selected_timesteps=selected_timesteps,
+ prob=prob,
+ conditioning_images_keys=conditioning_images_keys,
+ conditioning_masks_keys=conditioning_masks_keys,
+ bridge_noise_sigma=bridge_noise_sigma,
+ )
+
+ data_module = get_data_module(
+ train_shards=train_shards,
+ validation_shards=validation_shards,
+ batch_size=batch_size,
+ )
+
+ train_parameters = ["denoiser.*"]
+
+ # Training Config
+ training_config = TrainingConfig(
+ learning_rate=learning_rate,
+ lr_scheduler_name=learning_rate_scheduler,
+ lr_scheduler_kwargs=learning_rate_scheduler_kwargs,
+ log_keys=["image", "normal", "mask"],
+ trainable_params=train_parameters,
+ optimizer_name=optimizer,
+ optimizer_kwargs=optimizer_kwargs,
+ log_samples_model_kwargs={
+ "input_shape": None,
+ "num_steps": num_steps,
+ },
+ )
+ if (
+ os.path.exists(save_ckpt_path)
+ and resume_from_checkpoint
+ and "last.ckpt" in os.listdir(save_ckpt_path)
+ ):
+ start_ckpt = f"{save_ckpt_path}/last.ckpt"
+ print(f"Resuming from checkpoint: {start_ckpt}")
+
+ else:
+ start_ckpt = None
+
+ pipeline = TrainingPipeline(model=model, pipeline_config=training_config)
+
+ pipeline.save_hyperparameters(
+ {
+ f"embedder_{i}": embedder.config.to_dict()
+ for i, embedder in enumerate(model.conditioner.conditioners)
+ }
+ )
+
+ pipeline.save_hyperparameters(
+ {
+ "denoiser": model.denoiser.config,
+ "vae": model.vae.config.to_dict(),
+ "config_yaml": config_yaml,
+ "training": training_config.to_dict(),
+ "training_noise_scheduler": model.training_noise_scheduler.config,
+ "sampling_noise_scheduler": model.sampling_noise_scheduler.config,
+ }
+ )
+
+ training_signature = (
+ datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
+ + "-LBM-Surface"
+ + f"{os.environ['SLURM_JOB_ID']}"
+ + f"_{os.environ.get('SLURM_ARRAY_TASK_ID', 0)}"
+ )
+ dir_path = f"{save_ckpt_path}/logs/{training_signature}"
+ if os.environ["SLURM_PROCID"] == "0":
+ os.makedirs(dir_path, exist_ok=True)
+ if path_config is not None:
+ shutil.copy(path_config, f"{save_ckpt_path}/config.yaml")
+ run_name = training_signature
+
+ # Ignore parameters unused during training
+ ignore_states = []
+ for name, param in pipeline.model.named_parameters():
+ ignore = True
+ for regex in ["denoiser."]:
+ pattern = re.compile(regex)
+ if re.match(pattern, name):
+ ignore = False
+ if ignore:
+ ignore_states.append(param)
+
+ # FSDP Strategy
+ strategy = FSDPStrategy(
+ auto_wrap_policy=ModuleWrapPolicy(
+ [
+ UNet2DConditionModel,
+ BasicTransformerBlock,
+ ResnetBlock2D,
+ torch.nn.Conv2d,
+ ]
+ ),
+ activation_checkpointing_policy=ModuleWrapPolicy(
+ [
+ BasicTransformerBlock,
+ ResnetBlock2D,
+ ]
+ ),
+ sharding_strategy="SHARD_GRAD_OP",
+ ignored_states=ignore_states,
+ )
+
+ trainer = Trainer(
+ accelerator="gpu",
+ devices=int(os.environ["SLURM_NPROCS"]) // int(os.environ["SLURM_NNODES"]),
+ num_nodes=int(os.environ["SLURM_NNODES"]),
+ strategy=strategy,
+ default_root_dir="logs",
+ logger=loggers.WandbLogger(
+ project=wandb_project, offline=False, name=run_name, save_dir=save_ckpt_path
+ ),
+ callbacks=[
+ WandbSampleLogger(log_batch_freq=log_interval),
+ LearningRateMonitor(logging_interval="step"),
+ ModelCheckpoint(
+ dirpath=save_ckpt_path,
+ every_n_train_steps=save_interval,
+ save_last=True,
+ ),
+ ],
+ num_sanity_val_steps=0,
+ precision="bf16-mixed",
+ limit_val_batches=2,
+ val_check_interval=1000,
+ max_epochs=max_epochs,
+ )
+
+ trainer.fit(pipeline, data_module, ckpt_path=start_ckpt)
+
+
+def main_from_config(path_config: str = None):
+ with open(path_config, "r") as file:
+ config = yaml.safe_load(file)
+ logging.info(
+ f"Running main with config: {yaml.dump(config, default_flow_style=False)}"
+ )
+ main(**config, config_yaml=config, path_config=path_config)
+
+
+if __name__ == "__main__":
+ fire.Fire(main_from_config)
diff --git a/frpc_linux_amd64_v0.3 b/frpc_linux_amd64_v0.3
new file mode 100644
index 0000000000000000000000000000000000000000..8f0e467b5313aba1138c91ba7a4919dab2a68815
--- /dev/null
+++ b/frpc_linux_amd64_v0.3
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c791d1f047b41ff5885772fc4bf20b797c6059bbd82abb9e31de15e55d6a57c4
+size 11907224
diff --git a/img/input_img/1.jpg b/img/input_img/1.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..851ce04c7ff0a7e4d90cb12e18f6af0aa60065cb
Binary files /dev/null and b/img/input_img/1.jpg differ
diff --git a/img/output_img/output_image.jpg b/img/output_img/output_image.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d571a3bcf961879d9955d54ea1760697eb7ee9db
Binary files /dev/null and b/img/output_img/output_image.jpg differ
diff --git a/img/output_img/source_image.jpg b/img/output_img/source_image.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..dd818739f2d970e50de92489ba9fa32eef464d27
Binary files /dev/null and b/img/output_img/source_image.jpg differ
diff --git a/pyproject.toml b/pyproject.toml
new file mode 100644
index 0000000000000000000000000000000000000000..0ff5b2321340a44ec711bc672fdff91b404222bc
--- /dev/null
+++ b/pyproject.toml
@@ -0,0 +1,39 @@
+[build-system]
+requires = ["hatchling", "hatch-requirements-txt"]
+build-backend = "hatchling.build"
+
+[project]
+name = "lbm"
+dynamic = ["dependencies", "optional-dependencies"]
+description = "LBM: Latent Bridge Matching for Fast Image-to-Image Translation"
+readme = "README.md"
+requires-python = ">=3.10"
+authors = [
+ { name = "Clement Chadebec", email = "clement.chadebec@jasper.ai" },
+ { name = "Benjamin Aubin", email = "benjamin.aubin@jasper.ai" },
+]
+maintainers = [
+ { name = "Clement Chadebec", email = "clement.chadebec@jasper.ai" },
+]
+classifiers = [
+ "Programming Language :: Python :: 3",
+ "Programming Language :: Python :: 3.10",
+ "Programming Language :: Python :: 3.11",
+ "Programming Language :: Python :: 3.12",
+ "License :: OSI Approved :: Apache Software License",
+ "Operating System :: OS Independent",
+]
+version = "0.1"
+
+[project.urls]
+Homepage = "https://github.com/gojasper/LBM"
+Repository = "https://github.com/gojasper/LBM"
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+[tool.hatch.metadata.hooks.requirements_txt]
+files = ["requirements.txt"]
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/lbm"]
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000000000000000000000000000000000000..72b32babfd2137d8041819f932b9c42962183b88
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,24 @@
+# accelerate==1.4.0
+diffusers==0.32.2
+torch==2.7.0
+torchvision>=0.20.0
+black==24.2.0
+einops==0.7.0
+fire>=0.5.0
+gradio==5.29.0
+isort==5.13.2
+lightning==2.5.0
+lpips==0.1.4
+opencv-python==4.9.0.80
+peft==0.9.0
+pydantic>=2.6.1
+scipy>=1.12.0
+sentencepiece>=0.2.0
+timm==0.9.16
+tokenizers>=0.15.2
+torch-fidelity>=0.3.0
+torchmetrics>=1.3.1
+transformers==4.42.3
+wandb==0.16.2
+webdataset>=0.2.86
+kornia==0.8.0
\ No newline at end of file
diff --git a/src/lbm/config.py b/src/lbm/config.py
new file mode 100644
index 0000000000000000000000000000000000000000..4f69788bc87daabfacae31d63087179357ee94a8
--- /dev/null
+++ b/src/lbm/config.py
@@ -0,0 +1,141 @@
+import json
+import os
+import warnings
+from dataclasses import asdict, field
+from typing import Any, Dict, Union
+
+import yaml
+from pydantic import ValidationError
+from pydantic.dataclasses import dataclass
+from yaml import safe_load
+
+
+@dataclass
+class BaseConfig:
+ """This is the BaseConfig class which defines all the useful loading and saving methods
+ of the configs"""
+
+ name: str = field(init=False)
+
+ def __post_init__(self):
+ self.name = self.__class__.__name__
+
+ @classmethod
+ def from_dict(cls, config_dict: Dict[str, Any]) -> "BaseConfig":
+ """Creates a BaseConfig instance from a dictionnary
+
+ Args:
+ config_dict (dict): The Python dictionnary containing all the parameters
+
+ Returns:
+ :class:`BaseConfig`: The created instance
+ """
+ try:
+ config = cls(**config_dict)
+ except (ValidationError, TypeError) as e:
+ raise e
+ return config
+
+ @classmethod
+ def _dict_from_json(cls, json_path: Union[str, os.PathLike]) -> Dict[str, Any]:
+ try:
+ with open(json_path) as f:
+ try:
+ config_dict = json.load(f)
+ return config_dict
+
+ except (TypeError, json.JSONDecodeError) as e:
+ raise TypeError(
+ f"File {json_path} not loadable. Maybe not json ? \n"
+ f"Catch Exception {type(e)} with message: " + str(e)
+ ) from e
+
+ except FileNotFoundError:
+ raise FileNotFoundError(
+ f"Config file not found. Please check path '{json_path}'"
+ )
+
+ @classmethod
+ def from_json(cls, json_path: str) -> "BaseConfig":
+ """Creates a BaseConfig instance from a JSON config file
+
+ Args:
+ json_path (str): The path to the json file containing all the parameters
+
+ Returns:
+ :class:`BaseConfig`: The created instance
+ """
+ config_dict = cls._dict_from_json(json_path)
+
+ config_name = config_dict.pop("name")
+
+ if cls.__name__ != config_name:
+ warnings.warn(
+ f"You are trying to load a "
+ f"`{ cls.__name__}` while a "
+ f"`{config_name}` is given."
+ )
+
+ return cls.from_dict(config_dict)
+
+ def to_dict(self) -> dict:
+ """Transforms object into a Python dictionnary
+
+ Returns:
+ (dict): The dictionnary containing all the parameters"""
+ return asdict(self)
+
+ def to_json_string(self):
+ """Transforms object into a JSON string
+
+ Returns:
+ (str): The JSON str containing all the parameters"""
+ return json.dumps(self.to_dict())
+
+ def save_json(self, file_path: str):
+ """Saves a ``.json`` file from the dataclass
+
+ Args:
+ file_path (str): path to the file
+ """
+ with open(os.path.join(file_path), "w", encoding="utf-8") as fp:
+ fp.write(self.to_json_string())
+
+ def save_yaml(self, file_path: str):
+ """Saves a ``.yaml`` file from the dataclass
+
+ Args:
+ file_path (str): path to the file
+ """
+ with open(os.path.join(file_path), "w", encoding="utf-8") as fp:
+ yaml.dump(self.to_dict(), fp)
+
+ @classmethod
+ def from_yaml(cls, yaml_path: str) -> "BaseConfig":
+ """Creates a BaseConfig instance from a YAML config file
+
+ Args:
+ yaml_path (str): The path to the yaml file containing all the parameters
+
+ Returns:
+ :class:`BaseConfig`: The created instance
+ """
+ with open(yaml_path, "r") as f:
+ try:
+ config_dict = safe_load(f)
+ except yaml.YAMLError as e:
+ raise yaml.YAMLError(
+ f"File {yaml_path} not loadable. Maybe not yaml ? \n"
+ f"Catch Exception {type(e)} with message: " + str(e)
+ ) from e
+
+ config_name = config_dict.pop("name")
+
+ if cls.__name__ != config_name:
+ warnings.warn(
+ f"You are trying to load a "
+ f"`{ cls.__name__}` while a "
+ f"`{config_name}` is given."
+ )
+
+ return cls.from_dict(config_dict)
diff --git a/src/lbm/data/__init__.py b/src/lbm/data/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e38f17398a2f654fe7da0a63a8df3ac954d9a721
--- /dev/null
+++ b/src/lbm/data/__init__.py
@@ -0,0 +1,62 @@
+"""
+This module contains a collection of data related classes and functions to train the :mod:`cr.models`.
+In a training loop a batch of data is struvtued as a dictionnary on which the modules :mod:`cr.data.datasets`
+and :mod:`cr.data.filters` allow to perform several operations.
+
+
+Examples
+########
+
+Create a DataModule to train a model
+
+.. code-block::python
+
+ from cr.data import DataModule, DataModuleConfig
+ from cr.data.filters import KeyFilter, KeyFilterConfig
+ from cr.data.mappers import KeyRenameMapper, KeyRenameMapperConfig
+
+ # Create the filters and mappers
+ filters_mappers = [
+ KeyFilter(KeyFilterConfig(keys=["image", "txt"])),
+ KeyRenameMapper(
+ KeyRenameMapperConfig(key_map={"jpg": "image", "txt": "text"})
+ )
+ ]
+
+ # Create the DataModule
+ data_module = DataModule(
+ train_config=DataModuleConfig(
+ shards_path_or_urls="your urls or paths",
+ decoder="pil",
+ shuffle_buffer_size=100,
+ per_worker_batch_size=32,
+ num_workers=4,
+ ),
+ train_filters_mappers=filters_mappers,
+ eval_config=DataModuleConfig(
+ shards_path_or_urls="your urls or paths",
+ decoder="pil",
+ shuffle_buffer_size=100,
+ per_worker_batch_size=32,
+ num_workers=4,
+ ),
+ eval_filters_mappers=filters_mappers,
+ )
+
+ # This can then be passed to a :mod:`pytorch_lightning.Trainer` to train a model
+
+
+
+
+
+The :mod:`cr.data` includes the following submodules:
+
+- :mod:`cr.data.datasets`: a collection of :mod:`pytorch_lightning.LightningDataModule` used to train the models. In particular,
+ they can used to create the dataloaders and setup the data pipelines.
+- :mod:`cr.data.filters`: a collection of filters used apply filters on a training batch of data/
+
+"""
+
+from .datasets import DataModule
+
+__all__ = ["DataModule"]
diff --git a/src/lbm/data/datasets/__init__.py b/src/lbm/data/datasets/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..5715c941f419125ea0eb925827bfb9f0ecc97e01
--- /dev/null
+++ b/src/lbm/data/datasets/__init__.py
@@ -0,0 +1,9 @@
+"""
+A collection of :mod:`pytorch_lightning.LightningDataModule` used to train the models. In particular,
+they can be used to create the dataloaders and setup the data pipelines.
+"""
+
+from .dataset import DataModule
+from .datasets_config import DataModuleConfig
+
+__all__ = ["DataModule", "DataModuleConfig"]
diff --git a/src/lbm/data/datasets/collation_fn.py b/src/lbm/data/datasets/collation_fn.py
new file mode 100644
index 0000000000000000000000000000000000000000..046309fc94a1498354f02b3b737705d7bfaf1fb2
--- /dev/null
+++ b/src/lbm/data/datasets/collation_fn.py
@@ -0,0 +1,41 @@
+from typing import Dict, List, Union
+
+import numpy as np
+import torch
+
+
+def custom_collation_fn(
+ samples: List[Dict[str, Union[int, float, np.ndarray, torch.Tensor]]],
+ combine_tensors: bool = True,
+ combine_scalars: bool = True,
+) -> dict:
+ """
+ Collate function for PyTorch DataLoader.
+
+ Args:
+ samples(List[Dict[str, Union[int, float, np.ndarray, torch.Tensor]]]): List of samples.
+ combine_tensors (bool): Whether to turn lists of tensors into a single tensor.
+ combine_scalars (bool): Whether to turn lists of scalars into a single ndarray.
+ """
+ keys = set.intersection(*[set(sample.keys()) for sample in samples])
+ batched = {key: [] for key in keys}
+ for s in samples:
+ [batched[key].append(s[key]) for key in batched]
+
+ result = {}
+ for key in batched:
+ if isinstance(batched[key][0], (int, float)):
+ if combine_scalars:
+ result[key] = np.array(list(batched[key]))
+ elif isinstance(batched[key][0], torch.Tensor):
+ if combine_tensors:
+ result[key] = torch.stack(list(batched[key]))
+ elif isinstance(batched[key][0], np.ndarray):
+ if combine_tensors:
+ result[key] = np.array(list(batched[key]))
+ else:
+ result[key] = list(batched[key])
+
+ del samples
+ del batched
+ return result
diff --git a/src/lbm/data/datasets/dataset.py b/src/lbm/data/datasets/dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..c89540653dfb41ef16c6762f7286fd8a22e382b6
--- /dev/null
+++ b/src/lbm/data/datasets/dataset.py
@@ -0,0 +1,243 @@
+from typing import Callable, List, Union
+
+import pytorch_lightning as pl
+import webdataset as wds
+from webdataset import DataPipeline
+
+from ..filters import BaseFilter, FilterWrapper
+from ..mappers import BaseMapper, MapperWrapper
+from .collation_fn import custom_collation_fn
+from .datasets_config import DataModuleConfig
+
+
+class DataPipeline:
+ """
+ DataPipeline class for creating a dataloader from a single configuration
+
+ Args:
+
+ config (DataModuleConfig):
+ Configuration for the dataset
+
+ filters_mappers (Union[List[Union[BaseMapper, BaseFilter, FilterWrapper, MapperWrapper]]):
+ List of filters and mappers for the dataset. These will be sequentially applied.
+
+ batched_filters_mappers (List[Union[BaseMapper, BaseFilter, FilterWrapper, MapperWrapper]]):
+ List of batched transforms for the dataset. These will be sequentially applied.
+ """
+
+ def __init__(
+ self,
+ config: DataModuleConfig,
+ filters_mappers: List[
+ Union[BaseMapper, BaseFilter, FilterWrapper, MapperWrapper]
+ ],
+ batched_filters_mappers: List[
+ Union[BaseMapper, BaseFilter, FilterWrapper, MapperWrapper]
+ ] = None,
+ ):
+ self.config = config
+ self.shards_path_or_urls = config.shards_path_or_urls
+ self.filters_mappers = filters_mappers
+ self.batched_filters_mappers = batched_filters_mappers or []
+
+ if filters_mappers is None:
+ filters_mappers = []
+
+ # set processing pipeline
+ self.processing_pipeline = [wds.decode(config.decoder, handler=config.handler)]
+ self.processing_pipeline.extend(
+ self._add_filters_mappers(
+ filters_mappers=filters_mappers,
+ handler=config.handler,
+ )
+ )
+
+ def _add_filters_mappers(
+ self,
+ filters_mappers: List[
+ Union[
+ FilterWrapper,
+ MapperWrapper,
+ ]
+ ],
+ handler: Callable = wds.warn_and_continue,
+ ) -> List[Union[FilterWrapper, MapperWrapper]]:
+ tmp_pipeline = []
+ for filter_mapper in filters_mappers:
+ if isinstance(filter_mapper, FilterWrapper) or isinstance(
+ filter_mapper, BaseFilter
+ ):
+ tmp_pipeline.append(wds.select(filter_mapper))
+ elif isinstance(filter_mapper, MapperWrapper) or isinstance(
+ filter_mapper, BaseMapper
+ ):
+ tmp_pipeline.append(wds.map(filter_mapper, handler=handler))
+ elif isinstance(filter_mapper) or isinstance(filter_mapper):
+ tmp_pipeline.append(wds.map(filter_mapper, handler=handler))
+ else:
+ raise ValueError("Unknown type of filter/mapper")
+ return tmp_pipeline
+
+ def setup(self):
+ pipeline = [wds.SimpleShardList(self.shards_path_or_urls)]
+
+ # shuffle before split by node
+ if self.config.shuffle_before_split_by_node_buffer_size is not None:
+ pipeline.append(
+ wds.shuffle(
+ self.config.shuffle_before_split_by_node_buffer_size,
+ handler=self.config.handler,
+ )
+ )
+ # split by node
+ pipeline.append(wds.split_by_node)
+
+ # shuffle before split by workers
+ if self.config.shuffle_before_split_by_workers_buffer_size is not None:
+ pipeline.append(
+ wds.shuffle(
+ self.config.shuffle_before_split_by_workers_buffer_size,
+ handler=self.config.handler,
+ )
+ )
+ # split by worker
+ pipeline.extend(
+ [
+ wds.split_by_worker,
+ wds.tarfile_to_samples(
+ handler=self.config.handler,
+ rename_files=self.config.rename_files_fn,
+ ),
+ ]
+ )
+
+ # shuffle before filter mappers
+ if self.config.shuffle_before_filter_mappers_buffer_size is not None:
+ pipeline.append(
+ wds.shuffle(
+ self.config.shuffle_before_filter_mappers_buffer_size,
+ handler=self.config.handler,
+ )
+ )
+
+ # apply filters and mappers
+ pipeline.extend(self.processing_pipeline)
+
+ # shuffle after filter mappers
+ if self.config.shuffle_after_filter_mappers_buffer_size is not None:
+ pipeline.append(
+ wds.shuffle(
+ self.config.shuffle_after_filter_mappers_buffer_size,
+ handler=self.config.handler,
+ ),
+ )
+
+ # batching
+ pipeline.append(
+ wds.batched(
+ self.config.per_worker_batch_size,
+ collation_fn=custom_collation_fn,
+ )
+ )
+
+ # apply batched transforms
+ pipeline.extend(
+ self._add_filters_mappers(
+ filters_mappers=self.batched_filters_mappers,
+ handler=self.config.handler,
+ )
+ )
+
+ # create the data pipeline
+ pipeline = wds.DataPipeline(*pipeline, handler=self.config.handler)
+
+ # set the pipeline
+ self.pipeline = pipeline
+
+ def dataloader(self):
+ # return the loader
+ return wds.WebLoader(
+ self.pipeline,
+ batch_size=None,
+ num_workers=self.config.num_workers,
+ )
+
+
+class DataModule(pl.LightningDataModule):
+ """
+ Main DataModule class for creating data loaders and training/evaluating models
+
+ Args:
+
+ train_config (DataModuleConfig):
+ Configuration for the training dataset
+
+ train_filters_mappers (Union[List[Union[BaseMapper, BaseFilter, FilterWrapper, MapperWrapper]]):
+ List of filters and mappers for the training dataset. These will be sequentially applied.
+
+ train_batched_filters_mappers (List[Union[BaseMapper, BaseFilter, FilterWrapper, MapperWrapper]]):
+ List of batched transforms for the training dataset. These will be sequentially applied.
+
+ eval_config (DataModuleConfig):
+ Configuration for the evaluation dataset
+
+ eval_filters_mappers (List[Union[FilterWrapper, MapperWrapper]]):
+ List of filters and mappers for the evaluation dataset.These will be sequentially applied.
+
+ eval_batched_filters_mappers (List[Union[BaseMapper, BaseFilter, FilterWrapper, MapperWrapper]]):
+ List of batched transforms for the evaluation dataset. These will be sequentially applied.
+ """
+
+ def __init__(
+ self,
+ train_config: DataModuleConfig,
+ train_filters_mappers: List[
+ Union[BaseMapper, BaseFilter, FilterWrapper, MapperWrapper]
+ ] = None,
+ train_batched_filters_mappers: List[
+ Union[BaseMapper, BaseFilter, FilterWrapper, MapperWrapper]
+ ] = None,
+ eval_config: DataModuleConfig = None,
+ eval_filters_mappers: List[Union[FilterWrapper, MapperWrapper]] = None,
+ eval_batched_filters_mappers: List[
+ Union[BaseMapper, BaseFilter, FilterWrapper, MapperWrapper]
+ ] = None,
+ ):
+ super().__init__()
+
+ self.train_config = train_config
+ self.train_filters_mappers = train_filters_mappers
+ self.train_batched_filters_mappers = train_batched_filters_mappers
+
+ self.eval_config = eval_config
+ self.eval_filters_mappers = eval_filters_mappers
+ self.eval_batched_filters_mappers = eval_batched_filters_mappers
+
+ def setup(self, stage=None):
+ """
+ Setup the data module and create the webdataset processing pipelines
+ """
+
+ # train pipeline
+ self.train_pipeline = DataPipeline(
+ config=self.train_config,
+ filters_mappers=self.train_filters_mappers,
+ batched_filters_mappers=self.train_batched_filters_mappers,
+ )
+ self.train_pipeline.setup()
+
+ # eval pipeline
+ if self.eval_config is not None:
+ self.eval_pipeline = DataPipeline(
+ config=self.eval_config,
+ filters_mappers=self.eval_filters_mappers,
+ batched_filters_mappers=self.eval_batched_filters_mappers,
+ )
+ self.eval_pipeline.setup()
+
+ def train_dataloader(self):
+ return self.train_pipeline.dataloader()
+
+ def val_dataloader(self):
+ return self.eval_pipeline.dataloader()
diff --git a/src/lbm/data/datasets/datasets_config.py b/src/lbm/data/datasets/datasets_config.py
new file mode 100644
index 0000000000000000000000000000000000000000..4ecddeedce4c9fc1695e4deca7f1f83ce33be25c
--- /dev/null
+++ b/src/lbm/data/datasets/datasets_config.py
@@ -0,0 +1,42 @@
+from typing import Callable, List, Optional, Union
+
+import webdataset as wds
+from pydantic.dataclasses import dataclass
+
+from ...config import BaseConfig
+
+
+@dataclass
+class DataModuleConfig(BaseConfig):
+ """
+ Configuration for the DataModule
+
+ Args:
+
+ shards_path_or_urls (Union[str, List[str]]): The path or url to the shards. Defaults to None.
+ per_worker_batch_size (int): The batch size for the dataset. Defaults to 16.
+ num_workers (int): The number of workers to use. Defaults to 1.
+ shuffle_before_split_by_node_buffer_size (Optional[int]): The buffer size for the shuffle before split by node. Defaults to 100.
+ shuffle_before_split_by_workers_buffer_size (Optional[int]): The buffer size for the shuffle before split by workers. Defaults to 100.
+ shuffle_before_filter_mappers_buffer_size (Optional[int]): The buffer size for the shuffle before filter mappers. Defaults to 1000.
+ shuffle_after_filter_mappers_buffer_size (Optional[int]): The buffer size for the shuffle after filter mappers. Defaults to 1000.
+ decoder (str): The decoder to use. Defaults to "pil".
+ handler (Callable): A callable to handle the warnings. Defaults to wds.warn_and_continue.
+ rename_files_fn (Optional[Callable[[str], str]]): A callable to rename the files. Defaults to None.
+ """
+
+ shards_path_or_urls: Union[str, List[str]] = None
+ per_worker_batch_size: int = 16
+ num_workers: int = 1
+ shuffle_before_split_by_node_buffer_size: Optional[int] = 100
+ shuffle_before_split_by_workers_buffer_size: Optional[int] = 100
+ shuffle_before_filter_mappers_buffer_size: Optional[int] = 1000
+ shuffle_after_filter_mappers_buffer_size: Optional[int] = 1000
+ decoder: str = "pil"
+ handler: Callable = wds.warn_and_continue
+ rename_files_fn: Optional[Callable[[str], str]] = None
+
+ def __post_init__(self):
+ super().__post_init__()
+ if self.rename_files_fn is not None:
+ assert callable(self.rename_files_fn), "rename_files must be a callable"
diff --git a/src/lbm/data/filters/__init__.py b/src/lbm/data/filters/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..6a97c41fad372d0ef464e7f5f093ceced2b4c96f
--- /dev/null
+++ b/src/lbm/data/filters/__init__.py
@@ -0,0 +1,12 @@
+from .base import BaseFilter
+from .filter_wrapper import FilterWrapper
+from .filters import KeyFilter
+from .filters_config import BaseFilterConfig, KeyFilterConfig
+
+__all__ = [
+ "BaseFilter",
+ "FilterWrapper",
+ "KeyFilter",
+ "BaseFilterConfig",
+ "KeyFilterConfig",
+]
diff --git a/src/lbm/data/filters/base.py b/src/lbm/data/filters/base.py
new file mode 100644
index 0000000000000000000000000000000000000000..357dc2e099a971a62a0eb60ea40f76b537f7f3e5
--- /dev/null
+++ b/src/lbm/data/filters/base.py
@@ -0,0 +1,21 @@
+from typing import Any, Dict
+
+from .filters_config import BaseFilterConfig
+
+
+class BaseFilter:
+ """
+ Base class for filters. This class should be subclassed to create a new filter.
+
+ Args:
+
+ config (BaseFilterConfig):
+ Configuration for the filter
+ """
+
+ def __init__(self, config: BaseFilterConfig):
+ self.verbose = config.verbose
+
+ def __call__(self, sample: Dict[str, Any]) -> bool:
+ """This function should be implemented by the subclass"""
+ raise NotImplementedError
diff --git a/src/lbm/data/filters/filter_wrapper.py b/src/lbm/data/filters/filter_wrapper.py
new file mode 100644
index 0000000000000000000000000000000000000000..e2477e110f63ae0c11345b12c20881cafdf8af55
--- /dev/null
+++ b/src/lbm/data/filters/filter_wrapper.py
@@ -0,0 +1,36 @@
+from typing import Any, Dict, List, Union
+
+from .base import BaseFilter
+
+
+class FilterWrapper:
+ """
+ Wrapper for multiple filters. This class allows to apply multiple filters to a batch of data.
+ The filters are applied in the order they are passed to the wrapper.
+
+ Args:
+
+ filters (List[BaseFilter]):
+ List of filters to apply to the batch of data
+ """
+
+ def __init__(
+ self,
+ filters: Union[List[BaseFilter], None] = None,
+ ):
+ self.filters = filters
+
+ def __call__(self, batch: Dict[str, Any]) -> None:
+ """
+ Forward pass through all filters
+
+ Args:
+
+ batch: batch of data
+ """
+ filter_output = True
+ for filter in self.filters:
+ filter_output = filter(batch)
+ if not filter_output:
+ return False
+ return True
diff --git a/src/lbm/data/filters/filters.py b/src/lbm/data/filters/filters.py
new file mode 100644
index 0000000000000000000000000000000000000000..41ab776916bc773482872456965eb40979dc4af8
--- /dev/null
+++ b/src/lbm/data/filters/filters.py
@@ -0,0 +1,33 @@
+import logging
+
+from .base import BaseFilter
+from .filters_config import KeyFilterConfig
+
+logging.basicConfig(level=logging.INFO)
+
+
+class KeyFilter(BaseFilter):
+ """
+ This filter checks if ALL the given keys are present in the sample
+
+ Args:
+
+ config (KeyFilterConfig): configuration for the filter
+ """
+
+ def __init__(self, config: KeyFilterConfig):
+ super().__init__(config)
+ keys = config.keys
+ if isinstance(keys, str):
+ keys = [keys]
+
+ self.keys = set(keys)
+
+ def __call__(self, batch: dict) -> bool:
+ try:
+ res = self.keys.issubset(set(batch.keys()))
+ return res
+ except Exception as e:
+ if self.verbose:
+ logging.error(f"Error in KeyFilter: {e}")
+ return False
diff --git a/src/lbm/data/filters/filters_config.py b/src/lbm/data/filters/filters_config.py
new file mode 100644
index 0000000000000000000000000000000000000000..731097c5e32c53fe60218dc327219907f42e25ae
--- /dev/null
+++ b/src/lbm/data/filters/filters_config.py
@@ -0,0 +1,32 @@
+from typing import List, Union
+
+from pydantic.dataclasses import dataclass
+
+from ...config import BaseConfig
+
+
+@dataclass
+class BaseFilterConfig(BaseConfig):
+ """
+ Base configuration for filters
+
+ Args:
+
+ verbose (bool):
+ If True, print debug information. Defaults to False"""
+
+ verbose: bool = False
+
+
+@dataclass
+class KeyFilterConfig(BaseFilterConfig):
+ """
+ This filter checks if the keys are present in a sample.
+
+ Args:
+
+ keys (Union[str, List[str]]):
+ Key or list of keys to check. Defaults to "txt"
+ """
+
+ keys: Union[str, List[str]] = "txt"
diff --git a/src/lbm/data/mappers/__init__.py b/src/lbm/data/mappers/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..92a893f5485a46c7635c12cedfba375053e1f38f
--- /dev/null
+++ b/src/lbm/data/mappers/__init__.py
@@ -0,0 +1,19 @@
+from .base import BaseMapper
+from .mappers import KeyRenameMapper, RescaleMapper, TorchvisionMapper
+from .mappers_config import (
+ KeyRenameMapperConfig,
+ RescaleMapperConfig,
+ TorchvisionMapperConfig,
+)
+from .mappers_wrapper import MapperWrapper
+
+__all__ = [
+ "BaseMapper",
+ "KeyRenameMapper",
+ "RescaleMapper",
+ "TorchvisionMapper",
+ "KeyRenameMapperConfig",
+ "RescaleMapperConfig",
+ "TorchvisionMapperConfig",
+ "MapperWrapper",
+]
diff --git a/src/lbm/data/mappers/base.py b/src/lbm/data/mappers/base.py
new file mode 100644
index 0000000000000000000000000000000000000000..49414b4b37673923045a61c12bf1716787f64bee
--- /dev/null
+++ b/src/lbm/data/mappers/base.py
@@ -0,0 +1,26 @@
+from typing import Any, Dict
+
+from .mappers_config import BaseMapperConfig
+
+
+class BaseMapper:
+ """
+ Base class for the mappers used to modify the samples in the data pipeline.
+
+ Args:
+
+ config (BaseMapperConfig):
+ Configuration for the mapper.
+ """
+
+ def __init__(self, config: BaseMapperConfig):
+ self.config = config
+ self.key = config.key
+
+ if config.output_key is None:
+ self.output_key = config.key
+ else:
+ self.output_key = config.output_key
+
+ def map(self, batch: Dict[str, Any], *args, **kwargs) -> Dict[str, Any]:
+ raise NotImplementedError
diff --git a/src/lbm/data/mappers/mappers.py b/src/lbm/data/mappers/mappers.py
new file mode 100644
index 0000000000000000000000000000000000000000..41df641d3775270ce0fbba7e0c3c5bc31a8e7d1d
--- /dev/null
+++ b/src/lbm/data/mappers/mappers.py
@@ -0,0 +1,135 @@
+from typing import Any, Dict
+
+from torchvision import transforms
+
+from .base import BaseMapper
+from .mappers_config import (
+ KeyRenameMapperConfig,
+ RescaleMapperConfig,
+ TorchvisionMapperConfig,
+)
+
+
+class KeyRenameMapper(BaseMapper):
+ """
+ Rename keys in a sample according to a key map
+
+ Args:
+
+ config (KeyRenameMapperConfig): Configuration for the mapper
+
+ Examples
+ ########
+
+ 1. Rename keys in a sample according to a key map
+
+ .. code-block:: python
+
+ from cr.data.mappers import KeyRenameMapper, KeyRenameMapperConfig
+
+ config = KeyRenameMapperConfig(
+ key_map={"old_key": "new_key"}
+ )
+
+ mapper = KeyRenameMapper(config)
+
+ sample = {"old_key": 1}
+ new_sample = mapper(sample)
+ print(new_sample) # {"new_key": 1}
+
+ 2. Rename keys in a sample according to a key map and a condition key
+
+ .. code-block:: python
+
+ from cr.data.mappers import KeyRenameMapper, KeyRenameMapperConfig
+
+ config = KeyRenameMapperConfig(
+ key_map={"old_key": "new_key"},
+ condition_key="condition",
+ condition_fn=lambda x: x == 1
+ )
+
+ mapper = KeyRenameMapper(config)
+
+ sample = {"old_key": 1, "condition": 1}
+ new_sample = mapper(sample)
+ print(new_sample) # {"new_key": 1}
+
+ sample = {"old_key": 1, "condition": 0}
+ new_sample = mapper(sample)
+ print(new_sample) # {"old_key": 1}
+
+ ```
+ """
+
+ def __init__(self, config: KeyRenameMapperConfig):
+ super().__init__(config)
+ self.key_map = config.key_map
+ self.condition_key = config.condition_key
+ self.condition_fn = config.condition_fn
+ self.else_key_map = config.else_key_map
+
+ def __call__(self, batch: Dict[str, Any], *args, **kwrags):
+ if self.condition_key is not None:
+ condition_key = batch[self.condition_key]
+ if self.condition_fn(condition_key):
+ for old_key, new_key in self.key_map.items():
+ if old_key in batch:
+ batch[new_key] = batch.pop(old_key)
+
+ elif self.else_key_map is not None:
+ for old_key, new_key in self.else_key_map.items():
+ if old_key in batch:
+ batch[new_key] = batch.pop(old_key)
+
+ else:
+ for old_key, new_key in self.key_map.items():
+ if old_key in batch:
+ batch[new_key] = batch.pop(old_key)
+ return batch
+
+
+class TorchvisionMapper(BaseMapper):
+ """
+ Apply torchvision transforms to a sample
+
+ Args:
+
+ config (TorchvisionMapperConfig): Configuration for the mapper
+ """
+
+ def __init__(self, config: TorchvisionMapperConfig):
+ super().__init__(config)
+ chained_transforms = []
+ for transform, kwargs in zip(config.transforms, config.transforms_kwargs):
+ transform = getattr(transforms, transform)
+ chained_transforms.append(transform(**kwargs))
+ self.transforms = transforms.Compose(chained_transforms)
+
+ def __call__(self, batch: Dict[str, Any], *args, **kwrags) -> Dict[str, Any]:
+ if self.key in batch:
+ batch[self.output_key] = self.transforms(batch[self.key])
+ return batch
+
+
+class RescaleMapper(BaseMapper):
+ """
+ Rescale a sample from [0, 1] to [-1, 1]
+
+ Args:
+
+ config (RescaleMapperConfig): Configuration for the mapper
+ """
+
+ def __init__(self, config: RescaleMapperConfig):
+ super().__init__(config)
+
+ def __call__(self, batch: Dict[str, Any], *args, **kwrags) -> Dict[str, Any]:
+ if isinstance(batch[self.key], list):
+ tmp = []
+ for i, image in enumerate(batch[self.key]):
+ tmp.append(2 * image - 1)
+ batch[self.output_key] = tmp
+ else:
+ batch[self.output_key] = 2 * batch[self.key] - 1
+ return batch
diff --git a/src/lbm/data/mappers/mappers_config.py b/src/lbm/data/mappers/mappers_config.py
new file mode 100644
index 0000000000000000000000000000000000000000..6f0b843b8d77374dcaca8d1947a6b41183335693
--- /dev/null
+++ b/src/lbm/data/mappers/mappers_config.py
@@ -0,0 +1,109 @@
+from typing import Any, Callable, Dict, List, Optional
+
+from pydantic.dataclasses import dataclass
+
+from ...config import BaseConfig
+
+
+@dataclass
+class BaseMapperConfig(BaseConfig):
+ """
+ Base configuration for mappers.
+
+ Args:
+
+ verbose (bool):
+ If True, print debug information. Defaults to False
+
+ key (Optional[str]):
+ Key to apply the mapper to. Defaults to None
+
+ output_key (Optional[str]):
+ Key to store the output of the mapper. Defaults to None
+ """
+
+ verbose: bool = False
+ key: Optional[str] = None
+ output_key: Optional[str] = None
+
+
+@dataclass
+class KeyRenameMapperConfig(BaseMapperConfig):
+ """
+ Rename keys in a sample according to a key map
+
+ Args:
+
+ key_map (Dict[str, str]): Dictionary with the old keys as keys and the new keys as values
+ condition_key (Optional[str]): Key to use for the condition. Defaults to None
+ condition_fn (Optional[Callable[[Any], bool]]): Function to use for the condition to be met so
+ the key map is applied. Defaults to None.
+ else_key_map (Optional[Dict[str, str]]): Dictionary with the old keys as keys and the new keys as values
+ if the condition is not met. Defaults to None *i.e.* the original key will be used.
+ """
+
+ key_map: Dict[str, str] = None
+ condition_key: Optional[str] = None
+ condition_fn: Optional[Callable[[Any], bool]] = None
+ else_key_map: Optional[Dict[str, str]] = None
+
+ def __post_init__(self):
+ super().__post_init__()
+ assert self.key_map is not None, "key_map should be provided"
+ assert all(
+ isinstance(old_key, str) and isinstance(new_key, str)
+ for old_key, new_key in self.key_map.items()
+ ), "key_map should be a dictionary with string keys and values"
+ if self.condition_key is not None:
+ assert self.condition_fn is not None, "condition_fn should be provided"
+ assert callable(self.condition_fn), "condition_fn should be callable"
+ if self.condition_fn is not None:
+ assert self.condition_key is not None, "condition_key should be provided"
+ assert isinstance(
+ self.condition_key, str
+ ), "condition_key should be a string"
+ if self.else_key_map is not None:
+ assert all(
+ isinstance(old_key, str) and isinstance(new_key, str)
+ for old_key, new_key in self.else_key_map.items()
+ ), "else_key_map should be a dictionary with string keys and values"
+
+
+@dataclass
+class TorchvisionMapperConfig(BaseMapperConfig):
+ """
+ Apply torchvision transforms to a sample
+
+ Args:
+
+ key (str): Key to apply the transforms to
+ transforms (torchvision.transforms): List of torchvision transforms to apply
+ transforms_kwargs (Dict[str, Any]): List of kwargs for the transforms
+ """
+
+ key: str = "image"
+ transforms: List[str] = None
+ transforms_kwargs: List[Dict[str, Any]] = None
+
+ def __post_init__(self):
+ super().__post_init__()
+ if self.transforms is None:
+ self.transforms = []
+ if self.transforms_kwargs is None:
+ self.transforms_kwargs = []
+ assert len(self.transforms) == len(
+ self.transforms_kwargs
+ ), "Number of transforms and kwargs should be same"
+
+
+@dataclass
+class RescaleMapperConfig(BaseMapperConfig):
+ """
+ Rescale a sample from [0, 1] to [-1, 1]
+
+ Args:
+
+ key (str): Key to rescale
+ """
+
+ key: str = "image"
diff --git a/src/lbm/data/mappers/mappers_wrapper.py b/src/lbm/data/mappers/mappers_wrapper.py
new file mode 100644
index 0000000000000000000000000000000000000000..7c2373c6cfcb5af4cf7e3430f6b5cee449c8510b
--- /dev/null
+++ b/src/lbm/data/mappers/mappers_wrapper.py
@@ -0,0 +1,31 @@
+from typing import Any, Dict, List, Union
+
+from .base import BaseMapper
+
+
+class MapperWrapper:
+ """
+ Wrapper for the mappers to allow iterating over several mappers in one go.
+
+ Args:
+
+ mappers (Union[List[BaseMapper], None]): List of mappers to apply to the batch
+ """
+
+ def __init__(
+ self,
+ mappers: Union[List[BaseMapper], None] = None,
+ ):
+ self.mappers = mappers
+
+ def __call__(self, batch: Dict[str, Any]) -> Dict[str, Any]:
+ """
+ Forward pass through all mappers
+
+ Args:
+
+ batch (Dict[str, Any]): batch of data
+ """
+ for mapper in self.mappers:
+ batch = mapper(batch)
+ return batch
diff --git a/src/lbm/inference/__init__.py b/src/lbm/inference/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..925781d48cbed28c7431bce9b7d3c74593707b97
--- /dev/null
+++ b/src/lbm/inference/__init__.py
@@ -0,0 +1,4 @@
+from .inference import evaluate
+from .utils import get_model
+
+__all__ = ["evaluate", "get_model"]
diff --git a/src/lbm/inference/inference.py b/src/lbm/inference/inference.py
new file mode 100644
index 0000000000000000000000000000000000000000..a63f47e47824b06e524b05a57c28434bad963e91
--- /dev/null
+++ b/src/lbm/inference/inference.py
@@ -0,0 +1,70 @@
+import logging
+
+import PIL
+import torch
+from torchvision.transforms import ToPILImage, ToTensor
+
+from lbm.models.lbm import LBMModel
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+ASPECT_RATIOS = {
+ str(512 / 2048): (512, 2048),
+ str(1024 / 1024): (1024, 1024),
+ str(2048 / 512): (2048, 512),
+ str(896 / 1152): (896, 1152),
+ str(1152 / 896): (1152, 896),
+ str(512 / 1920): (512, 1920),
+ str(640 / 1536): (640, 1536),
+ str(768 / 1280): (768, 1280),
+ str(1280 / 768): (1280, 768),
+ str(1536 / 640): (1536, 640),
+ str(1920 / 512): (1920, 512),
+}
+
+
+@torch.no_grad()
+def evaluate(
+ model: LBMModel,
+ source_image: PIL.Image.Image,
+ num_sampling_steps: int = 1,
+):
+ """
+ Evaluate the model on an image coming from the source distribution and generate a new image from the target distribution.
+
+ Args:
+ model (LBMModel): The model to evaluate.
+ source_image (PIL.Image.Image): The source image to evaluate the model on.
+ num_sampling_steps (int): The number of sampling steps to use for the model.
+
+ Returns:
+ PIL.Image.Image: The generated image.
+ """
+
+ ori_h_bg, ori_w_bg = source_image.size
+ ar_bg = ori_h_bg / ori_w_bg
+ closest_ar_bg = min(ASPECT_RATIOS, key=lambda x: abs(float(x) - ar_bg))
+ source_dimensions = ASPECT_RATIOS[closest_ar_bg]
+
+ source_image = source_image.resize(source_dimensions)
+
+ img_pasted_tensor = ToTensor()(source_image).unsqueeze(0) * 2 - 1
+ batch = {
+ "source_image": img_pasted_tensor.cuda().to(torch.bfloat16),
+ }
+
+ z_source = model.vae.encode(batch[model.source_key])
+
+ output_image = model.sample(
+ z=z_source,
+ num_steps=num_sampling_steps,
+ conditioner_inputs=batch,
+ max_samples=1,
+ ).clamp(-1, 1)
+
+ output_image = (output_image[0].float().cpu() + 1) / 2
+ output_image = ToPILImage()(output_image)
+ output_image.resize((ori_h_bg, ori_w_bg))
+
+ return output_image
diff --git a/src/lbm/inference/utils.py b/src/lbm/inference/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..0697d9612cd1fcf12a44ea6992c6621f9635b4be
--- /dev/null
+++ b/src/lbm/inference/utils.py
@@ -0,0 +1,222 @@
+import logging
+import os
+from typing import List, Optional
+
+import torch
+import yaml
+from diffusers import FlowMatchEulerDiscreteScheduler
+from huggingface_hub import snapshot_download
+from safetensors.torch import load_file
+
+from lbm.models.embedders import (
+ ConditionerWrapper,
+ LatentsConcatEmbedder,
+ LatentsConcatEmbedderConfig,
+)
+from lbm.models.lbm import LBMConfig, LBMModel
+from lbm.models.unets import DiffusersUNet2DCondWrapper
+from lbm.models.vae import AutoencoderKLDiffusers, AutoencoderKLDiffusersConfig
+
+
+def get_model(
+ model_dir: str,
+ save_dir: Optional[str] = None,
+ torch_dtype: torch.dtype = torch.bfloat16,
+ device: str = "cuda",
+) -> LBMModel:
+ """Download the model from the model directory using either a local path or a path to HuggingFace Hub
+
+ Args:
+ model_dir (str): The path to the model directory containing the model weights and config, can be a local path or a path to HuggingFace Hub
+ save_dir (Optional[str]): The local path to save the model if downloading from HuggingFace Hub. Defaults to None.
+ torch_dtype (torch.dtype): The torch dtype to use for the model. Defaults to torch.bfloat16.
+ device (str): The device to use for the model. Defaults to "cuda".
+
+ Returns:
+ LBMModel: The loaded model
+ """
+ if not os.path.exists(model_dir):
+ local_dir = snapshot_download(
+ model_dir,
+ local_dir=save_dir,
+ )
+ model_dir = local_dir
+
+ model_files = os.listdir(model_dir)
+
+ # check yaml config file is present
+ yaml_file = [f for f in model_files if f.endswith(".yaml")]
+ if len(yaml_file) == 0:
+ raise ValueError("No yaml file found in the model directory.")
+
+ # check safetensors weights file is present
+ safetensors_files = sorted([f for f in model_files if f.endswith(".safetensors")])
+ ckpt_files = sorted([f for f in model_files if f.endswith(".ckpt")])
+ if len(safetensors_files) == 0 and len(ckpt_files) == 0:
+ raise ValueError("No safetensors or ckpt file found in the model directory")
+
+ if len(model_files) == 0:
+ raise ValueError("No model files found in the model directory")
+
+ with open(os.path.join(model_dir, yaml_file[0]), "r") as f:
+ config = yaml.safe_load(f)
+
+ model = _get_model_from_config(**config, torch_dtype=torch_dtype)
+
+ if len(safetensors_files) > 0:
+ logging.info(f"Loading safetensors file: {safetensors_files[-1]}")
+ sd = load_file(os.path.join(model_dir, safetensors_files[-1]))
+ model.load_state_dict(sd, strict=True)
+ elif len(ckpt_files) > 0:
+ logging.info(f"Loading ckpt file: {ckpt_files[-1]}")
+ sd = torch.load(
+ os.path.join(model_dir, ckpt_files[-1]),
+ map_location="cpu",
+ )["state_dict"]
+ sd = {k[6:]: v for k, v in sd.items() if k.startswith("model.")}
+ model.load_state_dict(
+ sd,
+ strict=True,
+ )
+ model.to(device).to(torch_dtype)
+
+ model.eval()
+
+ return model
+
+
+def _get_model_from_config(
+ backbone_signature: str = "stabilityai/stable-diffusion-xl-base-1.0",
+ vae_num_channels: int = 4,
+ unet_input_channels: int = 4,
+ timestep_sampling: str = "log_normal",
+ selected_timesteps: Optional[List[float]] = None,
+ prob: Optional[List[float]] = None,
+ conditioning_images_keys: Optional[List[str]] = [],
+ conditioning_masks_keys: Optional[List[str]] = [],
+ source_key: str = "source_image",
+ target_key: str = "source_image_paste",
+ bridge_noise_sigma: float = 0.0,
+ logit_mean: float = 0.0,
+ logit_std: float = 1.0,
+ pixel_loss_type: str = "lpips",
+ latent_loss_type: str = "l2",
+ latent_loss_weight: float = 1.0,
+ pixel_loss_weight: float = 0.0,
+ torch_dtype: torch.dtype = torch.bfloat16,
+ **kwargs,
+):
+
+ conditioners = []
+
+ denoiser = DiffusersUNet2DCondWrapper(
+ in_channels=unet_input_channels, # Add downsampled_image
+ out_channels=vae_num_channels,
+ center_input_sample=False,
+ flip_sin_to_cos=True,
+ freq_shift=0,
+ down_block_types=[
+ "DownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ ],
+ mid_block_type="UNetMidBlock2DCrossAttn",
+ up_block_types=["CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "UpBlock2D"],
+ only_cross_attention=False,
+ block_out_channels=[320, 640, 1280],
+ layers_per_block=2,
+ downsample_padding=1,
+ mid_block_scale_factor=1,
+ dropout=0.0,
+ act_fn="silu",
+ norm_num_groups=32,
+ norm_eps=1e-05,
+ cross_attention_dim=[320, 640, 1280],
+ transformer_layers_per_block=[1, 2, 10],
+ reverse_transformer_layers_per_block=None,
+ encoder_hid_dim=None,
+ encoder_hid_dim_type=None,
+ attention_head_dim=[5, 10, 20],
+ num_attention_heads=None,
+ dual_cross_attention=False,
+ use_linear_projection=True,
+ class_embed_type=None,
+ addition_embed_type=None,
+ addition_time_embed_dim=None,
+ num_class_embeds=None,
+ upcast_attention=None,
+ resnet_time_scale_shift="default",
+ resnet_skip_time_act=False,
+ resnet_out_scale_factor=1.0,
+ time_embedding_type="positional",
+ time_embedding_dim=None,
+ time_embedding_act_fn=None,
+ timestep_post_act=None,
+ time_cond_proj_dim=None,
+ conv_in_kernel=3,
+ conv_out_kernel=3,
+ projection_class_embeddings_input_dim=None,
+ attention_type="default",
+ class_embeddings_concat=False,
+ mid_block_only_cross_attention=None,
+ cross_attention_norm=None,
+ addition_embed_type_num_heads=64,
+ ).to(torch_dtype)
+
+ if conditioning_images_keys != [] or conditioning_masks_keys != []:
+
+ latents_concat_embedder_config = LatentsConcatEmbedderConfig(
+ image_keys=conditioning_images_keys,
+ mask_keys=conditioning_masks_keys,
+ )
+ latent_concat_embedder = LatentsConcatEmbedder(latents_concat_embedder_config)
+ latent_concat_embedder.freeze()
+ conditioners.append(latent_concat_embedder)
+
+ # Wrap conditioners and set to device
+ conditioner = ConditionerWrapper(
+ conditioners=conditioners,
+ )
+
+ ## VAE ##
+ # Get VAE model
+ vae_config = AutoencoderKLDiffusersConfig(
+ version=backbone_signature,
+ subfolder="vae",
+ tiling_size=(128, 128),
+ )
+ vae = AutoencoderKLDiffusers(vae_config).to(torch_dtype)
+ vae.freeze()
+ vae.to(torch_dtype)
+
+ ## Diffusion Model ##
+ # Get diffusion model
+ config = LBMConfig(
+ source_key=source_key,
+ target_key=target_key,
+ latent_loss_weight=latent_loss_weight,
+ latent_loss_type=latent_loss_type,
+ pixel_loss_type=pixel_loss_type,
+ pixel_loss_weight=pixel_loss_weight,
+ timestep_sampling=timestep_sampling,
+ logit_mean=logit_mean,
+ logit_std=logit_std,
+ selected_timesteps=selected_timesteps,
+ prob=prob,
+ bridge_noise_sigma=bridge_noise_sigma,
+ )
+
+ sampling_noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
+ backbone_signature,
+ subfolder="scheduler",
+ )
+
+ model = LBMModel(
+ config,
+ denoiser=denoiser,
+ sampling_noise_scheduler=sampling_noise_scheduler,
+ vae=vae,
+ conditioner=conditioner,
+ ).to(torch_dtype)
+
+ return model
diff --git a/src/lbm/models/__init__.py b/src/lbm/models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/src/lbm/models/base/__init__.py b/src/lbm/models/base/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..3850d04c3ba1ea3c19e9d7a6582a3541ae8ee199
--- /dev/null
+++ b/src/lbm/models/base/__init__.py
@@ -0,0 +1,4 @@
+from .base_model import BaseModel
+from .model_config import ModelConfig
+
+__all__ = ["BaseModel", "ModelConfig"]
diff --git a/src/lbm/models/base/base_model.py b/src/lbm/models/base/base_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..3302a653c3f127ea699e4cebe178e907615623a9
--- /dev/null
+++ b/src/lbm/models/base/base_model.py
@@ -0,0 +1,66 @@
+from typing import Any, Dict
+
+import torch
+import torch.nn as nn
+
+from .model_config import ModelConfig
+
+
+class BaseModel(nn.Module):
+ def __init__(self, config: ModelConfig):
+ nn.Module.__init__(self)
+ self.config = config
+ self.input_key = config.input_key
+ self.device = torch.device("cpu")
+ self.dtype = torch.float32
+
+ def on_fit_start(self, device: torch.device | None = None, *args, **kwargs):
+ """Called when the training starts
+
+ Args:
+ device (Optional[torch.device], optional): The device to use. Usefull to set
+ relevant parameters on the model and embedder to the right device only
+ once at the start of the training. Defaults to None.
+ """
+ if device is not None:
+ self.device = device
+ self.to(self.device)
+
+ def forward(self, batch: Dict[str, Any], *args, **kwargs):
+ raise NotImplementedError("forward method is not implemented")
+
+ def freeze(self):
+ """Freeze the model"""
+ self.eval()
+ for param in self.parameters():
+ param.requires_grad = False
+
+ def to(self, *args, **kwargs):
+ device, dtype, non_blocking, _ = torch._C._nn._parse_to(*args, **kwargs)
+ self = super().to(
+ device=device,
+ dtype=dtype,
+ non_blocking=non_blocking,
+ )
+
+ if device is not None:
+ self.device = device
+ if dtype is not None:
+ self.dtype = dtype
+ return self
+
+ def compute_metrics(self, batch: Dict[str, Any], *args, **kwargs):
+ """Compute the metrics"""
+ return {}
+
+ def sample(self, batch: Dict[str, Any], *args, **kwargs):
+ """Sample from the model"""
+ return {}
+
+ def log_samples(self, batch: Dict[str, Any], *args, **kwargs):
+ """Log the samples"""
+ return None
+
+ def on_train_batch_end(self, batch: Dict[str, Any], *args, **kwargs):
+ """Update the model an optimization is perforned on a batch."""
+ pass
diff --git a/src/lbm/models/base/model_config.py b/src/lbm/models/base/model_config.py
new file mode 100644
index 0000000000000000000000000000000000000000..0fd0d1b0c4f14d7a2ee57a8c1f6ea54807616893
--- /dev/null
+++ b/src/lbm/models/base/model_config.py
@@ -0,0 +1,8 @@
+from pydantic.dataclasses import dataclass
+
+from ...config import BaseConfig
+
+
+@dataclass
+class ModelConfig(BaseConfig):
+ input_key: str = "image"
diff --git a/src/lbm/models/embedders/__init__.py b/src/lbm/models/embedders/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..86813b2599b02e11c0fa49f087f077c192e3d2fc
--- /dev/null
+++ b/src/lbm/models/embedders/__init__.py
@@ -0,0 +1,4 @@
+from .conditioners_wrapper import ConditionerWrapper
+from .latents_concat import LatentsConcatEmbedder, LatentsConcatEmbedderConfig
+
+__all__ = ["LatentsConcatEmbedder", "LatentsConcatEmbedderConfig", "ConditionerWrapper"]
diff --git a/src/lbm/models/embedders/base/__init__.py b/src/lbm/models/embedders/base/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..6149ca48d14214b25884705424f50e6503344352
--- /dev/null
+++ b/src/lbm/models/embedders/base/__init__.py
@@ -0,0 +1,4 @@
+from .base_conditioner import BaseConditioner
+from .base_conditioner_config import BaseConditionerConfig
+
+__all__ = ["BaseConditioner", "BaseConditionerConfig"]
diff --git a/src/lbm/models/embedders/base/base_conditioner.py b/src/lbm/models/embedders/base/base_conditioner.py
new file mode 100644
index 0000000000000000000000000000000000000000..053a33b7488e2039cbd841cec5dd0b202b644f01
--- /dev/null
+++ b/src/lbm/models/embedders/base/base_conditioner.py
@@ -0,0 +1,60 @@
+from typing import Any, Dict, List, Optional, Union
+
+import torch
+
+from ...base.base_model import BaseModel
+from .base_conditioner_config import BaseConditionerConfig
+
+DIM2CONDITIONING = {
+ 2: "vector",
+ 3: "crossattn",
+ 4: "concat",
+}
+
+
+class BaseConditioner(BaseModel):
+ """This is the base class for all the conditioners. This absctacts the conditioning process
+
+ Args:
+
+ config (BaseConditionerConfig): The configuration of the conditioner
+
+ Examples
+ ########
+
+ To use the conditioner, you can import the class and use it as follows:
+
+ .. code-block:: python
+
+ from cr.models.embedders import BaseConditioner, BaseConditionerConfig
+
+ # Create the conditioner config
+ config = BaseConditionerConfig(
+ input_key="text", # The key for the input
+ unconditional_conditioning_rate=0.3, # Drops the conditioning with 30% probability during training
+ )
+
+ # Create the conditioner
+ conditioner = BaseConditioner(config)
+ """
+
+ def __init__(self, config: BaseConditionerConfig):
+ BaseModel.__init__(self, config)
+ self.config = config
+ self.input_key = config.input_key
+ self.dim2outputkey = DIM2CONDITIONING
+ self.ucg_rate = config.unconditional_conditioning_rate
+
+ def forward(
+ self, batch: Dict[str, Any], force_zero_embedding: bool = False, *args, **kwargs
+ ):
+ """
+ Forward pass of the embedder.
+
+ Args:
+
+ batch (Dict[str, Any]): A dictionary containing the input data.
+ force_zero_embedding (bool): Whether to force zero embedding.
+ This will return an embedding with all entries set to 0. Defaults to False.
+ """
+ raise NotImplementedError("Forward pass must be implemented in child class")
diff --git a/src/lbm/models/embedders/base/base_conditioner_config.py b/src/lbm/models/embedders/base/base_conditioner_config.py
new file mode 100644
index 0000000000000000000000000000000000000000..5f6eec24e2a6057dfa28c406e6862a815765d3be
--- /dev/null
+++ b/src/lbm/models/embedders/base/base_conditioner_config.py
@@ -0,0 +1,27 @@
+from typing import Literal
+
+from pydantic.dataclasses import dataclass
+
+from ....config import BaseConfig
+
+
+@dataclass
+class BaseConditionerConfig(BaseConfig):
+ """This is the ClipEmbedderConfig class which defines all the useful parameters to instantiate the model
+
+ Args:
+
+ input_key (str): The key for the input. Defaults to "text".
+ unconditional_conditioning_rate (float): Drops the conditioning with this probability during training. Defaults to 0.0.
+ """
+
+ input_key: str = "text"
+ unconditional_conditioning_rate: float = 0.0
+
+ def __post_init__(self):
+ super().__post_init__()
+
+ assert (
+ self.unconditional_conditioning_rate >= 0.0
+ and self.unconditional_conditioning_rate <= 1.0
+ ), "Unconditional conditioning rate should be between 0 and 1"
diff --git a/src/lbm/models/embedders/conditioners_wrapper.py b/src/lbm/models/embedders/conditioners_wrapper.py
new file mode 100644
index 0000000000000000000000000000000000000000..63d720d6ffe61d45cb3d528799d550af98fc5a96
--- /dev/null
+++ b/src/lbm/models/embedders/conditioners_wrapper.py
@@ -0,0 +1,114 @@
+import logging
+from typing import Any, Dict, List, Union
+
+import torch
+import torch.nn as nn
+
+from .base import BaseConditioner
+
+KEY2CATDIM = {
+ "vector": 1,
+ "crossattn": 2,
+ "concat": 1,
+}
+
+
+class ConditionerWrapper(nn.Module):
+ """
+ Wrapper for conditioners. This class allows to apply multiple conditioners in a single forward pass.
+
+ Args:
+
+ conditioners (List[BaseConditioner]): List of conditioners to apply in the forward pass.
+ """
+
+ def __init__(
+ self,
+ conditioners: Union[List[BaseConditioner], None] = None,
+ ):
+ nn.Module.__init__(self)
+ self.conditioners = nn.ModuleList(conditioners)
+ self.device = torch.device("cpu")
+ self.dtype = torch.float32
+
+ def conditioner_sanity_check(self):
+ cond_input_keys = []
+ for conditioner in self.conditioners:
+ cond_input_keys.append(conditioner.input_key)
+
+ assert all([key in set(cond_input_keys) for key in self.ucg_keys])
+
+ def on_fit_start(self, device: torch.device | None = None, *args, **kwargs):
+ """Called when the training starts"""
+ for conditioner in self.conditioners:
+ conditioner.on_fit_start(device=device, *args, **kwargs)
+
+ def forward(
+ self,
+ batch: Dict[str, Any],
+ ucg_keys: List[str] = None,
+ set_ucg_rate_zero=False,
+ *args,
+ **kwargs,
+ ):
+ """
+ Forward pass through all conditioners
+
+ Args:
+
+ batch: batch of data
+ ucg_keys: keys to use for ucg. This will force zero conditioning in all the
+ conditioners that have input_keys in ucg_keys
+ set_ucg_rate_zero: set the ucg rate to zero for all the conditioners except the ones in ucg_keys
+
+ Returns:
+
+ Dict[str, Any]: The output of the conditioner. The output of the conditioner is a dictionary with the main key "cond" and value
+ is a dictionary with the keys as the type of conditioning and the value as the conditioning tensor.
+ """
+ if ucg_keys is None:
+ ucg_keys = []
+ wrapper_outputs = dict(cond={})
+ for conditioner in self.conditioners:
+ if conditioner.input_key in ucg_keys:
+ force_zero_embedding = True
+ elif conditioner.ucg_rate > 0 and not set_ucg_rate_zero:
+ force_zero_embedding = bool(torch.rand(1) < conditioner.ucg_rate)
+ else:
+ force_zero_embedding = False
+
+ conditioner_output = conditioner.forward(
+ batch, force_zero_embedding=force_zero_embedding, *args, **kwargs
+ )
+ logging.debug(
+ f"conditioner:{conditioner.__class__.__name__}, input_key:{conditioner.input_key}, force_ucg_zero_embedding:{force_zero_embedding}"
+ )
+ for key in conditioner_output:
+ logging.debug(
+ f"conditioner_output:{key}:{conditioner_output[key].shape}"
+ )
+ if key in wrapper_outputs["cond"]:
+ wrapper_outputs["cond"][key] = torch.cat(
+ [wrapper_outputs["cond"][key], conditioner_output[key]],
+ KEY2CATDIM[key],
+ )
+ else:
+ wrapper_outputs["cond"][key] = conditioner_output[key]
+
+ return wrapper_outputs
+
+ def to(self, *args, **kwargs):
+ """
+ Move all conditioners to device and dtype
+ """
+ device, dtype, non_blocking, _ = torch._C._nn._parse_to(*args, **kwargs)
+ self = super().to(device=device, dtype=dtype, non_blocking=non_blocking)
+ for conditioner in self.conditioners:
+ conditioner.to(device=device, dtype=dtype, non_blocking=non_blocking)
+
+ if device is not None:
+ self.device = device
+ if dtype is not None:
+ self.dtype = dtype
+
+ return self
diff --git a/src/lbm/models/embedders/latents_concat/__init__.py b/src/lbm/models/embedders/latents_concat/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..edc22067821f6ed15d929cdca60c2e47bba2cb8f
--- /dev/null
+++ b/src/lbm/models/embedders/latents_concat/__init__.py
@@ -0,0 +1,4 @@
+from .latents_concat_embedder_config import LatentsConcatEmbedderConfig
+from .latents_concat_embedder_model import LatentsConcatEmbedder
+
+__all__ = ["LatentsConcatEmbedder", "LatentsConcatEmbedderConfig"]
diff --git a/src/lbm/models/embedders/latents_concat/latents_concat_embedder_config.py b/src/lbm/models/embedders/latents_concat/latents_concat_embedder_config.py
new file mode 100644
index 0000000000000000000000000000000000000000..7f8f0037357f38eded0ea6ef86378c12bf08d06b
--- /dev/null
+++ b/src/lbm/models/embedders/latents_concat/latents_concat_embedder_config.py
@@ -0,0 +1,31 @@
+from dataclasses import field
+from typing import List, Union
+
+from pydantic.dataclasses import dataclass
+
+from ..base import BaseConditionerConfig
+
+
+@dataclass
+class LatentsConcatEmbedderConfig(BaseConditionerConfig):
+ """
+ Configs for the LatentsConcatEmbedder embedder
+
+ Args:
+ image_keys (Union[List[str], None]): Keys of the images to compute the VAE embeddings
+ mask_keys (Union[List[str], None]): Keys of the masks to resize
+ """
+
+ image_keys: Union[List[str], None] = field(default_factory=lambda: ["image"])
+ mask_keys: Union[List[str], None] = field(default_factory=lambda: ["mask"])
+
+ def __post_init__(self):
+ super().__post_init__()
+
+ # Make sure that at least one of the image_keys or mask_keys is provided
+ assert (self.image_keys is not None) or (
+ self.mask_keys is not None
+ ), "At least one of the image_keys or mask_keys must be provided."
+
+ self.image_keys = self.image_keys if self.image_keys is not None else []
+ self.mask_keys = self.mask_keys if self.mask_keys is not None else []
diff --git a/src/lbm/models/embedders/latents_concat/latents_concat_embedder_model.py b/src/lbm/models/embedders/latents_concat/latents_concat_embedder_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..e91bd612e985e376ed35e18d231460be48d19f4a
--- /dev/null
+++ b/src/lbm/models/embedders/latents_concat/latents_concat_embedder_model.py
@@ -0,0 +1,80 @@
+from typing import Any, Dict
+
+import torch
+import torchvision.transforms.functional as F
+
+from lbm.models.vae import AutoencoderKLDiffusers
+
+from ..base import BaseConditioner
+from .latents_concat_embedder_config import LatentsConcatEmbedderConfig
+
+
+class LatentsConcatEmbedder(BaseConditioner):
+ """
+ Class computing VAE embeddings from given images and resizing the masks.
+ Then outputs are then concatenated to the noise in the latent space.
+
+ Args:
+ config (LatentsConcatEmbedderConfig): Configs to create the embedder
+ """
+
+ def __init__(self, config: LatentsConcatEmbedderConfig):
+ BaseConditioner.__init__(self, config)
+
+ def forward(
+ self, batch: Dict[str, Any], vae: AutoencoderKLDiffusers, *args, **kwargs
+ ) -> dict:
+ """
+ Args:
+ batch (dict): A batch of images to be processed by this embedder. In the batch,
+ the images must range between [-1, 1] and the masks range between [0, 1].
+ vae (AutoencoderKLDiffusers): VAE
+
+ Returns:
+ output (dict): outputs
+ """
+
+ # Check if image are of the same size
+ dims_list = []
+ for image_key in self.config.image_keys:
+ dims_list.append(batch[image_key].shape[-2:])
+ for mask_key in self.config.mask_keys:
+ dims_list.append(batch[mask_key].shape[-2:])
+ assert all(
+ dims == dims_list[0] for dims in dims_list
+ ), "All images and masks must have the same dimensions."
+
+ # Find the latent dimensions
+ if len(self.config.image_keys) > 0:
+ latent_dims = (
+ batch[self.config.image_keys[0]].shape[-2] // vae.downsampling_factor,
+ batch[self.config.image_keys[0]].shape[-1] // vae.downsampling_factor,
+ )
+ else:
+ latent_dims = (
+ batch[self.config.mask_keys[0]].shape[-2] // vae.downsampling_factor,
+ batch[self.config.mask_keys[0]].shape[-1] // vae.downsampling_factor,
+ )
+
+ outputs = []
+
+ # Resize the masks and concat them
+ for mask_key in self.config.mask_keys:
+ curr_latents = F.resize(
+ batch[mask_key],
+ size=latent_dims,
+ interpolation=F.InterpolationMode.BILINEAR,
+ )
+ outputs.append(curr_latents)
+
+ # Compute VAE embeddings from the images
+ for image_key in self.config.image_keys:
+ vae_embs = vae.encode(batch[image_key])
+ outputs.append(vae_embs)
+
+ # Concat all the outputs
+ outputs = torch.concat(outputs, dim=1)
+
+ outputs = {self.dim2outputkey[outputs.dim()]: outputs}
+
+ return outputs
diff --git a/src/lbm/models/lbm/__init__.py b/src/lbm/models/lbm/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..4b8e4d086c351e5d51ab4af159b81a4e41d65e16
--- /dev/null
+++ b/src/lbm/models/lbm/__init__.py
@@ -0,0 +1,4 @@
+from .lbm_config import LBMConfig
+from .lbm_model import LBMModel
+
+__all__ = ["LBMModel", "LBMConfig"]
diff --git a/src/lbm/models/lbm/lbm_config.py b/src/lbm/models/lbm/lbm_config.py
new file mode 100644
index 0000000000000000000000000000000000000000..5ff1886c97438af388319a666edb1e1091f781d4
--- /dev/null
+++ b/src/lbm/models/lbm/lbm_config.py
@@ -0,0 +1,101 @@
+from typing import List, Literal, Optional, Tuple
+
+from pydantic.dataclasses import dataclass
+
+from ..base import ModelConfig
+
+
+@dataclass
+class LBMConfig(ModelConfig):
+ """This is the Config for LBM Model class which defines all the useful parameters to be used in the model.
+
+ Args:
+
+ source_key (str):
+ Key for the source image. Defaults to "source_image"
+
+ target_key (str):
+ Key for the target image. Defaults to "target_image"
+
+ mask_key (Optional[str]):
+ Key for the mask showing the valid pixels. Defaults to None
+
+ latent_loss_type (str):
+ Loss type to use. Defaults to "l2". Choices are "l2", "l1"
+
+ pixel_loss_type (str):
+ Pixel loss type to use. Defaults to "l2". Choices are "l2", "l1", "lpips"
+
+ pixel_loss_max_size (int):
+ Maximum size of the image for pixel loss.
+ The image will be cropped to this size to reduce decoding computation cost. Defaults to 512
+
+ pixel_loss_weight (float):
+ Weight of the pixel loss. Defaults to 0.0
+
+ timestep_sampling (str):
+ Timestep sampling to use. Defaults to "uniform". Choices are "uniform"
+
+ input_key (str):
+ Key for the input. Defaults to "image"
+
+ controlnet_input_key (str):
+ Key for the controlnet conditioning. Defaults to "controlnet_conditioning"
+
+ adapter_input_key (str):
+ Key for the adapter conditioning. Defaults to "adapter_conditioning"
+
+ ucg_keys (Optional[List[str]]):
+ List of keys for which we enforce zero_conditioning during Classifier-free guidance. Defaults to None
+
+ prediction_type (str):
+ Type of prediction to use. Defaults to "epsilon". Choices are "epsilon", "v_prediction", "flow
+
+ logit_mean (Optional[float]):
+ Mean of the logit for the log normal distribution. Defaults to 0.0
+
+ logit_std (Optional[float]):
+ Standard deviation of the logit for the log normal distribution. Defaults to 1.0
+
+ guidance_scale (Optional[float]):
+ The guidance scale. Useful for finetunning guidance distilled diffusion models. Defaults to None
+
+ selected_timesteps (Optional[List[float]]):
+ List of selected timesteps to be sampled from if using `custom_timesteps` timestep sampling. Defaults to None
+
+ prob (Optional[List[float]]):
+ List of probabilities for the selected timesteps if using `custom_timesteps` timestep sampling. Defaults to None
+ """
+
+ source_key: str = "source_image"
+ target_key: str = "target_image"
+ mask_key: Optional[str] = None
+ latent_loss_weight: float = 1.0
+ latent_loss_type: Literal["l2", "l1"] = "l2"
+ pixel_loss_type: Literal["l2", "l1", "lpips"] = "l2"
+ pixel_loss_max_size: int = 512
+ pixel_loss_weight: float = 0.0
+ timestep_sampling: Literal["uniform", "log_normal", "custom_timesteps"] = "uniform"
+ logit_mean: Optional[float] = 0.0
+ logit_std: Optional[float] = 1.0
+ selected_timesteps: Optional[List[float]] = None
+ prob: Optional[List[float]] = None
+ bridge_noise_sigma: float = 0.001
+
+ def __post_init__(self):
+ super().__post_init__()
+ if self.timestep_sampling == "log_normal":
+ assert isinstance(self.logit_mean, float) and isinstance(
+ self.logit_std, float
+ ), "logit_mean and logit_std should be float for log_normal timestep sampling"
+
+ if self.timestep_sampling == "custom_timesteps":
+ assert isinstance(self.selected_timesteps, list) and isinstance(
+ self.prob, list
+ ), "timesteps and prob should be list for custom_timesteps timestep sampling"
+ assert len(self.selected_timesteps) == len(
+ self.prob
+ ), "timesteps and prob should be of same length for custom_timesteps timestep sampling"
+ assert (
+ sum(self.prob) == 1
+ ), "prob should sum to 1 for custom_timesteps timestep sampling"
diff --git a/src/lbm/models/lbm/lbm_model.py b/src/lbm/models/lbm/lbm_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..40f49603eb563e297ca7a1a14bbbb921dec747da
--- /dev/null
+++ b/src/lbm/models/lbm/lbm_model.py
@@ -0,0 +1,511 @@
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import lpips
+import numpy as np
+import torch
+import torch.nn as nn
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from tqdm import tqdm
+
+from ..base.base_model import BaseModel
+from ..embedders import ConditionerWrapper
+from ..unets import DiffusersUNet2DCondWrapper, DiffusersUNet2DWrapper
+from ..vae import AutoencoderKLDiffusers
+from .lbm_config import LBMConfig
+
+
+class LBMModel(BaseModel):
+ """This is the LBM class which defines the model.
+
+ Args:
+
+ config (LBMConfig):
+ Configuration for the model
+
+ denoiser (Union[DiffusersUNet2DWrapper, DiffusersTransformer2DWrapper]):
+ Denoiser to use for the diffusion model. Defaults to None
+
+ training_noise_scheduler (EulerDiscreteScheduler):
+ Noise scheduler to use for training. Defaults to None
+
+ sampling_noise_scheduler (EulerDiscreteScheduler):
+ Noise scheduler to use for sampling. Defaults to None
+
+ vae (AutoencoderKLDiffusers):
+ VAE to use for the diffusion model. Defaults to None
+
+ conditioner (ConditionerWrapper):
+ Conditioner to use for the diffusion model. Defaults to None
+ """
+
+ @classmethod
+ def load_from_config(cls, config: LBMConfig):
+ return cls(config=config)
+
+ def __init__(
+ self,
+ config: LBMConfig,
+ denoiser: Union[
+ DiffusersUNet2DWrapper,
+ DiffusersUNet2DCondWrapper,
+ ] = None,
+ training_noise_scheduler: FlowMatchEulerDiscreteScheduler = None,
+ sampling_noise_scheduler: FlowMatchEulerDiscreteScheduler = None,
+ vae: AutoencoderKLDiffusers = None,
+ conditioner: ConditionerWrapper = None,
+ ):
+ BaseModel.__init__(self, config)
+
+ self.vae = vae
+ self.denoiser = denoiser
+ self.conditioner = conditioner
+ self.sampling_noise_scheduler = sampling_noise_scheduler
+ self.training_noise_scheduler = training_noise_scheduler
+ self.timestep_sampling = config.timestep_sampling
+ self.latent_loss_type = config.latent_loss_type
+ self.latent_loss_weight = config.latent_loss_weight
+ self.pixel_loss_type = config.pixel_loss_type
+ self.pixel_loss_max_size = config.pixel_loss_max_size
+ self.pixel_loss_weight = config.pixel_loss_weight
+ self.logit_mean = config.logit_mean
+ self.logit_std = config.logit_std
+ self.prob = config.prob
+ self.selected_timesteps = config.selected_timesteps
+ self.source_key = config.source_key
+ self.target_key = config.target_key
+ self.mask_key = config.mask_key
+ self.bridge_noise_sigma = config.bridge_noise_sigma
+
+ self.num_iterations = nn.Parameter(
+ torch.tensor(0, dtype=torch.float32), requires_grad=False
+ )
+ if self.pixel_loss_type == "lpips" and self.pixel_loss_weight > 0:
+ self.lpips_loss = lpips.LPIPS(net="vgg")
+
+ else:
+ self.lpips_loss = None
+
+ def on_fit_start(self, device: torch.device | None = None, *args, **kwargs):
+ """Called when the training starts"""
+ super().on_fit_start(device=device, *args, **kwargs)
+ if self.vae is not None:
+ self.vae.on_fit_start(device=device, *args, **kwargs)
+ if self.conditioner is not None:
+ self.conditioner.on_fit_start(device=device, *args, **kwargs)
+
+ def forward(self, batch: Dict[str, Any], step=0, batch_idx=0, *args, **kwargs):
+
+ self.num_iterations += 1
+
+ # Get inputs/latents
+ if self.vae is not None:
+ vae_inputs = batch[self.target_key]
+ z = self.vae.encode(vae_inputs)
+ downsampling_factor = self.vae.downsampling_factor
+ else:
+ z = batch[self.target_key]
+ downsampling_factor = 1
+
+ if self.mask_key in batch:
+ valid_mask = batch[self.mask_key].bool()[:, 0, :, :].unsqueeze(1)
+ invalid_mask = ~valid_mask
+ valid_mask_for_latent = ~torch.max_pool2d(
+ invalid_mask.float(),
+ downsampling_factor,
+ downsampling_factor,
+ ).bool()
+ valid_mask_for_latent = valid_mask_for_latent.repeat((1, z.shape[1], 1, 1))
+
+ else:
+ valid_mask = torch.ones_like(batch[self.target_key]).bool()
+ valid_mask_for_latent = torch.ones_like(z).bool()
+
+ source_image = batch[self.source_key]
+ source_image = torch.nn.functional.interpolate(
+ source_image,
+ size=batch[self.target_key].shape[-2:],
+ mode="bilinear",
+ align_corners=False,
+ ).to(z.dtype)
+ if self.vae is not None:
+ z_source = self.vae.encode(source_image)
+
+ else:
+ z_source = source_image
+
+ # Get conditionings
+ conditioning = self._get_conditioning(batch, *args, **kwargs)
+
+ # Sample a timestep
+ timestep = self._timestep_sampling(n_samples=z.shape[0], device=z.device)
+ sigmas = None
+
+ # Create interpolant
+ sigmas = self._get_sigmas(
+ self.training_noise_scheduler, timestep, n_dim=4, device=z.device
+ )
+ noisy_sample = (
+ sigmas * z_source
+ + (1.0 - sigmas) * z
+ + self.bridge_noise_sigma
+ * (sigmas * (1.0 - sigmas)) ** 0.5
+ * torch.randn_like(z)
+ )
+
+ for i, t in enumerate(timestep):
+ if t.item() == self.training_noise_scheduler.timesteps[0]:
+ noisy_sample[i] = z_source[i]
+
+ # Predict noise level using denoiser
+ prediction = self.denoiser(
+ sample=noisy_sample,
+ timestep=timestep,
+ conditioning=conditioning,
+ *args,
+ **kwargs,
+ )
+
+ target = z_source - z
+ denoised_sample = noisy_sample - prediction * sigmas
+ target_pixels = batch[self.target_key]
+
+ # Compute loss
+ if self.latent_loss_weight > 0:
+ loss = self.latent_loss(prediction, target.detach(), valid_mask_for_latent)
+ latent_recon_loss = loss.mean()
+
+ else:
+ loss = torch.zeros(z.shape[0], device=z.device)
+ latent_recon_loss = torch.zeros_like(loss)
+
+ if self.pixel_loss_weight > 0:
+ denoised_sample = self._predicted_x_0(
+ model_output=prediction,
+ sample=noisy_sample,
+ sigmas=sigmas,
+ )
+ pixel_loss = self.pixel_loss(
+ denoised_sample, target_pixels.detach(), valid_mask
+ )
+ loss += self.pixel_loss_weight * pixel_loss
+
+ else:
+ pixel_loss = torch.zeros_like(latent_recon_loss)
+
+ return {
+ "loss": loss.mean(),
+ "latent_recon_loss": latent_recon_loss,
+ "pixel_recon_loss": pixel_loss.mean(),
+ "predicted_hr": denoised_sample,
+ "noisy_sample": noisy_sample,
+ }
+
+ def latent_loss(self, prediction, model_input, valid_latent_mask):
+ if self.latent_loss_type == "l2":
+ return torch.mean(
+ (
+ (prediction * valid_latent_mask - model_input * valid_latent_mask)
+ ** 2
+ ).reshape(model_input.shape[0], -1),
+ 1,
+ )
+ elif self.latent_loss_type == "l1":
+ return torch.mean(
+ torch.abs(
+ prediction * valid_latent_mask - model_input * valid_latent_mask
+ ).reshape(model_input.shape[0], -1),
+ 1,
+ )
+ else:
+ raise NotImplementedError(
+ f"Loss type {self.latent_loss_type} not implemented"
+ )
+
+ def pixel_loss(self, prediction, model_input, valid_mask):
+
+ latent_crop = self.pixel_loss_max_size // self.vae.downsampling_factor
+ input_crop = self.pixel_loss_max_size
+
+ crop_h = max((prediction.shape[2] - latent_crop), 0)
+ crop_w = max((prediction.shape[3] - latent_crop), 0)
+
+ input_crop_h = max((model_input.shape[2] - self.pixel_loss_max_size), 0)
+ input_crop_w = max((model_input.shape[3] - self.pixel_loss_max_size), 0)
+
+ # image random cropping
+ if crop_h == 0:
+ offset_h = 0
+ else:
+ offset_h = torch.randint(0, crop_h, (1,)).item()
+
+ if crop_w == 0:
+ offset_w = 0
+ else:
+ offset_w = torch.randint(0, crop_w, (1,)).item()
+ input_offset_h = offset_h * self.vae.downsampling_factor
+ input_offset_w = offset_w * self.vae.downsampling_factor
+
+ prediction = prediction[
+ :,
+ :,
+ crop_h
+ - offset_h : min(crop_h - offset_h + latent_crop, prediction.shape[2]),
+ crop_w
+ - offset_w : min(crop_w - offset_w + latent_crop, prediction.shape[3]),
+ ]
+
+ model_input = model_input[
+ :,
+ :,
+ input_crop_h
+ - input_offset_h : min(
+ input_crop_h - input_offset_h + input_crop, model_input.shape[2]
+ ),
+ input_crop_w
+ - input_offset_w : min(
+ input_crop_w - input_offset_w + input_crop, model_input.shape[3]
+ ),
+ ]
+
+ valid_mask = valid_mask[
+ :,
+ :,
+ input_crop_h
+ - input_offset_h : min(
+ input_crop_h - input_offset_h + input_crop, valid_mask.shape[2]
+ ),
+ input_crop_w
+ - input_offset_w : min(
+ input_crop_w - input_offset_w + input_crop, valid_mask.shape[3]
+ ),
+ ]
+
+ decoded_prediction = self.vae.decode(prediction).clamp(-1, 1)
+
+ if self.pixel_loss_type == "l2":
+ return torch.mean(
+ (
+ (decoded_prediction * valid_mask - model_input * valid_mask) ** 2
+ ).reshape(model_input.shape[0], -1),
+ 1,
+ )
+
+ elif self.pixel_loss_type == "l1":
+ return torch.mean(
+ torch.abs(
+ decoded_prediction * valid_mask - model_input * valid_mask
+ ).reshape(model_input.shape[0], -1),
+ 1,
+ )
+
+ elif self.pixel_loss_type == "lpips":
+ return self.lpips_loss(
+ decoded_prediction * valid_mask, model_input * valid_mask
+ ).mean()
+
+ def _get_conditioning(
+ self,
+ batch: Dict[str, Any],
+ ucg_keys: List[str] = None,
+ set_ucg_rate_zero=False,
+ *args,
+ **kwargs,
+ ):
+ """
+ Get the conditionings
+ """
+ if self.conditioner is not None:
+ return self.conditioner(
+ batch,
+ ucg_keys=ucg_keys,
+ set_ucg_rate_zero=set_ucg_rate_zero,
+ vae=self.vae,
+ *args,
+ **kwargs,
+ )
+ else:
+ return None
+
+ def _timestep_sampling(self, n_samples=1, device="cpu"):
+ if self.timestep_sampling == "uniform":
+ idx = torch.randint(
+ 0,
+ self.training_noise_scheduler.config.num_train_timesteps,
+ (n_samples,),
+ device="cpu",
+ )
+ return self.training_noise_scheduler.timesteps[idx].to(device=device)
+
+ elif self.timestep_sampling == "log_normal":
+ u = torch.normal(
+ mean=self.logit_mean,
+ std=self.logit_std,
+ size=(n_samples,),
+ device="cpu",
+ )
+ u = torch.nn.functional.sigmoid(u)
+ indices = (
+ u * self.training_noise_scheduler.config.num_train_timesteps
+ ).long()
+ return self.training_noise_scheduler.timesteps[indices].to(device=device)
+
+ elif self.timestep_sampling == "custom_timesteps":
+ idx = np.random.choice(len(self.selected_timesteps), n_samples, p=self.prob)
+
+ return torch.tensor(
+ self.selected_timesteps, device=device, dtype=torch.long
+ )[idx]
+
+ def _predicted_x_0(
+ self,
+ model_output,
+ sample,
+ sigmas=None,
+ ):
+ """
+ Predict x_0, the orinal denoised sample, using the model output and the timesteps depending on the prediction type.
+ """
+ pred_x_0 = sample - model_output * sigmas
+ return pred_x_0
+
+ def _get_sigmas(
+ self, scheduler, timesteps, n_dim=4, dtype=torch.float32, device="cpu"
+ ):
+ sigmas = scheduler.sigmas.to(device=device, dtype=dtype)
+ schedule_timesteps = scheduler.timesteps.to(device)
+ timesteps = timesteps.to(device)
+ step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < n_dim:
+ sigma = sigma.unsqueeze(-1)
+ return sigma
+
+ @torch.no_grad()
+ def sample(
+ self,
+ z: torch.Tensor,
+ num_steps: int = 20,
+ conditioner_inputs: Optional[Dict[str, Any]] = None,
+ max_samples: Optional[int] = None,
+ verbose: bool = False,
+ ):
+ self.sampling_noise_scheduler.set_timesteps(
+ sigmas=np.linspace(1, 1 / num_steps, num_steps)
+ )
+
+ sample = z
+
+ # Get conditioning
+ conditioning = self._get_conditioning(
+ conditioner_inputs, set_ucg_rate_zero=True, device=z.device
+ )
+
+ # If max_samples parameter is provided, limit the number of samples
+ if max_samples is not None:
+ sample = sample[:max_samples]
+
+ if conditioning:
+ conditioning["cond"] = {
+ k: v[:max_samples] for k, v in conditioning["cond"].items()
+ }
+
+ for i, t in tqdm(
+ enumerate(self.sampling_noise_scheduler.timesteps), disable=not verbose
+ ):
+ if hasattr(self.sampling_noise_scheduler, "scale_model_input"):
+ denoiser_input = self.sampling_noise_scheduler.scale_model_input(
+ sample, t
+ )
+
+ else:
+ denoiser_input = sample
+
+ # Predict noise level using denoiser using conditionings
+ pred = self.denoiser(
+ sample=denoiser_input,
+ timestep=t.to(z.device).repeat(denoiser_input.shape[0]),
+ conditioning=conditioning,
+ )
+
+ # Make one step on the reverse diffusion process
+ sample = self.sampling_noise_scheduler.step(
+ pred, t, sample, return_dict=False
+ )[0]
+ if i < len(self.sampling_noise_scheduler.timesteps) - 1:
+ timestep = (
+ self.sampling_noise_scheduler.timesteps[i + 1]
+ .to(z.device)
+ .repeat(sample.shape[0])
+ )
+ sigmas = self._get_sigmas(
+ self.sampling_noise_scheduler, timestep, n_dim=4, device=z.device
+ )
+ sample = sample + self.bridge_noise_sigma * (
+ sigmas * (1.0 - sigmas)
+ ) ** 0.5 * torch.randn_like(sample)
+ sample = sample.to(z.dtype)
+
+ if self.vae is not None:
+ decoded_sample = self.vae.decode(sample)
+
+ else:
+ decoded_sample = sample
+
+ return decoded_sample
+
+ def log_samples(
+ self,
+ batch: Dict[str, Any],
+ input_shape: Optional[Tuple[int, int, int]] = None,
+ max_samples: Optional[int] = None,
+ num_steps: Union[int, List[int]] = 20,
+ ):
+ if isinstance(num_steps, int):
+ num_steps = [num_steps]
+
+ logs = {}
+
+ N = max_samples if max_samples is not None else len(batch[self.source_key])
+
+ batch = {k: v[:N] for k, v in batch.items()}
+
+ # infer input shape based on VAE configuration if not passed
+ if input_shape is None:
+ if self.vae is not None:
+ # get input pixel size of the vae
+ input_shape = batch[self.target_key].shape[2:]
+ # rescale to latent size
+ input_shape = (
+ self.vae.latent_channels,
+ input_shape[0] // self.vae.downsampling_factor,
+ input_shape[1] // self.vae.downsampling_factor,
+ )
+ else:
+ raise ValueError(
+ "input_shape must be passed when no VAE is used in the model"
+ )
+
+ for num_step in num_steps:
+ source_image = batch[self.source_key]
+ source_image = torch.nn.functional.interpolate(
+ source_image,
+ size=batch[self.target_key].shape[2:],
+ mode="bilinear",
+ align_corners=False,
+ ).to(dtype=self.dtype)
+ if self.vae is not None:
+ z = self.vae.encode(source_image)
+
+ else:
+ z = source_image
+
+ with torch.autocast(dtype=self.dtype, device_type="cuda"):
+ logs[f"samples_{num_step}_steps"] = self.sample(
+ z,
+ num_steps=num_step,
+ conditioner_inputs=batch,
+ max_samples=N,
+ )
+
+ return logs
diff --git a/src/lbm/models/unets/__init__.py b/src/lbm/models/unets/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..380af60cb493c08f8f0d5915c13f50cb0c1a6d7b
--- /dev/null
+++ b/src/lbm/models/unets/__init__.py
@@ -0,0 +1,14 @@
+"""
+This module contains a collection of U-Net models.
+The :mod:`cr.models.unets` module includes the following classes:
+
+- :class:`DiffusersUNet2DWrapper`: A 2D U-Net model for diffusers.
+- :class:`DiffusersUNet2DCondWrapper`: A 2D U-Net model for diffusers with conditional input.
+"""
+
+from .unet import DiffusersUNet2DCondWrapper, DiffusersUNet2DWrapper
+
+__all__ = [
+ "DiffusersUNet2DWrapper",
+ "DiffusersUNet2DCondWrapper",
+]
diff --git a/src/lbm/models/unets/unet.py b/src/lbm/models/unets/unet.py
new file mode 100644
index 0000000000000000000000000000000000000000..1b4b7c261da390fc6cb628ce7f9c12981ea9cab4
--- /dev/null
+++ b/src/lbm/models/unets/unet.py
@@ -0,0 +1,148 @@
+from typing import Dict, List, Optional, Union
+
+import torch
+from diffusers.models import UNet2DConditionModel, UNet2DModel
+
+
+class DiffusersUNet2DWrapper(UNet2DModel):
+ """
+ Wrapper for the UNet2DModel from diffusers
+
+ See diffusers' UNet2DModel for more details
+ """
+
+ def __init__(self, *args, **kwargs):
+ UNet2DModel.__init__(self, *args, **kwargs)
+
+ def forward(
+ self,
+ sample: torch.Tensor,
+ timestep: Union[torch.Tensor, float, int],
+ conditioning: Dict[str, torch.Tensor] = None,
+ *args,
+ **kwargs,
+ ):
+ """
+ The forward pass of the model
+
+ Args:
+
+ sample (torch.Tensor): The input sample
+ timesteps (Union[torch.Tensor, float, int]): The number of timesteps
+ """
+ if conditioning is not None:
+ class_labels = conditioning["cond"].get("vector", None)
+ concat = conditioning["cond"].get("concat", None)
+
+ else:
+ class_labels = None
+ concat = None
+
+ if concat is not None:
+ sample = torch.cat([sample, concat], dim=1)
+
+ return super().forward(sample, timestep, class_labels).sample
+
+ def freeze(self):
+ """
+ Freeze the model
+ """
+ self.eval()
+ for param in self.parameters():
+ param.requires_grad = False
+
+
+class DiffusersUNet2DCondWrapper(UNet2DConditionModel):
+ """
+ Wrapper for the UNet2DConditionModel from diffusers
+
+ See diffusers' Unet2DConditionModel for more details
+ """
+
+ def __init__(self, *args, **kwargs):
+ UNet2DConditionModel.__init__(self, *args, **kwargs)
+ # BaseModel.__init__(self, config=ModelConfig())
+
+ def forward(
+ self,
+ sample: torch.Tensor,
+ timestep: Union[torch.Tensor, float, int],
+ conditioning: Dict[str, torch.Tensor],
+ ip_adapter_cond_embedding: Optional[List[torch.Tensor]] = None,
+ down_block_additional_residuals: torch.Tensor = None,
+ mid_block_additional_residual: torch.Tensor = None,
+ down_intrablock_additional_residuals: torch.Tensor = None,
+ *args,
+ **kwargs,
+ ):
+ """
+ The forward pass of the model
+
+ Args:
+
+ sample (torch.Tensor): The input sample
+ timesteps (Union[torch.Tensor, float, int]): The number of timesteps
+ conditioning (Dict[str, torch.Tensor]): The conditioning data
+ down_block_additional_residuals (List[torch.Tensor]): Residuals for the down blocks.
+ These residuals typically are used for the controlnet.
+ mid_block_additional_residual (List[torch.Tensor]): Residuals for the mid blocks.
+ These residuals typically are used for the controlnet.
+ down_intrablock_additional_residuals (List[torch.Tensor]): Residuals for the down intrablocks.
+ These residuals typically are used for the T2I adapters.middle block outputs. Defaults to False
+ """
+
+ assert isinstance(conditioning, dict), "conditionings must be a dictionary"
+ # assert "crossattn" in conditioning["cond"], "crossattn must be in conditionings"
+
+ class_labels = conditioning["cond"].get("vector", None)
+ crossattn = conditioning["cond"].get("crossattn", None)
+ concat = conditioning["cond"].get("concat", None)
+
+ # concat conditioning
+ if concat is not None:
+ sample = torch.cat([sample, concat], dim=1)
+
+ # down_intrablock_additional_residuals needs to be cloned, since unet will modify it
+ if down_intrablock_additional_residuals is not None:
+ down_intrablock_additional_residuals_clone = [
+ curr_residuals.clone()
+ for curr_residuals in down_intrablock_additional_residuals
+ ]
+ else:
+ down_intrablock_additional_residuals_clone = None
+
+ # Check diffusers.models.embeddings.py > MultiIPAdapterImageProjectionLayer > forward() for implementation
+ # Exepected format : List[torch.Tensor] of shape (batch_size, num_image_embeds, embed_dim)
+ # with length = number of ip_adapters loaded in the ip_adapter_wrapper
+ if ip_adapter_cond_embedding is not None:
+ added_cond_kwargs = {
+ "image_embeds": [
+ ip_adapter_embedding.unsqueeze(1)
+ for ip_adapter_embedding in ip_adapter_cond_embedding
+ ]
+ }
+ else:
+ added_cond_kwargs = None
+
+ return (
+ super()
+ .forward(
+ sample=sample,
+ timestep=timestep,
+ encoder_hidden_states=crossattn,
+ class_labels=class_labels,
+ added_cond_kwargs=added_cond_kwargs,
+ down_block_additional_residuals=down_block_additional_residuals,
+ mid_block_additional_residual=mid_block_additional_residual,
+ down_intrablock_additional_residuals=down_intrablock_additional_residuals_clone,
+ )
+ .sample
+ )
+
+ def freeze(self):
+ """
+ Freeze the model
+ """
+ self.eval()
+ for param in self.parameters():
+ param.requires_grad = False
diff --git a/src/lbm/models/utils.py b/src/lbm/models/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..adb59fbd24566f6f5cc7f68165a360776c6a8370
--- /dev/null
+++ b/src/lbm/models/utils.py
@@ -0,0 +1,377 @@
+import logging
+import math
+from copy import deepcopy
+from typing import List, Tuple
+
+import torch
+import torch.nn.functional as F
+
+TILING_METHODS = ["average", "gaussian", "linear"]
+
+
+class Tiler:
+ def get_tiles(
+ self,
+ input: torch.Tensor,
+ tile_size: tuple,
+ overlap_size: tuple,
+ scale: int = 1,
+ out_channels: int = 3,
+ ) -> List[List[torch.tensor]]:
+ """Get tiles
+ Args:
+ input (torch.Tensor): input array of shape (batch_size, channels, height, width)
+ tile_size (tuple): tile size
+ overlap_size (tuple): overlap size
+ scale (int): scaling factor of the output wrt input
+ out_channels (int): number of output channels
+ Returns:
+ List[List[torch.Tensor]]: List of tiles
+ """
+ # assert isinstance(scale, int)
+ assert (
+ overlap_size[0] <= tile_size[0]
+ ), f"Overlap size {overlap_size} must be smaller than tile size {tile_size}"
+ assert (
+ overlap_size[1] <= tile_size[1]
+ ), f"Overlap size {overlap_size} must be smaller than tile size {tile_size}"
+
+ B, C, H, W = input.shape
+ tile_size_H, tile_size_W = tile_size
+
+ # sets overlap to 0 if the input is smaller than the tile size (i.e. no overlap)
+ overlap_H, overlap_W = (
+ overlap_size[0] if H > tile_size_H else 0,
+ overlap_size[1] if W > tile_size_W else 0,
+ )
+
+ self.output_overlap_size = (
+ int(overlap_H * scale),
+ int(overlap_W * scale),
+ )
+ self.tile_size = tile_size
+ self.output_tile_size = (
+ int(tile_size_H * scale),
+ int(tile_size_W * scale),
+ )
+ self.output_shape = (
+ B,
+ out_channels,
+ int(H * scale),
+ int(W * scale),
+ )
+ tiles = []
+ logging.debug(f"(Tiler) Input shape: {(B, C, H, W)}")
+ logging.debug(f"(Tiler) Output shape: {self.output_shape}")
+ logging.debug(f"(Tiler) Tile size: {(tile_size_H, tile_size_W)}")
+ logging.debug(f"(Tiler) Overlap size: {(overlap_H, overlap_W)}")
+ # loop over all tiles in the image with overlap
+ for i in range(0, H, tile_size_H - overlap_H):
+ row = []
+ for j in range(0, W, tile_size_W - overlap_W):
+ tile = deepcopy(
+ input[
+ :,
+ :,
+ i : i + tile_size_H,
+ j : j + tile_size_W,
+ ]
+ )
+ row.append(tile)
+ tiles.append(row)
+ return tiles
+
+ def merge_tiles(
+ self, tiles: List[List[torch.tensor]], tiling_method: str = "gaussian"
+ ) -> torch.tensor:
+ """Merge tiles by averaging the overlaping regions
+ Args:
+ tiles (Dict[str, Tile]): dictionary of processed tiles
+ tiling_method (str): tiling method. Can be "average", "gaussian" or "linear"
+ Returns:
+ torch.tensor: output image
+ """
+ if tiling_method == "average":
+ return self._average_merge_tiles(tiles)
+ elif tiling_method == "gaussian":
+ return self._gaussian_merge_tiles(tiles)
+ elif tiling_method == "linear":
+ return self._linear_merge_tiles(tiles)
+ else:
+ raise ValueError(
+ f"Unknown tiling method {tiling_method}. Available methods are {TILING_METHODS}"
+ )
+
+ def _average_merge_tiles(self, tiles: List[List[torch.tensor]]) -> torch.tensor:
+ """Merge tiles by averaging the overlaping regions
+ Args:
+ tiles (Dict[str, Tile]): dictionary of processed tiles
+ Returns:
+ torch.tensor: output image
+ """
+
+ output = torch.zeros(self.output_shape)
+
+ # weights to store multiplicity
+ weights = torch.zeros(self.output_shape)
+
+ _, _, output_H, output_W = self.output_shape
+ output_overlap_size_H, output_overlap_size_W = self.output_overlap_size
+ output_tile_size_H, output_tile_size_W = self.output_tile_size
+
+ for id_i, i in enumerate(
+ range(
+ 0,
+ output_H,
+ output_tile_size_H - output_overlap_size_H,
+ )
+ ):
+ for id_j, j in enumerate(
+ range(
+ 0,
+ output_W,
+ output_tile_size_W - output_overlap_size_W,
+ )
+ ):
+ output[
+ :,
+ :,
+ i : i + output_tile_size_H,
+ j : j + output_tile_size_W,
+ ] += (
+ tiles[id_i][id_j] * 1
+ )
+ weights[
+ :,
+ :,
+ i : i + output_tile_size_H,
+ j : j + output_tile_size_W,
+ ] += 1
+
+ # outputs is summed up with this multiplicity
+ # so we need to divide by the weights wich is either 1, 2 or 4 depending on the region
+ output = output / weights
+ return output
+
+ def _gaussian_weights(
+ self, tile_width: int, tile_height: int, nbatches: int, channels: int
+ ):
+ """Generates a gaussian mask of weights for tile contributions.
+
+ Args:
+ tile_width (int): width of the tile
+ tile_height (int): height of the tile
+ nbatches (int): number of batches
+ channels (int): number of channels
+ Returns:
+ torch.tensor: weights
+ """
+ import numpy as np
+ from numpy import exp, pi, sqrt
+
+ latent_width = tile_width
+ latent_height = tile_height
+
+ var = 0.01
+ midpoint = (
+ latent_width - 1
+ ) / 2 # -1 because index goes from 0 to latent_width - 1
+ x_probs = [
+ exp(
+ -(x - midpoint)
+ * (x - midpoint)
+ / (latent_width * latent_width)
+ / (2 * var)
+ )
+ / sqrt(2 * pi * var)
+ for x in range(latent_width)
+ ]
+ midpoint = latent_height / 2
+ y_probs = [
+ exp(
+ -(y - midpoint)
+ * (y - midpoint)
+ / (latent_height * latent_height)
+ / (2 * var)
+ )
+ / sqrt(2 * pi * var)
+ for y in range(latent_height)
+ ]
+
+ weights = np.outer(y_probs, x_probs)
+ return torch.tile(
+ torch.tensor(weights, device="cpu"), (nbatches, channels, 1, 1)
+ )
+
+ def _gaussian_merge_tiles(self, tiles: List[List[torch.tensor]]) -> torch.tensor:
+ """Merge tiles by averaging the overlaping regions
+ Args:
+ List[List[torch.tensor]]: List of processed tiles
+ Returns:
+ torch.tensor: output image
+ """
+ B, output_C, output_H, output_W = self.output_shape
+ output_overlap_size_H, output_overlap_size_W = self.output_overlap_size
+ output_tile_size_H, output_tile_size_W = self.output_tile_size
+
+ output = torch.zeros(self.output_shape)
+ # weights to store multiplicity
+ weights = torch.zeros(self.output_shape)
+
+ for id_i, i in enumerate(
+ range(
+ 0,
+ output_H,
+ output_tile_size_H - output_overlap_size_H,
+ )
+ ):
+ for id_j, j in enumerate(
+ range(
+ 0,
+ output_W,
+ output_tile_size_W - output_overlap_size_W,
+ )
+ ):
+ w = self._gaussian_weights(
+ tiles[id_i][id_j].shape[3],
+ tiles[id_i][id_j].shape[2],
+ B,
+ output_C,
+ )
+ output[
+ :,
+ :,
+ i : i + output_tile_size_H,
+ j : j + output_tile_size_W,
+ ] += (
+ tiles[id_i][id_j] * w
+ )
+ weights[
+ :,
+ :,
+ i : i + output_tile_size_H,
+ j : j + output_tile_size_W,
+ ] += w
+
+ # outputs is summed up with this multiplicity
+ output = output / weights
+ return output
+
+ def _blend_v(
+ self, a: torch.Tensor, b: torch.Tensor, blend_extent: int
+ ) -> torch.Tensor:
+ blend_extent = min(a.shape[2], b.shape[2], blend_extent)
+ for y in range(blend_extent):
+ b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[
+ :, :, y, :
+ ] * (y / blend_extent)
+ return b
+
+ def _blend_h(
+ self, a: torch.Tensor, b: torch.Tensor, blend_extent: int
+ ) -> torch.Tensor:
+ blend_extent = min(a.shape[3], b.shape[3], blend_extent)
+ for x in range(blend_extent):
+ b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[
+ :, :, :, x
+ ] * (x / blend_extent)
+ return b
+
+ def _linear_merge_tiles(self, tiles: List[List[torch.tensor]]) -> torch.Tensor:
+ """Merge tiles by blending the overlaping regions
+ Args:
+ tiles (List[List[torch.tensor]]): List of processed tiles
+ Returns:
+ torch.Tensor: output image
+ """
+ output_overlap_size_H, output_overlap_size_W = self.output_overlap_size
+ output_tile_size_H, output_tile_size_W = self.output_tile_size
+
+ res_rows = []
+ tiles_copy = deepcopy(tiles)
+
+ # Cut the right and bottom overlap region
+ limit_i = output_tile_size_H - output_overlap_size_H
+ limit_j = output_tile_size_W - output_overlap_size_W
+ for i, tile_row in enumerate(tiles_copy):
+ res_row = []
+ for j, tile in enumerate(tile_row):
+ tile_val = tile
+ if j > 0:
+ tile_val = self._blend_h(
+ tile_row[j - 1], tile, output_overlap_size_W
+ )
+ tiles_copy[i][j] = tile_val
+ if i > 0:
+ tile_val = self._blend_v(
+ tiles_copy[i - 1][j], tile_val, output_overlap_size_H
+ )
+ tiles_copy[i][j] = tile_val
+ res_row.append(tile_val[:, :, :limit_i, :limit_j])
+ res_rows.append(torch.cat(res_row, dim=3))
+ output = torch.cat(res_rows, dim=2)
+ return output
+
+
+def extract_into_tensor(
+ a: torch.Tensor, t: torch.Tensor, x_shape: Tuple[int, ...]
+) -> torch.Tensor:
+ """
+ Extracts values from a tensor into a new tensor using indices from another tensor.
+
+ :param a: the tensor to extract values from.
+ :param t: the tensor containing the indices.
+ :param x_shape: the shape of the tensor to extract values into.
+ :return: a new tensor containing the extracted values.
+ """
+
+ b, *_ = t.shape
+ out = a.gather(-1, t)
+ return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def pad(x: torch.Tensor, base_h: int, base_w: int) -> torch.Tensor:
+ """
+ Pads a tensor to the nearest multiple of base_h and base_w.
+
+ :param x: the tensor to pad.
+ :param base_h: the base height.
+ :param base_w: the base width.
+ :return: the padded tensor.
+ """
+ h, w = x.shape[-2:]
+ h_ = math.ceil(h / base_h) * base_h
+ w_ = math.ceil(w / base_w) * base_w
+ if w_ != w:
+ x = F.pad(x, (0, abs(w_ - w), 0, 0))
+ if h_ != h:
+ x = F.pad(x, (0, 0, 0, abs(h_ - h)))
+ return x
+
+
+def append_dims(x: torch.Tensor, target_dims: int) -> torch.Tensor:
+ """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
+ dims_to_append = target_dims - x.ndim
+ if dims_to_append < 0:
+ raise ValueError(
+ f"input has {x.ndim} dims but target_dims is {target_dims}, which is less"
+ )
+ return x[(...,) + (None,) * dims_to_append]
+
+
+@torch.no_grad()
+def update_ema(
+ target_params: List[torch.Tensor],
+ source_params: List[torch.Tensor],
+ rate: float = 0.99,
+):
+ """
+ Update target parameters to be closer to those of source parameters using
+ an exponential moving average.
+
+ :param target_params: the target parameter sequence.
+ :param source_params: the source parameter sequence.
+ :param rate: the EMA rate (closer to 1 means slower).
+ """
+ for targ, src in zip(target_params, source_params):
+ targ.detach().mul_(rate).add_(src, alpha=1 - rate)
diff --git a/src/lbm/models/vae/__init__.py b/src/lbm/models/vae/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..9ad0c4b0c5eb9041b7ea15dea19a7e937a43daff
--- /dev/null
+++ b/src/lbm/models/vae/__init__.py
@@ -0,0 +1,4 @@
+from .autoencoderKL import AutoencoderKLDiffusers
+from .autoencoderKL_config import AutoencoderKLDiffusersConfig
+
+__all__ = ["AutoencoderKLDiffusers", "AutoencoderKLDiffusersConfig"]
diff --git a/src/lbm/models/vae/autoencoderKL.py b/src/lbm/models/vae/autoencoderKL.py
new file mode 100644
index 0000000000000000000000000000000000000000..ed0239f1758d6c10648cc9c03fd7edd668c7de3d
--- /dev/null
+++ b/src/lbm/models/vae/autoencoderKL.py
@@ -0,0 +1,136 @@
+import torch
+from diffusers.models import AutoencoderKL
+
+from ..base.base_model import BaseModel
+from ..utils import Tiler, pad
+from .autoencoderKL_config import AutoencoderKLDiffusersConfig
+
+
+class AutoencoderKLDiffusers(BaseModel):
+ """This is the VAE class used to work with latent models
+
+ Args:
+
+ config (AutoencoderKLDiffusersConfig): The config class which defines all the required parameters.
+ """
+
+ def __init__(self, config: AutoencoderKLDiffusersConfig):
+ BaseModel.__init__(self, config)
+ self.config = config
+ self.vae_model = AutoencoderKL.from_pretrained(
+ config.version,
+ subfolder=config.subfolder,
+ revision=config.revision,
+ )
+ self.tiling_size = config.tiling_size
+ self.tiling_overlap = config.tiling_overlap
+
+ # get downsampling factor
+ self._get_properties()
+
+ @torch.no_grad()
+ def _get_properties(self):
+ self.has_shift_factor = (
+ hasattr(self.vae_model.config, "shift_factor")
+ and self.vae_model.config.shift_factor is not None
+ )
+ self.shift_factor = (
+ self.vae_model.config.shift_factor if self.has_shift_factor else 0
+ )
+
+ # set latent channels
+ self.latent_channels = self.vae_model.config.latent_channels
+ self.has_latents_mean = (
+ hasattr(self.vae_model.config, "latents_mean")
+ and self.vae_model.config.latents_mean is not None
+ )
+ self.has_latents_std = (
+ hasattr(self.vae_model.config, "latents_std")
+ and self.vae_model.config.latents_std is not None
+ )
+ self.latents_mean = self.vae_model.config.latents_mean
+ self.latents_std = self.vae_model.config.latents_std
+
+ x = torch.randn(1, self.vae_model.config.in_channels, 32, 32)
+ z = self.encode(x)
+
+ # set downsampling factor
+ self.downsampling_factor = int(x.shape[2] / z.shape[2])
+
+ def encode(self, x: torch.tensor, batch_size: int = 8):
+ latents = []
+ for i in range(0, x.shape[0], batch_size):
+ latents.append(
+ self.vae_model.encode(x[i : i + batch_size]).latent_dist.sample()
+ )
+ latents = torch.cat(latents, dim=0)
+ latents = (latents - self.shift_factor) * self.vae_model.config.scaling_factor
+
+ return latents
+
+ def decode(self, z: torch.tensor):
+
+ if self.has_latents_mean and self.has_latents_std:
+ latents_mean = (
+ torch.tensor(self.latents_mean)
+ .view(1, self.latent_channels, 1, 1)
+ .to(z.device, z.dtype)
+ )
+ latents_std = (
+ torch.tensor(self.latents_std)
+ .view(1, self.latent_channels, 1, 1)
+ .to(z.device, z.dtype)
+ )
+ z = z * latents_std / self.vae_model.config.scaling_factor + latents_mean
+ else:
+ z = z / self.vae_model.config.scaling_factor + self.shift_factor
+
+ use_tiling = (
+ z.shape[2] > self.tiling_size[0] or z.shape[3] > self.tiling_size[1]
+ )
+
+ if use_tiling:
+ samples = []
+ for i in range(z.shape[0]):
+
+ z_i = z[i].unsqueeze(0)
+
+ tiler = Tiler()
+ tiles = tiler.get_tiles(
+ input=z_i,
+ tile_size=self.tiling_size,
+ overlap_size=self.tiling_overlap,
+ scale=self.downsampling_factor,
+ out_channels=3,
+ )
+
+ for i, tile_row in enumerate(tiles):
+ for j, tile in enumerate(tile_row):
+ tile_shape = tile.shape
+ # pad tile to inference size if tile is smaller than inference size
+ tile = pad(
+ tile,
+ base_h=self.tiling_size[0],
+ base_w=self.tiling_size[1],
+ )
+ tile_decoded = self.vae_model.decode(tile).sample
+ tiles[i][j] = (
+ tile_decoded[
+ 0,
+ :,
+ : int(tile_shape[2] * self.downsampling_factor),
+ : int(tile_shape[3] * self.downsampling_factor),
+ ]
+ .cpu()
+ .unsqueeze(0)
+ )
+
+ # merge tiles
+ samples.append(tiler.merge_tiles(tiles=tiles))
+
+ samples = torch.cat(samples, dim=0)
+
+ else:
+ samples = self.vae_model.decode(z).sample
+
+ return samples
diff --git a/src/lbm/models/vae/autoencoderKL_config.py b/src/lbm/models/vae/autoencoderKL_config.py
new file mode 100644
index 0000000000000000000000000000000000000000..f5c31f124dc10da40cf81e3c8f490d5d74b1a78d
--- /dev/null
+++ b/src/lbm/models/vae/autoencoderKL_config.py
@@ -0,0 +1,27 @@
+from typing import Tuple
+
+from pydantic.dataclasses import dataclass
+
+from ..base import ModelConfig
+
+
+@dataclass
+class AutoencoderKLDiffusersConfig(ModelConfig):
+ """This is the VAEConfig class which defines all the useful parameters to instantiate the model.
+
+ Args:
+
+ version (str): The version of the model. Defaults to "stabilityai/sdxl-vae".
+ subfolder (str): The subfolder of the model if loaded from another model. Defaults to "".
+ revision (str): The revision of the model. Defaults to "main".
+ input_key (str): The key of the input data in the batch. Defaults to "image".
+ tiling_size (Tuple[int, int]): The size of the tiling. Defaults to (64, 64).
+ tiling_overlap (Tuple[int, int]): The overlap of the tiling. Defaults to (16, 16).
+ """
+
+ version: str = "stabilityai/sdxl-vae"
+ subfolder: str = ""
+ revision: str = "main"
+ input_key: str = "image"
+ tiling_size: Tuple[int, int] = (64, 64)
+ tiling_overlap: Tuple[int, int] = (16, 16)
diff --git a/src/lbm/trainer/__init__.py b/src/lbm/trainer/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..506c801f80f19341b6dbb5806846d68d1e7cda03
--- /dev/null
+++ b/src/lbm/trainer/__init__.py
@@ -0,0 +1,85 @@
+"""
+This module contains the training pipeline and the training configuration along with all relevant parts
+of the training pipeline such as loggers and callbacks.
+
+The :mod:`cr.trainer` includes the following submodules:
+
+- :mod:`cr.trainer.trainer`: the main training pipeline class for ClipDrop.
+- :mod:`cr.trainer.training_config`: the configuration for the training pipeline.
+- :mod:`cr.trainer.loggers`: the loggers for logging samples to wandb.
+
+
+Examples
+########
+
+Train a model using the training pipeline
+
+.. code-block:: python
+
+ from cr.trainer import TrainingPipeline, TrainingConfig
+ from cr.data import DataPipeline, DataConfig
+ from pytorch_lightning import Trainer
+ from cr.data.datasets import DataModule, DataModuleConfig
+
+ # Create a model to train
+ model = DummyModel()
+
+ # Create a training configuration
+ config = TrainingConfig(
+ experiment_id="test",
+ optimizers_name=["AdamW"],
+ optimizers_kwargs=[{}],
+ learning_rates=[1e-3],
+ lr_schedulers_name=[None],
+ lr_schedulers_kwargs=[{}],
+ trainable_params=[["./*"]],
+ log_keys="txt",
+ log_samples_model_kwargs={
+ "max_samples": 8,
+ "num_steps": 20,
+ "input_shape": (4, 32, 32),
+ "guidance_scale": 7.5,
+ }
+ )
+
+ # Create a training pipeline
+ pipeline = TrainingPipeline(model=model, pipeline_config=config)
+
+ # Create a DataModule
+ data_module = DataModule(
+ train_config=DataModuleConfig(
+ shards_path_or_urls="your urls or paths",
+ decoder="pil",
+ shuffle_buffer_size=100,
+ per_worker_batch_size=32,
+ num_workers=4,
+ ),
+ train_filters_mappers=your_mappers_and_filters,
+ eval_config=DataModuleConfig(
+ shards_path_or_urls="your urls or paths",
+ decoder="pil",
+ shuffle_buffer_size=100,
+ per_worker_batch_size=32,
+ num_workers=4,
+ ),
+ eval_filters_mappers=your_mappers_and_filters,
+ )
+
+ # Create a trainer
+ trainer = Trainer(
+ accelerator="cuda",
+ max_epochs=1,
+ devices=1,
+ log_every_n_steps=1,
+ default_root_dir="your dir",
+ max_steps=2,
+ )
+
+ # Train the model
+ trainer.fit(pipeline, data_module)
+"""
+
+from .trainer import TrainingPipeline
+from .training_config import TrainingConfig
+
+__all__ = ["TrainingPipeline", "TrainingConfig"]
diff --git a/src/lbm/trainer/loggers.py b/src/lbm/trainer/loggers.py
new file mode 100644
index 0000000000000000000000000000000000000000..04472396702006daf3978f7cf2661d9884306cdc
--- /dev/null
+++ b/src/lbm/trainer/loggers.py
@@ -0,0 +1,324 @@
+import logging
+import math
+from typing import Any, Dict, List, Tuple
+
+import numpy as np
+import torch
+import wandb
+from PIL import Image, ImageDraw, ImageFont
+from pytorch_lightning import Trainer
+from pytorch_lightning.callbacks import Callback
+from pytorch_lightning.utilities import rank_zero_only
+from torchvision.utils import make_grid
+
+from ..trainer import TrainingPipeline
+
+logging.basicConfig(level=logging.INFO)
+
+
+def create_grid_texts(
+ texts: List[str],
+ n_cols: int = 4,
+ image_size: Tuple[int] = (512, 512),
+ font_size: int = 40,
+ margin: int = 5,
+ offset: int = 5,
+) -> Image.Image:
+ """
+ Create a grid of white images containing the given texts.
+
+ Args:
+ texts (List[str]): List of strings to be drawn on images.
+ n_cols (int): Number of columns in the grid.
+ image_size (tuple): Size of the generated images (width, height).
+ font_size (int): Font size of the text.
+ margin (int): Margin around the text.
+ offset (int): Offset between lines.
+
+ Returns:
+ PIL.Image: List of generated images as a grid
+ """
+
+ images = []
+ font = ImageFont.load_default(size=font_size)
+
+ for text in texts:
+ img = Image.new("RGB", image_size, color="white")
+ draw = ImageDraw.Draw(img)
+ margin_ = margin
+ offset_ = offset
+ for line in wrap_text(
+ text=text, draw=draw, max_width=image_size[0] - 2 * margin_, font=font
+ ):
+ draw.text((margin_, offset_), line, font=font, fill="black")
+ offset_ += font_size
+ images.append(img)
+
+ # create a pil grid
+ n_rows = math.ceil(len(images) / n_cols)
+ grid = Image.new(
+ "RGB", (n_cols * image_size[0], n_rows * image_size[1]), color="white"
+ )
+ for i, img in enumerate(images):
+ grid.paste(img, (i % n_cols * image_size[0], i // n_cols * image_size[1]))
+
+ return grid
+
+
+def wrap_text(
+ text: str, draw: ImageDraw.Draw, max_width: int, font: ImageFont
+) -> List[str]:
+ """
+ Wrap text to fit within a specified width when drawn.
+ It will return to the new line when the text is larger than the max_width.
+
+ Args:
+ text (str): The text to be wrapped.
+ draw (ImageDraw.Draw): The draw object to calculate text size.
+ max_width (int): The maximum width for the wrapped text.
+ font (ImageFont): The font used for the text.
+
+ Returns:
+ List[str]: List of wrapped lines.
+ """
+ lines = []
+ current_line = ""
+ for letter in text:
+ if draw.textbbox((0, 0), current_line + letter, font=font)[2] <= max_width:
+ current_line += letter
+ else:
+ lines.append(current_line)
+ current_line = letter
+ lines.append(current_line)
+ return lines
+
+
+class WandbSampleLogger(Callback):
+ """
+ Logger for logging samples to wandb. This logger is used to log images, text, and metrics to wandb.
+
+ Args:
+ log_batch_freq (int): The frequency of logging samples to wandb. Default is 100.
+ """
+
+ def __init__(self, log_batch_freq: int = 100):
+ super().__init__()
+ self.log_batch_freq = log_batch_freq
+
+ def on_train_batch_end(
+ self,
+ trainer: Trainer,
+ pl_module: TrainingPipeline,
+ outputs: Dict[str, Any],
+ batch: Any,
+ batch_idx: int,
+ ) -> None:
+ self.log_samples(trainer, pl_module, outputs, batch, batch_idx, split="train")
+ self._process_logs(trainer, outputs, split="train")
+
+ def on_validation_batch_end(
+ self,
+ trainer: Trainer,
+ pl_module: TrainingPipeline,
+ outputs: Dict[str, Any],
+ batch: Any,
+ batch_idx: int,
+ ) -> None:
+ self.log_samples(trainer, pl_module, outputs, batch, batch_idx, split="val")
+ self._process_logs(trainer, outputs, split="val")
+
+ @rank_zero_only
+ @torch.no_grad()
+ def log_samples(
+ self,
+ trainer: Trainer,
+ pl_module: TrainingPipeline,
+ outputs: Dict[str, Any],
+ batch: Dict[str, Any],
+ batch_idx: int,
+ split: str = "train",
+ ) -> None:
+ if hasattr(pl_module, "log_samples"):
+ if batch_idx % self.log_batch_freq == 0:
+ is_training = pl_module.training
+ if is_training:
+ pl_module.eval()
+
+ logs = pl_module.log_samples(batch)
+ logs = self._process_logs(trainer, logs, split=split)
+
+ if is_training:
+ pl_module.train()
+ else:
+ logging.warning(
+ "log_img method not found in LightningModule. Skipping image logging."
+ )
+
+ @rank_zero_only
+ def _process_logs(
+ self, trainer, logs: Dict[str, Any], rescale=True, split="train"
+ ) -> Dict[str, Any]:
+ for key, value in logs.items():
+ if isinstance(value, torch.Tensor):
+ value = value.detach().cpu()
+ if value.dim() == 4:
+ images = value
+ if rescale:
+ images = (images + 1.0) / 2.0
+ grid = make_grid(images, nrow=4)
+ grid = grid.permute(1, 2, 0)
+ grid = grid.mul(255).clamp(0, 255).to(torch.uint8)
+ logs[key] = grid.numpy()
+ trainer.logger.experiment.log(
+ {f"{key}/{split}": [wandb.Image(Image.fromarray(logs[key]))]},
+ step=trainer.global_step,
+ )
+
+ # Scalar tensor
+ if value.dim() == 1 or value.dim() == 0:
+ value = value.float().numpy()
+ trainer.logger.experiment.log(
+ {f"{key}/{split}": value}, step=trainer.global_step
+ )
+
+ # list of string (e.g. text)
+ if isinstance(value, list):
+ if isinstance(value[0], str):
+ pil_image_texts = create_grid_texts(value)
+ wandb_image = wandb.Image(pil_image_texts)
+ trainer.logger.experiment.log(
+ {f"{key}/{split}": [wandb_image]},
+ step=trainer.global_step,
+ )
+
+ # dict of tensors (e.g. metrics)
+ if isinstance(value, dict):
+ for k, v in value.items():
+ if isinstance(v, torch.Tensor):
+ value[k] = v.detach().cpu().numpy()
+ trainer.logger.experiment.log(
+ {f"{key}/{split}": value}, step=trainer.global_step
+ )
+
+ if isinstance(value, int) or isinstance(value, float):
+ trainer.logger.experiment.log(
+ {f"{key}/{split}": value}, step=trainer.global_step
+ )
+
+ return logs
+
+
+class TensorBoardSampleLogger(Callback):
+ """
+ Logger for logging samples to tensorboard. This logger is used to log images, text, and metrics to tensorboard.
+
+ Args:
+ log_batch_freq (int): The frequency of logging samples to tensorboard. Default is 100.
+ """
+
+ def __init__(self, log_batch_freq: int = 100):
+ super().__init__()
+ self.log_batch_freq = log_batch_freq
+
+ def on_train_batch_end(
+ self,
+ trainer: Trainer,
+ pl_module: TrainingPipeline,
+ outputs: Dict[str, Any],
+ batch: Any,
+ batch_idx: int,
+ ) -> None:
+ self.log_samples(trainer, pl_module, outputs, batch, batch_idx, split="train")
+ self._process_logs(trainer, outputs, split="train")
+
+ def on_validation_batch_end(
+ self,
+ trainer: Trainer,
+ pl_module: TrainingPipeline,
+ outputs: Dict[str, Any],
+ batch: Any,
+ batch_idx: int,
+ ) -> None:
+ self.log_samples(trainer, pl_module, outputs, batch, batch_idx, split="val")
+ self._process_logs(trainer, outputs, split="val")
+
+ @rank_zero_only
+ @torch.no_grad()
+ def log_samples(
+ self,
+ trainer: Trainer,
+ pl_module: TrainingPipeline,
+ outputs: Dict[str, Any],
+ batch: Dict[str, Any],
+ batch_idx: int,
+ split: str = "train",
+ ) -> None:
+ if hasattr(pl_module, "log_samples"):
+ if batch_idx % self.log_batch_freq == 0:
+ is_training = pl_module.training
+ if is_training:
+ pl_module.eval()
+
+ logs = pl_module.log_samples(batch)
+ logs = self._process_logs(trainer, logs, split=split)
+
+ if is_training:
+ pl_module.train()
+ else:
+ logging.warning(
+ "log_img method not found in LightningModule. Skipping image logging."
+ )
+
+ @rank_zero_only
+ def _process_logs(
+ self, trainer, logs: Dict[str, Any], rescale=True, split="train"
+ ) -> Dict[str, Any]:
+ for key, value in logs.items():
+ if isinstance(value, torch.Tensor):
+ value = value.detach().cpu()
+ if value.dim() == 4:
+ images = value
+ if rescale:
+ images = (images + 1.0) / 2.0
+ grid = make_grid(images, nrow=4)
+ # grid = grid.permute(1, 2, 0)
+ grid = grid.mul(255).clamp(0, 255).to(torch.uint8)
+ logs[key] = grid.numpy()
+ trainer.logger.experiment.add_image(
+ f"{key}/{split}",
+ logs[key],
+ trainer.global_step,
+ )
+
+ # Scalar tensor
+ if value.dim() == 1 or value.dim() == 0:
+ value = value.float().numpy()
+ trainer.logger.experiment.add_scalar(
+ f"{key}/{split}", value, trainer.global_step
+ )
+
+ # list of string (e.g. text)
+ if isinstance(value, list):
+ if isinstance(value[0], str):
+ pil_image_texts = create_grid_texts(value)
+ trainer.logger.experiment.add_image(
+ f"{key}/{split}",
+ np.transpose(np.array(pil_image_texts), (2, 0, 1)),
+ trainer.global_step,
+ )
+
+ # dict of tensors (e.g. metrics)
+ if isinstance(value, dict):
+ for k, v in value.items():
+ if isinstance(v, torch.Tensor):
+ value[k] = v.detach().cpu().numpy()
+ trainer.logger.experiment.add_scalar(
+ f"{key}/{split}", value, trainer.global_step
+ )
+
+ if isinstance(value, int) or isinstance(value, float):
+ trainer.logger.experiment.add_scalar(
+ f"{key}/{split}", value, trainer.global_step
+ )
+
+ return logs
diff --git a/src/lbm/trainer/trainer.py b/src/lbm/trainer/trainer.py
new file mode 100644
index 0000000000000000000000000000000000000000..035fe3aa4535e5c79fb1373ecb9b92b2d3035327
--- /dev/null
+++ b/src/lbm/trainer/trainer.py
@@ -0,0 +1,199 @@
+import importlib
+import logging
+import re
+import time
+from typing import Any, Dict
+
+import pytorch_lightning as pl
+import torch
+
+from ..models.base.base_model import BaseModel
+from .training_config import TrainingConfig
+
+logging.basicConfig(level=logging.INFO)
+
+
+class TrainingPipeline(pl.LightningModule):
+ """
+ Main Training Pipeline class
+
+ Args:
+
+ model (BaseModel): The model to train
+ pipeline_config (TrainingConfig): The configuration for the training pipeline
+ verbose (bool): Whether to print logs in the console. Default is False.
+ """
+
+ def __init__(
+ self,
+ model: BaseModel,
+ pipeline_config: TrainingConfig,
+ verbose: bool = False,
+ **kwargs,
+ ):
+ super().__init__()
+
+ self.model = model
+ self.pipeline_config = pipeline_config
+ self.log_samples_model_kwargs = pipeline_config.log_samples_model_kwargs
+
+ # save hyperparameters.
+ self.save_hyperparameters(ignore="model")
+ self.save_hyperparameters({"model_config": model.config.to_dict()})
+
+ # logger.
+ self.verbose = verbose
+
+ # setup logging.
+ log_keys = pipeline_config.log_keys
+
+ if isinstance(log_keys, str):
+ log_keys = [log_keys]
+
+ if log_keys is None:
+ log_keys = []
+
+ self.log_keys = log_keys
+
+ def on_fit_start(self) -> None:
+ self.model.on_fit_start(device=self.device)
+ if self.global_rank == 0:
+ self.timer = time.perf_counter()
+
+ def on_train_batch_end(
+ self, outputs: Dict[str, Any], batch: Any, batch_idx: int
+ ) -> None:
+ if self.global_rank == 0:
+ logging.debug("on_train_batch_end")
+ self.model.on_train_batch_end(batch)
+
+ average_time_frequency = 10
+ if self.global_rank == 0 and batch_idx % average_time_frequency == 0:
+ delta = time.perf_counter() - self.timer
+ logging.info(
+ f"Average time per batch {batch_idx} took {delta / (batch_idx + 1)} seconds"
+ )
+
+ def configure_optimizers(self) -> torch.optim.Optimizer:
+ """
+ Setup optimizers and learning rate schedulers.
+ """
+ optimizers = []
+ lr = self.pipeline_config.learning_rate
+ param_list = []
+ n_params = 0
+ param_list_ = {"params": []}
+ for name, param in self.model.named_parameters():
+ for regex in self.pipeline_config.trainable_params:
+ pattern = re.compile(regex)
+ if re.match(pattern, name):
+ if param.requires_grad:
+ param_list_["params"].append(param)
+ n_params += param.numel()
+
+ param_list.append(param_list_)
+
+ logging.info(f"Number of trainable parameters: {n_params}")
+
+ optimizer_cls = getattr(
+ importlib.import_module("torch.optim"),
+ self.pipeline_config.optimizer_name,
+ )
+ optimizer = optimizer_cls(
+ param_list, lr=lr, **self.pipeline_config.optimizer_kwargs
+ )
+ optimizers.append(optimizer)
+
+ self.optims = optimizers
+ schedulers_config = self.configure_lr_schedulers()
+
+ for name, param in self.model.named_parameters():
+ set_grad_false = True
+ for regex in self.pipeline_config.trainable_params:
+ pattern = re.compile(regex)
+ if re.match(pattern, name):
+ if param.requires_grad:
+ set_grad_false = False
+ if set_grad_false:
+ param.requires_grad = False
+
+ num_trainable_params = sum(
+ p.numel() for p in self.model.parameters() if p.requires_grad
+ )
+
+ logging.info(f"Number of trainable parameters: {num_trainable_params}")
+
+ schedulers_config = self.configure_lr_schedulers()
+
+ if schedulers_config is None:
+ return optimizers
+
+ return optimizers, [
+ schedulers_config_ for schedulers_config_ in schedulers_config
+ ]
+
+ def configure_lr_schedulers(self):
+ schedulers_config = []
+ if self.pipeline_config.lr_scheduler_name is None:
+ scheduler = None
+ schedulers_config.append(scheduler)
+ else:
+ scheduler_cls = getattr(
+ importlib.import_module("torch.optim.lr_scheduler"),
+ self.pipeline_config.lr_scheduler_name,
+ )
+ scheduler = scheduler_cls(
+ self.optims[0],
+ **self.pipeline_config.lr_scheduler_kwargs,
+ )
+ lr_scheduler_config = {
+ "scheduler": scheduler,
+ "interval": self.pipeline_config.lr_scheduler_interval,
+ "monitor": "val_loss",
+ "frequency": self.pipeline_config.lr_scheduler_frequency,
+ }
+ schedulers_config.append(lr_scheduler_config)
+
+ if all([scheduler is None for scheduler in schedulers_config]):
+ return None
+
+ return schedulers_config
+
+ def training_step(self, train_batch: Dict[str, Any], batch_idx: int) -> dict:
+ model_output = self.model(train_batch)
+ loss = model_output["loss"]
+ logging.info(f"loss: {loss}")
+ return {
+ "loss": loss,
+ "batch_idx": batch_idx,
+ }
+
+ def validation_step(self, val_batch: Dict[str, Any], val_idx: int) -> dict:
+ loss = self.model(val_batch, device=self.device)["loss"]
+
+ metrics = self.model.compute_metrics(val_batch)
+
+ return {"loss": loss, "metrics": metrics}
+
+ def log_samples(self, batch: Dict[str, Any]):
+ logging.debug("log_samples")
+ logs = self.model.log_samples(
+ batch,
+ **self.log_samples_model_kwargs,
+ )
+
+ if logs is not None:
+ N = min([logs[keys].shape[0] for keys in logs])
+ else:
+ N = 0
+
+ # Log inputs
+ if self.log_keys is not None:
+ for key in self.log_keys:
+ if key in batch:
+ if N > 0:
+ logs[key] = batch[key][:N]
+ else:
+ logs[key] = batch[key]
+
+ return logs
diff --git a/src/lbm/trainer/training_config.py b/src/lbm/trainer/training_config.py
new file mode 100644
index 0000000000000000000000000000000000000000..1c8663305aac6f8c8960f04d731846e6fae437b9
--- /dev/null
+++ b/src/lbm/trainer/training_config.py
@@ -0,0 +1,82 @@
+from dataclasses import field
+from typing import List, Literal, Optional, Union
+
+from pydantic.dataclasses import dataclass
+
+from ..config import BaseConfig
+
+
+@dataclass
+class TrainingConfig(BaseConfig):
+ """
+ Configuration for the training pipeline
+
+ Args:
+
+ experiment_id (str):
+ The experiment id for the training run. If not provided, a random id will be generated.
+ optimizer_name (str):
+ The optimizer to use. Default is "AdamW". Choices are "Adam", "AdamW", "Adadelta", "Adagrad", "RMSprop", "SGD"
+ optimizer_kwargs (Dict[str, Any])
+ The optimizer kwargs. Default is [{}]
+ learning_rate (float):
+ The learning rate to use. Default is 1e-3
+ lr_scheduler_name (str):
+ The learning rate scheduler to use. Default is None. Choices are "StepLR", "CosineAnnealingLR",
+ "CosineAnnealingWarmRestarts", "ReduceLROnPlateau", "ExponentialLR"
+ lr_scheduler_kwargs (Dict[str, Any])
+ The learning rate scheduler kwargs. Default is [{}]
+ lr_scheduler_interval (str):
+ The learning rate scheduler interval. Default is ["step"]. Choices are "step", "epoch"
+ lr_scheduler_frequency (int):
+ The learning rate scheduler frequency. Default is 1
+ metrics (List[str])
+ The metrics to use. Default is None
+ tracking_metrics: Optional[List[str]]
+ The metrics to track. Default is None
+ backup_every (int):
+ The frequency to backup the model. Default is 50.
+ trainable_params (Union[str, List[str]]):
+ Regexes indicateing the parameters to train.
+ Default is [["./*"]] (i.e. all parameters are trainable)
+ log_keys: Union[str, List[str]]:
+ The keys to log when sampling from the model. Default is "txt"
+ log_samples_model_kwargs (Dict[str, Any]):
+ The kwargs for logging samples from the model. Default is {
+ "max_samples": 4,
+ "num_steps": 20,
+ "input_shape": None,
+ }
+ """
+
+ experiment_id: Optional[str] = None
+ optimizer_name: Literal[
+ "Adam", "AdamW", "Adadelta", "Adagrad", "RMSprop", "SGD"
+ ] = field(default_factory=lambda: "AdamW")
+ optimizer_kwargs: Optional[dict] = field(default_factory=lambda: {})
+ learning_rate: float = field(default_factory=lambda: 1e-3)
+ lr_scheduler_name: Optional[
+ Literal[
+ "StepLR",
+ "CosineAnnealingLR",
+ "CosineAnnealingWarmRestarts",
+ "ReduceLROnPlateau",
+ "ExponentialLR",
+ None,
+ ]
+ ] = None
+ lr_scheduler_kwargs: Optional[dict] = field(default_factory=lambda: {})
+ lr_scheduler_interval: Optional[Literal["step", "epoch", None]] = "step"
+ lr_scheduler_frequency: Optional[int] = 1
+ metrics: Optional[List[str]] = None
+ tracking_metrics: Optional[List[str]] = None
+ backup_every: int = 50
+ trainable_params: List[str] = field(default_factory=lambda: ["./*"])
+ log_keys: Optional[Union[str, List[str]]] = "txt"
+ log_samples_model_kwargs: Optional[dict] = field(
+ default_factory=lambda: {
+ "max_samples": 4,
+ "num_steps": 20,
+ "input_shape": None,
+ }
+ )
diff --git a/src/lbm/trainer/utils.py b/src/lbm/trainer/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..1e827edcc1547b48c1cc72c6d00af08ec5ef0226
--- /dev/null
+++ b/src/lbm/trainer/utils.py
@@ -0,0 +1,193 @@
+import logging
+import os
+import re
+import time
+from typing import Dict, List, Literal, Optional, Tuple
+
+import torch
+
+
+class StateDictAdapter:
+ """
+ StateDictAdapter for adapting the state dict of a model to a checkpoint state dict.
+
+ This class will iterate over all keys in the checkpoint state dict and filter them by a list of regex keys.
+ For each matching key, the class will adapt the checkpoint state dict to the model state dict.
+ Depending on the target size, the class will add missing blocks or cut the block.
+ When adding missing blocks, the class will use a strategy to fill the missing blocks: either adding zeros or normal random values.
+
+ Example:
+
+ ```
+ adapter = StateDictAdapter()
+ new_state_dict = adapter(
+ model_state_dict=model.state_dict(),
+ checkpoint_state_dict=state_dict,
+ regex_keys=[
+ r"class_embedding.linear_1.weight",
+ r"conv_in.weight",
+ r"(down_blocks|up_blocks)\.\d+\.attentions\.\d+\.transformer_blocks\.\d+\.attn\d+\.(to_k|to_v)\.weight",
+ r"mid_block\.attentions\.\d+\.transformer_blocks\.\d+\.attn\d+\.(to_k|to_v)\.weight"
+ ]
+ )
+ ```
+
+ Args:
+ model_state_dict (Dict[str, torch.Tensor]): The model state dict.
+ checkpoint_state_dict (Dict[str, torch.Tensor]): The checkpoint state dict.
+ regex_keys (Optional[List[str]]): A list of regex keys to adapt the checkpoint state dict. Defaults to None.
+ Passing a list of regex will drastically reduce the latency.
+ If None, all keys in the checkpoint state dict will be adapted.
+ strategy (Literal["zeros", "normal"], optional): The strategy to fill the missing blocks. Defaults to "normal".
+
+ """
+
+ def _create_block(
+ self,
+ shape: List[int],
+ strategy: Literal["zeros", "normal"],
+ input: torch.Tensor = None,
+ ):
+ if strategy == "zeros":
+ return torch.zeros(shape)
+ elif strategy == "normal":
+ if input is not None:
+ mean = input.mean().item()
+ std = input.std().item()
+ return torch.randn(shape) * std + mean
+ else:
+ return torch.randn(shape)
+ else:
+ raise ValueError(f"Unknown strategy {strategy}")
+
+ def __call__(
+ self,
+ model_state_dict: Dict[str, torch.Tensor],
+ checkpoint_state_dict: Dict[str, torch.Tensor],
+ regex_keys: Optional[List[str]] = None,
+ strategy: Literal["zeros", "normal"] = "normal",
+ ):
+ start = time.perf_counter()
+ # if no regex keys are provided, we use all keys in the model state dict
+ if regex_keys is None:
+ regex_keys = list(model_state_dict.keys())
+
+ # iterate over all keys in the checkpoint state dict
+ for checkpoint_key in list(checkpoint_state_dict.keys()):
+ # iterate over all regex keys
+ for regex_key in regex_keys:
+ if re.match(regex_key, checkpoint_key):
+ dst_shape = model_state_dict[checkpoint_key].shape
+ src_shape = checkpoint_state_dict[checkpoint_key].shape
+
+ ## Sizes adapter
+ # if length of shapes are different, we need to unsqueeze or squeeze the tensor
+ if len(dst_shape) != len(src_shape):
+ # in the case [a] vs [a, b] -> unsqueeze [a, 1]
+ if len(src_shape) == 1:
+ checkpoint_state_dict[checkpoint_key] = (
+ checkpoint_state_dict[checkpoint_key].unsqueeze(1)
+ )
+ logging.info(
+ f"Unsqueeze {checkpoint_key}: {src_shape} -> {checkpoint_state_dict[checkpoint_key].shape}"
+ )
+ # in the case [a, b] vs [a] -> squeeze [a]
+ elif len(dst_shape) == 1:
+ checkpoint_state_dict[checkpoint_key] = (
+ checkpoint_state_dict[checkpoint_key][:, 0]
+ )
+ logging.info(
+ f"Squeeze {checkpoint_key}: {src_shape} -> {checkpoint_state_dict[checkpoint_key].shape}"
+ )
+ # in the other cases, raise an error
+ else:
+ raise ValueError(
+ f"Shapes of {checkpoint_key} are different: {dst_shape} != {src_shape}"
+ )
+
+ # update the shapes
+ dst_shape = model_state_dict[checkpoint_key].shape
+ src_shape = checkpoint_state_dict[checkpoint_key].shape
+ assert len(dst_shape) == len(
+ src_shape
+ ), f"Shapes of {checkpoint_key} are different: {dst_shape} != {src_shape}"
+
+ ## Shapes adapter
+ # modify the checkpoint state dict only if the shapes are different
+ if dst_shape != src_shape:
+ # create a copy of the tensor
+ tmp = torch.clone(checkpoint_state_dict[checkpoint_key])
+
+ # iterate over all dimensions
+ for i in range(len(dst_shape)):
+ if dst_shape[i] != src_shape[i]:
+ diff = dst_shape[i] - src_shape[i]
+
+ # if the difference is greater than 0, we need to add missing blocks
+ if diff > 0:
+ missing_shape = list(tmp.shape)
+ missing_shape[i] = diff
+ missing = self._create_block(
+ shape=missing_shape,
+ strategy=strategy,
+ input=tmp,
+ )
+ tmp = torch.cat((tmp, missing), dim=i)
+ logging.info(
+ f"Adapting {checkpoint_key} with strategy:{strategy} from shape {src_shape} to {dst_shape}"
+ )
+ # if the difference is less than 0, we need to cut the block
+ else:
+ tmp = tmp.narrow(i, 0, dst_shape[i])
+ logging.info(
+ f"Adapting {checkpoint_key} by narrowing from shape {src_shape} to {dst_shape}"
+ )
+
+ checkpoint_state_dict[checkpoint_key] = tmp
+ end = time.perf_counter()
+ logging.info(f"StateDictAdapter took {end-start:.2f} seconds")
+ return checkpoint_state_dict
+
+
+class StateDictRenamer:
+ """
+ StateDictRenamer for renaming keys in a checkpoint state dict.
+ This class will iterate over all keys in the checkpoint state dict and rename them according to a rename dict.
+
+ Example:
+
+ ```
+ renamer = StateDictRenamer()
+ new_state_dict = renamer(
+ checkpoint_state_dict=state_dict,
+ rename_dict={
+ "add_embedding.linear_1.weight": "class_embedding.linear_1.weight",
+ "add_embedding.linear_1.bias": "class_embedding.linear_1.bias",
+ "add_embedding.linear_2.weight": "class_embedding.linear_2.weight",
+ "add_embedding.linear_2.bias": "class_embedding.linear_2.bias",
+ }
+ )
+ ```
+
+ Args:
+
+ checkpoint_state_dict (Dict[str, torch.Tensor]): The checkpoint state dict.
+ rename_dict (Dict[str, str]): The dictionary mapping the old keys to new keys
+ """
+
+ def __call__(
+ self,
+ checkpoint_state_dict: Dict[str, torch.Tensor],
+ rename_dict: Dict[str, str],
+ ) -> Dict[str, torch.Tensor]:
+ for old_key, new_key in rename_dict.items():
+ if old_key not in checkpoint_state_dict:
+ logging.warning(f"Key {old_key} not found in checkpoint state dict")
+ continue
+ else:
+ assert (
+ new_key not in checkpoint_state_dict
+ ), f"Key {new_key} already exists in checkpoint state dict"
+ checkpoint_state_dict[new_key] = checkpoint_state_dict.pop(old_key)
+ logging.info(f"Renaming {old_key} to {new_key}")
+ return checkpoint_state_dict
diff --git a/tests/README.md b/tests/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..f814f32582e0b6a9e726c2ad85ee94ea8421f20e
--- /dev/null
+++ b/tests/README.md
@@ -0,0 +1,13 @@
+# Tests
+
+## Setup
+
+```shell
+pip3 install -r requirements.txt
+```
+
+## Run the tests
+
+```shell
+python3 -m pytest .
+```
\ No newline at end of file
diff --git a/tests/requirements.txt b/tests/requirements.txt
new file mode 100644
index 0000000000000000000000000000000000000000..55b033e901cdda93a26ac64b418f260224260a39
--- /dev/null
+++ b/tests/requirements.txt
@@ -0,0 +1 @@
+pytest
\ No newline at end of file
diff --git a/tests/test_dataset/test_filters.py b/tests/test_dataset/test_filters.py
new file mode 100644
index 0000000000000000000000000000000000000000..3158f867af16ee39199bbac57f0902c8aeafc11b
--- /dev/null
+++ b/tests/test_dataset/test_filters.py
@@ -0,0 +1,41 @@
+from lbm.data.filters import FilterWrapper, KeyFilter, KeyFilterConfig
+
+
+class TestKeyFilter:
+ def test_key_filter(self):
+ filter = KeyFilter(KeyFilterConfig(keys=["a", "b"]))
+ assert filter({"a": 1, "b": 2, "c": 3})
+ assert not filter({"a": 1})
+ assert not filter({"b": 2})
+ assert not filter({"c": 3})
+
+ def test_key_filter_single_key(self):
+ filter = KeyFilter(KeyFilterConfig(keys="a"))
+ assert filter({"a": 1, "b": 2})
+ assert not filter({"b": 2})
+
+
+class TestFilterWrapper:
+ def test_filter_wrapper(self):
+ filter = FilterWrapper(
+ [
+ KeyFilter(KeyFilterConfig(keys=["a", "b"])),
+ KeyFilter(KeyFilterConfig(keys="c")),
+ ]
+ )
+ assert not filter({"a": 1, "b": 2, "c": 3})
+ assert not filter({"a": 1})
+ assert not filter({"b": 2})
+ assert not filter({"c": 3})
+
+ def test_filter_wrapper(self):
+ filter = FilterWrapper(
+ [
+ KeyFilter(KeyFilterConfig(keys=["a", "b"])),
+ KeyFilter(KeyFilterConfig(keys=["a", "c"])),
+ ]
+ )
+ assert filter({"a": 1, "b": 2, "c": 3})
+ assert not filter({"a": 1})
+ assert not filter({"b": 2})
+ assert not filter({"c": 3})
diff --git a/tests/test_dataset/test_mappers.py b/tests/test_dataset/test_mappers.py
new file mode 100644
index 0000000000000000000000000000000000000000..196d0ecc81c9ddd721199152b42cef1f282b907b
--- /dev/null
+++ b/tests/test_dataset/test_mappers.py
@@ -0,0 +1,196 @@
+import json
+
+import pytest
+import torch
+from PIL import Image
+
+from lbm.data.mappers import (
+ KeyRenameMapper,
+ KeyRenameMapperConfig,
+ MapperWrapper,
+ RescaleMapper,
+ RescaleMapperConfig,
+ TorchvisionMapper,
+ TorchvisionMapperConfig,
+)
+
+
+class TestKeyRenameMapper:
+ @pytest.fixture()
+ def dummy_batch(self):
+ return {"image": 1, "text": 2, "label": "dummy_label"}
+
+ @pytest.fixture()
+ def mapper(self):
+ return KeyRenameMapper(
+ KeyRenameMapperConfig(
+ key_map={"image": "image_tensor", "text": "text_tensor"}
+ )
+ )
+
+ def test_mapper(self, mapper, dummy_batch):
+ output_data = mapper(dummy_batch)
+ assert output_data["image_tensor"] == 1
+ assert output_data["text_tensor"] == 2
+ assert output_data["label"] == "dummy_label"
+ assert "image" not in output_data
+ assert "text" not in output_data
+
+
+class TestKeyRenameMapperWithCondition:
+ @pytest.fixture(params=[1, 2])
+ def dummy_batch(self, request):
+ return {"image": 1, "text": 2, "label": request.param}
+
+ @pytest.fixture(params=[{"image": "image_not_met", "text": "text_not_met"}, None])
+ def else_key_map(self, request):
+ return request.param
+
+ @pytest.fixture()
+ def mapper(self, else_key_map):
+ return KeyRenameMapper(
+ KeyRenameMapperConfig(
+ key_map={"image": "image_tensor", "text": "text_tensor"},
+ condition_key="label",
+ condition_fn=lambda x: x == 1,
+ else_key_map=else_key_map,
+ )
+ )
+
+ def test_mapper(self, mapper, dummy_batch, else_key_map):
+ output_data = mapper(dummy_batch)
+ if dummy_batch["label"] == 1:
+ assert output_data["image_tensor"] == 1
+ assert output_data["text_tensor"] == 2
+ assert output_data["label"] == 1
+ assert "image" not in output_data
+ assert "text" not in output_data
+ elif else_key_map is not None:
+ assert output_data["image_not_met"] == 1
+ assert output_data["text_not_met"] == 2
+ assert output_data["label"] == 2
+ assert "image" not in output_data
+ assert "text" not in output_data
+ else:
+ assert output_data["image"] == 1
+ assert output_data["text"] == 2
+ assert output_data["label"] == 2
+ assert "image_tensor" not in output_data
+ assert "text_tensor" not in output_data
+
+
+class TestMapperWrapper:
+ @pytest.fixture()
+ def dummy_batch(self):
+ return {"image": 1, "text": 2, "label": "dummy_label"}
+
+ @pytest.fixture()
+ def mapper(self):
+ return MapperWrapper(
+ mappers=[
+ KeyRenameMapper(
+ KeyRenameMapperConfig(
+ key_map={"image": "image_tensor", "text": "text_tensor"}
+ )
+ ),
+ KeyRenameMapper(
+ KeyRenameMapperConfig(
+ key_map={
+ "image_tensor": "image_array",
+ "text_tensor": "text_array",
+ }
+ )
+ ),
+ ]
+ )
+
+ def test_mapper(self, mapper, dummy_batch):
+ output_data = mapper(dummy_batch)
+ assert output_data["image_array"] == 1
+ assert output_data["text_array"] == 2
+ assert output_data["label"] == "dummy_label"
+ assert "image" not in output_data
+ assert "text" not in output_data
+ assert "image_tensor" not in output_data
+ assert "text_tensor" not in output_data
+
+
+class TestTorchvisionMapper:
+ @pytest.fixture()
+ def dummy_batch(self):
+ return {
+ "image": torch.randn(
+ 3,
+ 256,
+ 256,
+ ),
+ "text": 2,
+ "label": "dummy_label",
+ }
+
+ @pytest.fixture()
+ def mapper(self):
+ return TorchvisionMapper(
+ TorchvisionMapperConfig(
+ key="image",
+ transforms=["CenterCrop", "ToPILImage"],
+ transforms_kwargs=[{"size": 224}, {}],
+ )
+ )
+
+ def test_mapper(self, mapper, dummy_batch):
+ output_data = mapper(dummy_batch)
+ assert output_data["image"].size == (224, 224)
+ assert isinstance(output_data["image"], Image.Image)
+ assert output_data["text"] == 2
+ assert output_data["label"] == "dummy_label"
+
+ @pytest.fixture()
+ def mapper_with_output_key(self):
+ return TorchvisionMapper(
+ TorchvisionMapperConfig(
+ key="image",
+ output_key="image_transformed",
+ transforms=["CenterCrop", "ToPILImage"],
+ transforms_kwargs=[{"size": 224}, {}],
+ )
+ )
+
+ def test_mapper(self, mapper_with_output_key, dummy_batch):
+ output_data = mapper_with_output_key(dummy_batch)
+ assert output_data["image_transformed"].size == (224, 224)
+ assert isinstance(output_data["image_transformed"], Image.Image)
+ assert isinstance(output_data["image"], torch.Tensor)
+ assert output_data["image"].size() == (3, 256, 256)
+ assert output_data["text"] == 2
+ assert output_data["label"] == "dummy_label"
+
+
+class TestRescaleMapper:
+ @pytest.fixture()
+ def dummy_batch(self):
+ return {
+ "image": torch.rand(
+ 3,
+ 256,
+ 256,
+ ),
+ "text": 2,
+ "label": "dummy_label",
+ }
+
+ @pytest.fixture()
+ def mapper(self):
+ return RescaleMapper(
+ RescaleMapperConfig(
+ input_key="image",
+ output_key="image",
+ )
+ )
+
+ def test_mapper(self, mapper, dummy_batch):
+ output_data = mapper(dummy_batch)
+ assert torch.all(output_data["image"] <= 1)
+ assert torch.all(output_data["image"] >= -1)
+ assert output_data["text"] == 2
+ assert output_data["label"] == "dummy_label"
diff --git a/tests/test_lbm/test_lbm.py b/tests/test_lbm/test_lbm.py
new file mode 100644
index 0000000000000000000000000000000000000000..9c9e4ccc31868da9150ea210708e2b9b328a9fad
--- /dev/null
+++ b/tests/test_lbm/test_lbm.py
@@ -0,0 +1,101 @@
+from copy import deepcopy
+
+import pytest
+import torch
+import torch.nn as nn
+from diffusers import FlowMatchEulerDiscreteScheduler
+
+from lbm.models.embedders import ConditionerWrapper
+from lbm.models.lbm import LBMConfig, LBMModel
+from lbm.models.unets import DiffusersUNet2DCondWrapper
+from lbm.models.vae import AutoencoderKLDiffusers, AutoencoderKLDiffusersConfig
+
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+
+class TestLBM:
+ @pytest.fixture()
+ def denoiser(self):
+ return DiffusersUNet2DCondWrapper(
+ in_channels=4, # VAE channels
+ out_channels=4, # VAE channels
+ up_block_types=["CrossAttnUpBlock2D"],
+ down_block_types=[
+ "CrossAttnDownBlock2D",
+ ],
+ cross_attention_dim=[320],
+ block_out_channels=[320],
+ transformer_layers_per_block=[1],
+ attention_head_dim=[5],
+ norm_num_groups=32,
+ )
+
+ @pytest.fixture()
+ def conditioner(self):
+ return ConditionerWrapper([])
+
+ @pytest.fixture()
+ def vae(self):
+ return AutoencoderKLDiffusers(AutoencoderKLDiffusersConfig())
+
+ @pytest.fixture()
+ def sampling_noise_scheduler(self):
+ return FlowMatchEulerDiscreteScheduler()
+
+ @pytest.fixture()
+ def training_noise_scheduler(self):
+ return FlowMatchEulerDiscreteScheduler()
+
+ @pytest.fixture()
+ def model_config(self):
+ return LBMConfig(
+ source_key="source_image",
+ target_key="target_image",
+ )
+
+ @pytest.fixture()
+ def model_input(self):
+ return {
+ "source_image": torch.randn(2, 3, 256, 256).to(DEVICE),
+ "target_image": torch.randn(2, 3, 256, 256).to(DEVICE),
+ }
+
+ @pytest.fixture()
+ def model(
+ self,
+ model_config,
+ denoiser,
+ vae,
+ sampling_noise_scheduler,
+ training_noise_scheduler,
+ conditioner,
+ ):
+ return LBMModel(
+ config=model_config,
+ denoiser=denoiser,
+ vae=vae,
+ sampling_noise_scheduler=sampling_noise_scheduler,
+ training_noise_scheduler=training_noise_scheduler,
+ conditioner=conditioner,
+ ).to(DEVICE)
+
+ @torch.no_grad()
+ def test_model_forward(self, model, model_input):
+ model_output = model(
+ model_input,
+ )
+ assert model_output["loss"] > 0.0
+
+ def test_optimizers(self, model, model_input):
+ optimizer = torch.optim.Adam(model.denoiser.parameters(), lr=1e-4)
+
+ model.train()
+ model_init = deepcopy(model)
+ optimizer.zero_grad()
+ loss = model(model_input)["loss"]
+ loss.backward()
+ optimizer.step()
+ assert not torch.equal(
+ torch.cat([p.flatten() for p in model.denoiser.parameters()]),
+ torch.cat([p.flatten() for p in model_init.denoiser.parameters()]),
+ )
diff --git a/tests/test_unets/test_unets_wrappers.py b/tests/test_unets/test_unets_wrappers.py
new file mode 100644
index 0000000000000000000000000000000000000000..615702a01edf352892fede215660419118420acd
--- /dev/null
+++ b/tests/test_unets/test_unets_wrappers.py
@@ -0,0 +1,127 @@
+import pytest
+import torch
+
+from lbm.models.unets import DiffusersUNet2DCondWrapper, DiffusersUNet2DWrapper
+
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+
+class TestDiffusersUNet2DWrapper:
+ # simulates class conditioning
+ @pytest.fixture(params=[None, torch.randint(256, (2,)).to(DEVICE)])
+ def conditioning(self, request):
+ if request.param is not None:
+ return {"cond": {"vector": request.param}}
+ return None
+
+ # simulates a latent sample
+ @pytest.fixture()
+ def sample(self):
+ return torch.rand(2, 6, 32, 32).to(DEVICE)
+
+ # simulates a timestep
+ @pytest.fixture(
+ params=[10.0, torch.randint(1000, (2,), dtype=torch.float).to(DEVICE), 3]
+ )
+ def timesteps(self, request):
+ return request.param
+
+ def test_unet2d_wrapper(self, sample, timesteps, conditioning):
+ unet = DiffusersUNet2DWrapper(
+ sample_size=sample.shape[2:],
+ in_channels=sample.shape[1],
+ out_channels=3,
+ num_class_embeds=256 if conditioning else None,
+ ).to(DEVICE)
+ output = unet(sample, timesteps, conditioning)
+ assert output.shape == (
+ sample.shape[0],
+ 3,
+ sample.shape[2],
+ sample.shape[3],
+ )
+
+
+class TestDiffusersUNet2DCondWrapper:
+ # simulates class conditioning
+ @pytest.fixture(params=[None, torch.randn(2, 256).to(DEVICE)])
+ def vector_conditioning(self, request):
+ if request.param is not None:
+ return {"vector": request.param}
+ return None
+
+ # simulates crossattn conditioning '(always needed for conditional UNet2D)' (see diffusers/models/unet.py
+ @pytest.fixture()
+ def crossattn_conditioning(self):
+ return {"crossattn": torch.randn(2, 12, 123).to(DEVICE)}
+
+ # simulates concat conditioning
+ @pytest.fixture(params=[None, torch.randn(2, 2, 32, 32).to(DEVICE)])
+ def concat_conditioning(self, request):
+ if request.param is not None:
+ return {"concat": request.param}
+ return None
+
+ @pytest.fixture()
+ def conditioning(
+ self, vector_conditioning, crossattn_conditioning, concat_conditioning
+ ):
+ cond = dict(cond=crossattn_conditioning)
+ if vector_conditioning is not None:
+ cond["cond"].update(vector_conditioning)
+ if concat_conditioning is not None:
+ cond["cond"].update(concat_conditioning)
+ return cond
+
+ # simulates a latent sample
+ @pytest.fixture()
+ def sample(self):
+ return torch.rand(2, 6, 32, 32).to(DEVICE)
+
+ # simulates a timestep
+ @pytest.fixture(
+ params=[10.0, torch.randint(1000, (2,), dtype=torch.float).to(DEVICE), 3]
+ )
+ def timesteps(self, request):
+ return request.param
+
+ def test_unet2d_cond_wrapper(self, sample, timesteps, conditioning):
+ # for concat
+ in_channels = (
+ sample.shape[1] + conditioning["cond"]["concat"].shape[1]
+ if conditioning["cond"].get("concat", None) is not None
+ else sample.shape[1]
+ )
+
+ # for vector
+ class_embed_type = (
+ "projection" if conditioning["cond"].get("vector") is not None else None
+ )
+ projection_class_embeddings_input_dim = (
+ conditioning["cond"]["vector"].shape[1]
+ if conditioning["cond"].get("vector") is not None
+ else None
+ )
+
+ # for crossattn
+ cross_attention_dim = (
+ conditioning["cond"]["crossattn"].shape[2]
+ if conditioning["cond"].get("crossattn") is not None
+ else 1280
+ )
+
+ unet = DiffusersUNet2DCondWrapper(
+ sample_size=sample.shape[2:],
+ in_channels=in_channels,
+ out_channels=3,
+ class_embed_type=class_embed_type,
+ projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+ cross_attention_dim=cross_attention_dim,
+ ).to(DEVICE)
+ output = unet(sample, timesteps, conditioning)
+ assert output.shape == (
+ sample.shape[0],
+ 3,
+ sample.shape[2],
+ sample.shape[3],
+ )
diff --git a/tests/test_vaes/test_autoencoder.py b/tests/test_vaes/test_autoencoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..935478de2f794a90956b30a0abf8a3c580e0ff49
--- /dev/null
+++ b/tests/test_vaes/test_autoencoder.py
@@ -0,0 +1,44 @@
+import pytest
+import torch
+
+from lbm.models.vae import AutoencoderKLDiffusers, AutoencoderKLDiffusersConfig
+
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+
+class TestAutoencoderKLDiffusers:
+ @pytest.fixture(
+ params=[
+ dict(),
+ dict(
+ version="stabilityai/stable-diffusion-xl-base-1.0",
+ subfolder="vae",
+ ),
+ ]
+ )
+ def model_config(self, request):
+ return AutoencoderKLDiffusersConfig(
+ **request.param, tiling_size=(16, 16), tiling_overlap=(8, 8), batch_size=1
+ )
+
+ @pytest.fixture()
+ def model(self, model_config):
+ return AutoencoderKLDiffusers(model_config).to(DEVICE)
+
+ def test_model_initialization(self, model, model_config):
+ assert model.config == model_config
+
+ def test_encode(self, model):
+ x = torch.randn(2, 3, 32, 32).to(DEVICE)
+ z = model.encode(x)
+ assert z.shape == (2, 4, 4, 4)
+
+ def test_decode(self, model):
+ z = torch.randn(2, 4, 4, 4).to(DEVICE)
+ x = model.decode(z)
+ assert x.shape == (2, 3, 32, 32)
+
+ def test_decode_tiling(self, model):
+ z = torch.randn(2, 4, 32, 32).to(DEVICE)
+ x = model.decode(z)
+ assert x.shape == (2, 3, 256, 256)