Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ 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
+mae-teaser.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1 @@
+checkpoint-*/

README.md

@@ -0,0 +1,9 @@
+# Generative Models Enable Generalizable Instance Segmentation
+<img src='mae-teaser.png'/>
+This is the official model release for the ImageNet-1k pretrained Masked Autoencoder (MAE-H) variant of our `gen2seg` generative instance segmenter. It is the same checkpoint we used to generate figures in the paper.  
+
+**The sementations in the image above are different from those in the paper teaser because they are generated with this MAE-H model instead of our SD model.** If you are looking for our Stable Diffusion 2 (SD) variant, you can find that at https://huggingface.co/reachomk/gen2seg-sd
+
+Please see our website https://reachomk.github.io/gen2seg for our paper, demos, and additional qualitative samples. 
+
+You can run this model at our GitHub: https://github.com/reachomk/gen2seg or our Huggingface Space: https://huggingface.co/spaces/reachomk/gen2seg-mae-h

arguments.txt

@@ -0,0 +1,31 @@
+modality: instance
+noise_type: None
+lr_exp_warmup_steps: 100
+lr_total_iter_length: 40000
+pretrained_model_name_or_path: facebook/vit-mae-huge
+revision: None
+variant: None
+output_dir: model-finetuned/mae_full_e2e_ft_mean_mixed_sqrtsep_meansep_norm
+seed: 500
+train_batch_size: 2
+num_train_epochs: 15
+max_train_steps: 30000
+gradient_accumulation_steps: 4
+gradient_checkpointing: True
+learning_rate: 6e-05
+dataloader_num_workers: 0
+adam_beta1: 0.9
+adam_beta2: 0.999
+adam_weight_decay: 0.01
+adam_epsilon: 1e-08
+max_grad_norm: 1.0
+logging_dir: logs
+mixed_precision: no
+report_to: tensorboard
+local_rank: 0
+checkpointing_steps: 10000
+checkpoints_total_limit: None
+resume_from_checkpoint: None
+enable_xformers_memory_efficient_attention: True
+tracker_project_name: e2e-ft-diffusion
+random_state_file: /nfs_share3/om/diffusion-e2e-ft/model-finetuned/stable_diffusion_e2e_ft_instance_10k/checkpoint-10000/random_states_0.pkl

config.json

@@ -0,0 +1,28 @@
+{
+  "_name_or_path": "facebook/vit-mae-huge",
+  "architectures": [
+    "ViTMAEForPreTraining"
+  ],
+  "attention_probs_dropout_prob": 0.0,
+  "decoder_hidden_size": 512,
+  "decoder_intermediate_size": 2048,
+  "decoder_num_attention_heads": 16,
+  "decoder_num_hidden_layers": 8,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.0,
+  "hidden_size": 1280,
+  "image_size": 224,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-12,
+  "mask_ratio": 0.0,
+  "model_type": "vit_mae",
+  "norm_pix_loss": false,
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 32,
+  "patch_size": 14,
+  "qkv_bias": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.47.0"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8976428a93ecf9dbd823764be3ffab46dbc3bf0356d672adbbf41d65ab23a23e
+size 2628376456