qaihm-bot commited on
Commit
9511cdd
·
verified ·
1 Parent(s): ed92f41

See https://github.com/quic/ai-hub-models/releases/v0.30.5 for changelog.

Files changed (4) hide show
  1. Mobile-VIT.onnx +3 -0
  2. Mobile-VIT.tflite +3 -0
  3. Mobile-VIT_w8a16.onnx +3 -0
  4. README.md +270 -0
Mobile-VIT.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3d6a762787772c01f48b2558df319fd64c19eeb09c1e3f2294b1f874077d698e
3
+ size 22448772
Mobile-VIT.tflite ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:79dfe96dc20eff22356d4622464a78e2f908d38bffe2951369e954d26df2fd33
3
+ size 22400324
Mobile-VIT_w8a16.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6abc280771a0c3dcd6a07f8ebab46b89682df64fd9f9d65092af00bda5f08c0f
3
+ size 22899050
README.md ADDED
@@ -0,0 +1,270 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: pytorch
3
+ license: other
4
+ tags:
5
+ - backbone
6
+ - android
7
+ pipeline_tag: image-classification
8
+
9
+ ---
10
+
11
+ ![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/mobile_vit/web-assets/model_demo.png)
12
+
13
+ # Mobile-VIT: Optimized for Mobile Deployment
14
+ ## Imagenet classifier and general purpose backbone
15
+
16
+
17
+ MobileVit is a machine learning model that can classify images from the Imagenet dataset. It can also be used as a backbone in building more complex models for specific use cases.
18
+
19
+ This model is an implementation of Mobile-VIT found [here](https://github.com/apple/ml-cvnets).
20
+
21
+
22
+ This repository provides scripts to run Mobile-VIT on Qualcomm® devices.
23
+ More details on model performance across various devices, can be found
24
+ [here](https://aihub.qualcomm.com/models/mobile_vit).
25
+
26
+
27
+ ### Model Details
28
+
29
+ - **Model Type:** Model_use_case.image_classification
30
+ - **Model Stats:**
31
+ - Model checkpoint: Imagenet
32
+ - Input resolution: 224x224
33
+ - Number of parameters: None
34
+ - Model size (float): None
35
+
36
+ | Model | Precision | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit | Target Model
37
+ |---|---|---|---|---|---|---|---|---|
38
+ | Mobile-VIT | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | TFLITE | 42.749 ms | 0 - 43 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
39
+ | Mobile-VIT | float | QCS8275 (Proxy) | Qualcomm® QCS8275 (Proxy) | QNN | 41.881 ms | 1 - 11 MB | NPU | Use Export Script |
40
+ | Mobile-VIT | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | TFLITE | 6.725 ms | 0 - 52 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
41
+ | Mobile-VIT | float | QCS8450 (Proxy) | Qualcomm® QCS8450 (Proxy) | QNN | 6.479 ms | 1 - 48 MB | NPU | Use Export Script |
42
+ | Mobile-VIT | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | TFLITE | 4.795 ms | 0 - 70 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
43
+ | Mobile-VIT | float | QCS8550 (Proxy) | Qualcomm® QCS8550 (Proxy) | QNN | 4.334 ms | 1 - 4 MB | NPU | Use Export Script |
44
+ | Mobile-VIT | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | TFLITE | 5.624 ms | 0 - 44 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
45
+ | Mobile-VIT | float | QCS9075 (Proxy) | Qualcomm® QCS9075 (Proxy) | QNN | 5.144 ms | 1 - 12 MB | NPU | Use Export Script |
46
+ | Mobile-VIT | float | SA7255P ADP | Qualcomm® SA7255P | TFLITE | 42.749 ms | 0 - 43 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
47
+ | Mobile-VIT | float | SA7255P ADP | Qualcomm® SA7255P | QNN | 41.881 ms | 1 - 11 MB | NPU | Use Export Script |
48
+ | Mobile-VIT | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | TFLITE | 4.752 ms | 0 - 69 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
49
+ | Mobile-VIT | float | SA8255 (Proxy) | Qualcomm® SA8255P (Proxy) | QNN | 4.334 ms | 1 - 3 MB | NPU | Use Export Script |
50
+ | Mobile-VIT | float | SA8295P ADP | Qualcomm® SA8295P | TFLITE | 7.903 ms | 0 - 42 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
51
+ | Mobile-VIT | float | SA8295P ADP | Qualcomm® SA8295P | QNN | 7.332 ms | 1 - 17 MB | NPU | Use Export Script |
52
+ | Mobile-VIT | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | TFLITE | 4.753 ms | 0 - 71 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
53
+ | Mobile-VIT | float | SA8650 (Proxy) | Qualcomm® SA8650P (Proxy) | QNN | 4.32 ms | 1 - 3 MB | NPU | Use Export Script |
54
+ | Mobile-VIT | float | SA8775P ADP | Qualcomm® SA8775P | TFLITE | 5.624 ms | 0 - 44 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
55
+ | Mobile-VIT | float | SA8775P ADP | Qualcomm® SA8775P | QNN | 5.144 ms | 1 - 12 MB | NPU | Use Export Script |
56
+ | Mobile-VIT | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | TFLITE | 4.766 ms | 0 - 70 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
57
+ | Mobile-VIT | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | QNN | 4.345 ms | 1 - 13 MB | NPU | Use Export Script |
58
+ | Mobile-VIT | float | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | ONNX | 4.682 ms | 0 - 43 MB | NPU | [Mobile-VIT.onnx](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.onnx) |
59
+ | Mobile-VIT | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | TFLITE | 3.288 ms | 0 - 55 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
60
+ | Mobile-VIT | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | QNN | 2.918 ms | 0 - 55 MB | NPU | Use Export Script |
61
+ | Mobile-VIT | float | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 3.179 ms | 0 - 63 MB | NPU | [Mobile-VIT.onnx](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.onnx) |
62
+ | Mobile-VIT | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | TFLITE | 3.199 ms | 0 - 46 MB | NPU | [Mobile-VIT.tflite](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.tflite) |
63
+ | Mobile-VIT | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | QNN | 2.804 ms | 1 - 50 MB | NPU | Use Export Script |
64
+ | Mobile-VIT | float | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | ONNX | 2.547 ms | 0 - 51 MB | NPU | [Mobile-VIT.onnx](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.onnx) |
65
+ | Mobile-VIT | float | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 4.718 ms | 1 - 1 MB | NPU | Use Export Script |
66
+ | Mobile-VIT | float | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 5.012 ms | 12 - 12 MB | NPU | [Mobile-VIT.onnx](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT.onnx) |
67
+ | Mobile-VIT | w8a16 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 Mobile | ONNX | 13.426 ms | 16 - 65 MB | NPU | [Mobile-VIT.onnx](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT_w8a16.onnx) |
68
+ | Mobile-VIT | w8a16 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 Mobile | ONNX | 9.98 ms | 17 - 159 MB | NPU | [Mobile-VIT.onnx](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT_w8a16.onnx) |
69
+ | Mobile-VIT | w8a16 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite Mobile | ONNX | 8.743 ms | 16 - 131 MB | NPU | [Mobile-VIT.onnx](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT_w8a16.onnx) |
70
+ | Mobile-VIT | w8a16 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 13.048 ms | 29 - 29 MB | NPU | [Mobile-VIT.onnx](https://huggingface.co/qualcomm/Mobile-VIT/blob/main/Mobile-VIT_w8a16.onnx) |
71
+
72
+
73
+
74
+
75
+ ## Installation
76
+
77
+
78
+ Install the package via pip:
79
+ ```bash
80
+ pip install "qai-hub-models[mobile-vit]"
81
+ ```
82
+
83
+
84
+ ## Configure Qualcomm® AI Hub to run this model on a cloud-hosted device
85
+
86
+ Sign-in to [Qualcomm® AI Hub](https://app.aihub.qualcomm.com/) with your
87
+ Qualcomm® ID. Once signed in navigate to `Account -> Settings -> API Token`.
88
+
89
+ With this API token, you can configure your client to run models on the cloud
90
+ hosted devices.
91
+ ```bash
92
+ qai-hub configure --api_token API_TOKEN
93
+ ```
94
+ Navigate to [docs](https://app.aihub.qualcomm.com/docs/) for more information.
95
+
96
+
97
+
98
+ ## Demo off target
99
+
100
+ The package contains a simple end-to-end demo that downloads pre-trained
101
+ weights and runs this model on a sample input.
102
+
103
+ ```bash
104
+ python -m qai_hub_models.models.mobile_vit.demo
105
+ ```
106
+
107
+ The above demo runs a reference implementation of pre-processing, model
108
+ inference, and post processing.
109
+
110
+ **NOTE**: If you want running in a Jupyter Notebook or Google Colab like
111
+ environment, please add the following to your cell (instead of the above).
112
+ ```
113
+ %run -m qai_hub_models.models.mobile_vit.demo
114
+ ```
115
+
116
+
117
+ ### Run model on a cloud-hosted device
118
+
119
+ In addition to the demo, you can also run the model on a cloud-hosted Qualcomm®
120
+ device. This script does the following:
121
+ * Performance check on-device on a cloud-hosted device
122
+ * Downloads compiled assets that can be deployed on-device for Android.
123
+ * Accuracy check between PyTorch and on-device outputs.
124
+
125
+ ```bash
126
+ python -m qai_hub_models.models.mobile_vit.export
127
+ ```
128
+ ```
129
+ Profiling Results
130
+ ------------------------------------------------------------
131
+ Mobile-VIT
132
+ Device : cs_8275 (ANDROID 14)
133
+ Runtime : TFLITE
134
+ Estimated inference time (ms) : 42.7
135
+ Estimated peak memory usage (MB): [0, 43]
136
+ Total # Ops : 577
137
+ Compute Unit(s) : npu (577 ops) gpu (0 ops) cpu (0 ops)
138
+ ```
139
+
140
+
141
+ ## How does this work?
142
+
143
+ This [export script](https://aihub.qualcomm.com/models/mobile_vit/qai_hub_models/models/Mobile-VIT/export.py)
144
+ leverages [Qualcomm® AI Hub](https://aihub.qualcomm.com/) to optimize, validate, and deploy this model
145
+ on-device. Lets go through each step below in detail:
146
+
147
+ Step 1: **Compile model for on-device deployment**
148
+
149
+ To compile a PyTorch model for on-device deployment, we first trace the model
150
+ in memory using the `jit.trace` and then call the `submit_compile_job` API.
151
+
152
+ ```python
153
+ import torch
154
+
155
+ import qai_hub as hub
156
+ from qai_hub_models.models.mobile_vit import Model
157
+
158
+ # Load the model
159
+ torch_model = Model.from_pretrained()
160
+
161
+ # Device
162
+ device = hub.Device("Samsung Galaxy S24")
163
+
164
+ # Trace model
165
+ input_shape = torch_model.get_input_spec()
166
+ sample_inputs = torch_model.sample_inputs()
167
+
168
+ pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])
169
+
170
+ # Compile model on a specific device
171
+ compile_job = hub.submit_compile_job(
172
+ model=pt_model,
173
+ device=device,
174
+ input_specs=torch_model.get_input_spec(),
175
+ )
176
+
177
+ # Get target model to run on-device
178
+ target_model = compile_job.get_target_model()
179
+
180
+ ```
181
+
182
+
183
+ Step 2: **Performance profiling on cloud-hosted device**
184
+
185
+ After compiling models from step 1. Models can be profiled model on-device using the
186
+ `target_model`. Note that this scripts runs the model on a device automatically
187
+ provisioned in the cloud. Once the job is submitted, you can navigate to a
188
+ provided job URL to view a variety of on-device performance metrics.
189
+ ```python
190
+ profile_job = hub.submit_profile_job(
191
+ model=target_model,
192
+ device=device,
193
+ )
194
+
195
+ ```
196
+
197
+ Step 3: **Verify on-device accuracy**
198
+
199
+ To verify the accuracy of the model on-device, you can run on-device inference
200
+ on sample input data on the same cloud hosted device.
201
+ ```python
202
+ input_data = torch_model.sample_inputs()
203
+ inference_job = hub.submit_inference_job(
204
+ model=target_model,
205
+ device=device,
206
+ inputs=input_data,
207
+ )
208
+ on_device_output = inference_job.download_output_data()
209
+
210
+ ```
211
+ With the output of the model, you can compute like PSNR, relative errors or
212
+ spot check the output with expected output.
213
+
214
+ **Note**: This on-device profiling and inference requires access to Qualcomm®
215
+ AI Hub. [Sign up for access](https://myaccount.qualcomm.com/signup).
216
+
217
+
218
+
219
+ ## Run demo on a cloud-hosted device
220
+
221
+ You can also run the demo on-device.
222
+
223
+ ```bash
224
+ python -m qai_hub_models.models.mobile_vit.demo --eval-mode on-device
225
+ ```
226
+
227
+ **NOTE**: If you want running in a Jupyter Notebook or Google Colab like
228
+ environment, please add the following to your cell (instead of the above).
229
+ ```
230
+ %run -m qai_hub_models.models.mobile_vit.demo -- --eval-mode on-device
231
+ ```
232
+
233
+
234
+ ## Deploying compiled model to Android
235
+
236
+
237
+ The models can be deployed using multiple runtimes:
238
+ - TensorFlow Lite (`.tflite` export): [This
239
+ tutorial](https://www.tensorflow.org/lite/android/quickstart) provides a
240
+ guide to deploy the .tflite model in an Android application.
241
+
242
+
243
+ - QNN (`.so` export ): This [sample
244
+ app](https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-50/sample_app.html)
245
+ provides instructions on how to use the `.so` shared library in an Android application.
246
+
247
+
248
+ ## View on Qualcomm® AI Hub
249
+ Get more details on Mobile-VIT's performance across various devices [here](https://aihub.qualcomm.com/models/mobile_vit).
250
+ Explore all available models on [Qualcomm® AI Hub](https://aihub.qualcomm.com/)
251
+
252
+
253
+ ## License
254
+ * The license for the original implementation of Mobile-VIT can be found
255
+ [here](https://github.com/pytorch/vision/blob/main/LICENSE).
256
+ * The license for the compiled assets for on-device deployment can be found [here](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/Qualcomm+AI+Hub+Proprietary+License.pdf)
257
+
258
+
259
+
260
+ ## References
261
+ * [MOBILEVIT: LIGHT-WEIGHT, GENERAL-PURPOSE, AND MOBILE-FRIENDLY VISION TRANSFORMER](https://arxiv.org/abs/2110.02178)
262
+ * [Source Model Implementation](https://github.com/apple/ml-cvnets)
263
+
264
+
265
+
266
+ ## Community
267
+ * Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI.
268
+ * For questions or feedback please [reach out to us](mailto:[email protected]).
269
+
270
+