qaihm-bot commited on
Commit
634b722
·
verified ·
1 Parent(s): 11601ae

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +252 -0
README.md ADDED
@@ -0,0 +1,252 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: pytorch
3
+ license: gpl-3.0
4
+ pipeline_tag: object-detection
5
+ tags:
6
+ - real_time
7
+ - android
8
+
9
+ ---
10
+
11
+ ![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/yolov3/web-assets/model_demo.png)
12
+
13
+ # Yolo-v3: Optimized for Mobile Deployment
14
+ ## Real-time object detection optimized for mobile and edge
15
+
16
+
17
+ YoloV3 is a machine learning model that predicts bounding boxes and classes of objects in an image.
18
+
19
+ This model is an implementation of Yolo-v3 found [here](https://github.com/ultralytics/yolov3/tree/v8).
20
+
21
+
22
+ This repository provides scripts to run Yolo-v3 on Qualcomm® devices.
23
+ More details on model performance across various devices, can be found
24
+ [here](https://aihub.qualcomm.com/models/yolov3).
25
+
26
+
27
+ ### Model Details
28
+
29
+ - **Model Type:** Object detection
30
+ - **Model Stats:**
31
+ - Model checkpoint: YoloV3 Tiny
32
+ - Input resolution: 416p (416x416)
33
+ - Number of parameters: 8.85M
34
+ - Model size: 24.4 MB
35
+
36
+ | Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model
37
+ |---|---|---|---|---|---|---|---|---|
38
+ | Yolo-v3 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 24.935 ms | 0 - 17 MB | FP16 | NPU | [Yolo-v3.tflite](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.tflite) |
39
+ | Yolo-v3 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 10.405 ms | 5 - 21 MB | FP16 | NPU | [Yolo-v3.so](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.so) |
40
+ | Yolo-v3 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 14.65 ms | 10 - 23 MB | FP16 | NPU | [Yolo-v3.onnx](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.onnx) |
41
+ | Yolo-v3 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 17.648 ms | 0 - 79 MB | FP16 | NPU | [Yolo-v3.tflite](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.tflite) |
42
+ | Yolo-v3 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 7.427 ms | 5 - 38 MB | FP16 | NPU | [Yolo-v3.so](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.so) |
43
+ | Yolo-v3 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 11.103 ms | 0 - 62 MB | FP16 | NPU | [Yolo-v3.onnx](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.onnx) |
44
+ | Yolo-v3 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 15.003 ms | 0 - 76 MB | FP16 | NPU | [Yolo-v3.tflite](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.tflite) |
45
+ | Yolo-v3 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 7.731 ms | 5 - 33 MB | FP16 | NPU | Use Export Script |
46
+ | Yolo-v3 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 11.942 ms | 3 - 42 MB | FP16 | NPU | [Yolo-v3.onnx](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.onnx) |
47
+ | Yolo-v3 | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 24.289 ms | 0 - 15 MB | FP16 | NPU | [Yolo-v3.tflite](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.tflite) |
48
+ | Yolo-v3 | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 9.55 ms | 5 - 6 MB | FP16 | NPU | Use Export Script |
49
+ | Yolo-v3 | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 27.309 ms | 0 - 73 MB | FP16 | NPU | [Yolo-v3.tflite](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.tflite) |
50
+ | Yolo-v3 | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 12.115 ms | 5 - 34 MB | FP16 | NPU | Use Export Script |
51
+ | Yolo-v3 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 9.718 ms | 5 - 5 MB | FP16 | NPU | Use Export Script |
52
+ | Yolo-v3 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 15.732 ms | 5 - 5 MB | FP16 | NPU | [Yolo-v3.onnx](https://huggingface.co/qualcomm/Yolo-v3/blob/main/Yolo-v3.onnx) |
53
+
54
+
55
+
56
+
57
+ ## Installation
58
+
59
+ This model can be installed as a Python package via pip.
60
+
61
+ ```bash
62
+ pip install "qai-hub-models[yolov3]"
63
+ ```
64
+
65
+
66
+
67
+ ## Configure Qualcomm® AI Hub to run this model on a cloud-hosted device
68
+
69
+ Sign-in to [Qualcomm® AI Hub](https://app.aihub.qualcomm.com/) with your
70
+ Qualcomm® ID. Once signed in navigate to `Account -> Settings -> API Token`.
71
+
72
+ With this API token, you can configure your client to run models on the cloud
73
+ hosted devices.
74
+ ```bash
75
+ qai-hub configure --api_token API_TOKEN
76
+ ```
77
+ Navigate to [docs](https://app.aihub.qualcomm.com/docs/) for more information.
78
+
79
+
80
+
81
+ ## Demo off target
82
+
83
+ The package contains a simple end-to-end demo that downloads pre-trained
84
+ weights and runs this model on a sample input.
85
+
86
+ ```bash
87
+ python -m qai_hub_models.models.yolov3.demo
88
+ ```
89
+
90
+ The above demo runs a reference implementation of pre-processing, model
91
+ inference, and post processing.
92
+
93
+ **NOTE**: If you want running in a Jupyter Notebook or Google Colab like
94
+ environment, please add the following to your cell (instead of the above).
95
+ ```
96
+ %run -m qai_hub_models.models.yolov3.demo
97
+ ```
98
+
99
+
100
+ ### Run model on a cloud-hosted device
101
+
102
+ In addition to the demo, you can also run the model on a cloud-hosted Qualcomm®
103
+ device. This script does the following:
104
+ * Performance check on-device on a cloud-hosted device
105
+ * Downloads compiled assets that can be deployed on-device for Android.
106
+ * Accuracy check between PyTorch and on-device outputs.
107
+
108
+ ```bash
109
+ python -m qai_hub_models.models.yolov3.export
110
+ ```
111
+ ```
112
+ Profiling Results
113
+ ------------------------------------------------------------
114
+ Yolo-v3
115
+ Device : Samsung Galaxy S23 (13)
116
+ Runtime : TFLITE
117
+ Estimated inference time (ms) : 24.9
118
+ Estimated peak memory usage (MB): [0, 17]
119
+ Total # Ops : 163
120
+ Compute Unit(s) : NPU (163 ops)
121
+ ```
122
+
123
+
124
+ ## How does this work?
125
+
126
+ This [export script](https://aihub.qualcomm.com/models/yolov3/qai_hub_models/models/Yolo-v3/export.py)
127
+ leverages [Qualcomm® AI Hub](https://aihub.qualcomm.com/) to optimize, validate, and deploy this model
128
+ on-device. Lets go through each step below in detail:
129
+
130
+ Step 1: **Compile model for on-device deployment**
131
+
132
+ To compile a PyTorch model for on-device deployment, we first trace the model
133
+ in memory using the `jit.trace` and then call the `submit_compile_job` API.
134
+
135
+ ```python
136
+ import torch
137
+
138
+ import qai_hub as hub
139
+ from qai_hub_models.models.yolov3 import Model
140
+
141
+ # Load the model
142
+ torch_model = Model.from_pretrained()
143
+
144
+ # Device
145
+ device = hub.Device("Samsung Galaxy S23")
146
+
147
+ # Trace model
148
+ input_shape = torch_model.get_input_spec()
149
+ sample_inputs = torch_model.sample_inputs()
150
+
151
+ pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])
152
+
153
+ # Compile model on a specific device
154
+ compile_job = hub.submit_compile_job(
155
+ model=pt_model,
156
+ device=device,
157
+ input_specs=torch_model.get_input_spec(),
158
+ )
159
+
160
+ # Get target model to run on-device
161
+ target_model = compile_job.get_target_model()
162
+
163
+ ```
164
+
165
+
166
+ Step 2: **Performance profiling on cloud-hosted device**
167
+
168
+ After compiling models from step 1. Models can be profiled model on-device using the
169
+ `target_model`. Note that this scripts runs the model on a device automatically
170
+ provisioned in the cloud. Once the job is submitted, you can navigate to a
171
+ provided job URL to view a variety of on-device performance metrics.
172
+ ```python
173
+ profile_job = hub.submit_profile_job(
174
+ model=target_model,
175
+ device=device,
176
+ )
177
+
178
+ ```
179
+
180
+ Step 3: **Verify on-device accuracy**
181
+
182
+ To verify the accuracy of the model on-device, you can run on-device inference
183
+ on sample input data on the same cloud hosted device.
184
+ ```python
185
+ input_data = torch_model.sample_inputs()
186
+ inference_job = hub.submit_inference_job(
187
+ model=target_model,
188
+ device=device,
189
+ inputs=input_data,
190
+ )
191
+ on_device_output = inference_job.download_output_data()
192
+
193
+ ```
194
+ With the output of the model, you can compute like PSNR, relative errors or
195
+ spot check the output with expected output.
196
+
197
+ **Note**: This on-device profiling and inference requires access to Qualcomm®
198
+ AI Hub. [Sign up for access](https://myaccount.qualcomm.com/signup).
199
+
200
+
201
+
202
+ ## Run demo on a cloud-hosted device
203
+
204
+ You can also run the demo on-device.
205
+
206
+ ```bash
207
+ python -m qai_hub_models.models.yolov3.demo --on-device
208
+ ```
209
+
210
+ **NOTE**: If you want running in a Jupyter Notebook or Google Colab like
211
+ environment, please add the following to your cell (instead of the above).
212
+ ```
213
+ %run -m qai_hub_models.models.yolov3.demo -- --on-device
214
+ ```
215
+
216
+
217
+ ## Deploying compiled model to Android
218
+
219
+
220
+ The models can be deployed using multiple runtimes:
221
+ - TensorFlow Lite (`.tflite` export): [This
222
+ tutorial](https://www.tensorflow.org/lite/android/quickstart) provides a
223
+ guide to deploy the .tflite model in an Android application.
224
+
225
+
226
+ - QNN (`.so` export ): This [sample
227
+ app](https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-50/sample_app.html)
228
+ provides instructions on how to use the `.so` shared library in an Android application.
229
+
230
+
231
+ ## View on Qualcomm® AI Hub
232
+ Get more details on Yolo-v3's performance across various devices [here](https://aihub.qualcomm.com/models/yolov3).
233
+ Explore all available models on [Qualcomm® AI Hub](https://aihub.qualcomm.com/)
234
+
235
+
236
+ ## License
237
+ * The license for the original implementation of Yolo-v3 can be found [here](https://github.com/ultralytics/yolov3/blob/v8/LICENSE).
238
+ * The license for the compiled assets for on-device deployment can be found [here](https://github.com/ultralytics/yolov3/blob/v8/LICENSE)
239
+
240
+
241
+
242
+ ## References
243
+ * [YOLOv3: An Incremental Improvement](https://arxiv.org/abs/1804.02767)
244
+ * [Source Model Implementation](https://github.com/ultralytics/yolov3/tree/v8)
245
+
246
+
247
+
248
+ ## Community
249
+ * Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI.
250
+ * For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com).
251
+
252
+