shreyajn commited on
Commit
dd3cc3d
1 Parent(s): d7aa5c7

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +237 -0
README.md ADDED
@@ -0,0 +1,237 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: pytorch
3
+ license: agpl-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/yolov8_det/web-assets/banner.png)
12
+
13
+ # Yolo-v8-Detection: Optimized for Mobile Deployment
14
+ ## Real-time object detection optimized for mobile and edge
15
+
16
+ YoloV8 is a machine learning model that predicts bounding boxes and classes of objects in an image.
17
+
18
+ This model is an implementation of Yolo-v8-Detection found [here](https://github.com/ultralytics/ultralytics/tree/main/ultralytics/models/yolo/detect).
19
+ This repository provides scripts to run Yolo-v8-Detection on Qualcomm® devices.
20
+ More details on model performance across various devices, can be found
21
+ [here](https://aihub.qualcomm.com/models/yolov8_det).
22
+
23
+
24
+ ### Model Details
25
+
26
+ - **Model Type:** Object detection
27
+ - **Model Stats:**
28
+ - Model checkpoint: YoloV8-N
29
+ - Input resolution: 640x640
30
+ - Number of parameters: 3.18M
31
+ - Model size: 12.2 MB
32
+
33
+
34
+ | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model
35
+ | ---|---|---|---|---|---|---|---|
36
+ | Samsung Galaxy S23 Ultra (Android 13) | Snapdragon® 8 Gen 2 | TFLite | 9.251 ms | 0 - 3 MB | FP16 | NPU | [Yolo-v8-Detection.tflite](https://huggingface.co/qualcomm/Yolo-v8-Detection/blob/main/Yolo-v8-Detection.tflite)
37
+ | Samsung Galaxy S23 Ultra (Android 13) | Snapdragon® 8 Gen 2 | QNN Model Library | 7.043 ms | 5 - 19 MB | FP16 | NPU | [Yolo-v8-Detection.so](https://huggingface.co/qualcomm/Yolo-v8-Detection/blob/main/Yolo-v8-Detection.so)
38
+
39
+
40
+ ## Installation
41
+
42
+ This model can be installed as a Python package via pip.
43
+
44
+ ```bash
45
+ pip install "qai-hub-models[yolov8_det]"
46
+ ```
47
+
48
+
49
+
50
+ ## Configure Qualcomm® AI Hub to run this model on a cloud-hosted device
51
+
52
+ Sign-in to [Qualcomm® AI Hub](https://app.aihub.qualcomm.com/) with your
53
+ Qualcomm® ID. Once signed in navigate to `Account -> Settings -> API Token`.
54
+
55
+ With this API token, you can configure your client to run models on the cloud
56
+ hosted devices.
57
+ ```bash
58
+ qai-hub configure --api_token API_TOKEN
59
+ ```
60
+ Navigate to [docs](https://app.aihub.qualcomm.com/docs/) for more information.
61
+
62
+
63
+
64
+ ## Demo off target
65
+
66
+ The package contains a simple end-to-end demo that downloads pre-trained
67
+ weights and runs this model on a sample input.
68
+
69
+ ```bash
70
+ python -m qai_hub_models.models.yolov8_det.demo
71
+ ```
72
+
73
+ The above demo runs a reference implementation of pre-processing, model
74
+ inference, and post processing.
75
+
76
+ **NOTE**: If you want running in a Jupyter Notebook or Google Colab like
77
+ environment, please add the following to your cell (instead of the above).
78
+ ```
79
+ %run -m qai_hub_models.models.yolov8_det.demo
80
+ ```
81
+
82
+
83
+ ### Run model on a cloud-hosted device
84
+
85
+ In addition to the demo, you can also run the model on a cloud-hosted Qualcomm®
86
+ device. This script does the following:
87
+ * Performance check on-device on a cloud-hosted device
88
+ * Downloads compiled assets that can be deployed on-device for Android.
89
+ * Accuracy check between PyTorch and on-device outputs.
90
+
91
+ ```bash
92
+ python -m qai_hub_models.models.yolov8_det.export
93
+ ```
94
+
95
+ ```
96
+ Profile Job summary of Yolo-v8-Detection
97
+ --------------------------------------------------
98
+ Device: Samsung Galaxy S23 Ultra (13)
99
+ Estimated Inference Time: 9.25 ms
100
+ Estimated Peak Memory Range: 0.22-2.53 MB
101
+ Compute Units: NPU (300) | Total (300)
102
+
103
+ Profile Job summary of Yolo-v8-Detection
104
+ --------------------------------------------------
105
+ Device: Samsung Galaxy S23 Ultra (13)
106
+ Estimated Inference Time: 7.04 ms
107
+ Estimated Peak Memory Range: 4.71-18.66 MB
108
+ Compute Units: NPU (294) | Total (294)
109
+
110
+
111
+ ```
112
+ ## How does this work?
113
+
114
+ This [export script](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/Yolo-v8-Detection/export.py)
115
+ leverages [Qualcomm® AI Hub](https://aihub.qualcomm.com/) to optimize, validate, and deploy this model
116
+ on-device. Lets go through each step below in detail:
117
+
118
+ Step 1: **Compile model for on-device deployment**
119
+
120
+ To compile a PyTorch model for on-device deployment, we first trace the model
121
+ in memory using the `jit.trace` and then call the `submit_compile_job` API.
122
+
123
+ ```python
124
+ import torch
125
+
126
+ import qai_hub as hub
127
+ from qai_hub_models.models.yolov8_det import Model
128
+
129
+ # Load the model
130
+ torch_model = Model.from_pretrained()
131
+ torch_model.eval()
132
+
133
+ # Device
134
+ device = hub.Device("Samsung Galaxy S23")
135
+
136
+ # Trace model
137
+ input_shape = torch_model.get_input_spec()
138
+ sample_inputs = torch_model.sample_inputs()
139
+
140
+ pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])
141
+
142
+ # Compile model on a specific device
143
+ compile_job = hub.submit_compile_job(
144
+ model=pt_model,
145
+ device=device,
146
+ input_specs=torch_model.get_input_spec(),
147
+ )
148
+
149
+ # Get target model to run on-device
150
+ target_model = compile_job.get_target_model()
151
+
152
+ ```
153
+
154
+
155
+ Step 2: **Performance profiling on cloud-hosted device**
156
+
157
+ After compiling models from step 1. Models can be profiled model on-device using the
158
+ `target_model`. Note that this scripts runs the model on a device automatically
159
+ provisioned in the cloud. Once the job is submitted, you can navigate to a
160
+ provided job URL to view a variety of on-device performance metrics.
161
+ ```python
162
+ profile_job = hub.submit_profile_job(
163
+ model=target_model,
164
+ device=device,
165
+ )
166
+
167
+ ```
168
+
169
+ Step 3: **Verify on-device accuracy**
170
+
171
+ To verify the accuracy of the model on-device, you can run on-device inference
172
+ on sample input data on the same cloud hosted device.
173
+ ```python
174
+ input_data = torch_model.sample_inputs()
175
+ inference_job = hub.submit_inference_job(
176
+ model=target_model,
177
+ device=device,
178
+ inputs=input_data,
179
+ )
180
+
181
+ on_device_output = inference_job.download_output_data()
182
+
183
+ ```
184
+ With the output of the model, you can compute like PSNR, relative errors or
185
+ spot check the output with expected output.
186
+
187
+ **Note**: This on-device profiling and inference requires access to Qualcomm®
188
+ AI Hub. [Sign up for early access](https://aihub.qualcomm.com/sign-up).
189
+
190
+
191
+ ## Run demo on a cloud-hosted device
192
+
193
+ You can also run the demo on-device.
194
+
195
+ ```bash
196
+ python -m qai_hub_models.models.yolov8_det.demo --on-device
197
+ ```
198
+
199
+ **NOTE**: If you want running in a Jupyter Notebook or Google Colab like
200
+ environment, please add the following to your cell (instead of the above).
201
+ ```
202
+ %run -m qai_hub_models.models.yolov8_det.demo -- --on-device
203
+ ```
204
+
205
+
206
+ ## Deploying compiled model to Android
207
+
208
+
209
+ The models can be deployed using multiple runtimes:
210
+ - TensorFlow Lite (`.tflite` export): [This
211
+ tutorial](https://www.tensorflow.org/lite/android/quickstart) provides a
212
+ guide to deploy the .tflite model in an Android application.
213
+
214
+
215
+ - QNN (`.so` export ): This [sample
216
+ app](https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-50/sample_app.html)
217
+ provides instructions on how to use the `.so` shared library in an Android application.
218
+
219
+
220
+ ## View on Qualcomm® AI Hub
221
+ Get more details on Yolo-v8-Detection's performance across various devices [here](https://aihub.qualcomm.com/models/yolov8_det).
222
+ Explore all available models on [Qualcomm® AI Hub](https://aihub.qualcomm.com/)
223
+
224
+ ## License
225
+ - The license for the original implementation of Yolo-v8-Detection can be found
226
+ [here](https://github.com/ultralytics/ultralytics/blob/main/LICENSE).
227
+ - 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).
228
+
229
+ ## References
230
+ * [Real-Time Flying Object Detection with YOLOv8](https://arxiv.org/abs/2305.09972)
231
+ * [Source Model Implementation](https://github.com/ultralytics/ultralytics/tree/main/ultralytics/models/yolo/detect)
232
+
233
+ ## Community
234
+ * Join [our AI Hub Slack community](https://join.slack.com/t/qualcomm-ai-hub/shared_invite/zt-2dgf95loi-CXHTDRR1rvPgQWPO~ZZZJg) to collaborate, post questions and learn more about on-device AI.
235
+ * For questions or feedback please [reach out to us](mailto:[email protected]).
236
+
237
+