Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -34,24 +34,38 @@ More details on model performance across various devices, can be found
|
|
| 34 |
|
| 35 |
| Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model
|
| 36 |
|---|---|---|---|---|---|---|---|---|
|
| 37 |
-
| Facial-Landmark-Detection-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 0.
|
| 38 |
-
| Facial-Landmark-Detection-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 |
|
| 39 |
-
| Facial-Landmark-Detection-Quantized | Samsung Galaxy
|
| 40 |
-
| Facial-Landmark-Detection-Quantized | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 |
|
| 41 |
-
| Facial-Landmark-Detection-Quantized |
|
| 42 |
-
| Facial-Landmark-Detection-Quantized |
|
| 43 |
-
| Facial-Landmark-Detection-Quantized |
|
| 44 |
-
| Facial-Landmark-Detection-Quantized |
|
| 45 |
-
| Facial-Landmark-Detection-Quantized |
|
| 46 |
-
| Facial-Landmark-Detection-Quantized |
|
| 47 |
-
| Facial-Landmark-Detection-Quantized |
|
| 48 |
-
| Facial-Landmark-Detection-Quantized |
|
| 49 |
-
| Facial-Landmark-Detection-Quantized |
|
| 50 |
-
| Facial-Landmark-Detection-Quantized |
|
| 51 |
-
| Facial-Landmark-Detection-Quantized |
|
| 52 |
-
| Facial-Landmark-Detection-Quantized |
|
| 53 |
-
| Facial-Landmark-Detection-Quantized |
|
| 54 |
-
| Facial-Landmark-Detection-Quantized |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
|
| 56 |
|
| 57 |
|
|
@@ -116,12 +130,88 @@ Facial-Landmark-Detection-Quantized
|
|
| 116 |
Device : Samsung Galaxy S23 (13)
|
| 117 |
Runtime : TFLITE
|
| 118 |
Estimated inference time (ms) : 0.2
|
| 119 |
-
Estimated peak memory usage (MB): [0,
|
| 120 |
-
Total # Ops :
|
| 121 |
-
Compute Unit(s) : NPU (
|
| 122 |
```
|
| 123 |
|
| 124 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 125 |
|
| 126 |
|
| 127 |
## Run demo on a cloud-hosted device
|
|
@@ -166,7 +256,6 @@ Explore all available models on [Qualcomm® AI Hub](https://aihub.qualcomm.com/)
|
|
| 166 |
|
| 167 |
|
| 168 |
## References
|
| 169 |
-
* [None](None)
|
| 170 |
* [Source Model Implementation](https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py)
|
| 171 |
|
| 172 |
|
|
|
|
| 34 |
|
| 35 |
| Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model
|
| 36 |
|---|---|---|---|---|---|---|---|---|
|
| 37 |
+
| Facial-Landmark-Detection-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 0.169 ms | 0 - 37 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 38 |
+
| Facial-Landmark-Detection-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 0.17 ms | 0 - 2 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.so](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.so) |
|
| 39 |
+
| Facial-Landmark-Detection-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 0.474 ms | 0 - 29 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.onnx](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.onnx) |
|
| 40 |
+
| Facial-Landmark-Detection-Quantized | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 0.137 ms | 0 - 27 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 41 |
+
| Facial-Landmark-Detection-Quantized | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 0.125 ms | 0 - 20 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.so](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.so) |
|
| 42 |
+
| Facial-Landmark-Detection-Quantized | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 0.347 ms | 0 - 35 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.onnx](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.onnx) |
|
| 43 |
+
| Facial-Landmark-Detection-Quantized | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 0.151 ms | 0 - 13 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 44 |
+
| Facial-Landmark-Detection-Quantized | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 0.128 ms | 0 - 16 MB | INT8 | NPU | Use Export Script |
|
| 45 |
+
| Facial-Landmark-Detection-Quantized | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 0.328 ms | 0 - 20 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.onnx](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.onnx) |
|
| 46 |
+
| Facial-Landmark-Detection-Quantized | SA7255P ADP | SA7255P | TFLITE | 1.121 ms | 0 - 11 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 47 |
+
| Facial-Landmark-Detection-Quantized | SA7255P ADP | SA7255P | QNN | 1.08 ms | 0 - 9 MB | INT8 | NPU | Use Export Script |
|
| 48 |
+
| Facial-Landmark-Detection-Quantized | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 0.174 ms | 0 - 37 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 49 |
+
| Facial-Landmark-Detection-Quantized | SA8255 (Proxy) | SA8255P Proxy | QNN | 0.17 ms | 0 - 2 MB | INT8 | NPU | Use Export Script |
|
| 50 |
+
| Facial-Landmark-Detection-Quantized | SA8295P ADP | SA8295P | TFLITE | 0.468 ms | 0 - 16 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 51 |
+
| Facial-Landmark-Detection-Quantized | SA8295P ADP | SA8295P | QNN | 0.442 ms | 0 - 18 MB | INT8 | NPU | Use Export Script |
|
| 52 |
+
| Facial-Landmark-Detection-Quantized | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 0.17 ms | 0 - 37 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 53 |
+
| Facial-Landmark-Detection-Quantized | SA8650 (Proxy) | SA8650P Proxy | QNN | 0.169 ms | 0 - 3 MB | INT8 | NPU | Use Export Script |
|
| 54 |
+
| Facial-Landmark-Detection-Quantized | SA8775P ADP | SA8775P | TFLITE | 0.371 ms | 0 - 11 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 55 |
+
| Facial-Landmark-Detection-Quantized | SA8775P ADP | SA8775P | QNN | 0.35 ms | 0 - 10 MB | INT8 | NPU | Use Export Script |
|
| 56 |
+
| Facial-Landmark-Detection-Quantized | RB3 Gen 2 (Proxy) | QCS6490 Proxy | TFLITE | 0.525 ms | 0 - 24 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 57 |
+
| Facial-Landmark-Detection-Quantized | RB3 Gen 2 (Proxy) | QCS6490 Proxy | QNN | 0.589 ms | 0 - 15 MB | INT8 | NPU | Use Export Script |
|
| 58 |
+
| Facial-Landmark-Detection-Quantized | RB5 (Proxy) | QCS8250 Proxy | TFLITE | 1.627 ms | 0 - 3 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 59 |
+
| Facial-Landmark-Detection-Quantized | QCS8275 (Proxy) | QCS8275 Proxy | TFLITE | 1.121 ms | 0 - 11 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 60 |
+
| Facial-Landmark-Detection-Quantized | QCS8275 (Proxy) | QCS8275 Proxy | QNN | 1.08 ms | 0 - 9 MB | INT8 | NPU | Use Export Script |
|
| 61 |
+
| Facial-Landmark-Detection-Quantized | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 0.172 ms | 0 - 37 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 62 |
+
| Facial-Landmark-Detection-Quantized | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 0.166 ms | 0 - 2 MB | INT8 | NPU | Use Export Script |
|
| 63 |
+
| Facial-Landmark-Detection-Quantized | QCS9075 (Proxy) | QCS9075 Proxy | TFLITE | 0.371 ms | 0 - 11 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 64 |
+
| Facial-Landmark-Detection-Quantized | QCS9075 (Proxy) | QCS9075 Proxy | QNN | 0.35 ms | 0 - 10 MB | INT8 | NPU | Use Export Script |
|
| 65 |
+
| Facial-Landmark-Detection-Quantized | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 0.282 ms | 0 - 29 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.tflite](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.tflite) |
|
| 66 |
+
| Facial-Landmark-Detection-Quantized | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 0.276 ms | 0 - 27 MB | INT8 | NPU | Use Export Script |
|
| 67 |
+
| Facial-Landmark-Detection-Quantized | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 0.229 ms | 1 - 1 MB | INT8 | NPU | Use Export Script |
|
| 68 |
+
| Facial-Landmark-Detection-Quantized | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 0.445 ms | 4 - 4 MB | INT8 | NPU | [Facial-Landmark-Detection-Quantized.onnx](https://huggingface.co/qualcomm/Facial-Landmark-Detection-Quantized/blob/main/Facial-Landmark-Detection-Quantized.onnx) |
|
| 69 |
|
| 70 |
|
| 71 |
|
|
|
|
| 130 |
Device : Samsung Galaxy S23 (13)
|
| 131 |
Runtime : TFLITE
|
| 132 |
Estimated inference time (ms) : 0.2
|
| 133 |
+
Estimated peak memory usage (MB): [0, 37]
|
| 134 |
+
Total # Ops : 39
|
| 135 |
+
Compute Unit(s) : NPU (39 ops)
|
| 136 |
```
|
| 137 |
|
| 138 |
|
| 139 |
+
## How does this work?
|
| 140 |
+
|
| 141 |
+
This [export script](https://aihub.qualcomm.com/models/facemap_3dmm_quantized/qai_hub_models/models/Facial-Landmark-Detection-Quantized/export.py)
|
| 142 |
+
leverages [Qualcomm® AI Hub](https://aihub.qualcomm.com/) to optimize, validate, and deploy this model
|
| 143 |
+
on-device. Lets go through each step below in detail:
|
| 144 |
+
|
| 145 |
+
Step 1: **Compile model for on-device deployment**
|
| 146 |
+
|
| 147 |
+
To compile a PyTorch model for on-device deployment, we first trace the model
|
| 148 |
+
in memory using the `jit.trace` and then call the `submit_compile_job` API.
|
| 149 |
+
|
| 150 |
+
```python
|
| 151 |
+
import torch
|
| 152 |
+
|
| 153 |
+
import qai_hub as hub
|
| 154 |
+
from qai_hub_models.models.facemap_3dmm_quantized import Model
|
| 155 |
+
|
| 156 |
+
# Load the model
|
| 157 |
+
torch_model = Model.from_pretrained()
|
| 158 |
+
|
| 159 |
+
# Device
|
| 160 |
+
device = hub.Device("Samsung Galaxy S24")
|
| 161 |
+
|
| 162 |
+
# Trace model
|
| 163 |
+
input_shape = torch_model.get_input_spec()
|
| 164 |
+
sample_inputs = torch_model.sample_inputs()
|
| 165 |
+
|
| 166 |
+
pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])
|
| 167 |
+
|
| 168 |
+
# Compile model on a specific device
|
| 169 |
+
compile_job = hub.submit_compile_job(
|
| 170 |
+
model=pt_model,
|
| 171 |
+
device=device,
|
| 172 |
+
input_specs=torch_model.get_input_spec(),
|
| 173 |
+
)
|
| 174 |
+
|
| 175 |
+
# Get target model to run on-device
|
| 176 |
+
target_model = compile_job.get_target_model()
|
| 177 |
+
|
| 178 |
+
```
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
Step 2: **Performance profiling on cloud-hosted device**
|
| 182 |
+
|
| 183 |
+
After compiling models from step 1. Models can be profiled model on-device using the
|
| 184 |
+
`target_model`. Note that this scripts runs the model on a device automatically
|
| 185 |
+
provisioned in the cloud. Once the job is submitted, you can navigate to a
|
| 186 |
+
provided job URL to view a variety of on-device performance metrics.
|
| 187 |
+
```python
|
| 188 |
+
profile_job = hub.submit_profile_job(
|
| 189 |
+
model=target_model,
|
| 190 |
+
device=device,
|
| 191 |
+
)
|
| 192 |
+
|
| 193 |
+
```
|
| 194 |
+
|
| 195 |
+
Step 3: **Verify on-device accuracy**
|
| 196 |
+
|
| 197 |
+
To verify the accuracy of the model on-device, you can run on-device inference
|
| 198 |
+
on sample input data on the same cloud hosted device.
|
| 199 |
+
```python
|
| 200 |
+
input_data = torch_model.sample_inputs()
|
| 201 |
+
inference_job = hub.submit_inference_job(
|
| 202 |
+
model=target_model,
|
| 203 |
+
device=device,
|
| 204 |
+
inputs=input_data,
|
| 205 |
+
)
|
| 206 |
+
on_device_output = inference_job.download_output_data()
|
| 207 |
+
|
| 208 |
+
```
|
| 209 |
+
With the output of the model, you can compute like PSNR, relative errors or
|
| 210 |
+
spot check the output with expected output.
|
| 211 |
+
|
| 212 |
+
**Note**: This on-device profiling and inference requires access to Qualcomm®
|
| 213 |
+
AI Hub. [Sign up for access](https://myaccount.qualcomm.com/signup).
|
| 214 |
+
|
| 215 |
|
| 216 |
|
| 217 |
## Run demo on a cloud-hosted device
|
|
|
|
| 256 |
|
| 257 |
|
| 258 |
## References
|
|
|
|
| 259 |
* [Source Model Implementation](https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py)
|
| 260 |
|
| 261 |
|