Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -38,47 +38,46 @@ More details on model performance across various devices, can be found
|
|
| 38 |
|
| 39 |
| Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model
|
| 40 |
|---|---|---|---|---|---|---|---|---|
|
| 41 |
-
|
|
| 42 |
-
|
|
| 43 |
-
|
|
| 44 |
-
|
|
| 45 |
-
|
|
| 46 |
-
|
|
| 47 |
-
|
|
| 48 |
-
|
|
| 49 |
-
|
|
| 50 |
-
|
|
| 51 |
-
|
|
| 52 |
-
|
|
| 53 |
-
|
|
| 54 |
-
|
|
| 55 |
-
|
|
| 56 |
-
|
|
| 57 |
-
|
|
| 58 |
-
|
|
| 59 |
-
|
|
| 60 |
-
| WhisperDecoder |
|
| 61 |
-
|
|
| 62 |
-
|
|
| 63 |
-
|
|
| 64 |
-
|
|
| 65 |
-
|
|
| 66 |
-
|
|
| 67 |
-
|
|
| 68 |
-
|
|
| 69 |
-
|
|
| 70 |
-
|
|
| 71 |
-
|
|
| 72 |
-
|
|
| 73 |
-
|
|
| 74 |
-
|
|
| 75 |
-
|
|
| 76 |
-
|
|
| 77 |
-
|
|
| 78 |
-
|
|
| 79 |
-
|
|
| 80 |
-
|
|
| 81 |
-
| WhisperDecoder | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 52.917 ms | 232 - 232 MB | FP16 | NPU | [Whisper-Small-En.onnx](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.onnx) |
|
| 82 |
|
| 83 |
|
| 84 |
|
|
@@ -139,23 +138,23 @@ python -m qai_hub_models.models.whisper_small_en.export
|
|
| 139 |
```
|
| 140 |
```
|
| 141 |
Profiling Results
|
| 142 |
-
------------------------------------------------------------
|
| 143 |
-
WhisperEncoder
|
| 144 |
-
Device : Samsung Galaxy S23 (13)
|
| 145 |
-
Runtime : TFLITE
|
| 146 |
-
Estimated inference time (ms) : 722.7
|
| 147 |
-
Estimated peak memory usage (MB): [69, 449]
|
| 148 |
-
Total # Ops : 911
|
| 149 |
-
Compute Unit(s) : GPU (900 ops) CPU (11 ops)
|
| 150 |
-
|
| 151 |
------------------------------------------------------------
|
| 152 |
WhisperDecoder
|
| 153 |
Device : Samsung Galaxy S23 (13)
|
| 154 |
Runtime : TFLITE
|
| 155 |
-
Estimated inference time (ms) :
|
| 156 |
-
Estimated peak memory usage (MB): [
|
| 157 |
Total # Ops : 2573
|
| 158 |
Compute Unit(s) : NPU (2573 ops)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 159 |
```
|
| 160 |
|
| 161 |
|
|
@@ -178,42 +177,42 @@ from qai_hub_models.models.whisper_small_en import Model
|
|
| 178 |
|
| 179 |
# Load the model
|
| 180 |
model = Model.from_pretrained()
|
| 181 |
-
encoder_model = model.encoder
|
| 182 |
decoder_model = model.decoder
|
|
|
|
| 183 |
|
| 184 |
# Device
|
| 185 |
device = hub.Device("Samsung Galaxy S23")
|
| 186 |
|
| 187 |
# Trace model
|
| 188 |
-
|
| 189 |
-
|
| 190 |
|
| 191 |
-
|
| 192 |
|
| 193 |
# Compile model on a specific device
|
| 194 |
-
|
| 195 |
-
model=
|
| 196 |
device=device,
|
| 197 |
-
input_specs=
|
| 198 |
)
|
| 199 |
|
| 200 |
# Get target model to run on-device
|
| 201 |
-
|
| 202 |
# Trace model
|
| 203 |
-
|
| 204 |
-
|
| 205 |
|
| 206 |
-
|
| 207 |
|
| 208 |
# Compile model on a specific device
|
| 209 |
-
|
| 210 |
-
model=
|
| 211 |
device=device,
|
| 212 |
-
input_specs=
|
| 213 |
)
|
| 214 |
|
| 215 |
# Get target model to run on-device
|
| 216 |
-
|
| 217 |
|
| 218 |
```
|
| 219 |
|
|
@@ -225,14 +224,14 @@ After compiling models from step 1. Models can be profiled model on-device using
|
|
| 225 |
provisioned in the cloud. Once the job is submitted, you can navigate to a
|
| 226 |
provided job URL to view a variety of on-device performance metrics.
|
| 227 |
```python
|
| 228 |
-
encoder_profile_job = hub.submit_profile_job(
|
| 229 |
-
model=encoder_target_model,
|
| 230 |
-
device=device,
|
| 231 |
-
)
|
| 232 |
decoder_profile_job = hub.submit_profile_job(
|
| 233 |
model=decoder_target_model,
|
| 234 |
device=device,
|
| 235 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 236 |
|
| 237 |
```
|
| 238 |
|
|
@@ -241,13 +240,6 @@ Step 3: **Verify on-device accuracy**
|
|
| 241 |
To verify the accuracy of the model on-device, you can run on-device inference
|
| 242 |
on sample input data on the same cloud hosted device.
|
| 243 |
```python
|
| 244 |
-
encoder_input_data = encoder_model.sample_inputs()
|
| 245 |
-
encoder_inference_job = hub.submit_inference_job(
|
| 246 |
-
model=encoder_target_model,
|
| 247 |
-
device=device,
|
| 248 |
-
inputs=encoder_input_data,
|
| 249 |
-
)
|
| 250 |
-
encoder_inference_job.download_output_data()
|
| 251 |
decoder_input_data = decoder_model.sample_inputs()
|
| 252 |
decoder_inference_job = hub.submit_inference_job(
|
| 253 |
model=decoder_target_model,
|
|
@@ -255,6 +247,13 @@ decoder_inference_job = hub.submit_inference_job(
|
|
| 255 |
inputs=decoder_input_data,
|
| 256 |
)
|
| 257 |
decoder_inference_job.download_output_data()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 258 |
|
| 259 |
```
|
| 260 |
With the output of the model, you can compute like PSNR, relative errors or
|
|
|
|
| 38 |
|
| 39 |
| Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model
|
| 40 |
|---|---|---|---|---|---|---|---|---|
|
| 41 |
+
| WhisperDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 29.81 ms | 11 - 96 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) |
|
| 42 |
+
| WhisperDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 11.818 ms | 59 - 139 MB | FP16 | NPU | [Whisper-Small-En.so](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.so) |
|
| 43 |
+
| WhisperDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 23.422 ms | 16 - 148 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) |
|
| 44 |
+
| WhisperDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 9.293 ms | 61 - 168 MB | FP16 | NPU | [Whisper-Small-En.so](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.so) |
|
| 45 |
+
| WhisperDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 19.441 ms | 15 - 176 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) |
|
| 46 |
+
| WhisperDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 7.329 ms | 47 - 181 MB | FP16 | NPU | Use Export Script |
|
| 47 |
+
| WhisperDecoder | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 28.585 ms | 16 - 98 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) |
|
| 48 |
+
| WhisperDecoder | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 12.218 ms | 39 - 40 MB | FP16 | NPU | Use Export Script |
|
| 49 |
+
| WhisperDecoder | SA7255P ADP | SA7255P | QNN | 75.01 ms | 53 - 64 MB | FP16 | NPU | Use Export Script |
|
| 50 |
+
| WhisperDecoder | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 29.634 ms | 16 - 96 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) |
|
| 51 |
+
| WhisperDecoder | SA8255 (Proxy) | SA8255P Proxy | QNN | 11.938 ms | 61 - 62 MB | FP16 | NPU | Use Export Script |
|
| 52 |
+
| WhisperDecoder | SA8295P ADP | SA8295P | TFLITE | 31.029 ms | 16 - 163 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) |
|
| 53 |
+
| WhisperDecoder | SA8295P ADP | SA8295P | QNN | 14.525 ms | 57 - 63 MB | FP16 | NPU | Use Export Script |
|
| 54 |
+
| WhisperDecoder | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 28.772 ms | 16 - 98 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) |
|
| 55 |
+
| WhisperDecoder | SA8650 (Proxy) | SA8650P Proxy | QNN | 12.051 ms | 61 - 62 MB | FP16 | NPU | Use Export Script |
|
| 56 |
+
| WhisperDecoder | SA8775P ADP | SA8775P | TFLITE | 33.281 ms | 16 - 175 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) |
|
| 57 |
+
| WhisperDecoder | SA8775P ADP | SA8775P | QNN | 14.774 ms | 51 - 61 MB | FP16 | NPU | Use Export Script |
|
| 58 |
+
| WhisperDecoder | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 33.302 ms | 16 - 141 MB | FP16 | NPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperDecoder.tflite) |
|
| 59 |
+
| WhisperDecoder | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 16.624 ms | 53 - 171 MB | FP16 | NPU | Use Export Script |
|
| 60 |
+
| WhisperDecoder | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 10.781 ms | 61 - 61 MB | FP16 | NPU | Use Export Script |
|
| 61 |
+
| WhisperEncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 700.75 ms | 67 - 440 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) |
|
| 62 |
+
| WhisperEncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 823.812 ms | 0 - 209 MB | FP16 | NPU | [Whisper-Small-En.so](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.so) |
|
| 63 |
+
| WhisperEncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 532.216 ms | 0 - 87 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) |
|
| 64 |
+
| WhisperEncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 586.956 ms | 0 - 839 MB | FP16 | NPU | [Whisper-Small-En.so](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.so) |
|
| 65 |
+
| WhisperEncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 552.324 ms | 111 - 139 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) |
|
| 66 |
+
| WhisperEncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 513.813 ms | 0 - 906 MB | FP16 | NPU | Use Export Script |
|
| 67 |
+
| WhisperEncoder | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 677.595 ms | 92 - 467 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) |
|
| 68 |
+
| WhisperEncoder | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 671.038 ms | 1 - 2 MB | FP16 | NPU | Use Export Script |
|
| 69 |
+
| WhisperEncoder | SA7255P ADP | SA7255P | TFLITE | 4432.265 ms | 108 - 142 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) |
|
| 70 |
+
| WhisperEncoder | SA7255P ADP | SA7255P | QNN | 3212.772 ms | 0 - 10 MB | FP16 | NPU | Use Export Script |
|
| 71 |
+
| WhisperEncoder | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 714.664 ms | 42 - 434 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) |
|
| 72 |
+
| WhisperEncoder | SA8255 (Proxy) | SA8255P Proxy | QNN | 682.786 ms | 1 - 3 MB | FP16 | NPU | Use Export Script |
|
| 73 |
+
| WhisperEncoder | SA8295P ADP | SA8295P | TFLITE | 657.129 ms | 110 - 142 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) |
|
| 74 |
+
| WhisperEncoder | SA8295P ADP | SA8295P | QNN | 701.786 ms | 0 - 6 MB | FP16 | NPU | Use Export Script |
|
| 75 |
+
| WhisperEncoder | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 694.432 ms | 110 - 448 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) |
|
| 76 |
+
| WhisperEncoder | SA8650 (Proxy) | SA8650P Proxy | QNN | 684.938 ms | 1 - 3 MB | FP16 | NPU | Use Export Script |
|
| 77 |
+
| WhisperEncoder | SA8775P ADP | SA8775P | TFLITE | 1290.958 ms | 100 - 132 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) |
|
| 78 |
+
| WhisperEncoder | SA8775P ADP | SA8775P | QNN | 604.972 ms | 0 - 10 MB | FP16 | NPU | Use Export Script |
|
| 79 |
+
| WhisperEncoder | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 973.335 ms | 109 - 207 MB | FP16 | GPU | [Whisper-Small-En.tflite](https://huggingface.co/qualcomm/Whisper-Small-En/blob/main/WhisperEncoder.tflite) |
|
| 80 |
+
| WhisperEncoder | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 504.643 ms | 0 - 0 MB | FP16 | NPU | Use Export Script |
|
|
|
|
| 81 |
|
| 82 |
|
| 83 |
|
|
|
|
| 138 |
```
|
| 139 |
```
|
| 140 |
Profiling Results
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
------------------------------------------------------------
|
| 142 |
WhisperDecoder
|
| 143 |
Device : Samsung Galaxy S23 (13)
|
| 144 |
Runtime : TFLITE
|
| 145 |
+
Estimated inference time (ms) : 29.8
|
| 146 |
+
Estimated peak memory usage (MB): [11, 96]
|
| 147 |
Total # Ops : 2573
|
| 148 |
Compute Unit(s) : NPU (2573 ops)
|
| 149 |
+
|
| 150 |
+
------------------------------------------------------------
|
| 151 |
+
WhisperEncoder
|
| 152 |
+
Device : Samsung Galaxy S23 (13)
|
| 153 |
+
Runtime : TFLITE
|
| 154 |
+
Estimated inference time (ms) : 700.8
|
| 155 |
+
Estimated peak memory usage (MB): [67, 440]
|
| 156 |
+
Total # Ops : 911
|
| 157 |
+
Compute Unit(s) : GPU (900 ops) CPU (11 ops)
|
| 158 |
```
|
| 159 |
|
| 160 |
|
|
|
|
| 177 |
|
| 178 |
# Load the model
|
| 179 |
model = Model.from_pretrained()
|
|
|
|
| 180 |
decoder_model = model.decoder
|
| 181 |
+
encoder_model = model.encoder
|
| 182 |
|
| 183 |
# Device
|
| 184 |
device = hub.Device("Samsung Galaxy S23")
|
| 185 |
|
| 186 |
# Trace model
|
| 187 |
+
decoder_input_shape = decoder_model.get_input_spec()
|
| 188 |
+
decoder_sample_inputs = decoder_model.sample_inputs()
|
| 189 |
|
| 190 |
+
traced_decoder_model = torch.jit.trace(decoder_model, [torch.tensor(data[0]) for _, data in decoder_sample_inputs.items()])
|
| 191 |
|
| 192 |
# Compile model on a specific device
|
| 193 |
+
decoder_compile_job = hub.submit_compile_job(
|
| 194 |
+
model=traced_decoder_model ,
|
| 195 |
device=device,
|
| 196 |
+
input_specs=decoder_model.get_input_spec(),
|
| 197 |
)
|
| 198 |
|
| 199 |
# Get target model to run on-device
|
| 200 |
+
decoder_target_model = decoder_compile_job.get_target_model()
|
| 201 |
# Trace model
|
| 202 |
+
encoder_input_shape = encoder_model.get_input_spec()
|
| 203 |
+
encoder_sample_inputs = encoder_model.sample_inputs()
|
| 204 |
|
| 205 |
+
traced_encoder_model = torch.jit.trace(encoder_model, [torch.tensor(data[0]) for _, data in encoder_sample_inputs.items()])
|
| 206 |
|
| 207 |
# Compile model on a specific device
|
| 208 |
+
encoder_compile_job = hub.submit_compile_job(
|
| 209 |
+
model=traced_encoder_model ,
|
| 210 |
device=device,
|
| 211 |
+
input_specs=encoder_model.get_input_spec(),
|
| 212 |
)
|
| 213 |
|
| 214 |
# Get target model to run on-device
|
| 215 |
+
encoder_target_model = encoder_compile_job.get_target_model()
|
| 216 |
|
| 217 |
```
|
| 218 |
|
|
|
|
| 224 |
provisioned in the cloud. Once the job is submitted, you can navigate to a
|
| 225 |
provided job URL to view a variety of on-device performance metrics.
|
| 226 |
```python
|
|
|
|
|
|
|
|
|
|
|
|
|
| 227 |
decoder_profile_job = hub.submit_profile_job(
|
| 228 |
model=decoder_target_model,
|
| 229 |
device=device,
|
| 230 |
)
|
| 231 |
+
encoder_profile_job = hub.submit_profile_job(
|
| 232 |
+
model=encoder_target_model,
|
| 233 |
+
device=device,
|
| 234 |
+
)
|
| 235 |
|
| 236 |
```
|
| 237 |
|
|
|
|
| 240 |
To verify the accuracy of the model on-device, you can run on-device inference
|
| 241 |
on sample input data on the same cloud hosted device.
|
| 242 |
```python
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 243 |
decoder_input_data = decoder_model.sample_inputs()
|
| 244 |
decoder_inference_job = hub.submit_inference_job(
|
| 245 |
model=decoder_target_model,
|
|
|
|
| 247 |
inputs=decoder_input_data,
|
| 248 |
)
|
| 249 |
decoder_inference_job.download_output_data()
|
| 250 |
+
encoder_input_data = encoder_model.sample_inputs()
|
| 251 |
+
encoder_inference_job = hub.submit_inference_job(
|
| 252 |
+
model=encoder_target_model,
|
| 253 |
+
device=device,
|
| 254 |
+
inputs=encoder_input_data,
|
| 255 |
+
)
|
| 256 |
+
encoder_inference_job.download_output_data()
|
| 257 |
|
| 258 |
```
|
| 259 |
With the output of the model, you can compute like PSNR, relative errors or
|