Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -35,48 +35,47 @@ More details on model performance across various devices, can be found
|
|
| 35 |
|
| 36 |
| Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model
|
| 37 |
|---|---|---|---|---|---|---|---|---|
|
| 38 |
-
| Midas-V2-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 1.
|
| 39 |
| Midas-V2-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 1.439 ms | 0 - 64 MB | INT8 | NPU | [Midas-V2-Quantized.so](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.so) |
|
| 40 |
-
| Midas-V2-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 9.
|
| 41 |
-
| Midas-V2-Quantized | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 0.
|
| 42 |
-
| Midas-V2-Quantized | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 1.
|
| 43 |
-
| Midas-V2-Quantized | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 7.
|
| 44 |
-
| Midas-V2-Quantized | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 0.
|
| 45 |
-
| Midas-V2-Quantized | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN |
|
| 46 |
-
| Midas-V2-Quantized | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 7.
|
| 47 |
-
| Midas-V2-Quantized | RB3 Gen 2 (Proxy) | QCS6490 Proxy | TFLITE | 3.
|
| 48 |
-
| Midas-V2-Quantized | RB3 Gen 2 (Proxy) | QCS6490 Proxy | QNN | 5.
|
| 49 |
-
| Midas-V2-Quantized | RB5 (Proxy) | QCS8250 Proxy | TFLITE | 15.
|
| 50 |
-
| Midas-V2-Quantized | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 1.
|
| 51 |
-
| Midas-V2-Quantized | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 1.
|
| 52 |
-
| Midas-V2-Quantized | SA7255P ADP | SA7255P | TFLITE |
|
| 53 |
-
| Midas-V2-Quantized | SA7255P ADP | SA7255P | QNN | 12.
|
| 54 |
-
| Midas-V2-Quantized | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 1.
|
| 55 |
-
| Midas-V2-Quantized | SA8255 (Proxy) | SA8255P Proxy | QNN | 1.
|
| 56 |
-
| Midas-V2-Quantized | SA8295P ADP | SA8295P | TFLITE | 1.
|
| 57 |
-
| Midas-V2-Quantized | SA8295P ADP | SA8295P | QNN | 2.
|
| 58 |
-
| Midas-V2-Quantized | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 1.
|
| 59 |
-
| Midas-V2-Quantized | SA8650 (Proxy) | SA8650P Proxy | QNN | 1.
|
| 60 |
-
| Midas-V2-Quantized | SA8775P ADP | SA8775P | TFLITE | 1.
|
| 61 |
-
| Midas-V2-Quantized | SA8775P ADP | SA8775P | QNN | 2.
|
| 62 |
-
| Midas-V2-Quantized | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 1.
|
| 63 |
-
| Midas-V2-Quantized | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 1.
|
| 64 |
-
| Midas-V2-Quantized | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 1.
|
| 65 |
-
| Midas-V2-Quantized | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 8.
|
| 66 |
|
| 67 |
|
| 68 |
|
| 69 |
|
| 70 |
## Installation
|
| 71 |
|
| 72 |
-
This model can be installed as a Python package via pip.
|
| 73 |
|
|
|
|
| 74 |
```bash
|
| 75 |
-
pip install "qai-hub-models[
|
| 76 |
```
|
| 77 |
|
| 78 |
|
| 79 |
-
|
| 80 |
## Configure Qualcomm® AI Hub to run this model on a cloud-hosted device
|
| 81 |
|
| 82 |
Sign-in to [Qualcomm® AI Hub](https://app.aihub.qualcomm.com/) with your
|
|
@@ -128,12 +127,88 @@ Midas-V2-Quantized
|
|
| 128 |
Device : Samsung Galaxy S23 (13)
|
| 129 |
Runtime : TFLITE
|
| 130 |
Estimated inference time (ms) : 1.1
|
| 131 |
-
Estimated peak memory usage (MB): [0,
|
| 132 |
Total # Ops : 145
|
| 133 |
Compute Unit(s) : NPU (145 ops)
|
| 134 |
```
|
| 135 |
|
| 136 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 137 |
|
| 138 |
|
| 139 |
## Run demo on a cloud-hosted device
|
|
@@ -171,7 +246,8 @@ Explore all available models on [Qualcomm® AI Hub](https://aihub.qualcomm.com/)
|
|
| 171 |
|
| 172 |
|
| 173 |
## License
|
| 174 |
-
* The license for the original implementation of Midas-V2-Quantized can be found
|
|
|
|
| 175 |
* 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)
|
| 176 |
|
| 177 |
|
|
|
|
| 35 |
|
| 36 |
| Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model
|
| 37 |
|---|---|---|---|---|---|---|---|---|
|
| 38 |
+
| Midas-V2-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 1.097 ms | 0 - 60 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 39 |
| Midas-V2-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 1.439 ms | 0 - 64 MB | INT8 | NPU | [Midas-V2-Quantized.so](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.so) |
|
| 40 |
+
| Midas-V2-Quantized | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 9.111 ms | 0 - 80 MB | INT8 | NPU | [Midas-V2-Quantized.onnx](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.onnx) |
|
| 41 |
+
| Midas-V2-Quantized | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 0.76 ms | 0 - 35 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 42 |
+
| Midas-V2-Quantized | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 1.012 ms | 0 - 27 MB | INT8 | NPU | [Midas-V2-Quantized.so](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.so) |
|
| 43 |
+
| Midas-V2-Quantized | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 7.0 ms | 0 - 144 MB | INT8 | NPU | [Midas-V2-Quantized.onnx](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.onnx) |
|
| 44 |
+
| Midas-V2-Quantized | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 0.723 ms | 0 - 32 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 45 |
+
| Midas-V2-Quantized | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 0.976 ms | 0 - 29 MB | INT8 | NPU | Use Export Script |
|
| 46 |
+
| Midas-V2-Quantized | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 7.494 ms | 0 - 130 MB | INT8 | NPU | [Midas-V2-Quantized.onnx](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.onnx) |
|
| 47 |
+
| Midas-V2-Quantized | RB3 Gen 2 (Proxy) | QCS6490 Proxy | TFLITE | 3.796 ms | 0 - 27 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 48 |
+
| Midas-V2-Quantized | RB3 Gen 2 (Proxy) | QCS6490 Proxy | QNN | 5.825 ms | 0 - 12 MB | INT8 | NPU | Use Export Script |
|
| 49 |
+
| Midas-V2-Quantized | RB5 (Proxy) | QCS8250 Proxy | TFLITE | 15.706 ms | 0 - 3 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 50 |
+
| Midas-V2-Quantized | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 1.097 ms | 0 - 59 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 51 |
+
| Midas-V2-Quantized | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 1.318 ms | 0 - 3 MB | INT8 | NPU | Use Export Script |
|
| 52 |
+
| Midas-V2-Quantized | SA7255P ADP | SA7255P | TFLITE | 10.93 ms | 0 - 22 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 53 |
+
| Midas-V2-Quantized | SA7255P ADP | SA7255P | QNN | 12.132 ms | 0 - 10 MB | INT8 | NPU | Use Export Script |
|
| 54 |
+
| Midas-V2-Quantized | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 1.099 ms | 0 - 48 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 55 |
+
| Midas-V2-Quantized | SA8255 (Proxy) | SA8255P Proxy | QNN | 1.342 ms | 0 - 2 MB | INT8 | NPU | Use Export Script |
|
| 56 |
+
| Midas-V2-Quantized | SA8295P ADP | SA8295P | TFLITE | 1.948 ms | 0 - 29 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 57 |
+
| Midas-V2-Quantized | SA8295P ADP | SA8295P | QNN | 2.531 ms | 0 - 14 MB | INT8 | NPU | Use Export Script |
|
| 58 |
+
| Midas-V2-Quantized | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 1.096 ms | 0 - 60 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 59 |
+
| Midas-V2-Quantized | SA8650 (Proxy) | SA8650P Proxy | QNN | 1.327 ms | 0 - 2 MB | INT8 | NPU | Use Export Script |
|
| 60 |
+
| Midas-V2-Quantized | SA8775P ADP | SA8775P | TFLITE | 1.611 ms | 0 - 22 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 61 |
+
| Midas-V2-Quantized | SA8775P ADP | SA8775P | QNN | 2.122 ms | 0 - 10 MB | INT8 | NPU | Use Export Script |
|
| 62 |
+
| Midas-V2-Quantized | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 1.427 ms | 0 - 33 MB | INT8 | NPU | [Midas-V2-Quantized.tflite](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.tflite) |
|
| 63 |
+
| Midas-V2-Quantized | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 1.859 ms | 0 - 30 MB | INT8 | NPU | Use Export Script |
|
| 64 |
+
| Midas-V2-Quantized | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 1.465 ms | 0 - 0 MB | INT8 | NPU | Use Export Script |
|
| 65 |
+
| Midas-V2-Quantized | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 8.898 ms | 23 - 23 MB | INT8 | NPU | [Midas-V2-Quantized.onnx](https://huggingface.co/qualcomm/Midas-V2-Quantized/blob/main/Midas-V2-Quantized.onnx) |
|
| 66 |
|
| 67 |
|
| 68 |
|
| 69 |
|
| 70 |
## Installation
|
| 71 |
|
|
|
|
| 72 |
|
| 73 |
+
Install the package via pip:
|
| 74 |
```bash
|
| 75 |
+
pip install "qai-hub-models[midas-quantized]"
|
| 76 |
```
|
| 77 |
|
| 78 |
|
|
|
|
| 79 |
## Configure Qualcomm® AI Hub to run this model on a cloud-hosted device
|
| 80 |
|
| 81 |
Sign-in to [Qualcomm® AI Hub](https://app.aihub.qualcomm.com/) with your
|
|
|
|
| 127 |
Device : Samsung Galaxy S23 (13)
|
| 128 |
Runtime : TFLITE
|
| 129 |
Estimated inference time (ms) : 1.1
|
| 130 |
+
Estimated peak memory usage (MB): [0, 60]
|
| 131 |
Total # Ops : 145
|
| 132 |
Compute Unit(s) : NPU (145 ops)
|
| 133 |
```
|
| 134 |
|
| 135 |
|
| 136 |
+
## How does this work?
|
| 137 |
+
|
| 138 |
+
This [export script](https://aihub.qualcomm.com/models/midas_quantized/qai_hub_models/models/Midas-V2-Quantized/export.py)
|
| 139 |
+
leverages [Qualcomm® AI Hub](https://aihub.qualcomm.com/) to optimize, validate, and deploy this model
|
| 140 |
+
on-device. Lets go through each step below in detail:
|
| 141 |
+
|
| 142 |
+
Step 1: **Compile model for on-device deployment**
|
| 143 |
+
|
| 144 |
+
To compile a PyTorch model for on-device deployment, we first trace the model
|
| 145 |
+
in memory using the `jit.trace` and then call the `submit_compile_job` API.
|
| 146 |
+
|
| 147 |
+
```python
|
| 148 |
+
import torch
|
| 149 |
+
|
| 150 |
+
import qai_hub as hub
|
| 151 |
+
from qai_hub_models.models.midas_quantized import Model
|
| 152 |
+
|
| 153 |
+
# Load the model
|
| 154 |
+
torch_model = Model.from_pretrained()
|
| 155 |
+
|
| 156 |
+
# Device
|
| 157 |
+
device = hub.Device("Samsung Galaxy S24")
|
| 158 |
+
|
| 159 |
+
# Trace model
|
| 160 |
+
input_shape = torch_model.get_input_spec()
|
| 161 |
+
sample_inputs = torch_model.sample_inputs()
|
| 162 |
+
|
| 163 |
+
pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])
|
| 164 |
+
|
| 165 |
+
# Compile model on a specific device
|
| 166 |
+
compile_job = hub.submit_compile_job(
|
| 167 |
+
model=pt_model,
|
| 168 |
+
device=device,
|
| 169 |
+
input_specs=torch_model.get_input_spec(),
|
| 170 |
+
)
|
| 171 |
+
|
| 172 |
+
# Get target model to run on-device
|
| 173 |
+
target_model = compile_job.get_target_model()
|
| 174 |
+
|
| 175 |
+
```
|
| 176 |
+
|
| 177 |
+
|
| 178 |
+
Step 2: **Performance profiling on cloud-hosted device**
|
| 179 |
+
|
| 180 |
+
After compiling models from step 1. Models can be profiled model on-device using the
|
| 181 |
+
`target_model`. Note that this scripts runs the model on a device automatically
|
| 182 |
+
provisioned in the cloud. Once the job is submitted, you can navigate to a
|
| 183 |
+
provided job URL to view a variety of on-device performance metrics.
|
| 184 |
+
```python
|
| 185 |
+
profile_job = hub.submit_profile_job(
|
| 186 |
+
model=target_model,
|
| 187 |
+
device=device,
|
| 188 |
+
)
|
| 189 |
+
|
| 190 |
+
```
|
| 191 |
+
|
| 192 |
+
Step 3: **Verify on-device accuracy**
|
| 193 |
+
|
| 194 |
+
To verify the accuracy of the model on-device, you can run on-device inference
|
| 195 |
+
on sample input data on the same cloud hosted device.
|
| 196 |
+
```python
|
| 197 |
+
input_data = torch_model.sample_inputs()
|
| 198 |
+
inference_job = hub.submit_inference_job(
|
| 199 |
+
model=target_model,
|
| 200 |
+
device=device,
|
| 201 |
+
inputs=input_data,
|
| 202 |
+
)
|
| 203 |
+
on_device_output = inference_job.download_output_data()
|
| 204 |
+
|
| 205 |
+
```
|
| 206 |
+
With the output of the model, you can compute like PSNR, relative errors or
|
| 207 |
+
spot check the output with expected output.
|
| 208 |
+
|
| 209 |
+
**Note**: This on-device profiling and inference requires access to Qualcomm®
|
| 210 |
+
AI Hub. [Sign up for access](https://myaccount.qualcomm.com/signup).
|
| 211 |
+
|
| 212 |
|
| 213 |
|
| 214 |
## Run demo on a cloud-hosted device
|
|
|
|
| 246 |
|
| 247 |
|
| 248 |
## License
|
| 249 |
+
* The license for the original implementation of Midas-V2-Quantized can be found
|
| 250 |
+
[here](https://github.com/isl-org/MiDaS/blob/master/LICENSE).
|
| 251 |
* 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)
|
| 252 |
|
| 253 |
|