Midas-V2-Quantized: Optimized for Mobile Deployment
Quantized Deep Convolutional Neural Network model for depth estimation
Midas is designed for estimating depth at each point in an image.
This model is an implementation of Midas-V2-Quantized found here.
This repository provides scripts to run Midas-V2-Quantized on Qualcomm® devices. More details on model performance across various devices, can be found here.
Model Details
- Model Type: Depth estimation
- Model Stats:
- Model checkpoint: MiDaS_small
- Input resolution: 256x256
- Number of parameters: 16.6M
- Model size: 16.6 MB
| Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model |
|---|---|---|---|---|---|---|---|---|
| Midas-V2 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 1.062 ms | 0 - 133 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 1.267 ms | 0 - 11 MB | INT8 | NPU | Midas-V2-Quantized.so |
| Midas-V2 | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 119.018 ms | 10 - 116 MB | INT8 | NPU | Midas-V2-Quantized.onnx |
| Midas-V2 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 0.753 ms | 0 - 56 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 0.9 ms | 0 - 19 MB | INT8 | NPU | Midas-V2-Quantized.so |
| Midas-V2 | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 95.401 ms | 4 - 338 MB | INT8 | NPU | Midas-V2-Quantized.onnx |
| Midas-V2 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 0.689 ms | 0 - 33 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 0.743 ms | 0 - 35 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 96.251 ms | 22 - 343 MB | INT8 | NPU | Midas-V2-Quantized.onnx |
| Midas-V2 | SA7255P ADP | SA7255P | TFLITE | 10.379 ms | 0 - 27 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | SA7255P ADP | SA7255P | QNN | 11.603 ms | 0 - 10 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 1.066 ms | 0 - 133 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | SA8255 (Proxy) | SA8255P Proxy | QNN | 1.284 ms | 0 - 3 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | SA8295P ADP | SA8295P | TFLITE | 1.917 ms | 0 - 29 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | SA8295P ADP | SA8295P | QNN | 2.192 ms | 0 - 18 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 1.058 ms | 0 - 133 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | SA8650 (Proxy) | SA8650P Proxy | QNN | 1.278 ms | 0 - 3 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | SA8775P ADP | SA8775P | TFLITE | 1.562 ms | 0 - 28 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | SA8775P ADP | SA8775P | QNN | 1.814 ms | 0 - 10 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | RB3 Gen 2 (Proxy) | QCS6490 Proxy | TFLITE | 3.593 ms | 0 - 44 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | RB3 Gen 2 (Proxy) | QCS6490 Proxy | QNN | 5.559 ms | 0 - 14 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | RB5 (Proxy) | QCS8250 Proxy | TFLITE | 16.112 ms | 0 - 3 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | QCS8275 (Proxy) | QCS8275 Proxy | TFLITE | 10.379 ms | 0 - 27 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | QCS8275 (Proxy) | QCS8275 Proxy | QNN | 11.603 ms | 0 - 10 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 1.06 ms | 0 - 135 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 1.277 ms | 0 - 2 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | QCS9075 (Proxy) | QCS9075 Proxy | TFLITE | 1.562 ms | 0 - 28 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | QCS9075 (Proxy) | QCS9075 Proxy | QNN | 1.814 ms | 0 - 10 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 1.541 ms | 0 - 46 MB | INT8 | NPU | Midas-V2-Quantized.tflite |
| Midas-V2 | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 1.797 ms | 0 - 44 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 1.402 ms | 0 - 0 MB | INT8 | NPU | Use Export Script |
| Midas-V2 | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 111.778 ms | 66 - 66 MB | INT8 | NPU | Midas-V2-Quantized.onnx |
Installation
Install the package via pip:
pip install "qai-hub-models[midas-quantized]"
Configure Qualcomm® AI Hub to run this model on a cloud-hosted device
Sign-in to Qualcomm® AI Hub with your
Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.
With this API token, you can configure your client to run models on the cloud hosted devices.
qai-hub configure --api_token API_TOKEN
Navigate to docs for more information.
Demo off target
The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.
python -m qai_hub_models.models.midas_quantized.demo
The above demo runs a reference implementation of pre-processing, model inference, and post processing.
NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).
%run -m qai_hub_models.models.midas_quantized.demo
Run model on a cloud-hosted device
In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:
- Performance check on-device on a cloud-hosted device
- Downloads compiled assets that can be deployed on-device for Android.
- Accuracy check between PyTorch and on-device outputs.
python -m qai_hub_models.models.midas_quantized.export
Profiling Results
------------------------------------------------------------
Midas-V2
Device : Samsung Galaxy S23 (13)
Runtime : TFLITE
Estimated inference time (ms) : 1.1
Estimated peak memory usage (MB): [0, 133]
Total # Ops : 140
Compute Unit(s) : NPU (140 ops)
How does this work?
This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:
Step 1: Compile model for on-device deployment
To compile a PyTorch model for on-device deployment, we first trace the model
in memory using the jit.trace and then call the submit_compile_job API.
import torch
import qai_hub as hub
from qai_hub_models.models.midas_quantized import Model
# Load the model
torch_model = Model.from_pretrained()
# Device
device = hub.Device("Samsung Galaxy S24")
# Trace model
input_shape = torch_model.get_input_spec()
sample_inputs = torch_model.sample_inputs()
pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])
# Compile model on a specific device
compile_job = hub.submit_compile_job(
model=pt_model,
device=device,
input_specs=torch_model.get_input_spec(),
)
# Get target model to run on-device
target_model = compile_job.get_target_model()
Step 2: Performance profiling on cloud-hosted device
After compiling models from step 1. Models can be profiled model on-device using the
target_model. Note that this scripts runs the model on a device automatically
provisioned in the cloud. Once the job is submitted, you can navigate to a
provided job URL to view a variety of on-device performance metrics.
profile_job = hub.submit_profile_job(
model=target_model,
device=device,
)
Step 3: Verify on-device accuracy
To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.
input_data = torch_model.sample_inputs()
inference_job = hub.submit_inference_job(
model=target_model,
device=device,
inputs=input_data,
)
on_device_output = inference_job.download_output_data()
With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.
Note: This on-device profiling and inference requires access to Qualcomm® AI Hub. Sign up for access.
Deploying compiled model to Android
The models can be deployed using multiple runtimes:
TensorFlow Lite (
.tfliteexport): This tutorial provides a guide to deploy the .tflite model in an Android application.QNN (
.soexport ): This sample app provides instructions on how to use the.soshared library in an Android application.
View on Qualcomm® AI Hub
Get more details on Midas-V2-Quantized's performance across various devices here. Explore all available models on Qualcomm® AI Hub
License
- The license for the original implementation of Midas-V2-Quantized can be found here.
- The license for the compiled assets for on-device deployment can be found here
References
- Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer
- Source Model Implementation
Community
- Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
- For questions or feedback please reach out to us.