Depth-Anything-SNPE

Running Depth Anything model on Qualcomm SNPE SDK.

Generate ONNX model

We will be using Depth-Anything-ONNX to generate ONNX weight files of the model.

Environment setup

# clone into Depth-Anything-ONNX
conda create -n dav2 python=3.11
conda activate dav2
pip install -r requirements.txt

Generation

python dynamo.py export --encoder vits -o depth_anything_vits_op16.onnx -b 1 --opset 16

Running SNPE SDK

This step assumes you have installed SNPE SDK successfully. You should see the following message if sourced correctly.

[INFO] AISW SDK environment set
[INFO] QNN_SDK_ROOT: /mnt/d/Downloads_F/v2.22.6.240515/qairt/2.22.6.240515
[INFO] SNPE_ROOT: /mnt/d/Downloads_F/v2.22.6.240515/qairt/2.22.6.240515
SDK environment now set up; additionally you may now run devtool to perform development tasks.
Run devtool --help for further details.

Python Environment

# clone the repository
conda create -f environment_snpe.yml
conda activate snpe

Generate DLC from ONNX

snpe-onnx-to-dlc --input_network assets/onnx/depth_anything_vits_op16.onnx -o assets/dlc/depth_anything_vits_op16.dlc

Run DLC

snpe-net-run --container assets/dlc/depth_anything_vits_op16.dlc --input_list assets/input_list.txt

A successful run should give the following output

-------------------------------------------------------------------------------
Model String: N/A
SNPE v2.22.6.240515184619_92920
-------------------------------------------------------------------------------

Processing graph : depth_anything_vits_op16
Processing DNN input(s):
./input/input.raw
Processing DNN input(s):
./input/input2.raw
Successfully executed graph depth_anything_vits_op16

Optimizations

Refer to OPTIMIZATION.md for further improving the performance of the DLC model.

References

• Qualcomm SNPE SDK