About:

Quantization on weed detection with PyTorch on deepweeds dataset for performance improvement. Tested on RPi, Mobile processors and PC.

Code and model details:

1. Clone this repository locally move into the project folder
2. In the local env run pip install -r requirements.txt

Training the model:

Download and place the image and lables from the deepweeds dataset into prepare_data/deepweeds/ and execute the train_model.py file. dataset-github-repo

Inference on the model:

Download the trained models from the Google-Drive link and place them in models/ and execute the inference.py file.

Hardware testing

1. Android: Tested on android device using Android Debug Bridge which enable the use of mobile processor (present on your debug device) without the need to build an actual application. However, few extra steps are needed to convert the current trained models ref .

Steps: First, clone the pytorch official repository

git clone https://github.com/pytorch/pytorch.git

cd pytorch

git submodule update --init --recursive

export ANDROID_ABI=arm64-v8a

# Change accordingly
export ANDROID_NDK=/path/to/Android/Sdk/ndk/21.0.6113669/

./scripts/build_android.sh \
-DBUILD_BINARY=ON \
-DBUILD_CAFFE2_MOBILE=OFF \
-DCMAKE_PREFIX_PATH=$(python -c 'from distutils.sysconfig import get_python_lib; print(get_python_lib())') \
-DPYTHON_EXECUTABLE=$(python -c 'import sys; print(sys.executable)') \

pass the downloaded models through save_for_RPI_inference present in inference.py

adb shell mkdir /data/local/tmp/pt
adb push build_android/install/bin/speed_benchmark_torch /data/local/tmp/pt
adb push <model.pt> /data/local/tmp/pt

adb shell  /data/local/tmp/pt/speed_benchmark_torch \
--model  /data/local/tmp/pt/<model.pt> --input_dims="1,3,224,224" \
--input_type=float --warmup=5 --iter 100

Sample Result:

Starting benchmark.
Running warmup runs.

Main runs.
Main run finished. Microseconds per iter: xxx. Iters per second: xxx

2. Raspberry Pi: Since RPi is ARM based, few instrutions are specific for this family of architectures which are different from the x86 ones. PyTorch handels most of them, just add torch.backends.quantized.engine = 'qnnpack' in inference file and run on RPi hardware.

3. PC: No extra updates needed, run the code present in inference on your hardware.

Citation

If QWID helps your research, we appreciate your citations. Here is the BibTeX entry:

@misc{rathore2023qwid,
      title={QWID: Quantized Weed Identification Deep neural network},
      author={Parikshit Singh Rathore},
      year={2023},
      eprint={2310.18921},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}