NeuroPilot Converter Tool

This page provides a guide for converting YOLOv5s PyTorch models to TFLite format using the NeuroPilot Converter Tool. It covers the environment setup, patch application, and conversion procedures for both INT8 and FP32 formats.

Set Up the YOLOv5 Environment

Note

It is recommended to use Python 3.7 for optimal compatibility with the required libraries and frameworks.

Install and Verify the NeuroPilot Converter Tool

Download the NeuroPilot SDK All-In-One Bundle (Version 7.X.X): Visit the NeuroPilot Downloads page.

Extract the bundle:

tar zxvf neuropilot-sdk-basic-<version>.tar.gz

Install the NeuroPilot Converter Tool for Python 3.7:

pip3 install /path/to/neuropilot-sdk-basic-<version>/offline_tool/mtk_converter-<version>+release-cp37-cp37-manylinux_2_5_x86_64.manylinux1_x86_64.whl

Verify the installation: Run the following script to check the installed package version:
```
python3 -c 'import mtk_converter; print(mtk_converter.__version__)'
```

Prepare and Export the YOLOv5 Model

Clone the repository:

git clone http://github.com/ultralytics/yolov5
cd yolov5
git reset --hard 485da42

Install dependencies:
```
pip3 install -r requirements.txt
pip3 install torch==1.9.0 torchvision==0.10.0
```
Note

The mtk_converter.PyTorchConverter requires PyTorch versions between 1.3.0 and 2.0.0. The developer must install compatible versions of torch and torchvision to avoid runtime errors.

Apply the patch:

wget https://mediatek-aiot.s3.ap-southeast-1.amazonaws.com/aiot/download/model-zoo/scripts/model_conversion_YOLOv5s_example_20240916.zip
unzip -j model_conversion_YOLOv5s_example_20240916.zip
git apply Fix_yolov5_mtk_tflite_issue.patch

Note

The Fix_yolov5_mtk_tflite_issue.patch adds support for MediaTek TFLite (MTK TFLite) in the YOLOv5 export script by modifying the Detect module and post-processing logic.

Export the PyTorch model to TorchScript format:

python3 export.py --weight yolov5s.pt --img-size 640 640 --include torchscript

Convert the PyTorch Model to TFLite

Convert to INT8 Format

Prepare calibration data: Create a script named prepare_calibration_data.py to generate images for quantization calibration.

python3 prepare_calibration_data.py

import os
import numpy as np
from utils.dataloaders import LoadImagesAndLabels
from utils.general import check_dataset

data = 'data/coco128.yaml'
num_batches = 100
calib_dir = 'calibration_dataset'
os.makedirs(calib_dir)

dataset = LoadImagesAndLabels(check_dataset(data)['train'], batch_size=1)

for idx, (im, _target, _path, _shape) in enumerate(dataset):
    if idx >= num_batches:
        break
    im = np.expand_dims(im, axis=0).astype(np.float32)
    im /= 255
    np.save(os.path.join(calib_dir, 'batch-{:05d}.npy'.format(idx)), im)

Perform INT8 conversion: Create a script named convert_to_quant_tflite.py to use the calibration data for conversion.

python3 convert_to_quant_tflite.py

import os
import numpy as np
import mtk_converter

calib_dir = 'calibration_dataset'
converter = mtk_converter.PyTorchConverter.from_script_module_file(
    'yolov5s.torchscript', input_shapes=[(1, 3, 640, 640)]
)

def data_gen():
    for fn in sorted(os.listdir(calib_dir)):
        yield [np.load(os.path.join(calib_dir, fn))]

converter.quantize = True
converter.calibration_data_gen = data_gen
converter.convert_to_tflite('yolov5s_int8_mtk.tflite')

Convert to FP32 Format

Create a script named convert_to_tflite.py to generate a full-precision TFLite model.

python3 convert_to_tflite.py

import mtk_converter

converter = mtk_converter.PyTorchConverter.from_script_module_file(
    'yolov5s.torchscript', input_shapes=[(1, 3, 640, 640)]
)
converter.convert_to_tflite('yolov5s_mtk.tflite')

Next Steps

After generating the .tflite model, refer to Visualizing AI Models to inspect the model structure and tensor information required for application development. Otherwise, proceed to Compile TFLite Models to DLA to generate a compiled binary for direclty NPU execution.