Introduction

A Conversion tool to convert YOLO v3 Darknet weights to TF Lite model (YOLO v3 PyTorch > ONNX > TensorFlow > TF Lite), and to TensorRT model (dynamic_axes branch).

Prerequisites

• python3
• torch==1.3.1
• torchvision==0.4.2
• onnx==1.6.0
• onnx-tf==1.5.0
• onnxruntime-gpu==1.0.0
• tensorflow-gpu==1.15.0

Docker

docker pull zldrobit/onnx:10.0-cudnn7-devel

Usage

• 1. Download pretrained Darknet weights:

cd weights
wget https://pjreddie.com/media/files/yolov3.weights 

• 2. Convert YOLO v3 model from Darknet weights to ONNX model: Change ONNX_EXPORT to True in models.py. Run

python3 detect.py --cfg cfg/yolov3.cfg --weights weights/yolov3.weights

The output ONNX file is weights/export.onnx.

• 3. Convert ONNX model to TensorFlow model:

python3 onnx2tf.py

The output file is weights/yolov3.pb.

• 4. Preprocess pb file to avoid NCHW conv, 5-D ops, and Int64 ops:

python3 prep.py

The output file is weights/yolov3_prep.pb.

• 5. Use TOCO to convert pb -> tflite:

toco --graph_def_file weights/yolov3_prep.pb \
    --output_file weights/yolov3.tflite \
    --output_format TFLITE \
    --inference_type FLOAT \
    --inference_input_type FLOAT \
    --input_arrays input.1 \
    --output_arrays concat_84

The output file is weights/yolov3.tflite. Now, you can run python3 tflite_detect.py --weights weights/yolov3.tflite to detect objects in an image.

Quantization

• 1. Install flatbuffers: Please refer to flatbuffers.

• 2. Download TFLite schema:

wget https://github.com/tensorflow/tensorflow/raw/r1.15/tensorflow/lite/schema/schema.fbs

• 3. Run TOCO to convert and quantize pb -> tflite:

toco --graph_def_file weights/yolov3_prep.pb \
    --output_file weights/yolov3_quant.tflite \
    --output_format TFLITE  \
    --input_arrays input.1 \
    --output_arrays concat_84 \
    --post_training_quantize

The output file is weights/yolov3_quant.tflite.

• 4. Convert tflite -> json:

flatc -t --strict-json --defaults-json -o weights schema.fbs  -- weights/yolov3_quant.tflite

The output file is weights/yolov3_quant.json.

• 5. Fix ReshapeOptions:

python3 fix_reshape.py

The output file is weights/yolov3_quant_fix_reshape.json.

• 6. Convert json -> tflite:

flatc -b -o weights schema.fbs weights/yolov3_quant_fix_reshape.json

The output file is weights/yolov3_quant_fix_reshape.tflite. Now, you can run

python3 tflite_detect.py --weights weights/yolov3_quant_fix_reshape.tflite

to detect objects in an image.

Auxiliary Files

• ONNX inference and detection: onnx_infer.py and onnx_detect.py.
• TensorFlow inference and detection: tf_infer.py and tf_detect.py.
• TF Lite inference, detection and debug: tflite_infer.py, tflite_detect.py and tflite_debug.py.

Known Issues

• The conversion code does not work with tensorflow==1.14.0: Running prep.py cause protobuf error (Channel order issue in Conv2D).
• fix_reshape.py does not fix shape attributes in TFLite tensors, which may cause unknown side effects.

TODO

• support tflite quantized model
• use dynamic_axes for ONNX export to support dynamic batching and TensorRT conversion (dynamic_axes branch)
• add TensorRT NMS support (trt_nms branch)

Acknowledgement

We borrow PyTorch code from ultralytics/yolov3, and TensorFlow low-level API conversion code from paulbauriegel/tensorflow-tools.