tensorflow-yolov4-tflite

YOLOv4, YOLOv4-tiny Implemented in Tensorflow 2.0. Convert YOLO v4, YOLOv3, YOLO tiny .weights to .pb, .tflite and trt format for tensorflow, tensorflow lite, tensorRT.

Download yolov4.weights file: https://drive.google.com/open?id=1cewMfusmPjYWbrnuJRuKhPMwRe_b9PaT

Prerequisites

Performance

Demo

# Convert darknet weights to tensorflow
## yolov4
python save_model.py --weights ./data/yolov4.weights --output ./checkpoints/yolov4-416 --input_size 416 --model yolov4 

## yolov4-tiny
python save_model.py --weights ./data/yolov4-tiny.weights --output ./checkpoints/yolov4-tiny-416 --input_size 416 --model yolov4 --tiny

# Run demo tensorflow
python detect.py --weights ./checkpoints/yolov4-416 --size 416 --model yolov4 --image ./data/kite.jpg

python detect.py --weights ./checkpoints/yolov4-tiny-416 --size 416 --model yolov4 --image ./data/kite.jpg --tiny

Output

Yolov4 original weight

Yolov4 tflite int8

Convert to tflite

# Save tf model for tflite converting
python save_model.py --weights ./data/yolov4.weights --output ./checkpoints/yolov4-416 --input_size 416 --model yolov4 --framework tflite

# yolov4
python convert_tflite.py --weights ./checkpoints/yolov4-416 --output ./checkpoints/yolov4-416.tflite

# yolov4 quantize float16
python convert_tflite.py --weights ./checkpoints/yolov4-416 --output ./checkpoints/yolov4-416-fp16.tflite --quantize_mode float16

# yolov4 quantize int8
python convert_tflite.py --weights ./checkpoints/yolov4-416 --output ./checkpoints/yolov4-416-int8.tflite --quantize_mode int8 --dataset ./coco_dataset/coco/val207.txt

# Run demo tflite model
python detect.py --weights ./checkpoints/yolov4-416.tflite --size 416 --model yolov4 --image ./data/kite.jpg --framework tflite

Yolov4 and Yolov4-tiny int8 quantization have some issues. I will try to fix that. You can try Yolov3 and Yolov3-tiny int8 quantization

Convert to TensorRT

```bash# yolov3 python save_model.py –weights ./data/yolov3.weights –output ./checkpoints/yolov3.tf –input_size 416 –model yolov3 python convert_trt.py –weights ./checkpoints/yolov3.tf –quantize_mode float16 –output ./checkpoints/yolov3-trt-fp16-416

yolov3-tiny

python save_model.py –weights ./data/yolov3-tiny.weights –output ./checkpoints/yolov3-tiny.tf –input_size 416 –tiny python convert_trt.py –weights ./checkpoints/yolov3-tiny.tf –quantize_mode float16 –output ./checkpoints/yolov3-tiny-trt-fp16-416

yolov4

python save_model.py –weights ./data/yolov4.weights –output ./checkpoints/yolov4.tf –input_size 416 –model yolov4 python convert_trt.py –weights ./checkpoints/yolov4.tf –quantize_mode float16 –output ./checkpoints/yolov4-trt-fp16-416


### Evaluate on COCO 2017 Dataset
```bash
# run script in /script/get_coco_dataset_2017.sh to download COCO 2017 Dataset
# preprocess coco dataset
cd data
mkdir dataset
cd ..
cd scripts
python coco_convert.py --input ./coco/annotations/instances_val2017.json --output val2017.pkl
python coco_annotation.py --coco_path ./coco 
cd ..

# evaluate yolov4 model
python evaluate.py --weights ./data/yolov4.weights
cd mAP/extra
python remove_space.py
cd ..
python main.py --output results_yolov4_tf

mAP50 on COCO 2017 Dataset

Benchmark

Detection	512x512	416x416	320x320
YoloV3	55.43	52.32
YoloV4	61.96	57.33

python benchmarks.py --size 416 --model yolov4 --weights ./data/yolov4.weights

TensorRT performance

Tesla P100

Tesla K80

Tesla T4

Tesla P4

Macbook Pro 15 (2.3GHz i7)

Traning your own model

YoloV4 416 images/s	FP32	FP16	INT8
Batch size 1	55	116
Batch size 8	70	152

Detection	512x512	416x416	320x320
YoloV3 FPS	40.6	49.4	61.3
YoloV4 FPS	33.4	41.7	50.0

Detection	512x512	416x416	320x320
YoloV3 FPS	10.8	12.9	17.6
YoloV4 FPS	9.6	11.7	16.0

Detection	512x512	416x416	320x320
YoloV3 FPS	27.6	32.3	45.1
YoloV4 FPS	24.0	30.3	40.1

Detection	512x512	416x416	320x320
YoloV3 FPS	20.2	24.2	31.2
YoloV4 FPS	16.2	20.2	26.5

# Prepare your dataset
# If you want to train from scratch:
In config.py set FISRT_STAGE_EPOCHS=0 
# Run script:
python train.py

# Transfer learning: 
python train.py --weights ./data/yolov4.weights

The training performance is not fully reproduced yet, so I recommended to use Alex’s Darknet to train your own data, then convert the .weights to tensorflow or tflite.