pra-dan/Converted-YOLO-Models

The repository contains YOLO models converted into TensorFlow and other frameworks and the process involved

The repo is a coalesced work of the original [darkflow] repo and the Currently the repo contains only Darknet-based YOLO models converted to TensorFlow using Darkflow and a hands-on tutorial by Zoltan Szabo.

Some of the already converted model files are linked below

YOLO Model	Train Dataset	mAP	FPS	Frozen Model (.pb)	Metafile
Darknet19 (448x448)	VOC-1000	76.4	-	pb	meta
v2-tiny	COCO	-	200	pb	meta
v2-tiny	VOC	57.1	207	failed	failed

The conversion for the YOLO v2-tiny (COCO dataset) is done as:

1. Install Dependencies

!apt-get update
!pip3 install numpy
!apt-get install python-opencv -y
!pip install cython

2. Clone and install Darkflow

# Clone the repository
!git clone https://github.com/thtrieu/darkflow
# Change into the darkflow dir and install darkflow with pip
%cd darkflow
!python setup.py build_ext --inplace

3. Fix assertion issue by modifying the loader.py Note: may not always be needed but it won't cause any issue

#Change directory to darkflow
%cd /content/darkflow
!sed -i "s/self.offset = 16/self.offset = 20/g"  ./darkflow/utils/loader.py

4. Download the model files (cfg file + weights file)

!wget -O "cfg/yolo-v2-tiny-coco.cfg" "https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolov2-tiny.cfg"
!wget -O "weights/yolo-v2-tiny-coco.weights" "https://pjreddie.com/media/files/yolov2-tiny.weights"

5. Convert using flow and generate frozen graph (.pb) Note: We keep cfg files in the darkflow/cfg directory and the weights in the darkflow/weights directory

#syntax: $ !./flow --model path_to_cfg_file --load path_to_weights --savepb
!./flow --model cfg/yolo-v2-tiny-coco.cfg --load weights/yolo-v2-tiny-coco.weights --savepb

Done ! You should be able to find your converted frozen graph in the darkflow/built_graph directory