YOLOv3 Tensorflow

This is an implementation of YOLOv3 in tensorflow. Although this can work for other variants of YOLO as well with minute changes.

YAD2K and darknet has been a great help.

And obviously Markdown Cheatsheet for this beautiful README

TODO

YOLOv3-darknet to tensorflow conversion (inference only)
Inference script
Training pipeline
Speed evaluation
Training on raccoon dataset
Subdivisions implentation - aids in training big batches with small GPU vRAM
SGD with burn-in
mAP evaluation
Training on VOC

Suggestions and pull requests are most welcome

Some Results

When running YOLOv3-608 model on a 1080Ti I am getting around 19FPS

When running YOLOv3-608 model on a 1050Ti I am getting around 6FPS

Even the tensorboard is beautiful

How to run the model you ask?

Make a virtual environment and install all the dependencies

I am using virtualenv.

Run the following commands

virtualenv env -p python3
source env\bin\activate

pip install -r requirements.txt
python inference.py path_to_the_image_directory path_for_saving_the_results --darknet_model 1`

The above command runs the pretrained model.

Training the model

Now this will be some work..

1. Data prepration

Try to find an object detection dataset which is having annotations in VOC format as my annotation parsing script works for only that. (Sorry COCO lovers)

I suggest raccoon-dataset provided by experiencor. (Who is having a great implementation of YOLOv3 in keras.)

Make a directory and put your dataset with two folders, one holding images and the other holding the corresponding annotations. There must be a one-to-one correspondence by file name between images and annotations.

I suggest the following structure.

+ dataset
|
|___
	+ dataset_name
	|
	|___
	|	+ images
	|	|
	|	|___ + all the image files
	|___
		+ annotations
		|
		|___ + all the annotation files

2. Telling the model what the new classes are

Make a dataset-name_classes.txt file inside the model_data folder, having one class of the dataset in one line. A sample file named sample_classes.txt has been provided for reference.

3. Updating configuration file

The configuration file is a python .py file containing variables, updation of which is pretty self explainatory by reading the comments provided.

4. Generating the anchors for your dataset (recommended but not mandatory)

If you are doing this step, update the variable anchors_path in your config.py file to point to the location where you want to save the newly generated anchors.

Run the following command

python k-means.py

5. Actually running the training. (Finally, yes I know the feel!!)

Running the following command will do some darkmagic as done in darknet (pun intended) and will start training the model.

python train.py

You might get an error while training for the first time, I am still working to resolve it. (Suggestions are welcome, I am still a noob.)

Run the command again and you are good to go.

6. Making detections using the trained weights

python inference.py path_to_the_image_directory path_for_saving_the_results`

kevintandi/YOLO