nam157 / hero-name-recognition

HERO NAME Recognition

Describe

In the problem of hero identification through the game league of legends and the requirements of the problem and the data provided. I realize this is a common classification problem and belongs to the type of multi-label classification. The problem belongs to the group of supervised math problems, to handle this problem I use CNN to solve this problem and use the pytorch framework to deploy. Use densenet121 backbone to train the model and use adam algorithm to optimize and use CrossEntropyLoss function to find out the model cost.

As for image processing, I receive a very small number of images and different image sizes. So I also want to handle it simply, resize it to 128*128 and use more data augmentation including methods: [Resize,RandomHorizontalFlip,RandomVerticalFlip,Normalize] Meaningful functions such as resizing to the same size, flipping random images horizontally, flipping random images vertically, and normalizing the data to a normal distribution. Also not much time, the next work is to need more data and more techniques to denoise image

Problems exist:

• The topic does not provide train data
• Using a fairly large architecture (Not recommended)
• The model is being overfitting
• Train and evaluate on a single episode. Although sharing data

Base environment

The environment is only available for Python3.8 and upcoming versions of Python3.9

Following packages are available in requirements.txt

# Install dependencies

pip install -r requirements.txt

Running

usage: Hero Name Recognition [-h] {train,export,infer}

python main --help

Training the classification:

# @path_dataset: train dataset
python main.py train --epochs 10 --path_dataset ./test_data/test.txt 

Export model from ckpt_path to TorchScript format:

#@convert_model: Export model to TorchScript format
python main.py export --convert_model ./save_model/model_hero_jit.pt

Inference with exported model for making prediction. input_path is a directory which contains many images

#@torchjit_ck: torchscript model and @folder_img: folder exist images and @save_file: save results
python main.py infer --torchjit_ck ./save_model/model_hero_jit.pt \ 
                     --folder_img ./test_data/test_images/ \
                     --save_file output.txt

Conclusion

This documentation has demonstrated how to use related module. Before actually start working on anything, please read the whole document first. If you need any clarifications, please contact me. Thanks for reading and good luck on improving the model.

Happy Coding