https://github.com/logcreative/quickdraw-classifier

CS420 Project
https://github.com/logcreative/quickdraw-classifier

Last synced: 2 days ago
JSON representation

CS420 Project

• Host: GitHub
• URL: https://github.com/logcreative/quickdraw-classifier
• Owner: LogCreative
• Created: 2022-05-20T12:44:13.000Z (over 2 years ago)
• Default Branch: main
• Last Pushed: 2022-06-19T09:05:29.000Z (over 2 years ago)
• Last Synced: 2024-12-06T18:57:14.274Z (17 days ago)
• Language: Python
• Homepage:
• Size: 3.94 MB
• Stars: 0
• Watchers: 2
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# quickdraw-classifier

CS420 Project

## Data

Unzip the data into `dataset/seq` folder.

Use `RPCL-pix2seq` to covert the seq data into png data. (You may need to clone this repo by `--recursive` parameter to download the submodule.)

```cmd

cd RPCL-pix2seq

python seq2png.py --input_dir=../dataset/seq --output_dir=../dataset/png --png_width=28 --categories={'bear'}

```

> **NOTICE** You need to use python<=3.7 to install tensorflow 1.15.

## Train

Once the data is prepared, you could train the model by running python on one of the following scripts:

```

python train_cnn.py

python train_rnn.py

python train_cnnrnn.py

```

For CNN model, you may need to modify the type of the structure in `config_train_cnn.py`. The value of `model` could be `resnet18`, `ResNet`, or `sketchnet`. For RNN model, we use Bidirectional LSTM structure. For CNN-RNN model, we use Sketch-a-Net for CNN branch and BiLSTM for RNN branch.

The training process uses PyTorch. During training, the best model will be saved as `best_{model}.pth` in the root folder. The test accuracy could be viewed in [report/result.dat](report/result.dat).

## Report

Report (Chinese) could be found in [report/ML_CS420_Project_report.pdf](report/ML_CS420_Project_report.pdf).