# 1. Choose your workspace and download our repository.
cd ${CUSTOMIZED_WORKSPACE}
git clone https://github.com/PengBoXiangShang/multigraph_transformer

# 2. Enter the directory.
cd multigraph_transformer

# 3. Clone our environment, and activate it.
conda-env create --name ${CUSTOMIZED_ENVIRONMENT_NAME} --file ./MGT_environment.yml
conda activate ${CUSTOMIZED_ENVIRONMENT_NAME}

# 4. Download our training/evaluation/testing datasets and the associated URL lists from our Google Drive folder. Then extract them into './dataloader' folder. data.tar.gz is 118MB, and its MD5 checksum is 8ce7347dfcc9f02376319ce321bbdd31.
cd ./dataloader
chmod +x download.sh
./download.sh
# If this script 'download.sh' can not work for you, please manually download data.tar.gz to current path via this link https://drive.google.com/open?id=1I4XKajNP6wtCpek4ZoCoVr2roGSOYBbW .
tar -zxvf data.tar.gz
rm -f data.tar.gz
cd ..

# 5. Train our MGT. Please see details in our code annotations.
# Please set the input arguments based on your case.
# When the program starts running, a folder named 'experimental_results/${CUSTOMIZED_EXPERIMENT_NAME}' will be created automatically to save your log, checkpoint, and TensorBoard curves.
python train_gra_transf_inpt5_new_dropout_2layerMLP_2nn4nnjnn_early_stop.py 
    --exp ${CUSTOMIZED_EXPERIMENT_NAME}   
    --batch_size ${CUSTOMIZED_SIZE}   
    --num_workers ${CUSTOMIZED_NUMBER} 
    --gpu ${CUSTOMIZED_GPU_NUMBER}

# Actually, we got the performance of MGT #17 (reported in Table 3 in our paper) by running the following command.
python train_gra_transf_inpt5_new_dropout_2layerMLP_2nn4nnjnn_early_stop.py 
    --exp train_gra_transf_inpt5_new_dropout_2layerMLP_2nn4nnjnn_early_stop_001   
    --batch_size 192   
    --num_workers 12 
    --gpu 1

@article{xu2019multigraph,
  title={Multi-Graph Transformer for Free-Hand Sketch Recognition},
  author={Xu, Peng and Joshi, Chaitanya K and Bresson, Xavier},
  journal={arXiv preprint arXiv:1912.11258},
  year={2019}
}

@article{xu2021multigraph,
  title={Multigraph transformer for free-hand sketch recognition},
  author={Xu, Peng and Joshi, Chaitanya K and Bresson, Xavier},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2021},
  publisher={IEEE}
}

# Based on the aforementioned operations and environment configurations.
# For brevity, we only provide the code for the two CNN baselines with best performance, i.e., Inceptionv3 and MobileNetv2.

# 1. Enter the 'dataloader' directory.
cd ${CUSTOMIZED_WORKSPACE}/multigraph_transformer/dataloader/

# 2. Download the training/evaluation/testing datasets (.PNG files) and the associated URL lists from our Google Drive folder. Then extract them into 'dataloader' folder. data_4_cnnbaselines.tar.gz is 558MB, and its MD5 checksum is 8f1132b400eb2bd9186f7f02d5c4d501.
chmod +x download_4_cnnbaselines.sh
./download_4_cnnbaselines.sh
tar -zxvf data_4_cnnbaselines.tar.gz
rm -f data_4_cnnbaselines.tar.gz
cd data_4_cnnbaselines
tar -zxvf tiny_train_set.tar.gz
rm -f tiny_train_set.tar.gz
tar -zxvf tiny_val_set.tar.gz
rm -f tiny_val_set.tar.gz
tar -zxvf tiny_test_set.tar.gz
rm -f tiny_test_set.tar.gz

# 3. Switch directory and run the scripts.
# Please set the input arguments based on your case.
# When the program starts running, a folder named 'experimental_results/${CUSTOMIZED_EXPERIMENT_NAME}' will be created automatically into ${CUSTOMIZED_WORKSPACE}/multigraph_transformer/baselines/cnn_baselines/, to save your log, checkpoint, and TensorBoard curves.
python train_inceptionv3.py 
    --exp ${CUSTOMIZED_EXPERIMENT_NAME}   
    --batch_size ${CUSTOMIZED_SIZE}   
    --num_workers ${CUSTOMIZED_NUMBER} 
    --gpu ${CUSTOMIZED_GPU_NUMBER}
    
# Actually, we got the performance of Inceptionv3 (reported in Table 2 in our paper) by running the following command.
python train_inceptionv3.py 
    --exp train_inceptionv3_001   
    --batch_size 64   
    --num_workers 12 
    --gpu 0

# Based on the aforementioned operations and environment configurations.
# For brevity, we only provide the code for the RNN baseline with best performance, i.e., bidirectional GRU.

# 1. Switch directory and run the scripts.
cd ${CUSTOMIZED_WORKSPACE}/multigraph_transformer/baselines/rnn_baselines/
# Please set the input arguments based on your case.
# When the program starts running, a folder named 'experimental_results/${CUSTOMIZED_EXPERIMENT_NAME}' will be created automatically into ${CUSTOMIZED_WORKSPACE}/multigraph_transformer/baselines/rnn_baselines/, to save your log, checkpoint, and TensorBoard curves.
python train_bigru.py 
    --exp ${CUSTOMIZED_EXPERIMENT_NAME}   
    --batch_size ${CUSTOMIZED_SIZE}   
    --num_workers ${CUSTOMIZED_NUMBER} 
    --gpu ${CUSTOMIZED_GPU_NUMBER}

# Actually, we got the performance of Bi-directional GRU (reported in Table 2 in our paper) by running the following command.
python train_bigru.py 
    --exp train_bigru_001   
    --batch_size 256   
    --num_workers 12 
    --gpu 0