Deep Learning Final Project

Yi-Yi Chu and Omar Hammami

Dependencies

• PyTorch Geometric==1.6.0

Then, you need to create a directory for recoreding finetuned results to avoid errors:

mkdir logs

For different experiments, please check the python script to see what directory needs to be created.

Training & Evaluation

For GCL-AllAug, run the following script

./go.sh $GPU_ID $DATASET_NAME random4 0.1

For GCL-NoAug, run the following script

./go_noaug.sh $GPU_ID $DATASET_NAME random4 0.1

For Uniformity Loss, run the following script

./go_uniform.sh $GPU_ID $DATASET_NAME random4 0.1

For random initialized GIN, run the following script

./go_random.sh $GPU_ID $DATASET_NAME random4 0.1

For hard negative GCL-NoAug, run the following script

./go_beta_1.sh $GPU_ID $DATASET_NAME random4 0.1

For easy negative GCL-NoAug, run the following script

./go_beta_2.sh $GPU_ID $DATASET_NAME random4 0.1

For GCL-AllAug-Diffusion, run the following script

python gsimclr.py --DS NCI1 --lr 0.01 --local --num-gc-layers 3 --aug all1 --seed 1

t-SNE Visualization

To run t-SNE visulization, run the following script

python gsimclr_tsne.py $GPU_ID $DATASET_NAME random4 0.1

Acknowledgements

The backbone implementation is reference to https://github.com/Shen-Lab/GraphCL.