Reproducibility Project

• Our experiments using PyTorch and Hugging Face frameworks

Replication experiments

Replication, Resolution adjustment, Data sampling, Domain shifting

ViT

• Replication, Resolution adjustment
  • accelerate launch vit_ddp.py
    • (parser setting) resolution : 512~64
• Data sampling
  • accelerate launch vit_ddp_sampling.py
    • (parser setting) dataset_ex : True / dataset_ratio : 1.0~0.1
• Domain shifting
  • accelerate launch vit_ddp_domain.py
    • (parser setting) dataset_name

BiT(ResNet)

• Replication, Resolution adjustment
  • accelerate launch bit_ddp.py
    • (parser setting) resolution : 512~64
• Data sampling
  • accelerate launch bit_ddp_sampling.py
    • (parser setting) dataset_ex : True / dataset_ratio : 1.0~0.1
• Domain shifting
  • accelerate launch bit_ddp_domain.py
    • (parser setting) dataset_name

Ablation and Parameter Study

• study/vit_ablation.ipynb
• study/vit_parameter_stduy.ipynb
• study/visualization.ipynb
  • visualize : linear projection, Positional embedding, Tranformer attention matrix
• study/vit.py
  • main file which selects vit model and train the model.
  • vit2_1.py : SGD optimizer, cosine learning rate decay (Replication original model)
  • vit4.py, vit_4_1.py : adam optimizer
  • vit5.py: SGD optimizer, linear learning rate decay
  • vit6.py: SGD optimizer, exponential learning rate decay
  • vit7.py: SGD optimizer, ReduceLROnPlateau learning rate decay
• study/model_with_positional_embedding.py, study/model_without_positional_embedding.py
  • model files are separated with two kinds, one is the model with adding positional embedding and the other is the model without adding positional embedding.
• study/patchdata.py
  • patchdata file make patches according to size of image and patch.
• study/test.py
  • test file derives accuracy and loss from test data.