Skip to content

Latest commit

 

History

History
171 lines (123 loc) · 6.72 KB

GETTING_STARTED.md

File metadata and controls

171 lines (123 loc) · 6.72 KB
Raw

How to run the code

Run python main.py +experiment=dit for experiments using DiT as the diffusion backbone. Here are the arguments which you can adapt for different datasets and hyper-parameters:

  • Classifier backbone: set model.class_arch to convnext_large/convnext_tiny/resnet18/vit_b_32/vit_b_16/vit_l_14 for ImageNet-trained classifiers
  • Larger DiT: use +experiment=dit and set input.sd_img_res=512
  • Optimizer: set tta.gradient_descent.optimizer to adam/sgd
  • Learning reate: set tta.gradient_descent.base_learning_rate to any numerical values
  • Dataset: set input.dataset_name to ImageNetDataset/ImageNetv2Dataset/ImageNetRDataset/ImageNetCDataset/ImageNetADataset/ImageNetStyleDataset.
  • Total batch size: set input.batch_size and tta.gradient_descent.accum_iter where the total batch size is the multiplication of these two parameters

How to improve performance

Empirically, we found that using larger total batch size results in more stable classification improvement. However, it takes longer time for TTA with larger batch size. Also, we found that some backbones are better with sgd optimizer than adam optimizer.

Commands to get started

Clip/Stable Diffusion

Single Sample TTA on FGVC-Aircraft and other datasets ``` python main.py +experiment=sd model.class_arch=clipb32 input.dataset_name=FGVCAircraftSubset ```

ConvNext-Large/DiT

ConvNext-Tiny works better with adam optimizer

Online TTA on ImageNet-C 
python main.py +experiment=dit model.class_arch=convnext_large input.batch_size=15 tta.gradient_descent.accum_iter=12 input.dataset_name=ImageNetCDataset tta.gradient_descent.base_learning_rate=1e-5 tta.gradient_descent.optimizer=adam tta.online=True input.subsample=1 log_freq=1

ConvNext-Tiny/DiT

ConvNext-Tiny works better with adam optimizer

Single-sample TTA on ImageNet-R 
python main.py +experiment=dit model.class_arch=convnext_tiny input.batch_size=12 tta.gradient_descent.accum_iter=15 input.dataset_name=ImageNetRDataset tta.gradient_descent.base_learning_rate=1e-5 tta.gradient_descent.optimizer=adam
Single-sample TTA on ImageNet-C 
python main.py +experiment=dit model.class_arch=convnext_tiny input.batch_size=20 tta.gradient_descent.accum_iter=9 input.dataset_name=ImageNetCDataset tta.gradient_descent.base_learning_rate=1e-5 tta.gradient_descent.optimizer=adam input.subsample=null
Single-sample TTA on ImageNet-A 
python main.py +experiment=dit model.class_arch=convnext_tiny input.batch_size=15 tta.gradient_descent.accum_iter=12 input.dataset_name=ImageNetADataset tta.gradient_descent.base_learning_rate=1e-5 tta.gradient_descent.optimizer=adam input.subsample=null
Single-sample TTA on ImageNet-v2 
python main.py +experiment=dit model.class_arch=convnext_tiny input.batch_size=15 tta.gradient_descent.accum_iter=12 input.dataset_name=ImageNetv2Dataset tta.gradient_descent.base_learning_rate=1e-5 tta.gradient_descent.optimizer=adam
Single-sample TTA on ImageNet 
python main.py +experiment=dit model.class_arch=convnext_tiny input.batch_size=15 tta.gradient_descent.accum_iter=12 input.dataset_name=ImageNetDataset tta.gradient_descent.base_learning_rate=1e-5 tta.gradient_descent.optimizer=adam

ResNet-18/DiT

ResNet-18 works better with adam optimizer

Single-sample TTA on ImageNet-R 
python main.py +experiment=dit model.class_arch=resnet18 input.batch_size=12 tta.gradient_descent.accum_iter=15 input.dataset_name=ImageNetRDataset tta.gradient_descent.base_learning_rate=1e-5 tta.gradient_descent.optimizer=adam
Single-sample TTA on ImageNet-C 
python main.py +experiment=dit model.class_arch=resnet18 input.batch_size=20 tta.gradient_descent.accum_iter=9 input.dataset_name=ImageNetCDataset tta.gradient_descent.base_learning_rate=5e-3 tta.gradient_descent.optimizer=sgd
Single-sample TTA on ImageNet-A 
python main.py +experiment=dit model.class_arch=resnet18 input.batch_size=15 tta.gradient_descent.accum_iter=12 input.dataset_name=ImageNetADataset tta.gradient_descent.base_learning_rate=1e-5 tta.gradient_descent.optimizer=adam input.subsample=null
Single-sample TTA on ImageNet-v2 
python main.py +experiment=dit model.class_arch=resnet18 input.batch_size=15 tta.gradient_descent.accum_iter=12 input.dataset_name=ImageNetv2Dataset tta.gradient_descent.base_learning_rate=1e-5 tta.gradient_descent.optimizer=adam
Single-sample TTA on ImageNet 
python main.py +experiment=dit model.class_arch=resnet18 input.batch_size=15 tta.gradient_descent.accum_iter=12 input.dataset_name=ImageNetDataset tta.gradient_descent.base_learning_rate=1e-5 tta.gradient_descent.optimizer=adam

ViT-B-32/DiT

ViT-B-32 works better with sgd optimizer

Single-sample TTA on ImageNet-R 
python main.py +experiment=dit model.class_arch=vit_b_32 input.batch_size=20 tta.gradient_descent.accum_iter=9 input.dataset_name=ImageNetRDataset tta.gradient_descent.base_learning_rate=5e-3 tta.gradient_descent.optimizer=sgd
Single-sample TTA on ImageNet-C * ImageNet-C ``` python main.py +experiment=dit model.class_arch=vit_b_32 input.batch_size=20 tta.gradient_descent.accum_iter=9 input.dataset_name=ImageNetCDataset tta.gradient_descent.base_learning_rate=5e-3 tta.gradient_descent.optimizer=sgd ```
Single-sample TTA on ImageNet-A 
python main.py +experiment=dit model.class_arch=vit_b_32 input.batch_size=20 tta.gradient_descent.accum_iter=9 input.dataset_name=ImageNetADataset tta.gradient_descent.base_learning_rate=5e-3 tta.gradient_descent.optimizer=sgd input.subsample=null
Single-sample TTA on ImageNet-v2 
python main.py +experiment=dit model.class_arch=vit_b_32 input.batch_size=15 tta.gradient_descent.accum_iter=12 input.dataset_name=ImageNetv2Dataset tta.gradient_descent.base_learning_rate=5e-3 tta.gradient_descent.optimizer=sgd
Single-sample TTA on ImageNet 
python main.py +experiment=dit model.class_arch=vit_b_32 input.batch_size=20 tta.gradient_descent.accum_iter=9 input.dataset_name=ImageNetDataset tta.gradient_descent.base_learning_rate=5e-3 tta.gradient_descent.optimizer=sgd