README.md

Visual_Prompt_Tuning_PyTorch

In this project, I applied Visual Prompt Tuning(VPT) method to the ViT-Base/4 model.
I used a ViT-Base/16 model pre-trained on ImageNet-21k and used CIFAR-10 for fine-tuning.
I compared three methods which were:

• Prompt fine-tuneing (97.83% in epoch 5!)
• Full fine-tuneing (without Prompts)
• Full fine-tuning (with Prompts)

Architecture of VPT for ViT-Base/4

※ Layers that are colored in Red(🔥) are intended to be trained, layers that are colored in Blue(❄️) are frozen!

• Input (3, 32, 32) image
• Divide into 64 patches
• Transform each patches into patch embeddings (768-dim)
• Add positional embeddings to each patches and the classification token
• Add 50 prompts to the Transformer encoder's layers
• Apply Layer Normalization to the output of the encoder
• Input the classification token into the MLP head
• Output prediction!

How to use the code

Follow the steps below to use the code!

• Type in "python main.py" to use the default training strategy.
• Add arguments to change the training procedure (epoch, batch size, etc....)
  For example, "python main.py --epoch 50 --full False --batch_size 256"

Implementation details

ViT-Base/16 was used. I had to modify the embedding layer so that CIFAR-10 could used for training.
The modified model is called ViT-Base/4, where 4 denotes the patch size.
Below are the implementation details of the ViT-Base/4 model.

• Input image size: (3, 32, 32)
• Patch size: (3, 4, 4)
• Patch numbers: 64
• Prompt numbers: 50
• Embedding size: 768
• Hidden size: 768*4
• Head numbers: 12
• Layer depth: 12

Experiment results

The graphs below are training and testing accuracies.
Orange line: ViT-Base/4 (Prompt Fine-tuning = P)
Red line: ViT-Base/4 (Full Fine-tuning without Prompts = F)
Blue Line: ViT-Base/4 (Full Fine-tuning with Prompts = P+F)

↑ Training Accuracy

↑ Testing Accuracy

It took, 5m20s 4m22s, and 6m50s to train the P, F, and P+F model respectively.
Although I did not trian the models for sufficient amount of time, I could observe that the P's convergene speed was very fast compared to the other models, which did not converge during 5 epochs.

Conclusion

From the experiment results, I can conclude that VPT is a time and cost efficient method!