ECE-285-ViT

Final project of ECE 285

TODO List

• Build module :
  • patch embedding (need to be debugged and test) : divide the image and do the patch and position embedding, include the Linear Projection and Flatten-related work.
  • Transformer Encoder (need to be built from scratch): build 1 but reuse the block for 5~6 times
    • Norm
    • Multihead Attention
    • Norm
    • MLP
  • Head
• Pretrain: We can use dataset ImageNet (same as ViT paper)
  • dataset website: [https://www.image-net.org/download.php]
• Train and test on our dataset

The structure

• All the source code are in model_src folder
• All the notebooks we used are in the root folder
• Here is our presentation video link (you need a UCSD email account to access it): https://drive.google.com/file/d/1dHjMl5QP530RDQmXeosVvtFcKUR_e9LG/view?usp=drive_link
• I also uploaded our code to my google drive if this repo link thing does not work.