CRATE (Coding RAte reduction TransformEr)

This repository is the official PyTorch implementation of the papers:

White-Box Transformers via Sparse Rate Reduction [NeurIPS-2023, paper link]. By Yaodong Yu (UC Berkeley), Sam Buchanan (TTIC), Druv Pai (UC Berkeley), Tianzhe Chu (UC Berkeley), Ziyang Wu (UC Berkeley), Shengbang Tong (UC Berkeley), Benjamin D Haeffele (Johns Hopkins University), and Yi Ma (UC Berkeley).
Emergence of Segmentation with Minimalistic White-Box Transformers [CPAL-2024, paper link]. By Yaodong Yu* (UC Berkeley), Tianzhe Chu* (UC Berkeley & ShanghaiTech U), Shengbang Tong (UC Berkeley & NYU), Ziyang Wu (UC Berkeley), Druv Pai (UC Berkeley), Sam Buchanan (TTIC), and Yi Ma (UC Berkeley & HKU). 2023. (* equal contribution)
Masked Autoencoding via Structured Diffusion with White-Box Transformers [ICLR-2024, paper link]. By Druv Pai (UC Berkeley), Ziyang Wu (UC Berkeley), Sam Buchanan, Yaodong Yu (UC Berkeley), and Yi Ma (UC Berkeley).

Also, we have released a larger journal-length overview paper of this line of research, which contains a superset of all the results presented above, and also more results in NLP and vision SSL.

White-Box Transformers via Sparse Rate Reduction: Compression is All There Is? [paper link]. By Yaodong Yu (UC Berkeley), Sam Buchanan (TTIC), Druv Pai (UC Berkeley), Tianzhe Chu (UC Berkeley), Ziyang Wu (UC Berkeley), Shengbang Tong (UC Berkeley), Hao Bai (UIUC), Yuexiang Zhai (UC Berkeley), Benjamin D Haeffele (Johns Hopkins University), and Yi Ma (UC Berkeley).

CRATE (Coding RAte reduction TransformEr)
- Theoretical Background: What is CRATE?
- Autoencoding
Implementation and experiments
Reference

Theoretical Background: What is CRATE?

CRATE (Coding RAte reduction TransformEr) is a white-box (mathematically interpretable) transformer architecture, where each layer performs a single step of an alternating minimization algorithm to optimize the sparse rate reduction objective

where $R$ and $R^{c}$ are different coding rates for the input representations w.r.t.~different codebooks, and the $\ell^{0}$-norm promotes the sparsity of the final token representations $\boldsymbol{Z} = f(\boldsymbol{X})$. The function $f$ is defined as $$f=f^{L} \circ f^{L-1} \circ \cdots \circ f^{1} \circ f^{\mathrm{pre}},$$ where $f^{\mathrm{pre}}$ is the pre-processing mapping, and $f^{\ell}$ is the $\ell$-th layer forward mapping that transforms the token distribution to optimize the above sparse rate reduction objective incrementally. More specifically, $f^{\ell}$ transforms the $\ell$-th layer token representations $\boldsymbol{Z}^{\ell}$ to $\boldsymbol{Z}^{\ell+1}$ via the $\texttt{MSSA}$ (Multi-Head Subspace Self-Attention) block and the $\texttt{ISTA}$ (Iterative Shrinkage-Thresholding Algorithms) block, i.e., $$\boldsymbol{Z}^{\ell+1} = f^{\ell}(\boldsymbol{Z}^{\ell}) = \texttt{ISTA}(\boldsymbol{Z}^{\ell} + \texttt{MSSA}(\boldsymbol{Z}^{\ell})).$$

1. CRATE Architecture overview

The following figure presents an overview of the pipeline for our proposed CRATE architecture:

2. One layer/block of CRATE

The following figure shows the overall architecture of one layer of CRATE as the composition of $\texttt{MSSA}$ and $\texttt{ISTA}$ blocks.

3. Per-layer optimization in CRATE

In the following figure, we measure the compression term [ $R^{c}$ ($\boldsymbol{Z}^{\ell+1/2}$) ] and the sparsity term [ $||\boldsymbol{Z}^{\ell+1}||_0$ ] defined in the sparse rate reduction objective, and we find that each layer of CRATE indeed optimizes the targeted objectives, showing that our white-box theoretical design is predictive of practice.

4. Segmentation visualization of CRATE

In the following figure, we visualize self-attention maps from a supervised CRATE model with 8x8 patches (similar to the ones shown in DINO :t-rex:).

We also discover a surprising empirical phenomenon where each attention head in CRATE retains its own semantics.

Autoencoding

We can also use our theory to build a principled autoencoder, which has the following architecture.

It has many of the same empirical properties as the base CRATE model, such as segmented attention maps and amenability to layer-wise analysis. We train it on the masked autoencoding task (calling this model CRATE-MAE), and it achieves comparable performance in linear probing and reconstruction quality as the base ViT-MAE.

Implementation and Experiments

Constructing a CRATE model

A CRATE model can be defined using the following code, (the below parameters are specified for CRATE-Tiny)

from model.crate import CRATE
dim = 384
n_heads = 6
depth = 12
model = CRATE(image_size=224,
              patch_size=16,
              num_classes=1000,
              dim=dim,
              depth=depth,
              heads=n_heads,
              dim_head=dim // n_heads)

Pre-trained Checkpoints (ImageNet-1K)

| model | dim | n_heads | depth | pre-trained checkpoint | | -------- | -------- | -------- | -------- | -------- | | CRATE-T(iny) | 384 | 6 | 12 | TODO | | CRATE-S(mall) | 576 | 12 | 12 | download link | | CRATE-B(ase) | 768 | 12 | 12 | TODO | | CRATE-L(arge) | 1024 | 16 | 24 | TODO |

Training CRATE on ImageNet

To train a CRATE model on ImageNet-1K, run the following script (training CRATE-tiny)

As an example, we use the following command for training CRATE-tiny on ImageNet-1K:

python main.py 
  --arch CRATE_tiny 
  --batch-size 512 
  --epochs 200 
  --optimizer Lion 
  --lr 0.0002 
  --weight-decay 0.05 
  --print-freq 25 
  --data DATA_DIR

and replace DATA_DIR with [imagenet-folder with train and val folders].

Finetuning pretrained / training random initialized CRATE on CIFAR10

python finetune.py 
  --bs 256 
  --net CRATE_tiny 
  --opt adamW  
  --lr 5e-5 
  --n_epochs 200 
  --randomaug 1 
  --data cifar10 
  --ckpt_dir CKPT_DIR 
  --data_dir DATA_DIR

Replace CKPT_DIR with the path for the pretrained CRATE weight, and replace DATA_DIR with the path for the CIFAR10 dataset. If CKPT_DIR is None, then this script is for training CRATE from random initialization on CIFAR10.

Demo: Emergent segmentation in CRATE

CRATE models exhibit emergent segmentation in their self-attention maps solely through supervised training. We provide a Colab Jupyter notebook to visualize the emerged segmentations from a supervised CRATE model. The demo provides visualizations which match the segmentation figures above.

Link: crate-emergence.ipynb (in colab)

CRATE

Install / Use

README