Awesome

Consistency driven Sequential Transformers Attention Model for Partially Observable Scenes

Authors: Samrudhdhi Rangrej, Chetan Srinidhi, James Clark Accepted to: CVPR'22 Paper  

An overview of our Sequential Transformers Attention Model (STAM). The STAM consists of a core T, classifiers G and D, an actor A, and a critic C (only used during training). Each training iteration consists of three steps: Step 1 (green path): Given a complete image X, the teacher model predicts a soft pseudo-label q(y|X). Step 2 (blue path): Given glimpses g_0:t, STAM predicts class distributions p_g(y_t|f^g_t) and p_d(y_t|f^d_t), value V(s_t), and attention policy π(l_t+1|s_t). Step 3 (orange path): An additional glimpse is sensed. Step 2 is repeated using all glimpses (including the additional glimpse) and the losses are computed. The model parameters are updated using the gradients from Step 2.

Requirements

• torch==1.8.1
• torchvision==0.9.1
• tensorboard==2.5.0
• timm==0.4.9
• fire==0.4.0

Datasets

• ImageNet
• fMoW (Download and prepare the dataset following the instructions provided in PatchDrop repository)

Training

• Adapt paths.py
• Teacher models:
  • ImageNet: Download weights for DeiT-small distilled model from deit.
  • fMoW: Finetune DeiT-small model from deit on the fMoW dataset for 100 epochs using the default hyperparameter setting and vertical flip augmentation.
• Student models (STAM):
  • ImageNet: run run_imagenet.sh
  • fMoW: run run_fmow.sh

Evaluation

• ImageNet: run eval_imagenet.sh
• fMoW: run eval_fmow.sh

Visualization

Visualization of glimpses selected by STAM on example images from t=0 to 15. (top) ImageNet; (bottom) fMoW. Complete images are shown for reference only. STAM does not observe a complete image.

Acknowledgement

Our code is based on deit.