LearnToPayAttention

PyTorch implementation of ICLR 2018 paper Learn To Pay Attention

My implementation is based on "(VGG-att3)-concat-pc" in the paper, and I trained the model on CIFAR-100 DATASET.
I implemented two version of the model, the only difference is whether to insert the attention module before or after the corresponding max-pooling layer.

(New!) Pre-trained models

Google drive link

Dependences

• PyTorch (>=0.4.1)
• OpenCV
• tensorboardX

NOTE If you are using PyTorch < 0.4.1, then replace torch.nn.functional.interpolate by torch.nn.Upsample. (Modify the code in utilities.py).

Training

1. Pay attention before max-pooling layers

python train.py --attn_mode before --outf logs_before --normalize_attn --log_images

2. Pay attention after max-pooling layers

python train.py --attn_mode after --outf logs_after --normalize_attn --log_images

Results

Training curve - loss

The x-axis is # iter

1. Pay attention before max-pooling layers
   

2. Pay attention after max-pooling layers
   

3. Plot in one figure
   

Training curve - accuracy on test data

The x-axis is # epoch

1. Pay attention before max-pooling layers
   

2. Pay attention after max-pooling layers
   

3. Plot in one figure
   

Quantitative results (on test data of CIFAR-100)

Attention map visualization (on test data of CIFAR-100)

From left to right: L1, L2, L3, original images

1. Pay attention before max-pooling layers
   

2. Pay attention after max-pooling layers