Suggestions of memory efficiency

Hi all.

I am using the following code from here that implements a Grad-CAM variant of Attention Rollout (a technique to visualize the learned attention maps inside vision transformer networks):

def grad_rollout(attentions, gradients, discard_ratio, output):
    result = torch.eye(attentions[0].size(-1))
    probs = torch.nn.Softmax(dim=-1)(output)
    pred = torch.argmax(probs, dim=-1)

    with torch.no_grad():
        for attention, grad in zip(attentions, gradients):                
            weights = grad
            attention_heads_fused = (attention*weights).mean(axis=1)
            attention_heads_fused[attention_heads_fused < 0] = 0

            # Drop the lowest attentions, but
            # don't drop the class token
            flat = attention_heads_fused.view(attention_heads_fused.size(0), -1)
            _, indices = flat.topk(int(flat.size(-1)*discard_ratio), -1, False)
            #indices = indices[indices != 0]
            flat[0, indices] = 0

            I = torch.eye(attention_heads_fused.size(-1))
            a = (attention_heads_fused + 1.0*I)/2
            a = a / a.sum(dim=-1)
            result = torch.matmul(a, result)
    # Look at the total attention between the class token,
    # and the image patches
    mask = result[0, 0 , 1 :]
    # In case of 224x224 image, this brings us from 196 to 14
    width = int(mask.size(-1)**0.5)
    mask = mask.reshape(width, width).numpy()
    mask = mask / np.max(mask)
    return mask, probs, pred

class VITAttentionGradRollout:
    def __init__(self, model, attention_layer_name='attn_drop',
        self.model = model
        self.discard_ratio = discard_ratio
        for name, module in self.model.named_modules():
            if attention_layer_name in name:

        self.attentions = []
        self.attention_gradients = []

    def get_attention(self, module, input, output):

    def get_attention_gradient(self, module, grad_input, grad_output):

    def __call__(self, input_tensor, category_index):
        output, _ = self.model(input_tensor)
        category_mask = torch.zeros(output.size()).to(DEVICE)
        category_mask[:, category_index] = 1
        loss = (output*category_mask).sum()

        return grad_rollout(self.attentions, self.attention_gradients,
            self.discard_ratio, output)

The code occupies all of my RAM pretty quickly and I am unable to figure out why. I have taken a look into this guide but could not make the necessary changes that might produce more efficiency in terms of memory usage.

Any help is appreciated :slight_smile: