Iterate over 5d tensor

vision
1

Hello,

I want to iterate over a 5d tensor ‘images’, extract 4d tensors and do some operation on the 4d tensors which is an individual bag of (batch size, channels, height, width).So in a 5d tensor, 1st dimension is batch size, 2nd dimension is bag size which is number of images in a bag (lets say 20),3rd dimension is RGB, 4th and 5th dimension are height and width.here is my for loop which I am using but this very slow.is there any other way to make this code faster? the for loop is very slow.

                new_output = []
                for i in range(args.bag_size):
                    individual_bag = images[:,i,:,:,:]
                    print('individual_bag shape',individual_bag.shape)
                    intermediate_output = model.get_intermediate_layers(individual_bag, n)
                    output = torch.cat([x[:, 0] for x in intermediate_output], dim=-1)
                    new_output.append(output)
                output = torch.mean(torch.stack(new_output),dim = 0)

Thanks,
Rohan

2

I don’t know what each “individual bag” represents, but since you are indexing each separately I assume they don’t have any dependency between each other. If so, would it be possible to flatten dim1 into dim0 and just increase the batch size for a single forward pass?

3

Each individual bag represents lets say 20 images, so instead of passing each image, I am passing a bag of images.I am extracting features from the network for each bag in a batch, and doing this for each batch.The purpose of taking a bag of images is to get mean representation but I think I would need to do it for each batch separately.

4

I’m unsure why you would need to perform this operation for each bag separately. You could still reshape the output back before applying the mean operation.