Skip to content

lesson5.md

Latest commit

 

History

History
43 lines (21 loc) · 3.01 KB

lesson5.md

File metadata and controls

43 lines (21 loc) · 3.01 KB

Convolutional Neural Networks

Q1: Why use a pooling layer?

To progressively reduce the spatial size of the representation to reduce the amount of parameters and computation in the network, and hence to also control overfitting.

Resource:

Q2: When trying to run the conv_visualization in my notebook, I'm getting the error ModuleNotFoundError: No module named 'cv2'

Possible solutions:

  • pip install opencv-python

  • conda install -c conda-forge opencv

Q3: Why am I getting this error: Invalid argument 0: Sizes of tensors must match except in dimension 0. Got 499 and 442 in dimension 2 at... ?

Your images are not all the same size.

Possible solutions:

  • transforms.RandomCrop(224)

  • transforms.RandomResizedCrop(224)

  • transforms.Resize(224)

  • transforms.CenterCrop(224)

Q4: While using transforms.Normalize, we pass in a list of means and a list of standard deviations (std) to normalize the input color channels. How do we define means and std values and why is it 0.5 in some cases?

Ideally, you should use the mean and standard deviation for each channel. In the case of Imagenet, you use these precalculated values normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]). The reason for using 0.5 in the case of mnist is to reduce complexity for the readers. Check Soumith Chintala’s comment here.

Q5: In weight=torch.from_numpy(filters).unsqueeze(1).type(torch.FloatTensor) what does unsqueeze(1) mean?

Answered by @Clement:

One of the reasons why we need to unsqueeze which adds in an additional dimension is due to how PyTorch passes in the CNN parameters.(N,Cin,H,W) is the input format to all CNNs defined in PyTorch, where N is the batch size, Cin is the number of channels i.e. grayscale = 1 and color = 3, H is the Height of the image, and W is the Width of the image. It is a neat trick to introduce a dimension of size 1 (using unsqueeze) and since if we are dealing with grayscale images where the number of input channels of grayscale is 1, we can just unsqueeze index 1 (which is the second position/dimension) of the tensor. With the help of @Mitch Deoudes for clarification on this particular question: In this particular case, if you look at the init() function, the "weight" variable is actually being used to directly set the weights of the conv layer. The weight variable is a (4,1,4,4) matrix defined as [output_channels, input_channels, height_of_filter, width_of_filter] as our input provided to the nn.Conv2d is in the format of [input_channels, output_channels, height_of_filter, width_of_filter]. Remember to multiply 2 matrices we have to ensure that the dimensions of the first column and first row to be matched i.e. (NxM x MxO) etc. Therefore, since we are dealing with grayscale we can use a .unsqueeze(1) to introduce a 1 in the second dimension - yet another neat trick.
⚡