In this exercise we will look more into ConvNets
As discussed in the lecture, the key ideas to convnets are local connectivity and weight sharing which lead us to convolutions.
Before trying to learn convolutional filters, let's manually construct them
Write a function that downloads an image at a given URL and converts it into greyscale and returns a 2D array.
Tip: use the requests and the PIL python libraries
Research how to access the parameters of the convolution kernel module torch.nn.Conv2D in PyTorch and make sure you understand where
kernel size, input channels and output channels come into definint the shape of the parameter tensors.
Create a Conv2D module with 1 input channel and 2 output channels with kernel size 3
- set the weights of of first output channel is a vertical edge detector
- set the weights of the second output channel is a horizontal edge detector
- create a size 1 input batch with a greyscale image of your choice and visualize the outputss
Create a ConvNet with the following structure
- Conv 5 x 5 ( 1 -> 16 channels) -> ReLU -> MaxPool 2 x 2
- Conv 3 x 3 ( 16 -> 16 channels) -> ReLU -> MaxPool 2 x 2
- Conv 2 x 2 ( 16 -> 32 channels)
Find out what the output on a random MNIST-like torch tensor is, i.e. x = torch.randn(123,1,28,28)
- Use
torch.nn.Unflattento make the model work with "flattened" MNIST tensors as we had them in the prior exercise, i.e. the model should work with a tensorx = torch.randn(123,784) - Use
torch.nn.Flattento flatten all the remaining dimensions after the three convolutions into
How big is this intermediate representation of an image?
- Take a note of the number of outputs after
Flattenand add a final lineary projection layer (i.e. a perceptron) to the network to predict 10 logits - Adapt your solution from Exercise 6 to train a CNN on mnist images
- Check out the following link, and try using the famour
inception_v3model to process an image of your choice - https://pytorch.org/hub/pytorch_vision_inception_v3/