- Feature learning: The CNN layer uses filters and pooling to extract + condense the input into a smaller but paramount features.
- Classification: It flatten out the CNN features, and use fully connected layer to do the classification task
CNN specification
-
filters:- purpose: filter is a weight metrix that learns to represent abstraction of the given matrix
- It will reduce the size of the original matrix as side effect, but can be dealt with using
padding - num_Kernels(filters) = num_input_channels * num_output_channels
- params:
- in_channels: For colour image, it's 3. For grey scale image, it's 1
- out_channels: Anything. The more channels, the more feature space representation
- kernel_size: the size of filter matrix
- stride: The number of pixel it shifts each time, usually 1
-
pooling:- purpose: condense the matrix representation into a smaller matrix. ie. Max pooling
- benefit: This will force the neural network to represent information condensely (throw away the unimportant)
- params:
- kernel_size: the size of matrix
- stride: The number of pixel it shifts each time
# create local host for web UI to check experiment
mlflow ui
#