This code allows analyzing the effect of the pooling layers in a CNN. It implements a simple CNN model trained to distinguish between filled rectangles (class 0) and circles (class 1). The model takes a 224x224 image as input and outputs a value between 0 and 1.
Pooling layers cause the model to be shift variant: shifting the input image by 1px causes the output to be different, even though CNNs are often said to be shift invariant.
Analyze the effect of the single pooling layer in the pre-trained network:
- Go to the folder
src/ - Then, execute
python main.py --mode analyze --num_pooling_layers 1.
This command shifts an input shape (rectangle, class 0) to be classified all around the input image and then records the prediction (should be close to 0 because of predicting class 0) for all the tested image locations. The output should look as follows:
As can be seen, depending on the position of the shape in the input image the output slightly differs. This shows that CNNs with pooling layers are not shift invariant.
General usage:
python main.py --mode {train,infer,analyze} --num_pooling_layers {0,1,2}.
Parameters:
- mode: first, train the model, then, you can either infer a few samples or analyze the effect of the pooling layers
- num_pooling_layers: number of pooling layers in the model, either 0, 1 or 2
Example to train the model:
python main.py --mode train --num_pooling_layers 1.
Example to infer a few samples with the trained the model:
python main.py --mode infer --num_pooling_layers 1.
This article discusses the effect and shows suggestions on how to reduce it.