This repo contains VisuCNN.ipynb which is a good template for many useful tools that can help in understanding CNNs.
- Visualise the filters, the transition of the image as it passes through different convolutions. This is very useful in understanding how and what the CNN learns.
- Data added to tensorboard along with histograms of weights and biases, learning rates, batch accuracies, etc.
- Since the filter and image transitions are added after every epoch, it allows us to visualise the gradual transition of the model across epochs.
- The facility to properly catalogue different tensorbaord files in directories.
- Adding model checkpoints after every epoch, and saving model state after all epochs have been completed.
- The whole file can be modified to run with any model, even pre-trained.
Install the dependecies and check it out!