This is the implementation of The All Convolutional Net Paper where all layers of the neural net are conv layers no pool layer added. We compare three implementations of a convolutional neural net.
- Baseline, regular conv->pool->conv->pool... architecture
- Removing all pool layers from (1) but with increased stride in each of the conv layers right before the pool layers.
- Replace all pool layers with a conv layer with increased stride.
Version 0.1
Ciphar-10 and Ciphar-100 image datasets are used
To install the required libraries, run the following from the terminal
pip -r install requirements.txt
To run experiments, use the experiments.notebook.ipynb notebook and follow the examples to train and compare the all convolutional model with increased stride against the standard convolutional architectures with pool layers.
Elias Hussen and Dan Rosenthal contributed to the project