| Author: | Lukas Turcani |
|---|
This repo provides a number of network architectures, implemented in PyTorch, as well as scripts used to run them.
To train a network, run the appropriate run script. For example:
$ python -m dnn.train_scripts.mnist.progressive_growing_gan
Notice that the scripts are run from the module level using the
python -m flag.
Networks can be modified using command line arguments, for example:
$ python -m dnn.train_scripts.mnist.simple_cnn --conv_in_channels 1 20 50 --conv_out_channels 20 50 60 --conv_kernel_size 5 5 5 --conv_strides 1 1 1 --conv_paddings 0 0 0 --conv_dilations 1 1 1 --pool_kernel_sizes 2 2 2 --pool_strides 2 2 2 --pool_paddings 0 0 0 --pool-dilations 1 1 1 --train_batch_size 100 --label_smoothing 0.5 --epochs 10
changes the default values for training batch size, label smoothing and the number of training epochs and adds a new convolution layer and pooling layer vs the default network. Any number of layers can be added / removed via the command line in this fashion.
Each script can have options viewed by:
$ python -m path.to.train.script --help
This is a short summary of some of the nice results from this repo. Not all implemented architectures are listed here.
This is a somewhat complex GAN, largely due to a unique training procedure where the resolution of the GAN is slowly increased during training.
Run with:
$ python -m dnn.train_scripts.mnist.progressive_growing_gan
Results:
This is a task where the generator is provided with an image that is missing some pixels and it is asked to fill them in. In this example, I cover up the bottom half of MNIST images and get the generator to fill them in. The generator used an autoencoder architecture.
Run with:
$ python -m dnn.train_scripts.mnist.image_inpainting
Results:
A more advanced GAN architecture, which is fully convolutional.
Run with:
$ python -m dnn.train_scripts.mnist.dcgan
Results:
This is a vanilla GAN using feed-forward networks as both the generator and discriminator.
Run with:
$ python -m dnn.train_scripts.mnist.fcgan
Results:
