Image colorization
Structure
- TinyImageNet/train.lua can be used to train the neural network framework. Options to choose the input dataset, loading a pre-trained model, and other hyperparameters are included in the code itself. If running on a GPU, you may use train_AWS.lua
- TinyImageNet/test_checkpoint.ipynb can be used to view the results (or intermediate results) of the training procedure.
- TinyImageNet/Net3.lua Contains the Neural Network framework, which we used for majority of our experiments. Other frameworks are in Net1.lua, Net2.lua, Net4.lua, which we use for experimenting.
- TinyImageNet/VGG.lua Contains functions for loading, pruning the VGG net. Also contains VGG-specific utility functions including preprocessing data and extracting hypercolumns from VGG
- TinyImageNet/Utils.lua Contains some generic torch utility functions which we needed while working on the project.
Setup
After cloning the git repository. Download the following data, which includes pre-trained VGG-caffe model, TinyImageNet dataset.
Data
-
TinyImageNet: TinyImageNet (will contain the training. test, val examples)= http://cs231n.stanford.edu/tiny-imagenet-200.zip (Unzip it)p
-
VGG_Caffe (contains the deploy text and trained model) a. DeployText = https://gist.githubusercontent.com/ksimonyan/211839e770f7b538e2d8/raw/c3ba00e272d9f48594acef1f67e5fd12aff7a806/VGG_ILSVRC_16_layers_deploy.prototxt b. The Model = http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_16_layers.caffemodel
Links
- Torch for Numpy users: https://github.com/torch/torch7/wiki/Torch-for-Numpy-users
- Tensors in Torch: https://github.com/torch/tutorials/blob/master/2_Tensors.ipynb
- 60min Deep neural network tutorial: https://github.com/soumith/cvpr2015/blob/master/Deep%20Learning%20with%20Torch.ipynb
- Image package: https://github.com/torch/image
- nn module documentation: https://github.com/torch/nn/blob/master/doc/overview.md#nn.overview.dok
- https://github.com/torch/torch7 - The whole of Torch code.
- torch-hdf5: https://github.com/deepmind/torch-hdf5
- How to save weights and biases of a trained network: add the gitter link