A deep neural network for colorizing greyscale images. DTU Project.
The models used in our paper is (from model.py) ColorizerV3 as the reg. CNN and ColorizerUNetV1 as U-Net. We kept the other models
To use our program in a terminal run the command, followed by arguments
python mode [-h] [--output OUTPUT] [--images-path IMAGES_PATH]
[--model MODEL] [--load LOAD] [--save SAVE]
[--log-output LOG_OUTPUT] [--loss LOSS] [--no-l NO_L]
[--save-best SAVE_BEST] [--save-frequency SAVE_FREQUENCY]
[--device DEVICE] [--learning-rate LEARNING_RATE]
[--num-workers NUM_WORKERS] [--batch-size BATCH_SIZE]
[--epochs EPOCHS] [--shuffle SHUFFLE]| Argument | Usage |
|---|---|
| mode | Eval for evaluation or train for training mode |
| -h | Shows the avaliable arguments |
| --output OUTPUT | Output path when evaluating |
| --images-path IMAGES_PATH | Path for the image dataset, should be a folder with 2 sub-folders, validation and training |
| --model MODEL | Specify which model to load [v1 | v2 | v3 | v3 | unet-v1 | unet-v2 | unet-v3] |
| --load LOAD | Path for loading trained parameters |
| --save SAVE | Path for saving trained parameters |
| --log-output LOG_OUTPUT | Path for saving a .csv log over the training and validation loss |
| --loss LOSS | Specify the loss function to use either gan or mse, defaults to mse: Mean Squared Error |
| --no-l NO_L | Either True or False, when evaluating specify if you only want A&B channels |
| --save-best SAVE_BEST | Path for saving the best trained parameteres |
| --save-frequency SAVE_FREQUENCY | How often the program should save the parameters |
| --device DEVICE | Either cuda:0 or cpu, defaults to cpu |
| --learning-rate LEARNING_RATE | Specify a learning to use when training |
| --num-workers NUM_WORKERS | The number of workers to use (rule of thumb: use number of cpu-cores) |
| --batch-size BATCH_SIZE | Specify the batch size to use when trainig, defaults to 32 |
| --epochs EPOCHS | Number of epochs to train |
| --shuffle SHUFFLE | Either True or False, should the program shuffle the dataloading |
If there are multiple cuda devices available the program will automaticaly make use of them with the torch.nn.DataParallel wrapper.
Usage with DTU HPC GPU node1
To submit at job to DTU's HPC GPU gpuv100 node, login to one of the following
nodes: login2.hpc.dtu.dk, login3.hpc.dtu.dk or login3.gbar.dtu.dk via an
ssh connection as follows.
ssh user@login2.hpc.dtu.dkwhere user is either your credentials (for employees) or studentid.
From here use the command linuxsh to go from the login-node to a
general-purpose-node. Here you can execute one of the jobscripts with the following command
bsub < jobscript_train_wide.shThis will send the job to the node configured in the job script.
The job script is configured to save the outputs of the program it executes in
a file named gpu-id.out. id is the job id on the gpu-node.
We are 3 DTU Students from the course Introduction to Intelligent Systems
s174300 Anne Agathe Pedersen
s183903 Niels Kjær Ersbøll
s183926 Anders Bredgaard Thuesen
[1]: DTU's guide for Using GPUs under LSF10