menjarleev/eecs545

eecs545 project

PhotometricGAN

create image dataset from .mat dataset

python ./data/mat2np.py --dataset_root YOUR_DATASET_ROOT --mat_dir MAT_DIR --img_dir IMG_DIR

This script creates a grayscale jpg image dataset on YOUR_DATASET_ROOT/IMG_DIR using mat dataset YOUR_DATASET_ROOT/MAT_DIR. For example, your dataset is stored in /home/ubuntu/data/BottleData, then YOUR_DATASET_ROOT=/home/ubuntu/data, MAT_DIR=BottleData, and the new generated dataset is stored in /home/ubuntu/data/IMG_DIR

• You should have Training_Data_128, Test_Data_128, Validation_Data_128stored in MAT_DIR, namely BottleData.
• note: please remove redundant .mat files in folder YOUR_DATASET_ROOT/MAT_DIR/PHASE/. For example, you need to manually remove files in /home/ubuntu/data/BottleData/Training_Data_128/*.mat by your own. In this way, only /home/ubuntu/data/BottleData/Training_Data_128/Bottle_95/ contains input.mat and output_\d+ files.

train the model

main.py --test --test_step 50 --dataset_root YOUR_DATASET_ROOT --num_resblock 12 
--train --max_step 500 --validation_interval 100 --log_interval 100 
--lr 0.00001 --validation --optim_name Adam --save_result --batch_size 8 --batch_size_eval 8

Args can be found in option/option.py

configuration

• TODO add configuration files