ESPCN(https://arxiv.org/abs/1609.05158) implementation using TensorFlow 2.0
- tensorflow >= 2.0.0
- tensorflow_datasets
- The model architecture is slightly different from the paper.(Especially filter number)
- The phase-shifting code is from https://github.com/kweisamx/TensorFlow-ESPCN.
- The COCO dataset by tensorflow_datsets takes up really much space! (About 77G in my case...)
- TPU code is still in development.
- This code was written for studying, so the code may be hard to understand... I'll try my best to improve code readability.
python3 train.py -exp_name EXP_NAME [-lr LEARNING_RATE] [-batch_size BATCH_SIZE] [-save_dir SAVE_DIR]
Learning rate about 0.001 is recommended. The model won't converge correctly if lr is too large.
python3 tflite.py -exp_name EXP_NAME -model_epoch EPOCH [-saved_dir SAVED_DIR] [-tflite_dir TFLITE_DIR]
I recommend reading the code for more training / converting options.