Quantized-FSRCNN

TensorFlow implementation of FSRCNN with quantized version. This implements illustrates that FSRCNN adopted with 16-bit fixed-point representation delivers performance nearly identical as a full precision one.

Prerequisites

• Python 2.7
• TensorFlow version> 1.2
• numpy
• Scipy version > 0.18
• h5py
• PIL

Usage

• Download the caffe training code from here, place folder Train and Test into （$root).

• Open MATLAB and run generate_train.m and generate_test.m to generate training and test data. You can also run data_aug.m to do data augmentation first.

• Modify the flags data_dir and test_dir in model.py as paths to the directory.

• Set NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN in utils.py as the number of samples for training.

• To train a new model, run FSRCNN.py --train True --gpu 0,1 --quantize True for training. And simultaneously run FSRCNN.py --train False --gpu 0 --quantize True for testing. Note that reload flag can be set to True for reloading a pre-train model. More flags are available for different usages. Check FSRCNN.py for all the possible flags.

• After training, you can evaluate performance of the model with TensorBoard. Run tensorboard --logdir=/tmp/FSRCNN_eval. Also, you can extract parameters and save them in the format .mat by set the flag save to True.

References

• FSRCNN
• CIFAR10-tutorial
• drakelevy/FSRCNN-TensorFlow
• liliumao/Tensorflow-srcnn