seathiefwang / rdn-tensorflow

Reproduction of paper:Residual Dense Network for Image Super-Resolution

RDN-Tensorflow

Reproduction of paper:Residual Dense Network for Image Super-Resolution

Requirements

• python > 3.5
• tensorflow > 1.0
• scipy
• numpy
• pillow
• scipy
• tqdm

Train

Prepare training data

Download DIV2K training data.DIV2K

Begin to train

run sh run_train.sh

#!/usr/bin/env bash
CUDA_VISIBLE_DEVICES=0 python3 train.py \
                                 --dataset data/DIV2K \
                                 --imgsize 128 \
                                 --scale 4 \
                                 --globallayers 16 \
                                 --locallayers 8 \
                                 --featuresize 64 \
                                 --batchsize 10 \
                                 --savedir saved_models \
                                 --iterations 1000 \
                                 --usepre 0

Test

python3 test.py --dataset [image dir]

or

python3 test.py --image [single image path]

The architecture of residual dense network (RDN)

Figure 1. The architecture of residual dense network (RDN).

Figure 2. Residual dense block (RDB) architecture.

References

jmiller656/EDSR-Tensorflow

rdn-tensorflow's Issues

module 'data' has no attribute 'preprocess'

The test code is incomplete ，‘module 'data' has no attribute 'preprocess'’

trained model

hi @seathiefwang
thank you for you great work in this project, unfortunately i forced to train with cpu to avoid my gpu memory limitation, but the training will take long time to done, so any chance to share your RDN trained model please