Meta-SR: A Magnification-Arbitrary Network for Super-Resolution && RDN model implementation

Super Resolution RDN and Meta-SR implementation use tf.estimator

Code is running on tensorflow-gpu==1.12

In 2 gpus mode, the gpu utils can achieve 95%+, if your have more gpus,you can set num_gpus=xxx。

Step by Step tutorial

1.Download the DIV2K_valid_HR.zip DIV2K_train_HR.zip DIV2K_test_HR.zip
2.Set the train_filenames, test_filenames and eval_filenames in model.py
3.Set num_gpus=xxx in config.py to support multiple gpus training, run python model.py or model_new to train and evaluate RDN models
4.For training meta-SR model, just adjust the value in config.py D.model = 'meta-SR'
5.change D.mode = 'predict' in config.py to predict results
6.download the results folder and see the comparison pictures in personal computer.

models

链接: https://pan.baidu.com/s/1kIG0vEdxgS6WhRs0cJlMSg 提取码: 8btp RDN

链接:https://pan.baidu.com/s/1iEeAhN_CpXTkXpvjsNbDKw 密码:gxf1 RDN

链接:https://pan.baidu.com/s/1VRjDlnoNfOT7vXwoTSviVQ 密码:im2v metaSR

This is the training result of RDN model of 3x and 2x SR. rename or put it in proper place according to config.py, finally predict you own pictures.

The 3x model is trained with image scaled to (0, 1). The 2x model is trained with image scaled to (-1, 1).

The meta-SR model is trained with 0~1, but the result are not excellent enough. you can use it just for test.But now exists 2 problems. 1. the result is not as well as RDN. 2. because of use the graph mode, when image is large, weight_prediction part will need use large memory,which will cause OOM, ths solution is mentioned in 5.26UPDATE, but concat small picture into large picture will cause slits in the picture, I will change the code to support large image prediction.

5.8 UPDATE

add a new implementation of meta-SR,(called batch_conv in basemodel.py which support convolution for different kernels values) support batch_size=16 and image_size=50 for multiple scale training. Speed up training speed. model_new.py suppprts batch_size = 16, model.py only support batch_size = 1 in meta-SR mode.

now training mode and some config are all in config.py

5.23 UPDATE

add a new implementation to avoid while_loop in both training and evaluation.

5.24 UPDATE

Now training and evaluation support x.x scales rather than interger scale

5.26 UPDATE

Now training and evaluation all works well in meta-SR. The limitation is because of weight_prediction fc module, when HxW is very large,we dont't have enough gpu memory to store it. So when predict, you need to change 1 large picture into several smaller picture, and predict for each smaller picture, finally concat those results back to finally large picture.