Zhen Li

Yes, you're right.

When training RDN, please ensure the `rgb_range` in your `*.json` file is 1. If it is, the average training loss (9.177..) for `rgb_range=1` is too large. Smaller than 1 is...

I re-confirmed the training process of RDNx2. It is OK. My log is shown below: ```shell Method: RDN || Scale: 2 || Epoch Range: (1 ~ 1000) ===> Training Epoch:...

I found the reason causing your mentioned problems. Comment this [line](https://github.com/Paper99/SRFBN_CVPR19/blob/da28bcf985543f7c24f456d43125c3d6fb7badc0/solvers/SRSolver.py#L312) for training a model without **Kaiming initialization**. This is a very interesting phenomenon for image SR.

Maybe you need to modify your network options in `test_*.json` according to your training configures (network options in `train_*.json`).

Each batch will first split into N (depending on the value of `split_batch`) sub-batches before feeding to the network. The larger `split_batch` will lead to less GPU memory usage during...

When you set a large batchsize, the option `split_batch` make it possible to update the parameters only once.

Hello! Sorry for the inconvenience. We didn't release the code about curriculum learning. For data generation, you could modify this [matlab script](https://github.com/Paper99/SRFBN_CVPR19/blob/master/scripts/Prepare_TrainData_HR_LR.m) according to [the script about generating test data](https://github.com/Paper99/SRFBN_CVPR19/blob/master/results/Prepare_TestData_HR_LR.m)....

1: For fast data I/O. 2, 3: We will test the compatibility of multi-gpu and give your feedback in a few days.

Hi, 1. Yes, you can. 2. This is the weakness of pixel-wise losses (i.e. L2 loss and L1 loss). In your final loss function, the simplest way for generating perceptually...