Training super-resolution networks with ESRGAN

SRGAN is a landmark SR technique. It is quickly approaching "seminal" status because of how many papers use some or all of the techniques originally introduced in this paper. ESRGAN is a followup paper by the same authors which strictly improves the results of SRGAN.

After considerable trial and error, I recommend an additional set of modifications to ESRGAN to improve training performance and reduce artifacts:

Gradient penalty loss on the discriminator keeps the discriminator gradients to the generator in check.
Adding noise of 1/255 to the discriminator prevents it from using the fixed input range of HR images for discrimination. (e.g. - natural HR images can only have values in increments of 1/255, while the generator has continuous outputs. The discriminator can cheat by using this fact.)
Adding GroupNorm to the discriminator layers. This further stabilizes gradients without the downsides of BatchNorm.
Adding a translational loss to the generator term. This loss works by computing using the generator to compute two HQ images during each training pass from random sub-patches of the original image. A L1 loss is then computed across the shared region of the two outputs with a very high gain. I found this to be tremendously helpful in reducing GAN artifacts as it forces the generator to be self-consistent.
Use a vanilla GAN. The ESRGAN paper promotes the use of RAGAN but I found its effect on result qualit to be minimal with the above modifications. In some cases, it can actually be harmful because it drives strange training dynamics on the discriminator. For example, I've observed the output of the discriminator to sometimes "explode" when using RAGAN because it does not force a fixed output value. It is also more computationally expensive to compute.

The examples below have all of these modifications added. I've also provided a reference file that should be closer to the original ESRGAN implementation, train_div2k_esrgan_reference.yml.

Training ESRGAN

DLAS can train and use ESRGAN models end-to-end. These docs will show you how.

Dataset Preparation

Start by assembling your dataset. The ESRGAN paper uses the DIV2K and Flickr2K datasets. These include a small set of high-resolution images. ESRGAN is trained on small sub-patches of those images. Generate these patches using the instructions found in 'Generating a chunked dataset' here.

Consider creating a validation set at the same time. These can just be a few medium-resolution, high-quality images. DLAS will downsample them for you and send them through your network for validation.

Training the model

Use the train_div2k_esrgan.yml configuration file in this directory as a template to train your ESRGAN. Search the file for <-- to find options that will need to be adjusted for your installation.

Train with: python train.py -opt train_div2k_esrgan.yml

Note that this configuration trains an RRDB network with an L1 pixel loss only for the first 100k steps. I recommend you save the model at step 100k (this is done by default, just copy the file out of the experiments/train_div2k_esrgan/models directory once it hits step 100k) so that you do not need to repeat this training in future experiments.

Using an ESRGAN model

Image SR

You can apply a pre-trained ESRGAN model against a set of images using the code in test.py. Documentation for this script is forthcoming but basically you feed it your training configuration file with the pretrain_model_generator option set properly and your folder with test images pointed to in the datasets section in lieu of the validation set.

Video SR

I've put together a script that strips a video into its constituent frames, applies an ESRGAN model to each frame one a time, then recombines the frames back into videos (without sound). You will need to use ffmpeg to stitch the videos back together and add sound, but this is trivial.

This script is called process_video.py and it takes a special configuration file. A sample config is provided in rrdb_process_video.yml in this directory. Further documentation on this procedure is forthcoming.

Fun fact: the foundations of DLAS lie in the (now defunct) MMSR github repo, which was primarily an implementation of ESRGAN.

4.5 KiB Raw Blame History