Reproduce EnhanceNet: Single Image Super-Resolution Through Automated Texture Synthesis by Sajjadi et al.
Given an low-resolution image, the network is trained to produce a 4x resolution image using different loss functions.
- Left: input image (upscaled with bi-cubic interpolation).
- Middle: using the implementation of the author (only contains the generator)
- Right: this implementation (with training code)
- Download MS COCO dataset & VGG19 converted from caffe model zoo:
wget http://images.cocodataset.org/zips/train2017.zip
wget http://models.tensorpack.com/caffe/vgg19.npz- Train an EnhanceNet-PAT using:
python enet-pat.py --vgg19 /path/to/vgg19.npz --data train2017.zip
# or: convert to an lmdb first and train with lmdb:
python data_sampler.py --lmdb train2017.lmdb --input train2017.zip --create
python enet-pat.py --vgg19 /path/to/vgg19.npz --data train2017.lmdbTraining is highly unstable and does not often give good results. The pretrained model may also fail on different types of images. You can download and play with the pretrained model here.
- Inference on an image and output in current directory:
python enet-pat.py --apply --load /path/to/model --lowres input.png