Photographic Image Synthesis with Cascaded Refinement Networks-Pytorch (https://arxiv.org/abs/1707.09405)
This is a Pytorch implementation of cascaded refinement networks to synthesize photographic images from semantic layouts. Now the pretrained model and codes for training the network from scratch are available for 256x512 resolution. Thanks to Qifeng Chen for his tensorflow implementation which helped a lot in developing this pytorch version.
Testing
- Download this package and keep all the subsequent mentioned files in the same folder.
- Download the pretrained VGG19 Net from VGG19
- Download the pretrained weights for the CRN network for 256x512 CRN
- Keep the mode=test and mention the semantic image name to be tested in the Cascadaed_Network_LM_256.py
- The synthesized images will be saved in current folder.
Training
- Follow steps 1 to 3 from the testing steps.
- Resize all the training images to 256x512. Keep the semantic segmentated training images in Label256Full folder and
the RGB training images in RGB256Full (without any subfolders). - Set mode=train in Cascadaed_Network_LM_256.py and run it for desired epochs (default is 200).
Future Work
- Soon the pretrained weights for resolution 512x1024 and 1024x20148 will be available along with training scripts.
Note
- All the codes are written to run on GPU. Suitable changes should be done if you want to run on CPU. Also feel free to
customize it according to your need.