fanwuyang/neural-enhance-super-resolution
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
57 Commits 1 Branch 0 Tags
d18c08f1b59c95d15db03ba01141c2a7bb30d903
Commit Graph
57 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
Alex J. Champandard d18c08f1b5 Integrated reflection padding instead of zero padding for extra quality during training and inference. 2016-11-02 21:39:34 +01:00
Alex J. Champandard 90c0b7ea43 Fix padding code, more reliable for specific upscale/downscale combinations. 2016-11-02 13:03:00 +01:00
Alex J. Champandard 3b2a6b9d8d Add extra padding on input to avoid zero-padding. Experiment with training values from ENet (segmentation). 2016-11-02 12:08:07 +01:00
Alex J. Champandard 7924cc4a85 Improve display and filenames for saving output. 2016-11-02 10:16:31 +01:00
Alex J. Champandard 93e5a41d9a Fix and optimize pre-processing of images. 2016-11-01 20:40:10 +01:00
Alex J. Champandard 11ba505252 Fix for gradient clipping code. 2016-11-01 16:04:33 +01:00
Alex J. Champandard cf65207a2e Use full range of tanh output rather than [-0.5, +0.5], avoids clipping. 2016-11-01 16:02:18 +01:00
Alex J. Champandard c610623b11 Add gradient clipping, helpful for preventing problems with extreme parameters/architectures. 2016-10-31 21:38:23 +01:00
Alex J. Champandard f2494f8078 Add new downscale layers, separate from upscale steps. Renamed --scales to --zoom for inference. 2016-10-31 13:49:17 +01:00
Alex J. Champandard 064f9dd589 Add three image pre-processing options, improve loading code. 2016-10-31 06:21:52 +01:00
Alex J. Champandard 5ef872b876 Add warning for files that may be too large for 4x. 2016-10-30 20:45:10 +01:00
Alex J. ChampandardandGitHub a5ad2c25e6 Merge pull request #18 from alexjc/training 
Training Improvements
 2016-10-30 20:29:13 +01:00
Alex J. Champandard 17fcad8d28 Refactor of changes related to training. 2016-10-30 20:26:03 +01:00
Alex J. ChampandardandGitHub 2b67daedb6 Merge pull request #12 from dribnet/generic_seeds 
Move generation of seeds out of training network.
 2016-10-30 17:19:33 +01:00
Alex J. ChampandardandGitHub cad5eff572 Merge pull request #11 from dribnet/save_every_epoch 
Added --save-every-epoch option
 2016-10-30 17:18:33 +01:00
Alex J. ChampandardandGitHub a9b0cd9887 Merge pull request #9 from zuphilip/patch-1 
Fix some typos in README.
 2016-10-30 17:07:15 +01:00
Alex J. ChampandardandGitHub fcc5e87858 Merge pull request #4 from dribnet/valid_dir 
Add valid dir when necessary.
 2016-10-30 17:06:33 +01:00
Alex J. ChampandardandGitHub 1478977f18 Merge pull request #10 from OndraM/fix-duplicate-param 
Fix duplicate param definition.
 2016-10-30 16:53:48 +01:00
Tom White 8f5167d235 Fix enhancer.process to pass img, seed 2016-10-31 02:14:24 +13:00
Tom White 37cb208374 Move generation of seeds out of training network 
This moves the generation of the image seeds out of the
training network and into the DataLoader. Currently seeds
are computed as bilinear downsamplings of the original image.

This is almost functionly equivalent to the version
it replaces, but opens up new possibilities at training
time because the seeds are now decoupled from the netork.
For example, seeds could be made with different interpolations
or even with other transformations such as image compression.
 2016-10-31 01:25:55 +13:00
Tom White b05ee6ad08 Added --save-every-epoch option 2016-10-31 00:12:28 +13:00
Tom White c5053806bd Add valid dir when necessary 2016-10-31 00:05:37 +13:00
Ondřej Machulda eb25e737cf Fix duplicate param definition 2016-10-30 11:48:53 +01:00
Philipp ZumsteinandGitHub f83e69e96a Fix some typos in README 2016-10-30 11:11:54 +01:00
Alex J. Champandard f68f04fb1c Improve instructions to train custom models so a new file is output and existing one is not loaded. Use --model parameter! 2016-10-30 10:50:57 +01:00
Alex J. Champandard 203917d122 Switch default to small model to reduce memory usage. 2016-10-30 10:49:23 +01:00
Alex J. Champandard 2b5fc8f51d Add docker instructions, fix for slow compute in CPU image. 2016-10-30 02:04:28 +02:00
Alex J. Champandard 74cb95609e Fix for docker build using latest Miniconda and Python 3.5 explicitly. 2016-10-29 23:57:08 +02:00
Alex J. Champandard 99c767b7e2 Add docker configuration files for CPU and GPU. 2016-10-29 22:29:30 +02:00
Alex J. Champandard bf22450b8d Update documentation for new --train usage, minor improvements. 2016-10-29 14:15:24 +02:00
Alex J. Champandard 4c55c48f62 Add argument for specifying training images, cleaned up file handling. 2016-10-29 12:39:07 +02:00
Alex J. Champandard 1c38f2ca31 New meme-friendly image and link to demo. 2016-10-29 01:35:44 +02:00
Alex J. Champandard b1c054ce9f Improve the README for applying and training models. 2016-10-29 00:39:32 +02:00
Alex J. Champandard f868514be3 Improve code for simply applying super-resolution. 2016-10-29 00:18:44 +02:00
Alex J. Champandard 30534c6dd1 Add old station example, fix bank example. 2016-10-28 22:55:10 +02:00
Alex J. Champandard 801a4707f4 Improve README and the visual examples. 2016-10-28 16:29:37 +02:00
Alex J. Champandard 39e12ea205 Add example GIF to the README, tweak text. 2016-10-28 14:38:28 +02:00
Alex J. Champandard c456221cb5 Experiment with recursive super-resolution and weight reuse, mixed results. 2016-10-28 04:28:30 +02:00
Alex J. Champandard 87304c93a6 Use traditional learning rate decaying rather than fast-restarts, works better when training continuously adapting GAN. 2016-10-28 04:27:17 +02:00
Alex J. Champandard 3809e9b02a Add loading of images into a buffer, using multiple fragments per JPG loaded. Works well with larger datasets like OpenImages, fully GPU bound. 2016-10-28 04:25:10 +02:00
Alex J. Champandard 619fad7f3c Add loading parameters from saved models. Clean up learning-rate code. 2016-10-11 17:24:10 +02:00
Alex J. Champandard bb3b24b04c Add README and requirements file. 2016-10-11 17:16:06 +02:00
Alex J. Champandard f978f058dd Add descriptions for command-line arguments. 2016-10-11 11:12:49 +02:00
Alex J. Champandard d8c9292b80 Tweak code for generalized super-resolution. 2016-10-11 08:59:42 +02:00
Alex J. Champandard 9558a11397 Add command-line for super-resolving images without training. 2016-10-07 21:05:38 +02:00
Alex J. Champandard 03a6813e95 Enable adversarial loss during training. 2016-10-05 14:52:28 +02:00
Alex J. Champandard 6542f435b4 Stabilize the training, tuned the architecture. 2016-10-05 00:27:26 +02:00
Alex J. Champandard 4580e5531a Add .gitignore and removed symlink. 2016-10-04 13:19:53 +02:00
Alex J. Champandard b52530cf20 Add support for PRELU (default), optional load/save. 2016-10-04 11:25:03 +02:00
Alex J. Champandard 4d82c7f157 Improve console output. 2016-10-04 02:16:34 +02:00