25 Commits

Author	SHA1	Message	Date
igv	e3015e62dc	Adjustments ... - decrease stride by 1 to overlap patches - train only once on the same image per epoch - replace low quality image - default model is 8-0-4-1	2021-03-07 13:16:27 +02:00
Niklas Haas	3f7b9554ea	tensorflow v2 compat	2020-12-16 11:19:13 +01:00
igv	85c1912989	Always use multiple processes for pre-processing	2019-10-04 08:57:17 +03:00
igv	8f30c0f221	Support for scale=1	2019-10-04 08:57:17 +03:00
igv	45c629e5c7	Don't divide test image into patches, don't remove colors and calculate PSNR and SSIM.	2018-08-31 16:21:39 +03:00
igv	09d7d07496	Replace transposed convolution with sub-pixel convolution	2018-08-23 15:29:46 +03:00
igv	d6b13cce93	Optimizations ... Mainly to speed up preprocessing and reduce memory usage Also get rid of expand_data.py	2017-11-04 13:56:05 +02:00
igv	fa94583a94	Support for (MS-)LapSRN	2017-10-21 13:22:25 +03:00
igv	4a02e34490	Support for ESPCN	2017-10-05 10:30:35 +03:00
igv	e0cf791fdf	Support for different radiuses for the deconvolution input tensor ... Also append radius to the checkpoint directory name.	2017-09-10 09:37:39 +03:00
igv	7e828beb9a	Make it possible to disable shrinking/expanding layers	2017-09-03 18:42:33 +03:00
igv	5346bcf898	Adjust deconvolution filter size for different scaling factors	2017-09-01 14:14:27 +03:00
igv	f8af70e011	Append model shape to the checkpoint directory name	2017-08-21 14:06:21 +03:00
igv	15bc798701	Option to add distortions to random low-res images	2017-07-27 14:46:28 +03:00
igv	82ed39e177	Fixes	2017-07-23 15:32:28 +03:00
Niklas Haas	bbbaabe2c0	Switch to the adam optimizer instead of SGD+momentum ... Converges faster, and doesn't get stuck forever like momentum optimization does	2017-07-11 03:00:49 +02:00
Niklas Haas	fad61f249b	Use MS-SSIM instead of MSE (PSNR)	2017-07-11 02:59:34 +02:00
Niklas Haas	db2c179879	utils.py: dump weights in a more machine parsable way ... This dumps them as literal Haskell source code, great for writing a GLSL generator in Haskell - but it should also be valid source code in many other languages	2017-07-03 00:26:32 +02:00
Niklas Haas	d5bf29d6de	model.py: also dump alpha variables ... In addition, I've also fixed a bug where the last alpha was hard coded to alpha7, instead of being dynamic relative to m. This would have introduced a conflict when m was high enough.	2017-07-02 23:29:28 +02:00
Drake Levy	e66845dbe0	Misc changes + make code more pythonic	2017-03-14 13:51:11 -07:00
Drake Levy	007e421178	Almost ready	2017-03-01 17:59:31 -08:00
Drake Levy	8562f557b3	Added result images, README	2017-03-01 17:38:07 -08:00
Drake Levy	9fb554e820	README and is_train -> train	2017-03-01 17:29:13 -08:00
Drake Levy	da3ba8f92b	Updated README and pretty much everything	2017-03-01 17:23:13 -08:00
Drake Levy	9bf8cf669a	Initial commit	2017-03-01 17:16:44 -08:00