nagadomi
81df729a8a
Remove -loss option
2016-07-09 15:05:11 +09:00
nagadomi
edac608f18
Add support for user specified pairwise data for universal filter
2016-07-05 02:42:40 +09:00
nagadomi
de669d24a4
Change default -validation_crops to 200.
2016-06-20 15:56:39 +09:00
nagadomi
b8ff8c6787
Remove -gamma_correction option
2016-06-10 07:37:39 +09:00
nagadomi
01b2e6d441
Remove -upsampling_filter option
2016-06-10 07:34:11 +09:00
nagadomi
e5cfd3dfce
Add -resume option
2016-06-09 02:39:52 +09:00
nagadomi
6c758ec5c0
Correct messages
2016-06-08 07:52:38 +09:00
nagadomi
307ae40883
Add noise_scale training
2016-06-08 06:39:36 +09:00
nagadomi
d0630d3a20
individual filters and box-only support
2016-06-06 14:04:13 +09:00
nagadomi
c89fd7249a
Add learning_rate_decay
2016-06-02 10:11:15 +09:00
nagadomi
634046d5f0
Fix training mode
2016-05-29 05:50:53 +09:00
nagadomi
b96bc5d453
Use correct criterion
2016-05-28 10:56:15 +09:00
nagadomi
8088460a20
Add oracle_rate option
2016-05-27 16:54:29 +09:00
nagadomi
8fec6f1b5a
Change the learning rate decay rate
2016-05-21 09:56:26 +09:00
nagadomi
7814691cbf
Add resize_blur parameter
...
latest graphicsmagick is required
2016-05-21 09:54:12 +09:00
nagadomi
a210090033
Convert model files; Add new pretrained model
...
- Add new pretrained model to ./models/upconv_7
- Move old models to ./models/vgg_7
- Use nn.LeakyReLU instead of w2nn.LeakyReLU
- Add useful attribute to .json
New JSON attribute:
The first layer has `model_config` attribute.
It contains:
model_arch: architecture name of model. see `lib/srcnn.lua`
scale_factor: if scale_factor > 1, model:forward() changes image resolution with scale_factor.
channels: input/output channels. if channels == 3, model is RGB model.
offset: pixel size that is to be removed from output.
for example:
(scale_factor=1, offset=7, input=100x100) => output=(100-7)x(100-7)
(scale_factor=2, offset=12, input=100x100) => output=(100*2-12)x(100*2-12)
And each layer has `class_name` attribute.
2016-05-15 03:04:08 +09:00
nagadomi
51ae485cd1
Add new models
...
upconv_7 is 2.3x faster than previous model
2016-05-13 09:49:53 +09:00
nagadomi
b8088ca209
Remove the limit of learning_rate_decay
2016-04-30 13:48:24 +09:00
nagadomi
7af5c9443d
Add model option and 12 layers net
2016-04-23 09:19:03 +09:00
nagadomi
68a6d4cef5
Use MSE instead of PSNR
...
PSNR depends on the minibatch size and those group.
2016-04-17 02:08:38 +09:00
nagadomi
fa9355be7c
Fix validation metric
2016-04-16 03:52:05 +09:00
nagadomi
a92b2cb386
Fix progressbar
2016-04-16 03:45:21 +09:00
nagadomi
ea780f1871
Fix validation metric
2016-04-16 03:23:37 +09:00
nagadomi
30fe5db735
Add upsampling_filter option
2016-04-02 22:03:27 +09:00
nagadomi
13f702b968
Add support for resizing with gamma correction
2016-03-28 19:07:09 +09:00
nagadomi
d4833160c7
Optionalize downsampling filters
2016-03-17 17:58:37 +09:00
nagadomi
4d115e4bdb
Add support for plotting loss chart
2016-03-14 05:06:14 +09:00
nagadomi
459c7c5e18
Use similar distribution on train/validation
2016-03-12 08:50:55 +09:00
nagadomi
1900ac7500
Use PSNR for evaluation
2016-03-12 06:53:42 +09:00
nagadomi
664322fa97
Remove cleanup_model.lua; Change model format from binary to ascii
2016-03-12 05:20:19 +09:00
nagadomi
290a5f960b
Use Huber loss instead of MSE
2016-03-11 11:13:45 +09:00
nagadomi
9b238bd693
Use clearState()
2016-03-11 11:12:02 +09:00
nagadomi
9f935835dd
Add -save_history option
2015-12-04 18:49:34 +09:00
nagadomi
aaac6ed6e5
Refactor training loop
...
more shuffle
2015-11-30 17:18:52 +09:00
nagadomi
c72ec3112b
Add -random_unsharp_mask_rate option for photo
2015-11-27 18:36:36 +09:00
nagadomi
d8ba661d6b
Add -jpeg_chroma_subsampling_rate option for JPEG denoise training
2015-11-26 17:10:57 +09:00
nagadomi
42bd89151e
Add -gpu option in train.lua
2015-11-13 19:26:58 +09:00
nagadomi
5ab2c605c8
Fix header include order issues
2015-11-08 18:31:46 +09:00
nagadomi
797b45ae23
Use roundf-like clip for 8 bit-depth image
...
Maybe PSNR +0.03 improved by this commit
2015-11-08 05:44:14 +09:00
nagadomi
3ea16b3b86
tunable parameters
2015-11-07 07:18:22 +09:00
nagadomi
c773e18e59
Add trade-off parameter for noise reduction
2015-11-07 06:37:53 +09:00
nagadomi
903d945652
cleanup
2015-11-06 10:08:54 +09:00
nagadomi
b35a9ae7d7
tuning
2015-11-03 06:10:44 +09:00
nagadomi
490eb33a6b
Minimize the weighted huber loss instead of the weighted mean square error
...
Huber loss is less sensitive to outliers(i.e. noise) in data than the squared error loss.
2015-10-31 22:05:59 +09:00
nagadomi
da786e15ba
remove noise_scale training
2015-10-28 16:27:31 +09:00
nagadomi
3abc5a03e3
refactor
2015-10-28 16:01:07 +09:00
nagadomi
8dea362bed
sync from internal repo
...
- Memory compression by snappy (lua-csnappy)
- Use RGB-wise Weighted MSE(R*0.299, G*0.587, B*0.114) instead of MSE
- Aggressive cropping for edge region
and some change.
2015-10-26 09:23:52 +09:00
nagadomi
54580ba8c0
add data augmentation method that uses overlay
2015-08-02 22:02:14 +09:00
nagadomi
e3d3a8355c
change training script
...
- add AlexNet's color noise (default: false)
- add `photo` category for noise level setting
2015-07-11 21:57:04 +09:00
nagadomi
fd9dadd7a4
use settings.validation_crop
2015-06-26 20:12:51 +09:00