forked from idealo/image-super-resolution
-
Notifications
You must be signed in to change notification settings - Fork 0
/
CHANGELOG
30 lines (22 loc) · 2.19 KB
/
CHANGELOG
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
- 2019-03-14 1.9.0 beta: added deep features from VGG19 and a discriminator for GAN training.
Moved all non strictly architecture building operations outside of the model files.
The models are combined when needed in the Trainer class.
In order to allow for GAN training `fit_generator` function had to be replaced
with the more granular `train_on_batch`. Now the project relies on
custom data handlers and loggers instead of the custom Keras generator.
- 1.9.2 beta: Included tutorial notebooks, modified init files for less verbose import statements.
Comprehensive docstrings refactor.
- 2019-03-25 1.9.3 beta: now ISR models are child of the ImageModel class, whose purpose is to collect
common functions across models. At the moment it only contains a predict function that takes numpy arrays
as input and returns numpy arrays in a proper image format.
- 2019-03-26 2.0.0 beta: brought a number of training hyperparameters to the surface of trainer and the CLI assistant.
Removed most hard-coded variables and improved assistant. Weights saving now allows metric selection, with suggested metrics list in assistant.
- 2019-04-01 2.0.1 beta: Cleaner training session tracking: now weights are saved together with a yaml configuration file that contains hyperparameters and other relevant training details.
Partial migration to pathlib from os.
- 2019-04-02 2.0.2 beta: Predictor uses the model parameters and the configuration file rather than inferring details from the name of the weights. Minor changes to input/output folder structure. Partial migration to pathlib.
Added configuration settings to pre-trained weights.
- 2019-04-03 2.0.3 beta: Parametric metrics and losses in trainer. Custom weights initialization range with RandomUniform initializer.
Added PSNR evaluation on Y channel (for literature comparison). Automatic session config generation from trainer input parameters.
Uniform naming for feature extractor and name change for generator network.
- 2019-04-16 2.0.5 beta: Added flatness check scheduler. Refactored some trainer variables into dictionaries for a more compact configuration summary.
-2019-05-30 2.1 beta: Added large image inference. Minor fixes and added new non-artifact-removing GANS model.