Config Description
The following list can be accessed by the terminal
| FLAG |
Supported script |
Use |
Defaults |
Note |
-c |
ALL |
Specify configuration file to use |
Config.yaml |
The Config file used in initializing some parameters in the pipeline |
-a |
ALL |
override configuration arguments in the config file |
None |
Using -a has higher priority than the configuration file selected with -c. E.g: -a Global.model_name= Sauvola2 |
INTRODUCTION TO PARAMETERS OF CONFIGURATION FILE
Global
| Parameter |
Use |
Defaults |
Note |
| use_gpu |
Set using GPU or not |
True |
Used to indicate whether to use GPU or not |
| model_name |
Specify the name of the run to be using in naming the model |
Sauvalo_Finetune |
This will be used in naming the created saved model name besides to Wandb Initializations if Used |
| pretrained_model |
Set the path to pretrained model |
pretrained_models/Sauvola_demo.h5 |
If path is None or doesn’t exist, the model will start from scratch. If the path exists, parameters related to Architecture will be ommited and will be initialized from the saved model |
Train
| Parameter |
Use |
Defaults |
Note |
| Optimizer |
Set Optimizer class name |
adam |
Check keras Optimizers for more |
| Loss |
Set Loss function name |
hinge |
Check keras Losses for more |
| Epoch |
Set Epochs |
100 |
|
| batch_size |
Set Batch Size |
1 |
|
| dataset |
Set the Dataset path |
Dataset |
Datset Folder should contain all images with names=TRAIN_*, and for each image there should be ground truth and source having same name but one ending with _source.png and groundtruth with _target.png e.g. for one image: Bickely2010_H01_source.png, Bickely2010_H01_target.png |
| Callbacks |
Callbacks Class |
|
|
| callbacks |
Set callbacks to be used |
['ModelCheckpoint','TensorBoard',
'EarlyStopping','ReduceLROnPlateau'] |
Add WandbCallback to the list to enable wandb API Visualizations |
| patience |
Set patience to be used in ['EarlyStopping',
'ReduceLROnPlateau'] |
15 |
Note in ReduceLROnPlateau the patiance is divied by 2 |
Architecture
In Sauvolanet, the network is divided into four stages: SauvolaMultiWindow, Pixelwise Window Attention (PWA), and Adaptive Sauolva Threshold (AST)
| Parameter |
Use |
Defaults |
Note |
| SauvolaMultiWindow |
SauvolaMultiWindow Class |
|
|
| window_size_list |
Sets the windows list sizes |
[3,5,7,11,15,19] |
[int], the used window sizes to compute Sauvola based thresholds |
| norm_type |
SauvolaMultiWindow Class |
'bnorm' |
str, one of {'inorm', 'bnorm'}, the normalization layer used in the conv_blocks {inorm: InstanceNormalization, bnorm: BatchNormalization} |
| activation |
Set the activation class name |
'relu' |
str, the used activation function inside the SauvolaMultiWindow Convolutions |
| base_filters |
Sets the number of base filters |
4 |
the number of base filters used in conv_blocks, i.e. the 1st conv uses base_filter of filters the 2nd conv uses 2*base_filter of filters and Kth conv uses K*base_filter of filters |
| init_k |
Set param k in Sauvola binarization |
0.2 |
Initialize param k in Sauvola binarization |
| init_R |
Set param R in Sauvola binarization |
0.5 |
Initialize param R in Sauvola binarization |
| train_k |
Set param k training flag |
True |
whether or not train the param k in Sauvola binarization |
| train_R |
Set param R training flag |
True |
whether or not train the param R in Sauvola binarization |
| DifferenceThresh |
DifferenceThresh Class |
|
|
| init_alpha |
Set param alpha in Sauvola binarization |
16 |
Initialize param alpha in Sauvola binarization |
| train_alpha |
Set param alpha training flag |
True |
whether or not train the param alpha in Sauvola binarization |