modified training scripts

README.md

@@ -58,4 +58,25 @@ The dataset is composed of 2 folders:
 
 In order to train the model, you first have to divide the dataset into training, validation and test splits.  
 You can do this by running the [`notebook/Training dataset creation.ipynb`](notebooks/Training%20dataset%20creation.ipynb) notebook.
+**Please notice** that these splits and patches are the ones used in the paper, but you can create your own by modifying the notebook.  
 
+If you want to inspect the raw products, a starting point is the [Raw satellite products processing](notebooks/Raw%20satellite%20products%20processing.ipynb) notebook.  
+
+# The whole pipeline
+## Normalization strategies
+All the normalization strategies used in the paper are provided as classes in the [`isplutils/data.py`](isplutils/data.py) file.  
+Please notice that for the `MinPMax` strategy, we used the [RobustScaler](https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.RobustScaler.html) implementation from `sklearn`.  
+Statistics are learned from the training set, and then applied to the validation and test sets.  
+We provide the scalers used in the paper, one for each satellite product, inside the folders of `pristine_images/full_res_products`.
+
+## Model training
+The `train_fe.py` takes care of training the models.  
+You can find the network definition in the [`isplutils/network.py`](isplutils/network.py) file.  
+All the hyperparameters for training are listed in the file.  
+To replicate the models used in the paper, follow the [train_all.sh](bash_scripts/train_all.sh) bash script.
+
+## Model evaluation
+Inside the `data/spliced_images` folder are contained the two datasets used in the paper, i.e.:
+1. `Standard Generated Dataset (SGD)`: images generated by simply normalizing the dynamics between 0 and 1 using a maximum scaling;
+2. `Histogram Equalized Generated Dataset (HEGD)`: images generated by equalizing the histogram of the images using a uniform distribution.
+Inside each folder there is a Pandas DataFrame containing info on the images.

bash_scripts/train_all.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+
+# USER PARAMETERS (put your device configuration params here)
+DEVICE=0
+TRAIN_DIR=../data/pristine_images/train_patches # path to training patches (YOUR PATH MAY BE DIFFERENT IF YOU SAVED THE SPLITS IN ANOTHER FOLDER, CHECK IT)
+VAL_DIR=../data/pristine_images/val_patches # path to training patches (YOUR PATH MAY BE DIFFERENT IF YOU SAVED THE SPLITS IN ANOTHER FOLDER, CHECK IT)
+
+echo ""
+echo "-------------------------------------------------"
+echo "| Train with MinPMax 99th percentile threshold |"
+echo "-------------------------------------------------"
+python ../train.py --gpu $DEVICE --batch_size 10 --num_iteration 128 --learning_rate 0.0001 \
+--epochs 500 --train_dir $TRAIN_DIR --val_dir $VAL_DIR --num_tiles_peracq 200 --batch_num_num_tiles_peracq 10 \
+--batch_num_pos_pertile 6 --scaler_type 99th_percentile --mean_robust_scaling --input_fp_channels 3 \
+--output_fp_channels 3
+
+echo ""
+echo "-------------------------------------------------"
+echo "| Train with MinPMax 95th percentile threshold |"
+echo "-------------------------------------------------"
+python ../train.py --gpu $DEVICE --batch_size 10 --num_iteration 128 --learning_rate 0.0001 \
+--epochs 500 --train_dir $TRAIN_DIR --val_dir $VAL_DIR --num_tiles_peracq 200 --batch_num_num_tiles_peracq 10 \
+--batch_num_pos_pertile 6 --scaler_type 95th_percentile --mean_robust_scaling --input_fp_channels 3 \
+--output_fp_channels 3
+
+echo ""
+echo "-------------------------------------------------"
+echo "| Train with MaxAbs scaling |"
+echo "-------------------------------------------------"
+python ../train.py --gpu $DEVICE --batch_size 10 --num_iteration 128 --learning_rate 0.0001 \
+--epochs 500 --train_dir $TRAIN_DIR --val_dir $VAL_DIR --num_tiles_peracq 200 --batch_num_num_tiles_peracq 10 \
+--batch_num_pos_pertile 6 --scaler_type sat_tiles_scaler --input_norm max_scaling --input_fp_channels 3 \
+--output_fp_channels 3
+
+echo ""
+echo "-------------------------------------------------"
+echo "| Train with HistogramEqualization scaling |"
+echo "-------------------------------------------------"
+python ../train.py --gpu $DEVICE --batch_size 10 --num_iteration 128 --learning_rate 0.0001 \
+--epochs 500 --train_dir $TRAIN_DIR --val_dir $VAL_DIR --num_tiles_peracq 200 --batch_num_num_tiles_peracq 10 \
+--batch_num_pos_pertile 6 --scaler_type sat_tiles_scaler --input_norm uniform_scaling --input_fp_channels 3 \
+--output_fp_channels 3

isplutils/data.py

@@ -21,7 +21,6 @@
 from joblib import load
 import os
 from abc import ABC, abstractmethod
-import imgaug.augmenters as iaa
 import skimage.exposure
 from typing import List
 
@@ -354,24 +353,6 @@ def __init__(self, batch_size, patch_size, data_dir, num_iteration, num_pos_pert
         self.input_norm = input_norm
         self.he_norm = he_norm
         self.gray_scale = gray_scale
-        self.p_aug = p_aug
-        if p_aug > 0:
-            # Let's keep contrast augmentation aside for a moment
-            # self.augs = iaa.Sometimes(p_aug, iaa.OneOf([iaa.SigmoidContrast(gain=(5, 20), cutoff=(0.25, 0.75)),
-            #                                             iaa.LogContrast(gain=(0.6, 1.4)),
-            #                                             iaa.LinearContrast((0.4, 1.6))]))
-            # Let's tru adding uniform equalization
-            # self.augs = iaa.Sometimes(p_aug,
-            #                           iaa.Lambda(lambda x, random_state, parents, hooks:
-            #                                      [SatTilesScaler().normalize_product(image, 'uniform_scaling',
-            #                                                                         self.mean_scaling_strategy)
-            #                                       for image in x],
-            #                                      lambda x, random_state, parents, hooks: x))
-            self.augs = iaa.Sometimes(p_aug,
-                                      iaa.Lambda(lambda x, random_state, parents, hooks:
-                                                 [skimage.exposure.equalize_hist(image, nbins=50)
-                                                  for image in x],
-                                                 lambda x, random_state, parents, hooks: x))
 
         # Load the data (first, set the random split for loading the patches)
         self.split_seed = np.random.random_integers(42) if split_seed is None else split_seed
@@ -523,16 +504,6 @@ def __init__(self, batch_size, patch_size, data_dir, num_iteration, num_pos_pert
         self.input_norm = input_norm
         self.he_norm = he_norm
         self.gray_scale = gray_scale
-        self.p_aug = p_aug
-        if p_aug > 0:
-            # Let's keep contrast augmentation aside for a moment
-            # self.augs = iaa.Sometimes(p_aug, iaa.OneOf([iaa.SigmoidContrast(gain=(5, 20), cutoff=(0.25, 0.75)),
-            #                                             iaa.LogContrast(gain=(0.6, 1.4)),
-            #                                             iaa.LinearContrast((0.4, 1.6))]))
-            # Let's tru adding uniform equalization
-            self.augs = iaa.Sometimes(p_aug, iaa.Lambda(
-                lambda x: SatTilesScaler().normalize_product(x, 'uniform_scaling', self.mean_scaling_strategy),
-                None))
 
         # Load the data (first, set the random split for loading the patches)
         self.split_seed = np.random.random_integers(42) if split_seed is None else split_seed

train_fe.py

@@ -141,8 +141,6 @@ def main(config: argparse.Namespace):
     norm = config.input_norm
     pos_const = config.pos_const
     output_fp_channels = config.output_fp_channels
-    separable_fp = config.separable_fp
-    depthwise_fp = config.depthwise_fp
 
     # --- Instantiate the DataGenerators --- #
     train_data_generator = DBLDataGenerator(batch_size=train_batch_size, patch_size=patch_size,
@@ -153,8 +151,7 @@ def main(config: argparse.Namespace):
                                             batch_num_tiles_peracq=batch_num_tiles_peracq,
                                             scaler_type=config.scaler_type,
                                             mean_scaling_strategy=config.mean_robust_scaling,
-                                            input_norm=norm,
-                                            p_aug=config.p_aug)
+                                            input_norm=norm)
     valid_data_generator = DBLDataGenerator(batch_size=train_batch_size, patch_size=patch_size,
                                             data_dir=config.val_dir, split_seed=config.split_seed,
                                             num_iteration=config.num_iteration,
@@ -163,8 +160,7 @@ def main(config: argparse.Namespace):
                                             batch_num_tiles_peracq=batch_num_tiles_peracq,
                                             scaler_type=config.scaler_type,
                                             mean_scaling_strategy=config.mean_robust_scaling,
-                                            input_norm=norm,
-                                            p_aug=config.p_aug)
+                                            input_norm=norm)
 
     # --- TRAINING --- #
     print('Starting training')
@@ -223,8 +219,7 @@ def main(config: argparse.Namespace):
     parser.add_argument('--scaler_type', type=str, help='Choose the scaler for the data. Choices are: '
                                                         '99th percentile robust scaler;'
                                                         '95th percentile robust scaler;'
-                                                        'Maximum scaling using each band statistics;'
-                                                        'input norm scaling with histogram equalization.',
+                                                        'Maximum scaling using each band statistics.',
                         default='99th_percentile', choices=['99th_percentile', '95th_percentile', 'sat_max',
                                                             'sat_tiles_scaler'])
     parser.add_argument('--mean_robust_scaling', action='store_true',