Remove rolling window and use fixed instead

Author: paren8esis

compute_class_weights.py

@@ -22,6 +22,8 @@
                         help='The prefix to use for the class weights file. Default none.')
     parser.add_argument('--ignore_zero', default=False, action='store_true', required=False,
                         help='Ignore the zero class.')
+    parser.add_argument('--fixed_window', action='store_true', default=False, required=False,
+                            help='Use a fixed window including months 4 (April) to 9 (September).')
     args = parser.parse_args()
 
     # Define paths
@@ -53,6 +55,7 @@
             linear_encoder=LINEAR_ENCODER,
             saved_medians=True,
             window_len=12,
+            fixed_window=args.fixed_window,
             requires_norm=True,
             return_masks=False,
             clouds=False,

pad_experiments.py

@@ -149,6 +149,8 @@ def main():
                              help='Starting learning rate. Default 1e-1')
     parser.add_argument('--window_len', type=int, default=6, required=False,
                              help='The length of the rolling window to be used. Default 6')
+    parser.add_argument('--fixed_window', action='store_true', default=False, required=False,
+                            help='Use a fixed window including months 4 (April) to 9 (September).')
 
     parser.add_argument('--bands', nargs='+', default=sorted(list(BANDS.keys())),
                              help='The image bands to use. Must be space separated')
@@ -388,13 +390,13 @@ def main():
             root_path_coco=root_path_coco,
             path_train=path_train,
             path_val=path_val,
-            # transforms=transforms,
             group_freq=args.group_freq,
             prefix=prefix,
             bands=args.bands,
             linear_encoder=LINEAR_ENCODER,
             saved_medians=args.saved_medians,
             window_len=args.window_len,
+            fixed_window=args.fixed_window,
             requires_norm=args.requires_norm,
             return_masks=args.return_masks,
             clouds=args.clouds,
@@ -459,6 +461,7 @@ def main():
             linear_encoder=LINEAR_ENCODER,
             saved_medians=args.saved_medians,
             window_len=args.window_len,
+            fixed_window=args.fixed_window,
             requires_norm=args.requires_norm,
             return_masks=args.return_masks,
             clouds=args.clouds,

utils/PAD_datamodule.py

@@ -34,6 +34,7 @@ def __init__(
             linear_encoder: dict = None,
             prefix: str = None,
             window_len: int = 12,
+            fixed_window: bool = False,
             requires_norm: bool = True,
             return_masks: bool = False,
             clouds: bool = True,
@@ -81,6 +82,9 @@ def __init__(
             over the data. E.g. if `window_len` = 6 and `group_freq` = '1M', then
             a 6-month rolling window will be applied and each batch will contain
             6 months of training data and the corresponding label.
+        fixed_window: boolean, default False
+            If True, then a fixed window including months 4 (April) to 9 (September) is used
+            instead of a rolling one.
         requires_norm: boolean, default True
             If True, then it normalizes the dataset to [0, 1] range.
         return_masks: boolean, default False
@@ -126,6 +130,7 @@ def __init__(
         self.linear_encoder = linear_encoder
         self.saved_medians = saved_medians
         self.window_len = window_len
+        self.fixed_window = fixed_window
         self.requires_norm = requires_norm
         self.return_masks = return_masks
         self.clouds = clouds
@@ -190,6 +195,7 @@ def setup(self, stage=None):
                                                 linear_encoder=self.linear_encoder,
                                                 saved_medians=self.saved_medians,
                                                 window_len=self.window_len,
+                                                fixed_window=self.fixed_window,
                                                 requires_norm=self.requires_norm,
                                                 return_masks=self.return_masks,
                                                 clouds=self.clouds,
@@ -211,6 +217,7 @@ def setup(self, stage=None):
                                                linear_encoder=self.linear_encoder,
                                                saved_medians=self.saved_medians,
                                                window_len=self.window_len,
+                                               fixed_window=self.fixed_window,
                                                requires_norm=self.requires_norm,
                                                return_masks=self.return_masks,
                                                clouds=self.clouds,
@@ -236,6 +243,7 @@ def setup(self, stage=None):
                                                linear_encoder=self.linear_encoder,
                                                saved_medians=self.saved_medians,
                                                window_len=self.window_len,
+                                               fixed_window=self.fixed_window,
                                                requires_norm=self.requires_norm,
                                                return_masks=self.return_masks,
                                                clouds=self.clouds,

utils/PAD_dataset.py

@@ -54,6 +54,7 @@ def __init__(
             linear_encoder: dict = None,
             prefix: str = None,
             window_len: int = 12,
+            fixed_window: bool = False,
             requires_norm: bool = True,
             return_masks: bool = False,
             clouds: bool = True,
@@ -94,6 +95,9 @@ def __init__(
             over the data. E.g. if `window_len` = 6 and `group_freq` = '1M', then
             a 6-month rolling window will be applied and each batch will contain
             6 months of training data and the corresponding label.
+        fixed_window: boolean, default False
+            If True, then a fixed window including months 4 (April) to 9 (September)
+            is used instead of a rolling one.
         requires_norm: boolean, default True
             If True, then it normalizes the dataset to [0, 1] range.
         return_masks: boolean, default False
@@ -199,6 +203,7 @@ def __init__(
 
         self.group_freq = group_freq
         self.window_len = window_len
+        self.fixed_window = fixed_window
         self.transforms = transforms
         self.linear_encoder = linear_encoder
 
@@ -335,14 +340,25 @@ def load_medians(self, path: Path, subpatch_id: int, start_bin: int) -> Tuple[np
         Just load and return
         """
         # `medians` is a 4d numpy array (window length, bands, img_size, img_size)
-        medians = np.empty((self.window_len, self.num_bands, self.output_size[0], self.output_size[1]),
-                           dtype=self.medians_dtype)
+        if self.fixed_window:
+            medians = np.empty((6, self.num_bands, self.output_size[0], self.output_size[1]),
+                                dtype=self.medians_dtype)
+        else:
+            medians = np.empty((self.window_len, self.num_bands, self.output_size[0], self.output_size[1]),
+                                dtype=self.medians_dtype)
 
         padded_id = f'{str(subpatch_id).rjust(len(str(self.num_subpatches)), "0")}'
 
         median_files = sorted(path.glob(f'sub{padded_id}_bin*'))
 
-        for i, bin_idx in enumerate(range(start_bin, start_bin + self.window_len)):
+        if self.fixed_window:
+            start_month = 3
+            end_month = 9
+        else:
+            start_month = start_bin
+            end_month = start_bin + self.window_len
+
+        for i, bin_idx in enumerate(range(start_month, end_month)):
             median = np.load(median_files[bin_idx]).astype(self.medians_dtype)
             medians[i] = median.copy()
 

visualize_predictions.py

@@ -71,6 +71,8 @@ def get_window(idx, window_len, image_size, coco_file):
                                   'Check: https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#timeseries-offset-aliases ')
     parser.add_argument('--window_len', type=int, default=6, required=False,
                              help='The length of the rolling window to be used. Default 6')
+    parser.add_argument('--fixed_window', action='store_true', default=False, required=False,
+                            help='Use a fixed window including months 4 (April) to 9 (September).')
 
     parser.add_argument('--bands', nargs='+', default=sorted(list(BANDS.keys())),
                              help='The image bands to use. Must be space separated')
@@ -181,6 +183,7 @@ def get_window(idx, window_len, image_size, coco_file):
         linear_encoder=LINEAR_ENCODER,
         saved_medians=args.saved_medians,
         window_len=args.window_len,
+        fixed_window=args.fixed_window,
         requires_norm=args.requires_norm,
         return_masks=False,
         clouds=False,