add TTA support from geo-inference

config/inference/default_binary.yaml

@@ -9,7 +9,9 @@ inference:
   workers: 0
   prep_data_only: False
   heatmap_threshold: 0.3
-
+  flip: False
+  rotate: True
+
   # GPU parameters
   gpu: ${training.num_gpus}
   max_used_perc: ${training.max_used_perc}  # If GPU's usage exceeds this percentage, it will be ignored

config/inference/default_multiclass.yaml

@@ -9,6 +9,8 @@ inference:
   workers: 0
   prep_data_only: False
   heatmap_threshold: 0.3
+  flip: False
+  rotate: True
 
   # GPU parameters
   gpu: ${training.num_gpus}

docs/source/mode.rst

@@ -211,6 +211,11 @@ will be found in :ref:`configurationdefaultparam` under ``inference`` and this c
     enhancement is applied and equalized images save to disk.
 - ``heatmap_threshold`` (float)
     Prediction probability Threshold (fraction of 1) to use. Default is ``0.3``.
+- ``flip`` (bool)
+    If True, perform horizontal and vertical flips during inference.  
+- ``rotate`` (bool)
+    If True, perform 90 degree rotation at inference.
+
 - ``gpu`` (int)
     Number of gpus to use at inference. 
 - ``max_used_perc`` (int)

inference_segmentation.py

@@ -65,6 +65,8 @@ def main(params:Union[DictConfig, Dict]):
     input_stac_item = get_key_def('input_stac_item', params['inference'], expected_type=str, to_path=True,
                                   validate_path_exists=True)
     vectorize = get_key_def('ras2vec', params['inference'], expected_type=bool, default=False)
+    transform_flip = get_key_def('flip', params['inference'], expected_type=bool, default=False)
+    transform_rotate = get_key_def('rotate', params['inference'], expected_type=bool, default=False)
 
     if raw_data_csv and input_stac_item:
         raise ValueError(f"Input imagery should be either a csv of stac item. Got inputs from both \"raw_data_csv\" "
@@ -106,6 +108,9 @@ def main(params:Union[DictConfig, Dict]):
                                  device=device_str,
                                  gpu_id=gpu_index,
                                  prediction_threshold=prediction_threshold,
+                                 transformer=True,
+                                 transformer_flip=transform_flip,
+                                 transformer_rotate=transform_rotate,
                                  )
 
     # LOOP THROUGH LIST OF INPUT IMAGES