Reorganize configuration files (#352)

* Reorganize configuration files

* Undo changes to tests/conf files

conf/bigearthnet.yaml

@@ -1,19 +1,21 @@
 trainer:
-  gpus: 1  # single GPU training
+  gpus: 1
   min_epochs: 10
   max_epochs: 40
   benchmark: True
-
 experiment:
   task: "bigearthnet"
   module:
     loss: "bce"
     classification_model: "resnet18"
     learning_rate: 1e-3
     learning_rate_schedule_patience: 6
+    weights: "random"
     in_channels: 14
-  datamodule:
     num_classes: 19
-    batch_size: 128
-    num_workers: 6
+  datamodule:
+    root_dir: "data/bigearthnet"
     bands: "all"
+    num_classes: ${experiment.module.num_classes}
+    batch_size: 128
+    num_workers: 4

conf/byol.yaml

@@ -3,7 +3,6 @@ trainer:
   min_epochs: 20
   max_epochs: 100
   benchmark: True
-
 experiment:
   task: "ssl"
   name: "test_byol"
@@ -12,12 +11,15 @@ experiment:
     encoder: "resnet18"
     input_channels: 4
     imagenet_pretraining: True
+    learning_rate: 1e-3
+    learning_rate_schedule_patience: 6
   datamodule:
-    batch_size: 64
-    num_workers: 6
+    root_dir: "data/chesapeake/cvpr"
     train_splits:
       - "de-train"
     val_splits:
       - "de-val"
     test_splits:
       - "de-test"
+    batch_size: 64
+    num_workers: 4

conf/chesapeake_cvpr.yaml

@@ -1,29 +1,35 @@
 trainer:
-  gpus: 1  # single GPU training
+  gpus: 1
   min_epochs: 20
   max_epochs: 100
   benchmark: True
-
 experiment:
   task: "chesapeake_cvpr"
   name: "chesapeake_cvpr_example"
   module:
-    loss: "ce"  # cross entropy loss
+    loss: "ce"
     segmentation_model: "unet"
     encoder_name: "resnet18"
-    encoder_weights: null  # use ImageNet weight initialization
+    encoder_weights: null
     encoder_output_stride: 16
-    learning_rate: 1e-2
+    learning_rate: 1e-3
     learning_rate_schedule_patience: 6
     in_channels: 4
     num_classes: 7
     num_filters: 256
+    ignore_zeros: False
+    imagenet_pretraining: True
   datamodule:
-    batch_size: 64
-    num_workers: 6
+    root_dir: "data/chesapeake/cvpr"
     train_splits:
       - "de-train"
     val_splits:
       - "de-val"
     test_splits:
       - "de-test"
+    patches_per_tile: 200
+    patch_size: 256
+    batch_size: 64
+    num_workers: 4
+    class_set: ${experiment.module.num_classes}
+    use_prior_labels: False

conf/cowc_counting.yaml

@@ -1,18 +1,16 @@
-program:  # These are the arguments that define how the train.py script works
-  seed: 1337
-  overwrite: True
-
+trainer:
+  gpus: 1
+  min_epochs: 15
 experiment:
   task: cowc_counting
   name: cowc_counting_test
   module:
     model: resnet18
     learning_rate: 1e-3
     learning_rate_schedule_patience: 2
+    pretrained: True
   datamodule:
+    root_dir: "data/cowc_counting"
+    seed: 0
     batch_size: 32
     num_workers: 4
-
-trainer:
-  min_epochs: 15
-  gpus: 1

conf/cyclone.yaml

@@ -1,18 +1,16 @@
-program:  # These are the arguments that define how the train.py script works
-  seed: 1337
-  overwrite: True
-
+trainer:
+  gpus: 1
+  min_epochs: 15
 experiment:
   task: "cyclone"
-  name: cyclone_test
+  name: "cyclone_test"
   module:
     model: "resnet18"
     learning_rate: 1e-3
     learning_rate_schedule_patience: 2
+    pretrained: True
   datamodule:
+    root_dir: "data/cyclone"
+    seed: 0
     batch_size: 32
     num_workers: 4
-
-trainer:
-  min_epochs: 15
-  gpus: 1

conf/etci2021.yaml

@@ -0,0 +1,16 @@
+experiment:
+  task: "etci2021"
+  module:
+    loss: "ce"
+    segmentation_model: "unet"
+    encoder_name: "resnet18"
+    encoder_weights: "imagenet"
+    learning_rate: 1e-3
+    learning_rate_schedule_patience: 6
+    in_channels: 7
+    num_classes: 2
+    ignore_zeros: True
+  datamodule:
+    root_dir: "data/etci2021"
+    batch_size: 32
+    num_workers: 4

conf/eurosat.yaml

@@ -0,0 +1,14 @@
+experiment:
+  task: "eurosat"
+  module:
+    loss: "ce"
+    classification_model: "resnet18"
+    learning_rate: 1e-3
+    learning_rate_schedule_patience: 6
+    weights: "random"
+    in_channels: 13
+    num_classes: 10
+  datamodule:
+    root_dir: "data/eurosat"
+    batch_size: 128
+    num_workers: 4

conf/landcoverai.yaml

@@ -1,22 +1,22 @@
 trainer:
-  gpus: 1  # single GPU training
+  gpus: 1
   min_epochs: 20
   max_epochs: 100
   benchmark: True
-
 experiment:
   task: "landcoverai"
   module:
     loss: "ce"
-    segmentation_model: "deeplabv3+"
+    segmentation_model: "unet"
     encoder_name: "resnet18"
-    encoder_weights: null
-    encoder_output_stride: 16
+    encoder_weights: "imagenet"
     learning_rate: 1e-3
     learning_rate_schedule_patience: 6
     in_channels: 3
     num_classes: 6
     num_filters: 256
+    ignore_zeros: False
   datamodule:
+    root_dir: "data/landcoverai"
     batch_size: 32
-    num_workers: 6
+    num_workers: 4

conf/naipchesapeake.yaml

@@ -1,13 +1,25 @@
-program:  # These are experiment level arguments
-  experiment_name: naip_chesapeake_test
+program:
+  experiment_name: "naip_chesapeake_test"
   overwrite: True
-  naip_data_dir: data/naip
-  chesapeake_data_dir: data/chesapeake
-
-trainer:  # These are all the arguments that will be passed to the pl.Trainer
+trainer:
   min_epochs: 15
-
-task:  # These are all the arguments that will be used to create an appropriate task
-  name: naipchesapeake
-  learning_rate: 1e-3
-  learning_rate_schedule_patience: 2
+experiment:
+  task: "naipchesapeake"
+  module:
+    loss: "ce"
+    segmentation_model: "deeplabv3+"
+    encoder_name: "resnet34"
+    encoder_weights: "imagenet"
+    encoder_output_stride: 16
+    learning_rate: 1e-3
+    learning_rate_schedule_patience: 2
+    in_channels: 4
+    num_classes: 13
+    num_filters: 64
+    ignore_zeros: False
+  datamodule:
+    naip_root_dir: "data/naip"
+    chesapeake_root_dir: "data/chesapeake/BAYWIDE"
+    batch_size: 32
+    num_workers: 4
+    patch_size: 32

conf/oscd.yaml

@@ -3,7 +3,6 @@ trainer:
   min_epochs: 20
   max_epochs: 500
   benchmark: True
-
 experiment:
   task: "oscd"
   module:
@@ -16,11 +15,15 @@ experiment:
     verbose: false
     in_channels: 26
     num_classes: 2
-    num_filters: 128
+    num_filters: 256
     ignore_zeros: True
   datamodule:
-    train_batch_size: 2
-    num_workers: 6
+    root_dir: "data/oscd"
+    batch_size: 32
+    num_workers: 4
     val_split_pct: 0.1
     bands: "all"
+    pad_size:
+    - 1028
+    - 1028
     num_patches_per_tile: 128

conf/resisc45.yaml

@@ -1,9 +1,8 @@
 trainer:
-  gpus: 1  # single GPU training
+  gpus: 1
   min_epochs: 10
   max_epochs: 40
   benchmark: True
-
 experiment:
   task: "resisc45"
   module:
@@ -12,6 +11,9 @@ experiment:
     learning_rate: 1e-3
     learning_rate_schedule_patience: 6
     weights: "random"
+    in_channels: 3
+    num_classes: 45
   datamodule:
+    root_dir: "data/resisc45"
     batch_size: 128
-    num_workers: 6
+    num_workers: 4

conf/sen12ms.yaml

@@ -1,11 +1,24 @@
-program:  # These are experiment level arguments
+program:
   experiment_name: sen12ms_test
   overwrite: True
-
-trainer:  # These are all the arguments that will be passed to the pl.Trainer
+trainer:
   min_epochs: 15
-
-task:  # These are all the arguments that will be used to create an appropriate task
-  name: sen12ms
-  learning_rate: 1e-3
-  learning_rate_schedule_patience: 2
+experiment:
+  task: "sen12ms"
+  module:
+    loss: "ce"
+    segmentation_model: "unet"
+    encoder_name: "resnet18"
+    encoder_weights: null
+    encoder_output_stride: 16
+    learning_rate: 1e-3
+    learning_rate_schedule_patience: 2
+    in_channels: 15
+    num_classes: 11
+    ignore_zeros: False
+  datamodule:
+    root_dir: "data/sen12ms"
+    band_set: "all"
+    batch_size: 32
+    num_workers: 4
+    seed: 0

conf/so2sat.yaml

@@ -1,18 +1,21 @@
 trainer:
-  gpus: 1  # single GPU training
+  gpus: 1
   min_epochs: 10
   max_epochs: 40
   benchmark: True
-
 experiment:
   task: "so2sat"
   module:
     loss: "ce"
     classification_model: "resnet18"
     learning_rate: 1e-3
     learning_rate_schedule_patience: 6
+    weights: "random"
     in_channels: 3
+    num_classes: 17
   datamodule:
+    root_dir: "data/so2sat"
     batch_size: 128
-    num_workers: 6
+    num_workers: 4
     bands: "rgb"
+    unsupervised_mode: False

conf/ucmerced.yaml

@@ -0,0 +1,14 @@
+experiment:
+  task: "ucmerced"
+  module:
+    loss: "ce"
+    classification_model: "resnet18"
+    weights: null
+    learning_rate: 1e-3
+    learning_rate_schedule_patience: 6
+    in_channels: 3
+    num_classes: 21
+  datamodule:
+    root_dir: "data/ucmerced"
+    batch_size: 128
+    num_workers: 4

tests/datamodules/test_chesapeake.py

@@ -14,9 +14,7 @@
 class TestChesapeakeCVPRDataModule:
     @pytest.fixture(scope="class")
     def datamodule(self) -> ChesapeakeCVPRDataModule:
-        conf = OmegaConf.load(
-            os.path.join("conf", "task_defaults", "chesapeake_cvpr_5.yaml")
-        )
+        conf = OmegaConf.load(os.path.join("tests", "conf", "chesapeake_cvpr_5.yaml"))
         kwargs = OmegaConf.to_object(conf.experiment.datamodule)
         kwargs = cast(Dict[str, Any], kwargs)
 

tests/trainers/test_byol.py

@@ -43,7 +43,7 @@ class TestBYOLTask:
         ],
     )
     def test_trainer(self, name: str, classname: Type[LightningDataModule]) -> None:
-        conf = OmegaConf.load(os.path.join("conf", "task_defaults", name + ".yaml"))
+        conf = OmegaConf.load(os.path.join("tests", "conf", name + ".yaml"))
         conf_dict = OmegaConf.to_object(conf.experiment)
         conf_dict = cast(Dict[Any, Dict[Any, Any]], conf_dict)