Use hydra for instantiating models and optimizers (#295)

* fixes #293 #246
add tests for optimizer instantiation in test_optimizers.py
adapt our unet models (models/unet.py) to expect same parameter names as smp models

* minor typo fixes

* name model yamls as close as possible to upcoming naming convention

* fix model name

config/gdl_config_template.yaml

@@ -1,5 +1,5 @@
 defaults:
-  - model: unet
+  - model: gdl_unet
   - training: default_training
   - loss: binary/softbce
   - optimizer: adamw

config/model/gdl_deeplabv3-dualhead.yaml

@@ -0,0 +1,5 @@
+# @package _global_
+model:
+  _target_: models.deeplabv3_dualhead.DeepLabV3_dualhead
+  conc_point: conv1  # Choose concatenation point in resnet encoder: 'conv1', 'maxpool', 'layer2', 'layer3', 'layer4'
+  encoder_weights: imagenet

config/model/gdl_unet.yaml

@@ -0,0 +1,5 @@
+# @package _global_
+model:
+  _target_: models.unet.UNet
+  dropout: False    # (bool) Use dropout or not
+  prob: False    # (float) Set dropout probability, e.g. 0.5

config/model/gdl_unetsmall.yaml

@@ -0,0 +1,6 @@
+# @package _global_
+model:
+  _target_: models.unet.UNetSmall
+  dropout: False    # (bool) Use dropout or not
+  prob: False    # (float) Set dropout probability, e.g. 0.5
+

config/model/smp_deeplabv3.yaml

@@ -0,0 +1,5 @@
+# @package _global_
+model:
+  _target_: segmentation_models_pytorch.DeepLabV3
+  encoder_name: resnet101
+  encoder_weights: imagenet

config/model/smp_deeplabv3plus.yaml

@@ -0,0 +1,5 @@
+# @package _global_
+model:
+  _target_: segmentation_models_pytorch.DeepLabV3Plus
+  encoder_name: resnext50_32x4d
+  encoder_weights: imagenet

config/model/smp_manet.yaml

@@ -0,0 +1,5 @@
+# @package _global_
+model:
+  _target_: segmentation_models_pytorch.MAnet
+  encoder_name: resnext101_32x8d
+  encoder_weights: imagenet

config/model/smp_pan.yaml

@@ -0,0 +1,5 @@
+# @package _global_
+model:
+  _target_: segmentation_models_pytorch.PAN
+  encoder_name: se_resnext101_32x4d
+  encoder_weights: imagenet

config/model/smp_unet-resnext101.yaml

@@ -0,0 +1,7 @@
+# @package _global_
+model:
+  _target_: segmentation_models_pytorch.Unet
+  encoder_name: resnext101_32x8d
+  encoder_depth: 4
+  decoder_channels: [ 256, 128, 64, 32 ]
+  encoder_weights: imagenet

config/model/smp_unet-spacenet-baseline.yaml

@@ -0,0 +1,5 @@
+# @package _global_
+model:
+  _target_: segmentation_models_pytorch.Unet
+  encoder_name: vgg11
+  encoder_weights: imagenet

config/model/smp_unet-spacenet-efficientnetb5.yaml

@@ -0,0 +1,6 @@
+# @package _global_
+model:
+  _target_: segmentation_models_pytorch.Unet
+  encoder_name: efficientnet-b5
+  encoder_weights: imagenet
+

config/model/smp_unet-spacenet-senet154.yaml

@@ -0,0 +1,5 @@
+# @package _global_
+model:
+  _target_: segmentation_models_pytorch.Unet
+  encoder_name: senet154
+  encoder_weights: imagenet

config/model/smp_unet.yaml

@@ -0,0 +1,6 @@
+# @package _global_
+model:
+  _target_: segmentation_models_pytorch.Unet
+  encoder_name: resnext50_32x4d
+  encoder_depth: 5
+  encoder_weights: imagenet

config/model/smp_unetplusplus.yaml

@@ -0,0 +1,8 @@
+# @package _global_
+model:
+  _target_: segmentation_models_pytorch.UnetPlusPlus
+  encoder_name: se_resnext50_32x4d
+  encoder_depth: 4
+  decoder_channels: [256, 128, 64, 32]
+  decoder_attention_type: scse
+  encoder_weights: imagenet

config/optimizer/adabound.yaml

@@ -0,0 +1,7 @@
+# @package _global_
+optimizer:
+  _target_: utils.adabound.AdaBound
+  lr: ${training.lr}
+  gamma: 1e-3
+  eps: 1e-8
+  weight_decay: 4e-5

config/optimizer/adaboundw.yaml

@@ -0,0 +1,7 @@
+# @package _global_
+optimizer:
+  _target_: utils.adabound.AdaBoundW
+  lr: ${training.lr}
+  gamma: 1e-3
+  eps: 1e-8
+  weight_decay: 4e-5

config/optimizer/adam.yaml

@@ -1,6 +1,4 @@
 # @package _global_
 optimizer:
-  optimizer_name: 'adam'
-  class_name: torch.optim.Adam
-  params:
-    lr: ${training.lr}
+  _target_: torch.optim.Adam
+  lr: ${training.lr}

config/optimizer/adamw.yaml

@@ -1,7 +1,5 @@
 # @package _global_
 optimizer:
-  optimizer_name: 'adamw'
-  class_name: torch.optim.AdamW
-  params:
-    lr: ${training.lr}
-    weight_decay: 0.001
+  _target_: torch.optim.AdamW
+  lr: ${training.lr}
+  weight_decay: 4e-5

config/optimizer/sgd.yaml

@@ -1,8 +1,6 @@
 # @package _global_
 optimizer:
-  optimizer_name: 'sgd'
-  class_name: torch.optim.SGD
-  params:
-    lr: ${training.lr}
-    momentum: 0.9
-    weight_decay: 0.0005
+  _target_: torch.optim.SGD
+  lr: ${training.lr}
+  momentum: 0.9
+  weight_decay: 4e-5

inference_segmentation.py

@@ -292,7 +292,6 @@ def main(params: dict) -> None:
     :param params: (dict) Parameters inputted during execution.
     """
     # PARAMETERS
-    model_name = get_key_def('model_name', params['model'], expected_type=str).lower()
     num_classes = len(get_key_def('classes_dict', params['dataset']).keys())
     num_classes = num_classes + 1 if num_classes > 1 else num_classes  # multiclass account for background
     modalities = read_modalities(get_key_def('modalities', params['dataset'], expected_type=str))
@@ -348,10 +347,9 @@ def main(params: dict) -> None:
 
     # CONFIGURE MODEL
     model = define_model(
-        model_name=model_name,
-        num_bands=num_bands,
-        num_classes=num_classes,
-        conc_point=conc_point,
+        net_params=params.model,
+        in_channels=num_bands,
+        out_classes=num_classes,
         main_device=device,
         devices=[list(gpu_devices_dict.keys())],
         state_dict_path=state_dict,