update training structure

hrdae/__main__.py

@@ -1,15 +1,22 @@
+from datetime import datetime
+
 import hydra
 from omegaconf import DictConfig, OmegaConf
 
 from .dataloaders import create_dataloader
 from .models import create_model
-from .option import Option, TrainExpOption, process_options, save_options
+from .option import Option, TrainExpOption, save_options
 
 
 @hydra.main(version_base=None, config_path="conf", config_name="config")
 def main(cfg: DictConfig):
     opt: Option = OmegaConf.to_object(cfg)  # type: ignore
-    opt = process_options(opt)
+    opt.experiment.result_dir = (
+        opt.experiment.result_dir
+        / opt.experiment.dataloader.__class__.__name__
+        / opt.experiment.model.__class__.__name__
+        / datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
+    )
     save_options(opt, opt.experiment.result_dir)
     if (
         isinstance(opt.experiment, TrainExpOption)

hrdae/conf/config.yaml

@@ -1,4 +1,3 @@
-task_name: mmnist-hrdae2d-guided1d
 defaults:
   - config_schema
   - experiment: train

hrdae/conf/experiment/dataloader/dataset/moving_mnist.yaml

@@ -1,4 +1,4 @@
 root: data
-slice_index: [32]
+slice_index: [8, 16, 24, 32, 40, 48, 56]
 defaults:
   - /config/experiment/dataloader/dataset/moving_mnist@_here_

hrdae/conf/experiment/model/network/hrdae2d.yaml

@@ -17,4 +17,4 @@ activation: sigmoid
 debug_show_dim: false
 defaults:
   - /config/experiment/model/network/hrdae2d@_here_
-  - motion_encoder: guided1d
+  - motion_encoder: rnn1d

hrdae/conf/experiment/model/network/motion_encoder/rnn1d.yaml

@@ -1,4 +1,4 @@
-in_channels: 1
+in_channels: 7
 conv_params:
   - kernel_size: [3]
     stride: [2]
@@ -14,4 +14,4 @@ conv_params:
     output_padding: [1]
 defaults:
   - /config/experiment/model/network/motion_encoder/rnn1d@_here_
-  - rnn: tcn1d
+  - rnn: conv_lstm1d

hrdae/conf/experiment/model/scheduler/onecyclelr.yaml

@@ -1,3 +1,3 @@
-max_lr: 0.001
+max_lr: 0.1
 defaults:
   - /config/experiment/model/scheduler/onecyclelr@_here_

hrdae/conf/experiment/model/vr.yaml

@@ -1,12 +1,10 @@
 loss_coef:
-  wmse: 0.5
-  pjc2d: 0.5
+  wmse: 1
 phase: "0"
-pred_diff: false
+pred_diff: true
 defaults:
   - /config/experiment/model/vr@_here_
   - [email protected]: wmse
-  - [email protected]: pjc2d
   - network: hrdae2d
   - optimizer: adam
   - scheduler: onecyclelr

hrdae/conf/experiment/test.yaml

@@ -1,4 +1,3 @@
-run_name: 2024-06-16-0158
 result_dir: results
 defaults:
   - /config/experiment/test@_here_

hrdae/conf/experiment/train.yaml

@@ -1,6 +1,5 @@
-run_name: 2024-06-16-0158
 result_dir: results
-n_epoch: 30
+n_epoch: 100
 debug: false
 defaults:
   - /config/experiment/train@_here_

hrdae/models/basic_model.py

@@ -52,13 +52,15 @@ def train(
         n_epoch: int,
         result_dir: Path,
         debug: bool,
-    ) -> None:
+    ) -> float:
         max_iter = None
         if debug:
             max_iter = 5
 
         self.network.to(self.device)
 
+        least_val_loss = float("inf")
+
         for epoch in range(n_epoch):
             self.network.train()
             running_loss = 0.0
@@ -105,6 +107,9 @@ def train(
                 avg_val_loss = total_val_loss / len(val_loader)
                 print(f"Epoch: {epoch+1}, Val Loss: {avg_val_loss:.6f}")
 
+                if avg_val_loss < least_val_loss:
+                    least_val_loss = avg_val_loss
+
             if epoch % 10 == 0:
                 data = next(iter(val_loader))
 
@@ -120,6 +125,8 @@ def train(
                     result_dir / "logs" / "reconstructed",
                 )
 
+        return least_val_loss
+
 
 def create_basic_model(
     opt: BasicModelOption,

hrdae/models/typing.py

@@ -13,5 +13,5 @@ def train(
         n_epoch: int,
         result_dir: Path,
         debug: bool,
-    ) -> None:
+    ) -> float:
         pass

hrdae/models/vr_model.py

@@ -60,13 +60,15 @@ def train(
         n_epoch: int,
         result_dir: Path,
         debug: bool,
-    ) -> None:
+    ) -> float:
         max_iter = None
         if debug:
             max_iter = 5
 
         self.network.to(self.device)
 
+        least_val_loss = float("inf")
+
         for epoch in range(n_epoch):
             self.network.train()
             running_loss = 0.0
@@ -126,6 +128,9 @@ def train(
                 avg_val_loss = total_val_loss / len(val_loader)
                 print(f"Epoch: {epoch+1}, Val Loss: {avg_val_loss:.6f}")
 
+                if avg_val_loss < least_val_loss:
+                    least_val_loss = avg_val_loss
+
             if epoch % 10 == 0:
                 data = next(iter(val_loader))
 
@@ -147,6 +152,8 @@ def train(
                     result_dir / "logs" / "reconstructed",
                 )
 
+        return least_val_loss
+
 
 def create_vr_model(
     opt: VRModelOption,

hrdae/option.py

@@ -9,7 +9,6 @@
 
 @dataclass
 class ExpOption:
-    run_name: str = MISSING
     result_dir: Path = field(default_factory=Path)
     debug: bool = False
 
@@ -29,18 +28,9 @@ class TestExpOption(ExpOption):
 
 @dataclass
 class Option:
-    task_name: str = MISSING
-
     experiment: ExpOption = MISSING
 
 
-def process_options(opt: Option) -> Option:
-    opt.experiment.result_dir = (
-        opt.experiment.result_dir / opt.task_name / opt.experiment.run_name
-    )
-    return opt
-
-
 def save_options(opt: Option, save_dir: Path) -> None:
     config = OmegaConf.structured(opt)
     save_dir.mkdir(parents=True, exist_ok=True)