started to add Hydra, need to finish

configs/callbacks/hnet_checkpoint.yaml

@@ -0,0 +1,9 @@
+checkpoint:
+  _target_: lightning.pytorch.callbacks.ModelCheckpoint
+  save_last: True # additionally, always save model from last epoch
+  verbose: False
+  dirpath: "checkpoints/${now:%Y%m%d}/"
+  filename: "epoch_{epoch:01d}" #"hnet_model_DOA{max_len}_{'-'.join(map(str, sample_range_used))}_epoch_{epoch:01d}"
+  monitor: "validation_loss"
+  save_top_k: 1
+  mode: "min"

configs/callbacks/rich_model_summary.yaml

@@ -0,0 +1,3 @@
+model_summary:
+  _target_: lightning.pytorch.callbacks.RichModelSummary
+  max_depth: 3

configs/logging/tensorboard.yaml

@@ -0,0 +1,9 @@
+log_every_n_steps: 100
+
+logger:
+  _target_: lightning.pytorch.loggers.tensorboard.TensorBoardLogger
+  save_dir: "tb_logs/"
+  name: null
+  version: .
+  log_graph: False
+  default_hp_metric: False

configs/run.yaml

@@ -0,0 +1,24 @@
+### Top-level config file for run.py
+
+# Top-level variables available in all config files
+description: "${hydra:runtime.choices.lightning_datamodule}: ${lightning_datamodule.id}"
+
+# Hydra configuration
+version_base: "1.3"
+hydra:
+  job:
+    chdir: True # change working directory to the job directory
+  run:
+    dir: "outputs/run/${hydra:runtime.choices.lightning_datamodule}/${lightning_datamodule.id}/${now:%Y-%m-%d_%H-%M-%S}"
+
+# Composing configs
+defaults:
+  - lightning_datamodule: null # NEEDS TO BE OVERRIDDEN, will also determinate the metrics and the run directory
+  - lightning_module: null # NEEDS TO BE OVERRIDDEN
+  - trainer: ddp
+  - callbacks: # Dict of callbacks
+      # - bwe_checkpoint
+      - rich_model_summary
+  - logging: tensorboard
+  - metrics: ${lightning_datamodule}
+  - _self_ # priority is given to run.yaml for overrides

configs/trainer/ddp.yaml

@@ -0,0 +1,25 @@
+_target_: lightning.Trainer
+
+# Hardware
+accelerator: gpu
+num_nodes: 1
+devices: '0,'
+strategy: 'ddp_find_unused_parameters_true'
+
+# Epochs and batch sizes
+max_epochs: -1
+#accumulate_grad_batches: 2
+#limit_train_batches: 20
+#limit_val_batches: 10
+#limit_test_batches: 0.05
+#check_val_every_n_epoch: 15
+
+# Logging
+log_every_n_steps: ${logging.log_every_n_steps}
+#track_grad_norm: 1
+
+# Training
+#overfit_batches: 1
+#precision: 64
+#weights_summary: "full"
+#precision=16

run.py

@@ -14,19 +14,19 @@
 from hungarian_net.lightning_modules.hnet_gru_lightning import HNetGRULightning
 
 
-# @hydra.main(
-#     config_path="configs",
-#     config_name="run.yaml",
-#     version_base="1.3",
-# )
-def main(
-    batch_size=256,
+ @hydra.main(
+     config_path="configs",
+     config_name="run.yaml",
+     version_base="1.3",
+ )
+def main(cfg: DictConfig):
+    """ batch_size=256,
     nb_epochs=1000,
     max_len=2,
     sample_range_used=[3000, 5000, 15000],
     filename_train="data/reference/hung_data_train",
-    filename_test="data/reference/hung_data_test",
-):
+    filename_test="data/reference/hung_data_test", """
+
 
 
     # TODO: Réécriture/factorisation du code sur le modèle de VibraVox de Julien HAURET
@@ -55,29 +55,34 @@ def main(
         max_len=max_len,
     )
 
+    # Instantiate LightningModule
+    lightning_module: LightningModule = hydra.utils.instantiate(
+        cfg.lightning_module,
+        metrics=metrics,
+    )
+
     # Get current date
     current_date = datetime.datetime.now().strftime("%Y%m%d")
 
-    os.makedirs(f"models/{current_date}", exist_ok=True)
+    os.makedirs(f"checkpoints/{current_date}", exist_ok=True)
 
     # Human-readable filename
-    dirpath = f"models/{current_date}/"
+    dirpath = f"checkpoints/{current_date}/"
     out_filename = f"hnet_model_DOA{max_len}_{'-'.join(map(str, sample_range_used))}"
 
-    logger = TensorBoardLogger("tb_logs", name="hungarian_net")
-    checkpoint_callback = ModelCheckpoint(
+    """ checkpoint_callback = ModelCheckpoint(
         dirpath=dirpath,
         filename=out_filename,
         monitor="validation_loss",
         save_top_k=1,
         mode="min",
-    )
+    ) """
 
-    trainer = L.Trainer(
-        max_epochs=nb_epochs,
-        logger=logger,
-        callbacks=[checkpoint_callback],
-        # gpus=1 if use_cuda else 0
+    # Instantiate Trainer
+    callbacks: List[Callback] = list(hydra.utils.instantiate(cfg.callbacks).values())
+    logger: Logger = hydra.utils.instantiate(cfg.logging.logger)
+    trainer: Trainer = hydra.utils.instantiate(
+        cfg.trainer, callbacks=callbacks, logger=logger, _convert_="partial"
     )
 
     trainer.fit(lightning_module, datamodule=lightning_datamodule)