Merge pull request #434 from datamol-io/default_wandb

Disable wandb by default + complete integration tests

expts/hydra-configs/experiment/toymix_mpnn.yaml

@@ -0,0 +1,13 @@
+# @package _global_
+
+constants:
+  name: neurips2023_small_data_mpnn
+  entity: "multitask-gnn"
+  seed: 42
+  max_epochs: 100
+  data_dir: expts/data/neurips2023/small-dataset
+  raise_train_error: true
+
+trainer:
+  model_checkpoint:
+    dirpath: models_checkpoints/neurips2023-small-mpnn/

graphium/cli/train_finetune.py

@@ -44,13 +44,16 @@ def run_training_finetuning(cfg: DictConfig) -> None:
 
     st = timeit.default_timer()
 
+    # Disable wandb if the user is not logged in.
     wandb_cfg = cfg["constants"].get("wandb")
-    if wandb_cfg is not None:
-        wandb.init(
-            entity=wandb_cfg["entity"],
-            project=wandb_cfg["project"],
-            config=cfg,
+    if wandb.login() is False:
+        logger.info(
+            "Not logged in to wandb - disabling wandb logging.\n"
+            + "To enable wandb, run `wandb login` from the command line."
         )
+        wandb.init(mode="disabled")
+    elif wandb_cfg is not None:
+        wandb.init(config=cfg, **wandb_cfg)
 
     ## == Instantiate all required objects from their respective configs ==
     # Accelerator

tests/test_ipu_training.py

@@ -0,0 +1,50 @@
+import hydra
+from graphium.cli.train_finetune import run_training_finetuning
+import pytest
+
+FINETUNING_CONFIG_KEY = "finetuning"
+
+
+@pytest.mark.parametrize("acc_type, acc_prec", [("cpu", 32), ("ipu", 16)])
+@pytest.mark.ipu
+def test_cli(acc_type, acc_prec) -> None:
+    """
+    The main CLI endpoint for training and fine-tuning Graphium models.
+    """
+    with hydra.initialize(version_base=None, config_path="../expts/hydra-configs"):
+        # config is relative to a module
+        cfg = hydra.compose(
+            config_name="main",
+            overrides=[
+                f"accelerator={acc_type}",
+                "tasks=toymix",
+                "training=toymix",
+                # Reducing number of parameters in the toymix architecture
+                "architecture=toymix",
+                "architecture.pe_encoders.encoders.la_pos.hidden_dim=16",
+                "architecture.pe_encoders.encoders.la_pos.num_layers=1",
+                "architecture.pe_encoders.encoders.rw_pos.hidden_dim=16",
+                "architecture.pe_encoders.encoders.rw_pos.num_layers=1",
+                "architecture.pre_nn.hidden_dims=32",
+                "architecture.pre_nn.depth=1",
+                "architecture.pre_nn.out_dim=16",
+                "architecture.gnn.in_dim=16",
+                "architecture.gnn.out_dim=16",
+                "architecture.gnn.depth=2",
+                "architecture.task_heads.qm9.depth=1",
+                "architecture.task_heads.tox21.depth=1",
+                "architecture.task_heads.zinc.depth=1",
+                # Set the number of epochs
+                "constants.max_epochs=2",
+                "+datamodule.args.task_specific_args.qm9.sample_size=1000",
+                "+datamodule.args.task_specific_args.tox21.sample_size=1000",
+                "+datamodule.args.task_specific_args.zinc.sample_size=1000",
+                "trainer.trainer.check_val_every_n_epoch=1",
+                f"trainer.trainer.precision={acc_prec}",  # perhaps you can make this 32 for CPU and 16 for IPU
+            ],
+        )
+        if acc_type == "ipu":
+            cfg["accelerator"]["ipu_config"].append("useIpuModel(True)")
+            cfg["accelerator"]["ipu_inference_config"].append("useIpuModel(True)")
+
+        run_training_finetuning(cfg)