Adding Hydra based trainer target to fairseq in fbcode

Summary: Adding fairseq entrypoint section of e2e pipeline so FairseqConfig to hydra_main, runs smoothly

Reviewed By: jieru-hu

Differential Revision: D29714729

fbshipit-source-id: e3694e0037bb4c4f69208c1d6ec7df91d42fb588

fairseq/dataclass/initialize.py

@@ -47,7 +47,7 @@ def add_defaults(cfg: DictConfig) -> None:
                 field_cfg = DictConfig({"_name": field_cfg})
                 field_cfg.__dict__["_parent"] = field_cfg.__dict__["_parent"]
 
-            name = field_cfg.get("_name")
+            name = getattr(field_cfg, "_name", None)
 
             if k == "task":
                 dc = TASK_DATACLASS_REGISTRY.get(name)

fairseq_cli/hydra_train.py

@@ -11,6 +11,7 @@
 from fairseq_cli.train import main as pre_main
 from fairseq import distributed_utils, metrics
 from fairseq.dataclass.configs import FairseqConfig
+from fairseq.dataclass.utils import omegaconf_no_object_check
 from fairseq.utils import reset_logging
 
 import hydra
@@ -24,25 +25,32 @@
 
 @hydra.main(config_path=os.path.join("..", "fairseq", "config"), config_name="config")
 def hydra_main(cfg: FairseqConfig) -> float:
+    _hydra_main(cfg)
+
+
+def _hydra_main(cfg: FairseqConfig, **kwargs) -> float:
     add_defaults(cfg)
 
     if cfg.common.reset_logging:
         reset_logging()  # Hydra hijacks logging, fix that
     else:
-        with open_dict(cfg):
-            # make hydra logging work with ddp (see # see https://github.com/facebookresearch/hydra/issues/1126)
-            cfg.job_logging_cfg = OmegaConf.to_container(HydraConfig.get().job_logging, resolve=True)
-
-    cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=True, enum_to_str=True))
+        # check if directly called or called through hydra_main
+        if HydraConfig.initialized():
+            with open_dict(cfg):
+                # make hydra logging work with ddp (see # see https://github.com/facebookresearch/hydra/issues/1126)
+                cfg.job_logging_cfg = OmegaConf.to_container(HydraConfig.get().job_logging, resolve=True)
+
+    with omegaconf_no_object_check():
+        cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=True, enum_to_str=True))
     OmegaConf.set_struct(cfg, True)
 
     try:
         if cfg.common.profile:
             with torch.cuda.profiler.profile():
                 with torch.autograd.profiler.emit_nvtx():
-                    distributed_utils.call_main(cfg, pre_main)
+                    distributed_utils.call_main(cfg, pre_main, **kwargs)
         else:
-            distributed_utils.call_main(cfg, pre_main)
+            distributed_utils.call_main(cfg, pre_main, **kwargs)
     except BaseException as e:
         if not cfg.common.suppress_crashes:
             raise