riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-07-31 02:02:21 +06:00
Code Issues Actions 23 Packages Projects Releases Wiki Activity

Fix loading the best model on the last stage of training (#11718)

Browse Source
This commit is contained in:
Volodymyr Byno 2021-05-13 23:11:12 +03:00 committed by GitHub
parent 252082001d
commit 218d552f30
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 15 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
src/transformers/trainer.py
View File

@ -1059,18 +1059,7 @@ class Trainer:
# We load the model state dict on the CPU to avoid an OOM error.
state_dict = torch.load(os.path.join(resume_from_checkpoint, WEIGHTS_NAME), map_location="cpu")
# If the model is on the GPU, it still works!
load_result = self.model.load_state_dict(state_dict, strict=False)
if len(load_result.missing_keys) != 0:
if load_result.missing_keys == self.model._keys_to_ignore_on_save:
self.model.tie_weights()
else:
logger.warn(
f"There were missing keys in the checkpoint model loaded: {load_result.missing_keys}."
)
if len(load_result.unexpected_keys) != 0:
logger.warn(
f"There were unexpected keys in the checkpoint model loaded: {load_result.unexpected_keys}."
)
self._load_state_dict_in_model(state_dict)
# If model was re-initialized, put it on the right device and update self.model_wrapped
if model_reloaded:
@ -1363,7 +1352,7 @@ class Trainer:
# We load the model state dict on the CPU to avoid an OOM error.
state_dict = torch.load(os.path.join(self.state.best_model_checkpoint, WEIGHTS_NAME), map_location="cpu")
# If the model is on the GPU, it still works!
self.model.load_state_dict(state_dict)
self._load_state_dict_in_model(state_dict)
if self.deepspeed:
self.deepspeed.load_checkpoint(
@ -1385,6 +1374,17 @@ class Trainer:
return TrainOutput(self.state.global_step, self._total_loss_scalar / self.state.global_step, metrics)
def _load_state_dict_in_model(self, state_dict):
load_result = self.model.load_state_dict(state_dict, strict=False)
if len(load_result.missing_keys) != 0:
if set(load_result.missing_keys) == set(self.model._keys_to_ignore_on_save):
self.model.tie_weights()
else:
logger.warn(f"There were missing keys in the checkpoint model loaded: {load_result.missing_keys}.")
if len(load_result.unexpected_keys) != 0:
logger.warn(f"There were unexpected keys in the checkpoint model loaded: {load_result.unexpected_keys}.")
def _maybe_log_save_evaluate(self, tr_loss, model, trial, epoch):
if self.control.should_log:
logs: Dict[str, float] = {}

3
tests/test_modeling_common.py
View File

@ -180,7 +180,8 @@ class ModelTesterMixin:
# Test we can load the state dict in the model, necessary for the checkpointing API in Trainer.
load_result = model.load_state_dict(state_dict_saved, strict=False)
self.assertTrue(
len(load_result.missing_keys) == 0 or load_result.missing_keys == model._keys_to_ignore_on_save
len(load_result.missing_keys) == 0
or set(load_result.missing_keys) == set(model._keys_to_ignore_on_save)
)
self.assertTrue(len(load_result.unexpected_keys) == 0)