fixed jsonlogger and version bump

tests/datamodules/helpers.py

@@ -5,10 +5,10 @@
 from pytorch_lightning.utilities.distributed import distributed_available
 from transformers import BertTokenizer
 
-from tests.helpers import DummyDataModule, DummyTransformerModelWithOptim, standard_args
+from tests.helpers import DummyDataModule, DummyTransformerModel, standard_args
 
 
-class DummyTransformersModelIDCheck(DummyTransformerModelWithOptim):
+class DummyTransformersModelIDCheck(DummyTransformerModel):
 
     def __init__(self, hyperparameters, train_len: int, valid_len: int, test_len: int):
         super().__init__(hyperparameters)

tests/helpers.py

@@ -104,7 +104,6 @@ def __init__(self, hyperparameters):
         self.model = BertForSequenceClassification(config)
 
     def training_step(self, batch, batch_idx):
-        """ Training step on BertForSequenceClassification. """
         batch['labels'] = batch['labels'].to(dtype=torch.long)
         kwargs = {k: batch[k] for k in ["input_ids", "attention_mask", "token_type_ids", "labels"]}
         results = self(**kwargs)
@@ -122,9 +121,6 @@ def test_step(self, batch, batch_idx, dataset_idx):
         results = self(**kwargs)
         return {'loss': results.loss, 'ids': batch['ids']}
 
-
-class DummyTransformerModelWithOptim(DummyTransformerModel):
-
     def configure_optimizers(self):
         self.computed_steps = self.num_training_steps()
         return AdamW(self.model.parameters())

tests/models/helpers.py

@@ -3,7 +3,7 @@
 from pytorch_lightning import Trainer
 from transformers import BertTokenizer
 
-from tests.helpers import DummyDataModule, DummyTransformerModelWithOptim, standard_args
+from tests.helpers import DummyDataModule, DummyTransformerModel, standard_args
 
 
 def do_test_fix_max_steps(max_epochs, accumulate_grad_batches, batch_size, **kwargs):
@@ -26,7 +26,7 @@ def do_test_fix_max_steps(max_epochs, accumulate_grad_batches, batch_size, **kwa
 
     tokenizer = BertTokenizer('tests/data/vocab.txt')
     # not checking ids because sometimes the sampler will duplicate elements to fill all gpus
-    model = DummyTransformerModelWithOptim(hyperparameters)
+    model = DummyTransformerModel(hyperparameters)
 
     # Datasets
     datamodule = DummyDataModule(hyperparameters, length_train=40, length_test=40, length_valid=40, tokenizer=tokenizer)

tests/test_callbacks.py

@@ -8,7 +8,7 @@
 import pytorch_lightning as pl
 from transformers import BertTokenizer
 
-from tests.helpers import DummyDataModule, DummyTransformerModelWithOptim, standard_args
+from tests.helpers import DummyDataModule, DummyTransformerModel, standard_args
 from transformers_lightning.callbacks.transformers_model_checkpoint import TransformersModelCheckpointCallback
 
 
@@ -70,7 +70,7 @@ def test_model_checkpointing_callback(
     )
 
     # instantiate PL model
-    model = DummyTransformerModelWithOptim(hyperparameters)
+    model = DummyTransformerModel(hyperparameters)
 
     # Datasets
     datamodule = DummyDataModule(hyperparameters, length_train=96, length_valid=96, length_test=96, tokenizer=tokenizer)

tests/test_loggers.py

@@ -7,14 +7,32 @@
 import pytorch_lightning as pl
 from transformers import BertTokenizer
 
-from tests.helpers import DummyDataModule, DummyTransformerModelWithOptim, standard_args
+from tests.helpers import DummyDataModule, DummyTransformerModel, standard_args
 from transformers_lightning.loggers.jsonboard_logger import JsonBoardLogger
 
 
 def random_name():
     return ''.join(random.choices(string.ascii_uppercase + string.digits, k=20))
 
 
+class DummyTransformerModelWithLogging(DummyTransformerModel):
+
+    def training_step(self, batch, batch_idx):
+        res = super().training_step(batch, batch_idx)
+        self.log('training/loss', res['loss'], on_step=True)
+        return res
+
+    def validation_step(self, batch, batch_idx):
+        res = super().validation_step(batch, batch_idx)
+        self.log('validation/loss', res['loss'])
+        return res
+
+    def test_step(self, batch, batch_idx):
+        res = super().test_step(batch, batch_idx)
+        self.log('test/loss', res['loss'])
+        return res
+
+
 def test_jsonboard_logger():
 
     hyperparameters = Namespace(
@@ -30,6 +48,7 @@ def test_jsonboard_logger():
         jsonboard_dir='jsonboard',
         name=random_name(),
         val_check_interval=0.25,
+        log_every_n_steps=1,
         **standard_args,
     )
 
@@ -39,13 +58,12 @@ def test_jsonboard_logger():
     # instantiate PL trainer
     trainer = pl.Trainer.from_argparse_args(
         hyperparameters,
-        profiler='simple',
-        logger=[logger],
+        logger=logger,
         enable_checkpointing=False
     )
 
     # instantiate PL model
-    model = DummyTransformerModelWithOptim(hyperparameters)
+    model = DummyTransformerModelWithLogging(hyperparameters)
 
     # Datasets
     datamodule = DummyDataModule(hyperparameters, length_train=96, length_valid=96, length_test=96, tokenizer=tokenizer)

transformers_lightning/info.py

@@ -1,4 +1,4 @@
-__version__ = '0.7.10'
+__version__ = '0.7.11'
 __author__ = 'Luca Di Liello and Matteo Gabburo'
 __author_email__ = 'luca.diliello@unitn.it'
 __license__ = 'GNU GENERAL PUBLIC LICENSE v2'

transformers_lightning/loggers/jsonboard_logger.py

@@ -44,12 +44,9 @@ class JsonBoardLogger(LightningLoggerBase):
     def __init__(self, hyperparameters: Namespace):
         super().__init__()
         self.hyperparameters = hyperparameters
-        self._name = hyperparameters.name
         self._version = None
         self._fs = get_filesystem(hyperparameters.jsonboard_dir)
         self._experiment = None
-        self.hparams = {}
-        self.meta = {}
 
     def reset(self):
         r""" Reset experiment. """
@@ -119,12 +116,8 @@ def log_hyperparams(self, params: Union[Dict[str, Any], Namespace] = None) -> No
         """
 
         assert rank_zero_only.rank == 0, "tried to init log dirs in non global_rank=0"
-
         params = _convert_params(params)
 
-        # store params to output
-        self.hparams.update(params)
-
         # format params into the suitable for tensorboard
         params = _flatten_dict(params)
         params = self._sanitize_params(params)
@@ -150,12 +143,8 @@ def log_metadata(self, metadata: Union[Dict[str, Any], Namespace] = None) -> Non
         """
 
         assert rank_zero_only.rank == 0, "tried to init log dirs in non global_rank=0"
-
         metadata = _convert_params(metadata)
 
-        # store params to output
-        self.meta.update(metadata)
-
         # format params into the suitable for tensorboard
         metadata = _flatten_dict(metadata)
         metadata = self._sanitize_params(metadata)
@@ -176,7 +165,6 @@ def log_metadata(self, metadata: Union[Dict[str, Any], Namespace] = None) -> Non
     def log_metrics(self, metrics: Dict[str, float], step: Optional[int] = None) -> None:
         r""" Just write the metrics to disk. """
         assert rank_zero_only.rank == 0, "experiment tried to log from global_rank != 0"
-        metrics = _add_prefix(metrics)
         if metrics:
             try:
                 self._sanitize_and_write_metrics(metrics, step + 1)
@@ -200,7 +188,7 @@ def name(self) -> str:
         Returns:
             The name of the experiment.
         """
-        return self._name
+        return self.hyperparameters.name
 
     @property
     def version(self) -> int: