minor changes

pyproject.toml

@@ -13,6 +13,7 @@ classifiers = [
     "Programming Language :: Python :: 3.8",
     "Programming Language :: Python :: 3.9",
     "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
     "Environment :: Console",
 ]
 keywords = ["nlp", "transformers", "language-model"]

zeldarose/datasets/mbart.py

@@ -24,9 +24,7 @@
 from loguru import logger
 from torch.nn.utils.rnn import pad_sequence
 
-# Nouvo plan : un lecteur de jsonlines custom qui prédécoupe en source/target avec attribut src et
-# tgt, on charge ça dans dataset pour le système de cache, puis dans le dataloader on sample et
-# quelque part dans le trainmodule on ajoute le bruit utiliser
+# utiliser
 # <https://huggingface.co/docs/datasets/loading#python-generator> comme ça on peut streamer l'entrée
 
 
@@ -44,8 +42,9 @@ def extract_from_jsonline(
     source_langs: Collection[str],
     target_langs: Collection[str],
 ) -> Generator[DataRow, None, None]:
-    # We deal with both top-level tranlatifrdgggggggggggggggggggggggggggggggggggggggggwons and 🤗's
-    # conventional format for this task
+    # We deal with both top-level (`{fr: "J'ai chanté", "br": "Me m'eus kanet."}`) and 🤗's
+    # conventional format (`{"translation": {fr: "J'ai chanté", "br": "Me m'eus kanet."}}`) for this
+    # task.
     example = cast(Mapping[str, str], example.get("translation", example))
     for dns_lang in denoise_langs:
         if not (dns_str := example.get(dns_lang)):
@@ -314,8 +313,8 @@ def __init__(
         else:
             self.val_dataset_path = None
 
-        self.train_dataset = None
-        self.val_dataset = None
+        self.train_dataset: Optional[datasets.Dataset] = None
+        self.val_dataset: Optional[datasets.Dataset] = None
 
     def prepare_data(self):
         # NOTE (2021-08-12): This should'nt be needed since this method should only be called on

zeldarose/datasets/transform.py

@@ -180,8 +180,8 @@ def __init__(
         else:
             self.val_dataset_path = None
 
-        self.train_dataset = None
-        self.val_dataset = None
+        self.train_dataset: Optional[datasets.Dataset] = None
+        self.val_dataset: Optional[datasets.Dataset] = None
 
     def prepare_data(self):
         # NOTE(2021-08-12): This should'nt be needed since this method should only be called on rank

zeldarose/train_transformer.py

@@ -418,7 +418,7 @@ def main(
             f" loader batch size({device_batch_size} samples per device × {total_devices} devices)"
             " try using fewer devices"
         )
-    elif tuning_config.batch_size % (device_batch_size * total_devices):
+    elif tuning_config.batch_size % (device_batch_size * total_devices) != 0:
         remainder = tuning_config.batch_size % device_batch_size * total_devices
         logger.warning(
             f"Batch size ({tuning_config.batch_size}) is not a multiple of loader batch size"