DOC : Example for fine_tuning BERT models

Author: Thibeb

docs/examples/fine_tuning.md

@@ -0,0 +1,118 @@
+---
+jupytext:
+  text_representation:
+    extension: .md
+    format_name: myst
+    format_version: 0.13
+    jupytext_version: 1.16.1
+---
+
+```{code-cell}
+import torch
+from medkit.training import TrainerConfig, Trainer
+from medkit.text.metrics.ner import SeqEvalMetricsComputer
+from medkit.text.ner.hf_entity_matcher import HFEntityMatcher
+from medkit.io.medkit_json import load_text_documents
+import os
+import shutil
+
+train, val, test = [], [], []
+
+#Merge each corpus split into one to get a massive amount of data to fine-tune on
+for c in ['quaero','e3c', 'casm2']:
+    train += list(load_text_documents(f"/content/drive/MyDrive/datasets/{c}/train.jsonl"))
+    val += list(load_text_documents(f"/content/drive/MyDrive/datasets/{c}/val.jsonl"))
+    test += list(load_text_documents(f"/content/drive/MyDrive/datasets/{c}/test.jsonl"))
+```
+
+```{code-cell}
+CHECKPOINT_DIR = "checkpoints_drbert/"
+
+DEVICE = 0 if torch.cuda.is_available() else -1
+
+trainable_matcher = HFEntityMatcher.make_trainable(
+    model_name_or_path="Dr-BERT/DrBERT-4GB-CP-PubMedBERT",
+    labels=["ANAT","CHEM","DEVI","DISO","GEOG","LIVB","OBJC","PHEN","PHYS","PROC"],
+    tagging_scheme="iob2",
+    tokenizer_max_length=512,
+    device=DEVICE,
+    tag_subtokens=True
+)
+
+trainer_config = TrainerConfig(
+    output_dir=CHECKPOINT_DIR,
+    learning_rate=5e-5,
+    nb_training_epochs=10,
+    batch_size=16,
+)
+
+ner_metrics_computer = SeqEvalMetricsComputer(
+    id_to_label=trainable_matcher.id_to_label,
+    tagging_scheme='iob2',
+    return_metrics_by_label=False,
+    average='weighted'
+)
+
+trainer = Trainer(
+    config=trainer_config,
+    component=trainable_matcher,
+    train_data=train,
+    eval_data=val,
+    metrics_computer=ner_metrics_computer,
+)
+
+#Train model
+history = trainer.train()
+
+#Get best checkpoint, rename it and save it on my local drive
+checkpoint_paths = sorted(glob(CHECKPOINT_DIR + "/checkpoint_*"))
+checkpoint_path = checkpoint_paths[0]
+os.rename(checkpoint_path, f'{CHECKPOINT_DIR}/DrBert-Generalized')
+shutil.move(f'{CHECKPOINT_DIR}/DrBert-Generalized','/content/drive/MyDrive/models')
+```
+
+```{code-cell}
+CHECKPOINT_DIR = "checkpoints_cam/"
+
+DEVICE = 0 if torch.cuda.is_available() else -1
+
+trainable_matcher = HFEntityMatcher.make_trainable(
+    model_name_or_path="almanach/camembert-bio-base",
+    labels=["ANAT","CHEM","DEVI","DISO","GEOG","LIVB","OBJC","PHEN","PHYS","PROC"],
+    tagging_scheme="iob2",
+    tokenizer_max_length=512,
+    device=DEVICE,
+    tag_subtokens=True
+)
+
+trainer_config = TrainerConfig(
+    output_dir=CHECKPOINT_DIR,
+    learning_rate=5e-5,
+    nb_training_epochs=10,
+    batch_size=16
+)
+
+ner_metrics_computer = SeqEvalMetricsComputer(
+    id_to_label=trainable_matcher.id_to_label,
+    tagging_scheme='iob2',
+    return_metrics_by_label=False,
+    average='weighted'
+)
+
+trainer = Trainer(
+    config=trainer_config,
+    component=trainable_matcher,
+    train_data=train,
+    eval_data=val,
+    metrics_computer=ner_metrics_computer,
+)
+
+#Train model
+history = trainer.train()
+
+#Get best checkpoint, rename it and save it on my local drive
+checkpoint_paths = sorted(glob(CHECKPOINT_DIR + "/checkpoint_*"))
+checkpoint_path = checkpoint_paths[0]
+os.rename(checkpoint_path, f'{CHECKPOINT_DIR}/CamemBert-Bio-Generalized')
+shutil.move(f'{CHECKPOINT_DIR}/CamemBert-Bio-Generalized','/content/drive/MyDrive/models')
+```