scripts: add training and evaluation helpers

evaluator.py

@@ -0,0 +1,44 @@
+import sys
+
+from datasets import load_dataset
+from transformers import TrainingArguments
+from span_marker import SpanMarkerModel, Trainer
+
+
+# Load the dataset, ensure "tokens" and "ner_tags" columns, and get a list of labels
+dataset = load_dataset("gwlms/germeval2014")
+labels = dataset["train"].features["ner_tags"].feature.names
+
+# Initialize a SpanMarker model using a pretrained BERT-style encoder
+model_name = sys.argv[1]
+model = SpanMarkerModel.from_pretrained(
+    model_name,
+    labels=labels,
+    # SpanMarker hyperparameters:
+    model_max_length=256,
+    marker_max_length=128,
+    entity_max_length=8,
+)
+
+args = TrainingArguments(
+    output_dir="/tmp",
+    per_device_eval_batch_size=64,
+)
+
+# Initialize the trainer using our model, training args & dataset, and train
+trainer = Trainer(
+    model=model,
+    args=args,
+    train_dataset=dataset["train"],
+    eval_dataset=dataset["validation"],
+)
+
+
+print("Evaluating on development set...")
+dev_metrics = trainer.evaluate(dataset["validation"], metric_key_prefix="eval")
+print(dev_metrics)
+
+print("Evaluating on test set...")
+test_metrics = trainer.evaluate(dataset["test"], metric_key_prefix="test")
+print(test_metrics)
+

trainer.py

@@ -0,0 +1,62 @@
+from datasets import load_dataset
+from transformers import TrainingArguments
+from span_marker import SpanMarkerModel, Trainer
+
+def perform_training(learning_rate: float, seed: int) -> None:
+    # Load the dataset, ensure "tokens" and "ner_tags" columns, and get a list of labels
+    dataset = load_dataset("gwlms/germeval2014")
+    labels = dataset["train"].features["ner_tags"].feature.names
+
+    # Initialize a SpanMarker model using a pretrained BERT-style encoder
+    model_name = "deepset/gelectra-large"
+    model = SpanMarkerModel.from_pretrained(
+        model_name,
+        labels=labels,
+        # SpanMarker hyperparameters:
+        model_max_length=256,
+        marker_max_length=128,
+        entity_max_length=8,
+    )
+
+    # Prepare the 🤗 transformers training arguments
+    args = TrainingArguments(
+        output_dir=f"./span_marker-gelectra-large-bs16-lr{learning_rate}-{seed}",
+        # Training Hyperparameters:
+        learning_rate=learning_rate,
+        per_device_train_batch_size=16,
+        per_device_eval_batch_size=16,
+        num_train_epochs=3,
+        weight_decay=0.01,
+        warmup_ratio=0.1,
+        fp16=True,  # Replace `bf16` with `fp16` if your hardware can't use bf16.
+        # Other Training parameters
+        logging_first_step=True,
+        logging_steps=50,
+        evaluation_strategy="epoch",
+        save_strategy="epoch",
+        save_total_limit=11,
+        dataloader_num_workers=2,
+        seed=seed,
+        load_best_model_at_end=True,
+    )
+
+    # Initialize the trainer using our model, training args & dataset, and train
+    trainer = Trainer(
+        model=model,
+        args=args,
+        train_dataset=dataset["train"],
+        eval_dataset=dataset["validation"],
+    )
+    trainer.train()
+    trainer.save_model(f"./span_marker-gelectra-large-bs16-lr{learning_rate}-{seed}/best-checkpoint")
+
+    # Compute & save the metrics on the test set
+    metrics = trainer.evaluate(dataset["test"], metric_key_prefix="test")
+    trainer.save_metrics("test", metrics)
+
+
+if __name__ == "__main__":
+    for learning_rate in [5e-05]:
+        for seed in [1,2,3,4,5]:
+            perform_training(learning_rate, seed)
+