Upload train.py with huggingface_hub
Browse files
train.py
ADDED
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from pathlib import Path
|
2 |
+
import shutil
|
3 |
+
from datasets import load_dataset
|
4 |
+
from transformers import TrainingArguments
|
5 |
+
from span_marker import SpanMarkerModel, Trainer
|
6 |
+
from span_marker.model_card import SpanMarkerModelCardData
|
7 |
+
from huggingface_hub import upload_folder, upload_file
|
8 |
+
|
9 |
+
|
10 |
+
def main() -> None:
|
11 |
+
# Load the dataset, ensure "tokens" and "ner_tags" columns, and get a list of labels
|
12 |
+
dataset = load_dataset("ljvmiranda921/tlunified-ner")
|
13 |
+
labels = dataset["train"].features["ner_tags"].feature.names
|
14 |
+
|
15 |
+
# Initialize a SpanMarker model using a pretrained BERT-style encoder
|
16 |
+
encoder_id = "jcblaise/roberta-tagalog-base"
|
17 |
+
model_id = f"tomaarsen/span-marker-roberta-tagalog-base-tlunified"
|
18 |
+
model = SpanMarkerModel.from_pretrained(
|
19 |
+
encoder_id,
|
20 |
+
labels=labels,
|
21 |
+
# SpanMarker hyperparameters:
|
22 |
+
model_max_length=256,
|
23 |
+
marker_max_length=128,
|
24 |
+
entity_max_length=8,
|
25 |
+
# Model card variables
|
26 |
+
model_card_data=SpanMarkerModelCardData(
|
27 |
+
model_id=model_id,
|
28 |
+
encoder_id=encoder_id,
|
29 |
+
dataset_name="TLUnified",
|
30 |
+
license="gpl-3.0",
|
31 |
+
language=["tl"],
|
32 |
+
),
|
33 |
+
)
|
34 |
+
|
35 |
+
# Prepare the 🤗 transformers training arguments
|
36 |
+
output_dir = Path("models") / model_id
|
37 |
+
args = TrainingArguments(
|
38 |
+
output_dir=output_dir,
|
39 |
+
run_name=model_id,
|
40 |
+
# Training Hyperparameters:
|
41 |
+
learning_rate=5e-5,
|
42 |
+
per_device_train_batch_size=32,
|
43 |
+
per_device_eval_batch_size=32,
|
44 |
+
num_train_epochs=3,
|
45 |
+
weight_decay=0.01,
|
46 |
+
warmup_ratio=0.1,
|
47 |
+
bf16=True, # Replace `bf16` with `fp16` if your hardware can't use bf16.
|
48 |
+
# Other Training parameters
|
49 |
+
logging_first_step=True,
|
50 |
+
logging_steps=50,
|
51 |
+
evaluation_strategy="steps",
|
52 |
+
save_strategy="steps",
|
53 |
+
eval_steps=200,
|
54 |
+
save_total_limit=1,
|
55 |
+
dataloader_num_workers=4,
|
56 |
+
)
|
57 |
+
|
58 |
+
# Initialize the trainer using our model, training args & dataset, and train
|
59 |
+
trainer = Trainer(
|
60 |
+
model=model,
|
61 |
+
args=args,
|
62 |
+
train_dataset=dataset["train"],
|
63 |
+
eval_dataset=dataset["validation"],
|
64 |
+
)
|
65 |
+
trainer.train()
|
66 |
+
|
67 |
+
# Compute & save the metrics on the test set
|
68 |
+
metrics = trainer.evaluate(dataset["test"], metric_key_prefix="test")
|
69 |
+
trainer.save_metrics("test", metrics)
|
70 |
+
|
71 |
+
# Save the model & training script locally
|
72 |
+
trainer.save_model(output_dir / "checkpoint-final")
|
73 |
+
shutil.copy2(__file__, output_dir / "checkpoint-final" / "train.py")
|
74 |
+
|
75 |
+
# Upload everything to the Hub
|
76 |
+
# breakpoint()
|
77 |
+
model.push_to_hub(model_id, private=True)
|
78 |
+
upload_folder(folder_path=output_dir / "runs", path_in_repo="runs", repo_id=model_id)
|
79 |
+
upload_file(path_or_fileobj=__file__, path_in_repo="train.py", repo_id=model_id)
|
80 |
+
upload_file(path_or_fileobj=output_dir / "all_results.json", path_in_repo="all_results.json", repo_id=model_id)
|
81 |
+
upload_file(path_or_fileobj=output_dir / "emissions.csv", path_in_repo="emissions.csv", repo_id=model_id)
|
82 |
+
|
83 |
+
|
84 |
+
if __name__ == "__main__":
|
85 |
+
main()
|