Spaces:
Paused
Paused
akash commited on
Commit Β·
fa81b74
1
Parent(s): 01e7f9c
Fix dataset loading (stanfordnlp/imdb) and update requirements
Browse files- requirements.txt +3 -2
- train.py +38 -23
requirements.txt
CHANGED
|
@@ -1,9 +1,10 @@
|
|
| 1 |
gradio
|
| 2 |
transformers
|
| 3 |
-
datasets=
|
| 4 |
accelerate
|
| 5 |
torch
|
| 6 |
huggingface_hub
|
| 7 |
safetensors
|
| 8 |
tokenizers
|
| 9 |
-
scikit-learn
|
|
|
|
|
|
| 1 |
gradio
|
| 2 |
transformers
|
| 3 |
+
datasets>=3.0.0
|
| 4 |
accelerate
|
| 5 |
torch
|
| 6 |
huggingface_hub
|
| 7 |
safetensors
|
| 8 |
tokenizers
|
| 9 |
+
scikit-learn
|
| 10 |
+
numpy
|
train.py
CHANGED
|
@@ -5,21 +5,30 @@ This script downloads a pre-trained model and fine-tunes it on IMDb dataset
|
|
| 5 |
|
| 6 |
import os
|
| 7 |
import sys
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
from transformers import (
|
| 9 |
-
AutoTokenizer,
|
| 10 |
AutoModelForSequenceClassification,
|
|
|
|
| 11 |
Trainer,
|
| 12 |
TrainingArguments,
|
| 13 |
)
|
| 14 |
-
from datasets import load_dataset
|
| 15 |
-
from huggingface_hub import login
|
| 16 |
-
import torch
|
| 17 |
|
| 18 |
print(f"β
PyTorch version: {torch.__version__}")
|
| 19 |
print(f"β
Transformers version: {__import__('transformers').__version__}")
|
| 20 |
print(f"β
Datasets version: {__import__('datasets').__version__}")
|
| 21 |
|
| 22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
def setup_training():
|
| 24 |
"""Setup and run the fine-tuning process"""
|
| 25 |
|
|
@@ -30,16 +39,16 @@ def setup_training():
|
|
| 30 |
# 1. Load Dataset
|
| 31 |
print("\nπ Step 1: Loading dataset...")
|
| 32 |
try:
|
| 33 |
-
#
|
| 34 |
-
dataset = load_dataset("imdb")
|
| 35 |
print(f"β
Loaded {len(dataset['train'])} training samples")
|
| 36 |
print(f"β
Loaded {len(dataset['test'])} test samples")
|
| 37 |
-
|
| 38 |
# Use a subset for faster training
|
| 39 |
print("\nβ‘ Using subset for faster training (500 train, 100 test)")
|
| 40 |
train_subset = dataset["train"].select(range(500))
|
| 41 |
test_subset = dataset["test"].select(range(100))
|
| 42 |
-
|
| 43 |
except Exception as e:
|
| 44 |
print(f"β Failed to load dataset: {str(e)}")
|
| 45 |
return None
|
|
@@ -53,7 +62,9 @@ def setup_training():
|
|
| 53 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 54 |
model = AutoModelForSequenceClassification.from_pretrained(
|
| 55 |
model_name,
|
| 56 |
-
num_labels=2
|
|
|
|
|
|
|
| 57 |
)
|
| 58 |
print("β
Model and tokenizer loaded successfully")
|
| 59 |
except Exception as e:
|
|
@@ -94,13 +105,14 @@ def setup_training():
|
|
| 94 |
num_train_epochs=2,
|
| 95 |
per_device_train_batch_size=8,
|
| 96 |
per_device_eval_batch_size=8,
|
| 97 |
-
|
| 98 |
save_strategy="epoch",
|
| 99 |
-
logging_dir="./logs",
|
| 100 |
logging_steps=10,
|
| 101 |
load_best_model_at_end=True,
|
|
|
|
|
|
|
| 102 |
save_total_limit=2,
|
| 103 |
-
report_to=
|
| 104 |
push_to_hub=False,
|
| 105 |
)
|
| 106 |
print("β
Training arguments configured")
|
|
@@ -115,7 +127,8 @@ def setup_training():
|
|
| 115 |
args=training_args,
|
| 116 |
train_dataset=tokenized_train,
|
| 117 |
eval_dataset=tokenized_test,
|
| 118 |
-
|
|
|
|
| 119 |
)
|
| 120 |
|
| 121 |
trainer.train()
|
|
@@ -171,24 +184,26 @@ def push_to_hub():
|
|
| 171 |
model.push_to_hub(repo_name)
|
| 172 |
tokenizer.push_to_hub(repo_name)
|
| 173 |
|
| 174 |
-
print(
|
| 175 |
print(f"π https://huggingface.co/{repo_name}")
|
| 176 |
|
| 177 |
|
| 178 |
if __name__ == "__main__":
|
| 179 |
-
# Train the model
|
| 180 |
model_path = setup_training()
|
| 181 |
|
| 182 |
-
if model_path:
|
| 183 |
-
print("\n
|
| 184 |
-
|
| 185 |
|
| 186 |
-
|
|
|
|
|
|
|
|
|
|
| 187 |
push_to_hub()
|
| 188 |
else:
|
| 189 |
print("βοΈ Skipping upload. You can push later.")
|
| 190 |
-
|
| 191 |
-
print("\nβ
All done! Your model is ready.")
|
| 192 |
-
print("π Model location: ./my_finetuned_model")
|
| 193 |
else:
|
| 194 |
-
print("
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
|
| 6 |
import os
|
| 7 |
import sys
|
| 8 |
+
|
| 9 |
+
import numpy as np
|
| 10 |
+
import torch
|
| 11 |
+
from datasets import load_dataset
|
| 12 |
+
from huggingface_hub import login
|
| 13 |
+
from sklearn.metrics import accuracy_score
|
| 14 |
from transformers import (
|
|
|
|
| 15 |
AutoModelForSequenceClassification,
|
| 16 |
+
AutoTokenizer,
|
| 17 |
Trainer,
|
| 18 |
TrainingArguments,
|
| 19 |
)
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
print(f"β
PyTorch version: {torch.__version__}")
|
| 22 |
print(f"β
Transformers version: {__import__('transformers').__version__}")
|
| 23 |
print(f"β
Datasets version: {__import__('datasets').__version__}")
|
| 24 |
|
| 25 |
|
| 26 |
+
def compute_metrics(eval_pred):
|
| 27 |
+
logits, labels = eval_pred
|
| 28 |
+
predictions = np.argmax(logits, axis=-1)
|
| 29 |
+
return {"accuracy": float(accuracy_score(labels, predictions))}
|
| 30 |
+
|
| 31 |
+
|
| 32 |
def setup_training():
|
| 33 |
"""Setup and run the fine-tuning process"""
|
| 34 |
|
|
|
|
| 39 |
# 1. Load Dataset
|
| 40 |
print("\nπ Step 1: Loading dataset...")
|
| 41 |
try:
|
| 42 |
+
# Hub now requires namespace/name. "imdb" is stanfordnlp/imdb.
|
| 43 |
+
dataset = load_dataset("stanfordnlp/imdb")
|
| 44 |
print(f"β
Loaded {len(dataset['train'])} training samples")
|
| 45 |
print(f"β
Loaded {len(dataset['test'])} test samples")
|
| 46 |
+
|
| 47 |
# Use a subset for faster training
|
| 48 |
print("\nβ‘ Using subset for faster training (500 train, 100 test)")
|
| 49 |
train_subset = dataset["train"].select(range(500))
|
| 50 |
test_subset = dataset["test"].select(range(100))
|
| 51 |
+
|
| 52 |
except Exception as e:
|
| 53 |
print(f"β Failed to load dataset: {str(e)}")
|
| 54 |
return None
|
|
|
|
| 62 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 63 |
model = AutoModelForSequenceClassification.from_pretrained(
|
| 64 |
model_name,
|
| 65 |
+
num_labels=2,
|
| 66 |
+
id2label={0: "NEGATIVE", 1: "POSITIVE"},
|
| 67 |
+
label2id={"NEGATIVE": 0, "POSITIVE": 1},
|
| 68 |
)
|
| 69 |
print("β
Model and tokenizer loaded successfully")
|
| 70 |
except Exception as e:
|
|
|
|
| 105 |
num_train_epochs=2,
|
| 106 |
per_device_train_batch_size=8,
|
| 107 |
per_device_eval_batch_size=8,
|
| 108 |
+
eval_strategy="epoch",
|
| 109 |
save_strategy="epoch",
|
|
|
|
| 110 |
logging_steps=10,
|
| 111 |
load_best_model_at_end=True,
|
| 112 |
+
metric_for_best_model="eval_accuracy",
|
| 113 |
+
greater_is_better=True,
|
| 114 |
save_total_limit=2,
|
| 115 |
+
report_to="none",
|
| 116 |
push_to_hub=False,
|
| 117 |
)
|
| 118 |
print("β
Training arguments configured")
|
|
|
|
| 127 |
args=training_args,
|
| 128 |
train_dataset=tokenized_train,
|
| 129 |
eval_dataset=tokenized_test,
|
| 130 |
+
processing_class=tokenizer,
|
| 131 |
+
compute_metrics=compute_metrics,
|
| 132 |
)
|
| 133 |
|
| 134 |
trainer.train()
|
|
|
|
| 184 |
model.push_to_hub(repo_name)
|
| 185 |
tokenizer.push_to_hub(repo_name)
|
| 186 |
|
| 187 |
+
print("β
Model pushed successfully!")
|
| 188 |
print(f"π https://huggingface.co/{repo_name}")
|
| 189 |
|
| 190 |
|
| 191 |
if __name__ == "__main__":
|
|
|
|
| 192 |
model_path = setup_training()
|
| 193 |
|
| 194 |
+
if not model_path:
|
| 195 |
+
print("\nβ Training failed. Please check the errors above.")
|
| 196 |
+
sys.exit(1)
|
| 197 |
|
| 198 |
+
print("\n" + "=" * 60)
|
| 199 |
+
if sys.stdin.isatty():
|
| 200 |
+
choice = input("π Push model to Hugging Face Hub? (y/n): ")
|
| 201 |
+
if choice.lower() == "y":
|
| 202 |
push_to_hub()
|
| 203 |
else:
|
| 204 |
print("βοΈ Skipping upload. You can push later.")
|
|
|
|
|
|
|
|
|
|
| 205 |
else:
|
| 206 |
+
print("βοΈ Non-interactive run: skipping Hub upload prompt.")
|
| 207 |
+
|
| 208 |
+
print("\nβ
All done! Your model is ready.")
|
| 209 |
+
print("π Model location: ./my_finetuned_model")
|