feat: persist model and logs in Hugging Face Space + add model push to hub
Browse files- Updated paths to save model and logs to /home/user/app (persistent in Spaces)
- Modified logging to stream to both file and UI log buffer
- Updated model saving/loading to use MODEL_DIR inside the persistent path
- Added push_model_to_hub() to upload trained model/tokenizer to Hugging Face Hub
- Extended Gradio UI with 'Evaluate Model' and 'Push Model to Hub' buttons
- app.py +8 -5
- train_abuse_model.py +86 -15
app.py
CHANGED
@@ -1,17 +1,20 @@
|
|
1 |
import gradio as gr
|
2 |
-
from train_abuse_model import run_training
|
3 |
|
4 |
with gr.Blocks() as demo:
|
5 |
gr.Markdown("## π§ Abuse Detection Fine-Tuning App")
|
6 |
-
gr.Markdown(
|
7 |
-
|
8 |
-
)
|
9 |
with gr.Row():
|
10 |
start_btn = gr.Button("π Start Training")
|
|
|
|
|
11 |
|
12 |
-
output_box = gr.Textbox(label="
|
13 |
|
14 |
start_btn.click(fn=run_training, outputs=output_box)
|
|
|
|
|
15 |
|
16 |
if __name__ == "__main__":
|
17 |
demo.launch()
|
|
|
1 |
import gradio as gr
|
2 |
+
from train_abuse_model import run_training, evaluate_saved_model, push_model_to_hub
|
3 |
|
4 |
with gr.Blocks() as demo:
|
5 |
gr.Markdown("## π§ Abuse Detection Fine-Tuning App")
|
6 |
+
gr.Markdown("β οΈ Keep this tab open while training or evaluating.")
|
7 |
+
|
|
|
8 |
with gr.Row():
|
9 |
start_btn = gr.Button("π Start Training")
|
10 |
+
eval_btn = gr.Button("π Evaluate Trained Model")
|
11 |
+
push_btn = gr.Button("π€ Push Model to Hub")
|
12 |
|
13 |
+
output_box = gr.Textbox(label="Logs", lines=25, interactive=False)
|
14 |
|
15 |
start_btn.click(fn=run_training, outputs=output_box)
|
16 |
+
eval_btn.click(fn=evaluate_saved_model, outputs=output_box)
|
17 |
+
push_btn.click(fn=push_model_to_hub, outputs=output_box)
|
18 |
|
19 |
if __name__ == "__main__":
|
20 |
demo.launch()
|
train_abuse_model.py
CHANGED
@@ -5,6 +5,7 @@ import logging
|
|
5 |
import io
|
6 |
import os
|
7 |
import gradio as gr # β
required for progress bar
|
|
|
8 |
|
9 |
# Python standard + ML packages
|
10 |
import pandas as pd
|
@@ -30,18 +31,23 @@ from transformers import (
|
|
30 |
TrainingArguments
|
31 |
)
|
32 |
|
|
|
|
|
|
|
|
|
33 |
# configure logging
|
34 |
-
log_buffer = io.StringIO()
|
35 |
logging.basicConfig(
|
36 |
level=logging.INFO,
|
37 |
format="%(asctime)s - %(levelname)s - %(message)s",
|
38 |
handlers=[
|
39 |
-
logging.FileHandler(
|
40 |
-
logging.StreamHandler(log_buffer)
|
41 |
]
|
42 |
)
|
43 |
logger = logging.getLogger(__name__)
|
44 |
|
|
|
45 |
# Check versions
|
46 |
logger.info("Transformers version:", transformers.__version__)
|
47 |
|
@@ -50,6 +56,9 @@ logger.info("Transformers version: %s", torch.__version__)
|
|
50 |
logger.info("torch.cuda.is_available(): %s", torch.cuda.is_available())
|
51 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
52 |
|
|
|
|
|
|
|
53 |
# Custom Dataset class
|
54 |
|
55 |
class AbuseDataset(Dataset):
|
@@ -127,33 +136,81 @@ def tune_thresholds(probs, true_labels, verbose=True):
|
|
127 |
def evaluate_model_with_thresholds(trainer, test_dataset):
|
128 |
"""Run full evaluation with automatic threshold tuning."""
|
129 |
logger.info("\nπ Running model predictions...")
|
|
|
|
|
130 |
predictions = trainer.predict(test_dataset)
|
131 |
probs = torch.sigmoid(torch.tensor(predictions.predictions)).numpy()
|
132 |
true_soft = np.array(predictions.label_ids)
|
133 |
|
134 |
logger.info("\nπ Tuning thresholds...")
|
|
|
135 |
best_low, best_high, best_f1 = tune_thresholds(probs, true_soft)
|
136 |
|
137 |
logger.info(f"\nβ
Best thresholds: low={best_low:.2f}, high={best_high:.2f} (macro F1={best_f1:.3f})")
|
|
|
138 |
|
139 |
final_pred_soft = map_to_3_classes(probs, best_low, best_high)
|
140 |
final_pred_str = convert_to_label_strings(final_pred_soft)
|
141 |
true_str = convert_to_label_strings(true_soft)
|
142 |
|
143 |
logger.info("\nπ Final Evaluation Report (multi-class per label):\n")
|
|
|
144 |
logger.info(classification_report(
|
145 |
true_str,
|
146 |
final_pred_str,
|
147 |
labels=["no", "plausibly", "yes"],
|
|
|
148 |
zero_division=0
|
149 |
))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
150 |
|
151 |
-
return {
|
152 |
-
"thresholds": (best_low, best_high),
|
153 |
-
"macro_f1": best_f1,
|
154 |
-
"true_labels": true_str,
|
155 |
-
"pred_labels": final_pred_str
|
156 |
-
}
|
157 |
|
158 |
token = os.environ.get("HF_TOKEN") # Reads my token from a secure hf secret
|
159 |
|
@@ -202,6 +259,10 @@ train_texts, val_texts, train_labels, val_labels = train_test_split(
|
|
202 |
model_name = "microsoft/deberta-v3-base"
|
203 |
|
204 |
def run_training(progress=gr.Progress(track_tqdm=True)):
|
|
|
|
|
|
|
|
|
205 |
yield "π Starting training...\n"
|
206 |
try:
|
207 |
logger.info("Starting training run...")
|
@@ -269,11 +330,10 @@ def run_training(progress=gr.Progress(track_tqdm=True)):
|
|
269 |
trainer.train()
|
270 |
|
271 |
# Save the model and tokenizer
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
logger.info(" Training completed and model saved.")
|
278 |
yield "π Training complete! Model saved.\n"
|
279 |
|
@@ -284,7 +344,6 @@ def run_training(progress=gr.Progress(track_tqdm=True)):
|
|
284 |
# Evaluation
|
285 |
try:
|
286 |
if 'trainer' in locals():
|
287 |
-
label_map = {0.0: "no", 0.5: "plausibly", 1.0: "yes"}
|
288 |
evaluate_model_with_thresholds(trainer, test_dataset)
|
289 |
logger.info("Evaluation completed")
|
290 |
except Exception as e:
|
@@ -292,3 +351,15 @@ def run_training(progress=gr.Progress(track_tqdm=True)):
|
|
292 |
log_buffer.seek(0)
|
293 |
return log_buffer.read()
|
294 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
import io
|
6 |
import os
|
7 |
import gradio as gr # β
required for progress bar
|
8 |
+
from pathlib import Path
|
9 |
|
10 |
# Python standard + ML packages
|
11 |
import pandas as pd
|
|
|
31 |
TrainingArguments
|
32 |
)
|
33 |
|
34 |
+
PERSIST_DIR = Path("/home/user/app")
|
35 |
+
MODEL_DIR = PERSIST_DIR / "saved_model"
|
36 |
+
LOG_FILE = PERSIST_DIR / "training.log"
|
37 |
+
|
38 |
# configure logging
|
39 |
+
log_buffer = io.StringIO()
|
40 |
logging.basicConfig(
|
41 |
level=logging.INFO,
|
42 |
format="%(asctime)s - %(levelname)s - %(message)s",
|
43 |
handlers=[
|
44 |
+
logging.FileHandler(LOG_FILE),
|
45 |
+
logging.StreamHandler(log_buffer)
|
46 |
]
|
47 |
)
|
48 |
logger = logging.getLogger(__name__)
|
49 |
|
50 |
+
|
51 |
# Check versions
|
52 |
logger.info("Transformers version:", transformers.__version__)
|
53 |
|
|
|
56 |
logger.info("torch.cuda.is_available(): %s", torch.cuda.is_available())
|
57 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
58 |
|
59 |
+
# Label mapping for evaluation
|
60 |
+
label_map = {0.0: "no", 0.5: "plausibly", 1.0: "yes"}
|
61 |
+
|
62 |
# Custom Dataset class
|
63 |
|
64 |
class AbuseDataset(Dataset):
|
|
|
136 |
def evaluate_model_with_thresholds(trainer, test_dataset):
|
137 |
"""Run full evaluation with automatic threshold tuning."""
|
138 |
logger.info("\nπ Running model predictions...")
|
139 |
+
yield "\nπ Running model predictions..."
|
140 |
+
|
141 |
predictions = trainer.predict(test_dataset)
|
142 |
probs = torch.sigmoid(torch.tensor(predictions.predictions)).numpy()
|
143 |
true_soft = np.array(predictions.label_ids)
|
144 |
|
145 |
logger.info("\nπ Tuning thresholds...")
|
146 |
+
yield "\nπ Tuning thresholds..."
|
147 |
best_low, best_high, best_f1 = tune_thresholds(probs, true_soft)
|
148 |
|
149 |
logger.info(f"\nβ
Best thresholds: low={best_low:.2f}, high={best_high:.2f} (macro F1={best_f1:.3f})")
|
150 |
+
yield f"\nβ
Best thresholds: low={best_low:.2f}, high={best_high:.2f} (macro F1={best_f1:.3f})"
|
151 |
|
152 |
final_pred_soft = map_to_3_classes(probs, best_low, best_high)
|
153 |
final_pred_str = convert_to_label_strings(final_pred_soft)
|
154 |
true_str = convert_to_label_strings(true_soft)
|
155 |
|
156 |
logger.info("\nπ Final Evaluation Report (multi-class per label):\n")
|
157 |
+
yield "\nπ Final Evaluation Report (multi-class per label):\n "
|
158 |
logger.info(classification_report(
|
159 |
true_str,
|
160 |
final_pred_str,
|
161 |
labels=["no", "plausibly", "yes"],
|
162 |
+
digits=3,
|
163 |
zero_division=0
|
164 |
))
|
165 |
+
yield classification_report(
|
166 |
+
true_str,
|
167 |
+
final_pred_str,
|
168 |
+
labels=["no", "plausibly", "yes"],
|
169 |
+
digits=3,
|
170 |
+
zero_division=0
|
171 |
+
)
|
172 |
+
def load_saved_model_and_tokenizer():
|
173 |
+
tokenizer = DebertaV2Tokenizer.from_pretrained(MODEL_DIR)
|
174 |
+
model = AutoModelForSequenceClassification.from_pretrained(MODEL_DIR).to(device)
|
175 |
+
return tokenizer, model
|
176 |
+
|
177 |
+
def evaluate_saved_model(progress=gr.Progress(track_tqdm=True)):
|
178 |
+
if os.path.exists("saved_model/"):
|
179 |
+
yield "β
Trained model found! Skipping training...\n"
|
180 |
+
else:
|
181 |
+
yield "β No trained model found. Please train the model first.\n"
|
182 |
+
return
|
183 |
+
try:
|
184 |
+
logger.info("π Loading saved model for evaluation...")
|
185 |
+
yield "π Loading saved model for evaluation...\n"
|
186 |
+
|
187 |
+
tokenizer, model = load_saved_model_and_tokenizer()
|
188 |
+
test_dataset = AbuseDataset(test_texts, test_labels, tokenizer)
|
189 |
+
|
190 |
+
trainer = Trainer(
|
191 |
+
model=model,
|
192 |
+
args=TrainingArguments(
|
193 |
+
output_dir="./results_eval",
|
194 |
+
per_device_eval_batch_size=4,
|
195 |
+
logging_dir="./logs_eval",
|
196 |
+
disable_tqdm=True
|
197 |
+
),
|
198 |
+
eval_dataset=test_dataset
|
199 |
+
)
|
200 |
+
|
201 |
+
label_map = {0.0: "no", 0.5: "plausibly", 1.0: "yes"}
|
202 |
+
|
203 |
+
# Re-yield from generator
|
204 |
+
for line in evaluate_model_with_thresholds(trainer, test_dataset):
|
205 |
+
yield line
|
206 |
+
|
207 |
+
logger.info("β
Evaluation complete.\n")
|
208 |
+
yield "\nβ
Evaluation complete.\n"
|
209 |
+
|
210 |
+
except Exception as e:
|
211 |
+
logger.exception(f"β Evaluation failed: {e}")
|
212 |
+
yield f"β Evaluation failed: {e}\n"
|
213 |
|
|
|
|
|
|
|
|
|
|
|
|
|
214 |
|
215 |
token = os.environ.get("HF_TOKEN") # Reads my token from a secure hf secret
|
216 |
|
|
|
259 |
model_name = "microsoft/deberta-v3-base"
|
260 |
|
261 |
def run_training(progress=gr.Progress(track_tqdm=True)):
|
262 |
+
if os.path.exists("saved_model/"):
|
263 |
+
yield "β
Trained model found! Skipping training...\n"
|
264 |
+
yield evaluate_saved_model()
|
265 |
+
return
|
266 |
yield "π Starting training...\n"
|
267 |
try:
|
268 |
logger.info("Starting training run...")
|
|
|
330 |
trainer.train()
|
331 |
|
332 |
# Save the model and tokenizer
|
333 |
+
MODEL_DIR.mkdir(parents=True, exist_ok=True)
|
334 |
+
model.save_pretrained(MODEL_DIR)
|
335 |
+
tokenizer.save_pretrained(MODEL_DIR)
|
336 |
+
|
|
|
337 |
logger.info(" Training completed and model saved.")
|
338 |
yield "π Training complete! Model saved.\n"
|
339 |
|
|
|
344 |
# Evaluation
|
345 |
try:
|
346 |
if 'trainer' in locals():
|
|
|
347 |
evaluate_model_with_thresholds(trainer, test_dataset)
|
348 |
logger.info("Evaluation completed")
|
349 |
except Exception as e:
|
|
|
351 |
log_buffer.seek(0)
|
352 |
return log_buffer.read()
|
353 |
|
354 |
+
def push_model_to_hub():
|
355 |
+
try:
|
356 |
+
logger.info("π Pushing model to Hugging Face Hub...")
|
357 |
+
tokenizer, model = load_saved_model_and_tokenizer()
|
358 |
+
model.push_to_hub("rshakked/safe-talk", use_auth_token=token)
|
359 |
+
tokenizer.push_to_hub("rshakked/safe-talk", use_auth_token=token)
|
360 |
+
return "β
Model pushed to hub successfully!"
|
361 |
+
except Exception as e:
|
362 |
+
logger.exception("β Failed to push model to hub.")
|
363 |
+
return f"β Failed to push model: {e}"
|
364 |
+
|
365 |
+
|