fix: include missing predict_pipeline.py in repo with run_prediction_pipeline()
Browse files- predict_pipeline.py +55 -0
predict_pipeline.py
CHANGED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import zipfile
|
| 2 |
+
import tempfile
|
| 3 |
+
from pathlib import Path
|
| 4 |
+
import torch
|
| 5 |
+
from transformers import DebertaV2Tokenizer, AutoModelForSequenceClassification
|
| 6 |
+
|
| 7 |
+
from train_abuse_model import (
|
| 8 |
+
MODEL_DIR,
|
| 9 |
+
device,
|
| 10 |
+
load_saved_model_and_tokenizer,
|
| 11 |
+
map_to_3_classes,
|
| 12 |
+
convert_to_label_strings
|
| 13 |
+
)
|
| 14 |
+
|
| 15 |
+
def run_prediction_pipeline(desc_input, chat_zip):
|
| 16 |
+
try:
|
| 17 |
+
# Start with the base input
|
| 18 |
+
merged_input = desc_input.strip()
|
| 19 |
+
|
| 20 |
+
# If a chat zip was uploaded
|
| 21 |
+
if chat_zip:
|
| 22 |
+
with tempfile.TemporaryDirectory() as tmpdir:
|
| 23 |
+
with zipfile.ZipFile(chat_zip.name, 'r') as zip_ref:
|
| 24 |
+
zip_ref.extractall(tmpdir)
|
| 25 |
+
chat_texts = []
|
| 26 |
+
for file in Path(tmpdir).glob("*.txt"):
|
| 27 |
+
with open(file, encoding="utf-8", errors="ignore") as f:
|
| 28 |
+
chat_texts.append(f.read())
|
| 29 |
+
full_chat = "\n".join(chat_texts)
|
| 30 |
+
|
| 31 |
+
# π§ MOCK summarization
|
| 32 |
+
summary = "[Mock summary of Hebrew WhatsApp chat...]"
|
| 33 |
+
|
| 34 |
+
# π MOCK translation
|
| 35 |
+
translated_summary = "[Translated summary in English]"
|
| 36 |
+
|
| 37 |
+
merged_input = f"{desc_input.strip()}\n\n[Summary]: {translated_summary}"
|
| 38 |
+
|
| 39 |
+
# Load classifier
|
| 40 |
+
tokenizer, model = load_saved_model_and_tokenizer()
|
| 41 |
+
inputs = tokenizer(merged_input, truncation=True, padding=True, max_length=512, return_tensors="pt").to(device)
|
| 42 |
+
|
| 43 |
+
with torch.no_grad():
|
| 44 |
+
outputs = model(**inputs).logits
|
| 45 |
+
probs = torch.sigmoid(outputs).cpu().numpy()
|
| 46 |
+
|
| 47 |
+
# Static threshold values (or load from config later)
|
| 48 |
+
best_low, best_high = 0.35, 0.65
|
| 49 |
+
pred_soft = map_to_3_classes(probs, best_low, best_high)
|
| 50 |
+
pred_str = convert_to_label_strings(pred_soft)
|
| 51 |
+
|
| 52 |
+
return merged_input, ", ".join(pred_str)
|
| 53 |
+
|
| 54 |
+
except Exception as e:
|
| 55 |
+
return f"β Prediction failed: {e}", ""
|