APEXPRIT commited on
Commit
1fcb4a0
·
verified ·
1 Parent(s): 155810c

Upload folder using huggingface_hub

Browse files
Files changed (3) hide show
  1. Dockerfile +2 -1
  2. api.py +12 -4
  3. retrain_on_server.py +40 -0
Dockerfile CHANGED
@@ -24,8 +24,9 @@ RUN pip install --no-cache-dir -r requirements.txt
24
  # Copy the rest of the application code
25
  COPY . .
26
 
27
- # Ensure models directory exists
28
  RUN mkdir -p models
 
29
 
30
  # Expose the port the app runs on (Hugging Face default)
31
  EXPOSE 7860
 
24
  # Copy the rest of the application code
25
  COPY . .
26
 
27
+ # Ensure models directory exists and retrain model to match server environment
28
  RUN mkdir -p models
29
+ RUN python retrain_on_server.py
30
 
31
  # Expose the port the app runs on (Hugging Face default)
32
  EXPOSE 7860
api.py CHANGED
@@ -82,13 +82,21 @@ async def analyze_video_endpoint(
82
  raise HTTPException(status_code=400, detail=result["error"])
83
 
84
  # Format clean JSON response for the Flutter app
 
 
 
 
 
 
 
85
  response_data = {
86
  "status": "success",
87
- "prediction": result.get("prediction", "unknown"),
88
- "verdict": result.get("prediction", "unknown"), # added for Flutter
89
- "probability": result.get("deception_probability", 0.5),
 
90
  "confidence": result.get("confidence", 0.0),
91
- "is_deceptive": result.get("prediction") == "deceptive",
92
  }
93
 
94
  # Add temporal analysis (to draw the timeline graph in Flutter)
 
82
  raise HTTPException(status_code=400, detail=result["error"])
83
 
84
  # Format clean JSON response for the Flutter app
85
+ raw_prob = result.get("deception_probability", 0.5)
86
+
87
+ # BOOST SENSITIVITY: Apply a power curve (p^0.7) to make the AI more "sensitive"
88
+ # This shifts 0.4 -> 0.53 (making it deceptive) and 0.1 -> 0.2
89
+ boosted_prob = float(pow(raw_prob, 0.7))
90
+ verdict = "deceptive" if boosted_prob >= 0.5 else "truthful"
91
+
92
  response_data = {
93
  "status": "success",
94
+ "prediction": verdict,
95
+ "verdict": verdict,
96
+ "raw_probability": raw_prob,
97
+ "probability": boosted_prob,
98
  "confidence": result.get("confidence", 0.0),
99
+ "is_deceptive": verdict == "deceptive",
100
  }
101
 
102
  # Add temporal analysis (to draw the timeline graph in Flutter)
retrain_on_server.py ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import argparse
3
+ import pandas as pd
4
+ import numpy as np
5
+ from src.model.classifier import DeceptionClassifier
6
+ from src.utils.config import get_default_config
7
+
8
+ def main():
9
+ print("🚀 STARTING ON-SERVER MODEL RETRAINING...")
10
+
11
+ # 1. Generate local synthetic data (since we're on the server)
12
+ from create_pretrained_model import generate_synthetic_data, train_classifier, train_subject_calibrator
13
+
14
+ config = get_default_config()
15
+ output_dir = "./models"
16
+ os.makedirs(output_dir, exist_ok=True)
17
+
18
+ print("1. Generating fresh synthetic training data...")
19
+ df = generate_synthetic_data(num_samples=2000)
20
+
21
+ print("2. Training classifier with local scikit-learn version...")
22
+ # Use a slightly more "sensitive" training approach
23
+ classifier = DeceptionClassifier(config.model)
24
+ X = df.drop(['is_deceptive', 'subject_id'], axis=1)
25
+ y = df['is_deceptive']
26
+
27
+ classifier.train(X, y)
28
+
29
+ print("3. Saving perfectly matched model...")
30
+ classifier.save(os.path.join(output_dir, "deception_model.pkl"))
31
+
32
+ # Save feature names for reference
33
+ import json
34
+ with open(os.path.join(output_dir, "feature_names.json"), 'w') as f:
35
+ json.dump(classifier.feature_names, f)
36
+
37
+ print("✅ RETRAINING COMPLETE! Model is now perfectly synced with server versions.")
38
+
39
+ if __name__ == "__main__":
40
+ main()