File size: 11,956 Bytes
171eb01 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 |
from rest_framework.views import APIView
from rest_framework.response import Response
from rest_framework import status
from django.conf import settings
import torch
from transformers import AutoTokenizer, AutoModelForMaskedLM
import os
import math
# Model: roberta-base (Masked LM)
# Logic: Low Perplexity = High AI Probability (Predictable)
MODEL_NAME = "roberta-base"
print(f"Loading Model: {MODEL_NAME}...")
try:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using Device: {device}")
# Use MaskedLM to calculate loss/perplexity
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForMaskedLM.from_pretrained(MODEL_NAME)
model.to(device)
model.eval()
print("Model Loaded Successfully!")
except Exception as e:
print(f"Error loading model: {e}")
model = None
tokenizer = None
import time
def calculate_token_scores(text):
"""
Analyzes text at token level to identify AI-generated regions.
Returns: List of (start_char, end_char, score) tuples.
"""
if not tokenizer or not model:
return []
encodings = tokenizer(text, return_tensors="pt", truncation=True, max_length=512, return_offsets_mapping=True)
input_ids = encodings.input_ids.to(device)
offsets = encodings.offset_mapping[0].cpu().numpy() # [(0,0), (0,3), (3,4)...]
seq_len = input_ids.shape[1]
if seq_len < 4:
return []
# Calculate loss for every token (Stride=1 for max granularity)
nlls = []
# Reduced BATCH_SIZE to 4 to prevent OOM on 4GB GPU
BATCH_SIZE = 4
tensor_input_ids = input_ids.repeat(BATCH_SIZE, 1)
# Iterate all tokens (excluding CLS/SEP)
indices_to_mask = list(range(1, seq_len - 1))
token_losses = [0.0] * seq_len
loss_fct = torch.nn.CrossEntropyLoss(reduction='none')
torch.cuda.empty_cache() # Clear memory before loop
for i in range(0, len(indices_to_mask), BATCH_SIZE):
batch_indices = indices_to_mask[i : i + BATCH_SIZE]
current_curr_size = len(batch_indices)
if current_curr_size == 0: break
batch_input_ids = tensor_input_ids[:current_curr_size].clone()
batch_labels = torch.full(batch_input_ids.shape, -100).to(device)
for j, pos in enumerate(batch_indices):
batch_labels[j, pos] = batch_input_ids[j, pos].item()
batch_input_ids[j, pos] = tokenizer.mask_token_id
with torch.no_grad():
outputs = model(batch_input_ids)
predictions = outputs.logits # [B, L, V]
predictions = predictions.permute(0, 2, 1)
loss = loss_fct(predictions, batch_labels) # [B, L]
masked_losses = loss.sum(dim=1) # [B]
for j, pos in enumerate(batch_indices):
token_losses[pos] = masked_losses[j].item()
# Identify ranges to highlight
# Threshold: If loss is low (< approx 2.0 - 3.0), it's likely AI.
# Let's say Threshold = 2.5 (similar to global threshold)
# Adjusted to 1.0 (Strict) to avoid highlighting common human words.
HIGHLIGHT_THRESHOLD = 1.0
highlights = []
for i in range(1, seq_len - 1):
loss = token_losses[i]
# Lower loss = More likely AI
if loss < HIGHLIGHT_THRESHOLD:
start, end = offsets[i]
# Filter out special tokens or empty offsets
if start == end: continue
highlights.append((start, end, loss))
return highlights
def calculate_perplexity(text):
"""
Calculates Pseudo-Perplexity (PPL) for Masked Language Models (like RoBERTa).
Formula: PPL = exp( -1/N * sum( log(P(w_i | context)) ) )
OPTIMIZATION:
- Strided Masking (Stride=3): Mask every 3rd token. 3x Speedup.
- Batch Size: 16
"""
t0 = time.time()
# 1. Tokenize
encodings = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
input_ids = encodings.input_ids.to(device) # Shape: [1, seq_len]
seq_len = input_ids.shape[1]
# If sequence is too short, return default
if seq_len < 4:
return 100.0
nlls = [] # Negative Log Likelihoods
# 2. Batched Masking for Efficiency
BATCH_SIZE = 8
# STRIDE: How many tokens to skip?
# Stride 1 = All tokens (Slowest, Most Accurate)
# Stride 3 = Every 3rd token (3x Faster, Good Approx)
STRIDE = 3
tensor_input_ids = input_ids.repeat(BATCH_SIZE, 1) # [BATCH, seq_len]
# Iterate through tokens [1...seq_len-2] to avoid CLS/SEP
start_idx = 1
end_idx = seq_len - 1
loss_fct = torch.nn.CrossEntropyLoss(reduction='none')
# Create list of indices to mask
indices_to_mask = list(range(start_idx, end_idx, STRIDE))
total_steps = len(indices_to_mask)
for i in range(0, total_steps, BATCH_SIZE):
# valid batch current step
batch_indices = indices_to_mask[i : i + BATCH_SIZE]
current_batch_size = len(batch_indices)
if current_batch_size == 0: break
# Prepare inputs: Clone the repeated text
batch_input_ids = tensor_input_ids[:current_batch_size].clone()
# Create labels: -100 means ignore index for loss calculation
batch_labels = torch.full(batch_input_ids.shape, -100).to(device)
# Mask the target tokens
for j, token_pos in enumerate(batch_indices):
# Save the original token ID as the label
batch_labels[j, token_pos] = batch_input_ids[j, token_pos].item()
# Replace input with [MASK]
batch_input_ids[j, token_pos] = tokenizer.mask_token_id
# Forward pass
with torch.no_grad():
outputs = model(batch_input_ids)
predictions = outputs.logits # [batch, seq_len, vocab]
# Calculate loss only for the masked positions
predictions = predictions.permute(0, 2, 1)
loss = loss_fct(predictions, batch_labels)
# Get scalar loss for the masked token
masked_losses = loss.sum(dim=1) # [current_batch_size]
nlls.append(masked_losses)
if not nlls:
return 0.0
# Stack all NLLs
all_nlls = torch.cat(nlls)
# Mean NLL
mean_nll = all_nlls.mean()
# Perplexity = exp(mean_nll)
ppl = torch.exp(mean_nll)
print(f"Inference Time: {time.time() - t0:.2f}s")
return ppl.item()
class AnalyzeView(APIView):
def post(self, request):
if not model or not tokenizer:
return Response({"error": "Model not loaded"}, status=status.HTTP_503_SERVICE_UNAVAILABLE)
data = request.data
text = data.get('text', '')
if not text and 'file' in request.FILES:
text = "File content placeholder"
if not text:
return Response({"error": "No text provided"}, status=status.HTTP_400_BAD_REQUEST)
try:
ppl = calculate_perplexity(text)
# PPL < 2.5 -> AI
# PPL > 2.5 -> Human
# Sigmoid Curve calibrated at 2.5
# Formula: 100 / (1 + exp(3 * (ppl - 2.5)))
ai_score = 100 / (1 + math.exp(3 * (ppl - 2.5)))
# Clamp between 2 and 99
ai_score = max(2.5, min(99.5, ai_score))
label = "AI Generated" if ai_score > 50 else "Human Written"
return Response({
"score": round(ai_score, 1),
"label": label,
"perplexity": round(ppl, 2),
"device": str(device)
})
except Exception as e:
return Response({"error": str(e)}, status=status.HTTP_500_INTERNAL_SERVER_ERROR)
from django.http import FileResponse
from reportlab.lib.pagesizes import letter
from reportlab.lib import colors
from reportlab.platypus import SimpleDocTemplate, Paragraph, Spacer
from reportlab.lib.styles import getSampleStyleSheet, ParagraphStyle
from reportlab.lib.enums import TA_JUSTIFY
import io
class ReportView(APIView):
def post(self, request):
try:
data = request.data
text = data.get('text', '')
score = data.get('score', 0)
label = data.get('label', 'Unknown')
# Granular Highlighting Logic
# 1. Get highlight ranges
highlights = calculate_token_scores(text)
# 2. Sort highlights by start index
highlights.sort(key=lambda x: x[0])
# 3. Construct XML tagged string
# IMPORTANT: Escape text to prevent XML errors in ReportLab
import html
formatted_text = ""
current_idx = 0
# Simple greedy tagging
print(f"Generating PDF for text length: {len(text)} with {len(highlights)} highlights.")
for start, end, loss in highlights:
if start < current_idx: continue # Skip overlaps
# Append non-highlighted text
segment = text[current_idx:start]
formatted_text += html.escape(segment)
# Append highlighted text
segment = text[start:end]
# XML tag for yellow background
formatted_text += f'<font backColor="yellow">{html.escape(segment)}</font>'
current_idx = end
# Append remaining text
formatted_text += html.escape(text[current_idx:])
# Handle newlines for HTML
formatted_text = formatted_text.replace('\n', '<br/>')
# Create PDF
buffer = io.BytesIO()
doc = SimpleDocTemplate(buffer, pagesize=letter,
rightMargin=72, leftMargin=72,
topMargin=72, bottomMargin=18)
Story = []
styles = getSampleStyleSheet()
styles.add(ParagraphStyle(name='Justify', alignment=TA_JUSTIFY))
# Header
Story.append(Paragraph("DetectAI Analysis Report", styles["Heading1"]))
Story.append(Paragraph(f"Date: {time.strftime('%Y-%m-%d %H:%M:%S')}", styles["Normal"]))
Story.append(Spacer(1, 12))
# Result
# Calculate Highlighted Percentage
total_chars = len(text)
highlighted_chars = sum((h[1] - h[0]) for h in highlights)
highlight_ratio = (highlighted_chars / total_chars * 100) if total_chars > 0 else 0
# Display metrics
res_color = "red" if score > 50 else "green"
# Primary Metric: AI Probability
Story.append(Paragraph(f'AI Probability: <font color="{res_color}"><b>{score}%</b></font>', styles["Heading2"]))
# Secondary Metric: Highlighted Content
Story.append(Paragraph(f'Highlighted Content: <b>{highlight_ratio:.1f}%</b>', styles["Normal"]))
Story.append(Spacer(1, 12))
# Content Header
Story.append(Paragraph("Analyzed Content:", styles["Heading3"]))
Story.append(Spacer(1, 6))
# The Content (Highlighted)
Story.append(Paragraph(formatted_text, styles["Justify"]))
doc.build(Story)
buffer.seek(0)
return FileResponse(buffer, as_attachment=True, filename='detectAI_report.pdf')
except Exception as e:
print(f"Report Generation Error: {e}")
return Response({"error": str(e)}, status=status.HTTP_500_INTERNAL_SERVER_ERROR)
|