File size: 10,408 Bytes
1a7ee60 | 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 | """
Lightweight Document Classifier using MiniLM
Fine-tuned for invoice, receipt, and form classification
Optimized for CPU inference
"""
import torch
import torch.nn as nn
from transformers import AutoTokenizer, AutoModel, AutoConfig
from typing import Dict, List, Tuple
import logging
import numpy as np
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
class DocumentClassifier(nn.Module):
"""
Lightweight document classifier based on MiniLM
4 classes: INVOICE, RECEIPT, FORM, OTHER
"""
def __init__(
self,
model_name: str = "nreimers/MiniLM-L6-H384-uncased",
num_labels: int = 4,
dropout_prob: float = 0.1
):
super().__init__()
self.num_labels = num_labels
self.model_name = model_name
# Load pre-trained MiniLM
self.config = AutoConfig.from_pretrained(model_name)
self.backbone = AutoModel.from_pretrained(model_name, config=self.config)
# Classification head
self.dropout = nn.Dropout(dropout_prob)
self.classifier = nn.Linear(self.config.hidden_size, num_labels)
logger.info(f"Initialized DocumentClassifier with {model_name}")
logger.info(f"Model parameters: {sum(p.numel() for p in self.parameters()) / 1e6:.2f}M")
def forward(
self,
input_ids: torch.Tensor,
attention_mask: torch.Tensor,
labels: torch.Tensor = None,
**kwargs
) -> Dict[str, torch.Tensor]:
"""
Forward pass
Args:
input_ids: Token IDs (batch_size, seq_len)
attention_mask: Attention mask (batch_size, seq_len)
labels: Ground truth labels (batch_size,)
Returns:
Dictionary with loss (if labels provided) and logits
"""
# Get embeddings from backbone
outputs = self.backbone(
input_ids=input_ids,
attention_mask=attention_mask
)
# Use [CLS] token representation
pooled_output = outputs.last_hidden_state[:, 0, :] # (batch_size, hidden_size)
# Apply dropout and classification
pooled_output = self.dropout(pooled_output)
logits = self.classifier(pooled_output) # (batch_size, num_labels)
# Calculate loss if labels provided
loss = None
if labels is not None:
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(logits, labels)
return {
"loss": loss,
"logits": logits,
}
class DocumentClassifierInference:
"""
Inference wrapper for document classification
Handles tokenization and prediction
"""
def __init__(
self,
model_path: str,
model_name: str = "nreimers/MiniLM-L6-H384-uncased",
device: str = None,
use_fp16: bool = False
):
"""
Args:
model_path: Path to fine-tuned model weights
model_name: Base model name (for tokenizer)
device: Device to run on ('cpu', 'cuda', or None for auto)
use_fp16: Use FP16 precision (faster on GPU)
"""
self.device = device if device else ('cuda' if torch.cuda.is_available() else 'cpu')
logger.info(f"Loading model on device: {self.device}")
# Load tokenizer
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
# Load model
self.model = DocumentClassifier(model_name=model_name)
# Load fine-tuned weights
try:
state_dict = torch.load(model_path, map_location=self.device)
self.model.load_state_dict(state_dict)
logger.info(f"Loaded fine-tuned weights from {model_path}")
except Exception as e:
logger.warning(f"Could not load weights from {model_path}: {str(e)}")
logger.warning("Using pre-trained weights (not fine-tuned)")
self.model.to(self.device)
self.model.eval()
# Apply FP16 if requested
if use_fp16 and self.device == 'cuda':
self.model.half()
logger.info("Using FP16 precision")
# Label mapping
self.id2label = {
0: 'INVOICE',
1: 'RECEIPT',
2: 'FORM',
3: 'OTHER'
}
self.label2id = {v: k for k, v in self.id2label.items()}
@torch.no_grad()
def predict(
self,
text: str,
max_length: int = 512,
return_probabilities: bool = True
) -> Dict:
"""
Predict document class from text
Args:
text: Input text (OCR extracted)
max_length: Maximum sequence length
return_probabilities: Return class probabilities
Returns:
Dictionary with predicted class and probabilities
"""
# Tokenize
inputs = self.tokenizer(
text,
max_length=max_length,
padding='max_length',
truncation=True,
return_tensors='pt'
)
# Move to device
inputs = {k: v.to(self.device) for k, v in inputs.items()}
# Forward pass
outputs = self.model(**inputs)
logits = outputs['logits']
# Get predictions
probabilities = torch.softmax(logits, dim=-1)
predicted_class_id = torch.argmax(probabilities, dim=-1).item()
predicted_class = self.id2label[predicted_class_id]
confidence = probabilities[0, predicted_class_id].item()
result = {
'predicted_class': predicted_class,
'confidence': confidence,
}
if return_probabilities:
result['probabilities'] = {
self.id2label[i]: probabilities[0, i].item()
for i in range(len(self.id2label))
}
return result
@torch.no_grad()
def predict_batch(
self,
texts: List[str],
max_length: int = 512,
batch_size: int = 8
) -> List[Dict]:
"""
Predict document classes for multiple texts
Args:
texts: List of input texts
max_length: Maximum sequence length
batch_size: Batch size for processing
Returns:
List of prediction dictionaries
"""
all_results = []
for i in range(0, len(texts), batch_size):
batch_texts = texts[i:i+batch_size]
# Tokenize batch
inputs = self.tokenizer(
batch_texts,
max_length=max_length,
padding='max_length',
truncation=True,
return_tensors='pt'
)
# Move to device
inputs = {k: v.to(self.device) for k, v in inputs.items()}
# Forward pass
outputs = self.model(**inputs)
logits = outputs['logits']
# Get predictions
probabilities = torch.softmax(logits, dim=-1)
predicted_classes = torch.argmax(probabilities, dim=-1)
# Parse results
for j in range(len(batch_texts)):
pred_id = predicted_classes[j].item()
result = {
'predicted_class': self.id2label[pred_id],
'confidence': probabilities[j, pred_id].item(),
'probabilities': {
self.id2label[k]: probabilities[j, k].item()
for k in range(len(self.id2label))
}
}
all_results.append(result)
return all_results
def export_to_onnx(
model_path: str,
output_path: str,
model_name: str = "nreimers/MiniLM-L6-H384-uncased"
):
"""
Export model to ONNX format for faster inference
Args:
model_path: Path to PyTorch model weights
output_path: Path to save ONNX model
model_name: Base model name
"""
import torch.onnx
logger.info("Exporting model to ONNX...")
# Load model
model = DocumentClassifier(model_name=model_name)
state_dict = torch.load(model_path, map_location='cpu')
model.load_state_dict(state_dict)
model.eval()
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
# Create dummy input
dummy_text = "This is a sample invoice text for ONNX export"
dummy_input = tokenizer(
dummy_text,
max_length=512,
padding='max_length',
truncation=True,
return_tensors='pt'
)
# Export to ONNX
torch.onnx.export(
model,
(dummy_input['input_ids'], dummy_input['attention_mask']),
output_path,
input_names=['input_ids', 'attention_mask'],
output_names=['logits'],
dynamic_axes={
'input_ids': {0: 'batch_size'},
'attention_mask': {0: 'batch_size'},
'logits': {0: 'batch_size'}
},
opset_version=14,
)
logger.info(f"Model exported to {output_path}")
if __name__ == "__main__":
# Example usage
# Create a sample model (not trained)
model = DocumentClassifier()
print(f"\nModel architecture:")
print(model)
print(f"\nTotal parameters: {sum(p.numel() for p in model.parameters()) / 1e6:.2f}M")
# Test forward pass
batch_size = 2
seq_len = 128
dummy_input_ids = torch.randint(0, 1000, (batch_size, seq_len))
dummy_attention_mask = torch.ones(batch_size, seq_len)
dummy_labels = torch.tensor([0, 1])
outputs = model(dummy_input_ids, dummy_attention_mask, dummy_labels)
print(f"\nOutput logits shape: {outputs['logits'].shape}")
print(f"Loss: {outputs['loss'].item():.4f}")
|