File size: 9,128 Bytes
c863b63 | 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 | import torch
import torch.nn as nn
from torch.nn.utils.fusion import fuse_conv_bn_eval
from typing import Optional, Dict, Any
import numpy as np
import time
from pathlib import Path
from ..config import settings
from .rawnet import RawNet, RawNetConfig
class QuantizedModelWrapper:
"""
Wrapper for INT8 quantized RawNet model.
Handles dynamic quantization, inference, and model loading/saving.
"""
def __init__(
self,
model: Optional[RawNet] = None,
config: Optional[RawNetConfig] = None,
quantized: bool = settings.quantized,
):
self.config = config or RawNetConfig()
self.quantized = quantized
self._model = model
self._quantized_model: Optional[nn.Module] = None
self._device = torch.device("cpu") # Force CPU for on-premise
self._warmup_done = False
@property
def model(self) -> nn.Module:
if self.quantized and self._quantized_model is not None:
return self._quantized_model
return self._model
def load_model(self, model_path: str = None, strict: bool = True) -> None:
"""Load model weights from file."""
path = model_path or settings.model_path
path = Path(path)
if not path.exists():
path.parent.mkdir(parents=True, exist_ok=True)
if self._model is None:
self._model = RawNet(self.config)
self._model.to(self._device)
self._model.eval()
torch.save(self._model.state_dict(), path)
else:
if self._model is None:
self._model = RawNet(self.config)
state_dict = torch.load(path, map_location=self._device, weights_only=True)
self._model.load_state_dict(state_dict, strict=strict)
self._model.to(self._device)
self._model.eval()
if self.quantized:
self.quantize()
def quantize(self) -> None:
"""Apply Conv-BN fusion and INT8 Linear-only quantization for edge latency."""
if self._model is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
print("Applying Conv-BN fusion + INT8 Linear quantization...")
start = time.time()
# 1. Pre-cache sinc filters to eliminate torch.sinc() at runtime
self._model.sinc_conv.precompute_kernel()
# 2. Fuse BatchNorm into Conv1D layers (eliminates 4 BN forward passes)
self._fuse_conv_bn()
# 3. Quantize Linear layers only (GRU dequantization overhead
# exceeds INT8 compute savings on 2 vCPUs)
self._quantized_model = torch.quantization.quantize_dynamic(
self._model,
qconfig_spec={nn.Linear},
dtype=torch.qint8,
inplace=False,
)
self._quantized_model.to(self._device)
self._quantized_model.eval()
elapsed = (time.time() - start) * 1000
print(f"Optimization completed in {elapsed:.1f}ms")
# Verify quantization
self._verify_quantization()
def _fuse_conv_bn(self) -> None:
"""Fold BatchNorm parameters into Conv1D biases for inference speedup."""
model = self._model
for block in getattr(model, 'res_blocks', []):
if hasattr(block, 'conv1') and hasattr(block, 'bn1'):
block.conv1 = fuse_conv_bn_eval(block.conv1, block.bn1)
block.bn1 = nn.Identity()
if hasattr(block, 'conv2') and hasattr(block, 'bn2'):
block.conv2 = fuse_conv_bn_eval(block.conv2, block.bn2)
block.bn2 = nn.Identity()
# Detach fused params so they are leaf tensors (required by deepcopy)
for conv in [block.conv1, block.conv2]:
if hasattr(conv, 'weight'):
conv.weight = nn.Parameter(conv.weight.detach().clone())
if hasattr(conv, 'bias') and conv.bias is not None:
conv.bias = nn.Parameter(conv.bias.detach().clone())
def _verify_quantization(self) -> None:
"""Verify that quantization was applied to target layers."""
quantized_layers = 0
total_layers = 0
for name, module in self._quantized_model.named_modules():
if isinstance(module, (nn.Linear, nn.GRU)):
total_layers += 1
if hasattr(module, 'weight') and module.weight.dtype == torch.qint8:
quantized_layers += 1
print(f"Quantized {quantized_layers}/{total_layers} target layers")
def warmup(self, num_runs: int = 3) -> float:
"""Run warmup inferences to stabilize timing."""
if self._model is None:
raise RuntimeError("Model not loaded")
model = self.model
dummy_input = torch.randn(1, 1, 24000, device=self._device)
times = []
with torch.no_grad():
for _ in range(num_runs):
start = time.perf_counter()
_ = model(dummy_input)
times.append((time.perf_counter() - start) * 1000)
self._warmup_done = True
avg_time = sum(times) / len(times)
print(f"Warmup complete. Avg inference: {avg_time:.2f}ms")
return avg_time
def predict(self, audio: np.ndarray) -> Dict[str, Any]:
"""
Run inference on preprocessed audio chunk.
audio: [T] or [1, T] float32 numpy array
Returns dict with logits, probabilities, predicted class
"""
if self._model is None:
raise RuntimeError("Model not loaded. Call load_model() first.")
if not self._warmup_done:
self.warmup()
model = self.model
# Prepare input tensor
if audio.ndim == 1:
audio = audio[np.newaxis, :] # [1, T]
if audio.ndim == 2:
audio = audio[np.newaxis, :, :] # [1, 1, T]
input_tensor = torch.from_numpy(audio).to(self._device, dtype=torch.float32)
with torch.no_grad():
start = time.perf_counter()
logits = model(input_tensor)
inference_ms = (time.perf_counter() - start) * 1000
probs = torch.softmax(logits, dim=-1)
pred_class = torch.argmax(probs, dim=-1).item()
pred_prob = probs[0, pred_class].item()
return {
"logits": logits[0].cpu().numpy(),
"probabilities": probs[0].cpu().numpy(),
"predicted_class": pred_class,
"predicted_label": settings.class_labels[pred_class],
"confidence": pred_prob,
"inference_ms": inference_ms,
"is_bonafide": pred_class == 0,
}
def predict_batch(self, audio_batch: np.ndarray) -> list[Dict[str, Any]]:
"""Batch inference for multiple chunks."""
if audio_batch.ndim == 2:
audio_batch = audio_batch[:, np.newaxis, :] # [B, 1, T]
input_tensor = torch.from_numpy(audio_batch).to(self._device, dtype=torch.float32)
model = self.model
with torch.no_grad():
logits = model(input_tensor)
probs = torch.softmax(logits, dim=-1)
pred_classes = torch.argmax(probs, dim=-1).cpu().numpy()
pred_probs = torch.max(probs, dim=-1).values.cpu().numpy()
results = []
for i in range(len(audio_batch)):
results.append({
"logits": logits[i].cpu().numpy(),
"probabilities": probs[i].cpu().numpy(),
"predicted_class": int(pred_classes[i]),
"predicted_label": settings.class_labels[pred_classes[i]],
"confidence": float(pred_probs[i]),
"is_bonafide": pred_classes[i] == 0,
})
return results
def save_quantized(self, path: str) -> None:
"""Save quantized model state dict."""
if self._quantized_model is None:
raise RuntimeError("No quantized model to save")
torch.save(self._quantized_model.state_dict(), path)
def get_model_info(self) -> Dict[str, Any]:
"""Get model metadata."""
model = self.model
total_params = sum(p.numel() for p in model.parameters())
quantized_params = 0
if self.quantized:
for p in model.parameters():
if p.dtype == torch.qint8:
quantized_params += p.numel()
return {
"architecture": "RawNet1D",
"quantized": self.quantized,
"total_parameters": total_params,
"quantized_parameters": quantized_params,
"num_classes": self.config.num_classes,
"class_labels": settings.class_labels,
"input_shape": [1, 1, 24000],
"device": str(self._device),
}
def create_model(
model_path: str = None,
quantized: bool = settings.quantized,
config: RawNetConfig = None,
) -> QuantizedModelWrapper:
"""Factory function to create and load model."""
wrapper = QuantizedModelWrapper(config=config, quantized=quantized)
wrapper.load_model(model_path)
return wrapper |