Spaces:
Running
Running
File size: 14,128 Bytes
cd846d7 2a47411 cd846d7 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 | """
DeepECG Inference Module for HeartWatch AI
===========================================
This module provides CPU-optimized inference for 4 EfficientNetV2 models:
- 77-class ECG diagnosis
- LVEF <= 40% prediction
- LVEF < 50% prediction
- 5-year AFib risk prediction
The preprocessing exactly replicates DeepECG's pipeline:
1. Load signal as (samples, leads) = (2500, 12)
2. Transpose to (leads, samples) = (12, 2500)
3. Apply MHI factor scaling: signal *= (1/0.0048)
4. Apply sigmoid to model logits
Models are downloaded from HuggingFace Hub using HF_TOKEN from environment.
"""
import os
import json
import time
import logging
from typing import Dict, Optional, Any, Union
from pathlib import Path
import numpy as np
import torch
from huggingface_hub import snapshot_download
# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# CPU optimizations for HuggingFace Spaces (no GPU)
torch.set_num_threads(2)
torch.set_flush_denormal(True)
class DeepECGInference:
"""
CPU-optimized inference engine for DeepECG EfficientNetV2 models.
Loads 4 models from HuggingFace Hub:
- heartwise/EfficientNetV2_77_Classes: 77-class ECG diagnosis
- heartwise/EfficientNetV2_LVEF_40: LVEF <= 40% prediction
- heartwise/EfficientNetV2_LVEF_50: LVEF < 50% prediction
- heartwise/EfficientNetV2_AFIB_5y: 5-year AFib risk prediction
Attributes:
device: Always CPU for HF Spaces
models: Dict containing loaded TorchScript models
class_names: List of 77 ECG diagnosis class names
mhi_factor: Scaling factor for signal preprocessing (1/0.0048)
"""
# Model repository mappings
MODEL_REPOS = {
"diagnosis_77": "heartwise/EfficientNetV2_77_Classes",
"lvef_40": "heartwise/EfficientNetV2_LVEF_equal_under_40",
"lvef_50": "heartwise/EfficientNetV2_LVEF_under_50",
"afib_5y": "heartwise/EfficientNetV2_AFIB_5y",
}
# Expected input specifications
EXPECTED_LEADS = 12
EXPECTED_SAMPLES = 2500 # 10 seconds at 250 Hz
SAMPLING_RATE = 250 # Hz
# Preprocessing constants from DeepECG
MHI_FACTOR = 1 / 0.0048 # ~208.33
def __init__(self, cache_dir: Optional[str] = None):
"""
Initialize the inference engine.
Args:
cache_dir: Directory to cache downloaded models.
Defaults to ./weights
"""
self.device = torch.device("cpu")
self.cache_dir = cache_dir or os.path.join(os.getcwd(), "weights")
self.models: Dict[str, torch.jit.ScriptModule] = {}
self.class_names: list = []
self._load_class_names()
def _load_class_names(self) -> None:
"""Load the 77 ECG class names from class_names.json."""
class_names_path = os.path.join(
os.path.dirname(os.path.abspath(__file__)),
"class_names.json"
)
try:
with open(class_names_path, "r") as f:
self.class_names = json.load(f)
logger.info(f"Loaded {len(self.class_names)} class names")
except FileNotFoundError:
logger.warning(f"class_names.json not found at {class_names_path}")
self.class_names = []
def _get_hf_token(self) -> Optional[str]:
"""Get HuggingFace token from environment variable."""
token = os.environ.get("HF_TOKEN")
if not token:
logger.warning("HF_TOKEN environment variable not set")
return token
def _download_model(self, repo_id: str, model_name: str) -> str:
"""
Download model from HuggingFace Hub.
Args:
repo_id: HuggingFace repository ID
model_name: Local name for the model
Returns:
Path to the downloaded model directory
"""
local_dir = os.path.join(self.cache_dir, model_name)
if os.path.exists(local_dir):
logger.info(f"Model {model_name} already cached at {local_dir}")
return local_dir
logger.info(f"Downloading {repo_id} to {local_dir}")
os.makedirs(local_dir, exist_ok=True)
hf_token = self._get_hf_token()
local_dir = snapshot_download(
repo_id=repo_id,
local_dir=local_dir,
repo_type="model",
token=hf_token
)
logger.info(f"Downloaded {repo_id} to {local_dir}")
return local_dir
def _load_model_from_dir(self, model_dir: str) -> torch.jit.ScriptModule:
"""
Load a TorchScript model from a directory.
Args:
model_dir: Directory containing the .pt file
Returns:
Loaded TorchScript model
Raises:
ValueError: If no .pt file is found in the directory
"""
pt_file = next(
(f for f in os.listdir(model_dir) if f.endswith('.pt')),
None
)
if not pt_file:
raise ValueError(f"No .pt file found in {model_dir}")
model_path = os.path.join(model_dir, pt_file)
model = torch.jit.load(model_path, map_location=self.device)
model.eval()
return model
def load_models(self) -> None:
"""
Download and load all 4 models from HuggingFace Hub.
Uses HF_TOKEN from os.environ for authentication.
Models are loaded in eval mode on CPU.
"""
logger.info("Loading DeepECG models...")
for model_key, repo_id in self.MODEL_REPOS.items():
try:
model_dir = self._download_model(repo_id, model_key)
self.models[model_key] = self._load_model_from_dir(model_dir)
logger.info(f"Loaded model: {model_key} from {repo_id}")
except Exception as e:
logger.error(f"Failed to load {model_key}: {e}")
raise
logger.info(f"Successfully loaded {len(self.models)} models")
def preprocess_ecg(
self,
ecg_signal: Union[np.ndarray, torch.Tensor]
) -> torch.Tensor:
"""
Preprocess ECG signal to match DeepECG's exact preprocessing.
The preprocessing pipeline:
1. Ensure signal is numpy array with correct shape
2. Handle shape: expect (samples, leads) = (2500, 12) or (12, 2500)
3. Transpose to (leads, samples) = (12, 2500) if needed
4. Convert to float32 tensor
5. Add batch dimension: (1, 12, 2500)
6. Apply MHI factor scaling: signal *= (1/0.0048)
Args:
ecg_signal: Raw ECG signal, shape (samples, leads) or (leads, samples)
Expected: 12 leads, 2500 samples (10s at 250Hz)
Returns:
Preprocessed tensor ready for model inference, shape (1, 12, 2500)
Raises:
ValueError: If signal shape is invalid
"""
# Convert to numpy if tensor
if isinstance(ecg_signal, torch.Tensor):
ecg_signal = ecg_signal.numpy()
# Ensure float32
ecg_signal = ecg_signal.astype(np.float32)
# Handle shape - expect (samples, leads) = (2500, 12) or (12, 2500)
if ecg_signal.ndim != 2:
raise ValueError(
f"Expected 2D signal, got shape {ecg_signal.shape}"
)
# Determine orientation and transpose if needed
# If shape is (samples, leads) = (2500, 12), transpose to (12, 2500)
# If shape is (12, 2500), it's already correct
if ecg_signal.shape[0] == self.EXPECTED_SAMPLES and ecg_signal.shape[1] == self.EXPECTED_LEADS:
# Shape is (2500, 12) -> transpose to (12, 2500)
ecg_signal = ecg_signal.T
elif ecg_signal.shape[0] == self.EXPECTED_LEADS and ecg_signal.shape[1] == self.EXPECTED_SAMPLES:
# Shape is already (12, 2500)
pass
else:
# Try to infer orientation
if ecg_signal.shape[1] == self.EXPECTED_LEADS:
ecg_signal = ecg_signal.T
elif ecg_signal.shape[0] != self.EXPECTED_LEADS:
raise ValueError(
f"Invalid signal shape {ecg_signal.shape}. "
f"Expected (2500, 12) or (12, 2500)"
)
# Verify final shape
if ecg_signal.shape[0] != self.EXPECTED_LEADS:
raise ValueError(
f"Signal must have {self.EXPECTED_LEADS} leads, "
f"got {ecg_signal.shape[0]}"
)
# Convert to tensor and add batch dimension
signal_tensor = torch.from_numpy(ecg_signal).float()
signal_tensor = signal_tensor.unsqueeze(0) # (1, 12, samples)
# Move to device (CPU)
signal_tensor = signal_tensor.to(self.device)
# Apply MHI factor scaling (this is done in model __call__ in DeepECG)
signal_tensor = signal_tensor * self.MHI_FACTOR
return signal_tensor
def predict(
self,
ecg_signal: Union[np.ndarray, torch.Tensor]
) -> Dict[str, Any]:
"""
Run inference on an ECG signal using all 4 models.
Args:
ecg_signal: Raw ECG signal, shape (samples, leads) or (leads, samples)
Expected: 12 leads, 2500 samples (10s at 250Hz)
Returns:
Dictionary containing:
- diagnosis_77: Dict with 'probabilities' (77 floats) and 'class_names'
- lvef_40: Probability of LVEF <= 40%
- lvef_50: Probability of LVEF < 50%
- afib_5y: Probability of AFib within 5 years
- inference_time_ms: Total inference time in milliseconds
"""
if not self.models:
raise RuntimeError("Models not loaded. Call load_models() first.")
start_time = time.time()
# Preprocess the signal
signal_tensor = self.preprocess_ecg(ecg_signal)
results = {}
with torch.no_grad():
# 77-class diagnosis
if "diagnosis_77" in self.models:
logits = self.models["diagnosis_77"](signal_tensor)
probs = torch.sigmoid(logits)
probs_list = probs.squeeze().cpu().numpy().tolist()
results["diagnosis_77"] = {
"probabilities": probs_list,
"class_names": self.class_names if self.class_names else None,
}
# LVEF <= 40%
if "lvef_40" in self.models:
logits = self.models["lvef_40"](signal_tensor)
prob = torch.sigmoid(logits)
results["lvef_40"] = float(prob.squeeze().cpu().numpy())
# LVEF < 50%
if "lvef_50" in self.models:
logits = self.models["lvef_50"](signal_tensor)
prob = torch.sigmoid(logits)
results["lvef_50"] = float(prob.squeeze().cpu().numpy())
# 5-year AFib risk
if "afib_5y" in self.models:
logits = self.models["afib_5y"](signal_tensor)
prob = torch.sigmoid(logits)
results["afib_5y"] = float(prob.squeeze().cpu().numpy())
end_time = time.time()
results["inference_time_ms"] = (end_time - start_time) * 1000
return results
def predict_diagnosis_top_k(
self,
ecg_signal: Union[np.ndarray, torch.Tensor],
k: int = 5
) -> Dict[str, Any]:
"""
Get top-k diagnoses from the 77-class model.
Args:
ecg_signal: Raw ECG signal
k: Number of top predictions to return
Returns:
Dictionary with top-k predictions sorted by probability
"""
results = self.predict(ecg_signal)
if "diagnosis_77" not in results:
raise RuntimeError("77-class diagnosis model not loaded")
probs = results["diagnosis_77"]["probabilities"]
class_names = results["diagnosis_77"]["class_names"] or [f"Class_{i}" for i in range(77)]
# Get top-k indices
top_k_indices = np.argsort(probs)[::-1][:k]
top_k_predictions = [
{
"class_name": class_names[idx],
"probability": probs[idx],
"class_index": int(idx)
}
for idx in top_k_indices
]
return {
"top_k_predictions": top_k_predictions,
"inference_time_ms": results["inference_time_ms"]
}
def get_inference_engine(cache_dir: Optional[str] = None) -> DeepECGInference:
"""
Factory function to create and initialize a DeepECGInference instance.
Args:
cache_dir: Optional directory to cache models
Returns:
Initialized DeepECGInference with models loaded
"""
engine = DeepECGInference(cache_dir=cache_dir)
engine.load_models()
return engine
if __name__ == "__main__":
# Example usage / testing
print("DeepECG Inference Module")
print("=" * 50)
# Create inference engine
engine = DeepECGInference()
# Load models (requires HF_TOKEN environment variable)
try:
engine.load_models()
print("Models loaded successfully!")
# Create dummy signal for testing
dummy_signal = np.random.randn(2500, 12).astype(np.float32)
# Run inference
results = engine.predict(dummy_signal)
print(f"\nInference time: {results['inference_time_ms']:.2f} ms")
print(f"LVEF <= 40%: {results['lvef_40']:.4f}")
print(f"LVEF < 50%: {results['lvef_50']:.4f}")
print(f"5-year AFib risk: {results['afib_5y']:.4f}")
print(f"77-class diagnosis: {len(results['diagnosis_77']['probabilities'])} classes")
# Get top-5 diagnoses
top_5 = engine.predict_diagnosis_top_k(dummy_signal, k=5)
print("\nTop 5 diagnoses:")
for pred in top_5["top_k_predictions"]:
print(f" {pred['class_name']}: {pred['probability']:.4f}")
except Exception as e:
print(f"Error: {e}")
print("\nMake sure HF_TOKEN environment variable is set:")
print(" export HF_TOKEN='your_huggingface_token'")
|