Update modeling_sybil_wrapper.py
Browse files- modeling_sybil_wrapper.py +253 -95
modeling_sybil_wrapper.py
CHANGED
@@ -1,140 +1,298 @@
|
|
1 |
"""
|
2 |
-
|
3 |
-
This
|
4 |
"""
|
5 |
|
6 |
import os
|
7 |
-
import sys
|
8 |
import json
|
|
|
9 |
import torch
|
10 |
-
import
|
11 |
-
from typing import
|
12 |
-
from transformers import PreTrainedModel
|
13 |
from dataclasses import dataclass
|
14 |
from transformers.modeling_outputs import BaseModelOutput
|
|
|
15 |
|
16 |
-
# Add
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
|
21 |
try:
|
22 |
from .configuration_sybil import SybilConfig
|
|
|
|
|
23 |
except ImportError:
|
24 |
from configuration_sybil import SybilConfig
|
|
|
|
|
25 |
|
26 |
|
27 |
@dataclass
|
28 |
class SybilOutput(BaseModelOutput):
|
29 |
"""
|
30 |
-
Output class for Sybil model.
|
|
|
|
|
|
|
|
|
31 |
"""
|
32 |
risk_scores: torch.FloatTensor = None
|
33 |
attentions: Optional[Dict] = None
|
34 |
|
35 |
|
36 |
-
class SybilHFWrapper
|
37 |
"""
|
38 |
-
Hugging Face wrapper
|
39 |
-
|
40 |
"""
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
os.
|
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 |
Args:
|
87 |
-
|
88 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
return_attentions: Whether to return attention maps
|
90 |
|
91 |
Returns:
|
92 |
-
SybilOutput with risk scores and optional
|
93 |
"""
|
|
|
|
|
94 |
|
95 |
-
|
96 |
-
|
|
|
97 |
|
98 |
-
|
99 |
-
|
|
|
|
|
|
|
|
|
100 |
|
101 |
-
|
102 |
-
|
|
|
|
|
|
|
103 |
|
104 |
-
|
105 |
-
risk_scores = torch.tensor(prediction.scores[0])
|
106 |
|
107 |
-
|
108 |
-
|
109 |
-
attentions=prediction.attentions[0] if return_attentions else None
|
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 |
-
with open(os.path.join(save_directory, "model_info.json"), "w") as f:
|
140 |
-
json.dump(info, f, indent=2)
|
|
|
1 |
"""
|
2 |
+
Self-contained Hugging Face wrapper for Sybil lung cancer risk prediction model.
|
3 |
+
This version works directly from HF without requiring external Sybil package.
|
4 |
"""
|
5 |
|
6 |
import os
|
|
|
7 |
import json
|
8 |
+
import sys
|
9 |
import torch
|
10 |
+
import numpy as np
|
11 |
+
from typing import List, Dict, Optional
|
|
|
12 |
from dataclasses import dataclass
|
13 |
from transformers.modeling_outputs import BaseModelOutput
|
14 |
+
from safetensors.torch import load_file
|
15 |
|
16 |
+
# Add model path to sys.path for imports
|
17 |
+
current_dir = os.path.dirname(os.path.abspath(__file__))
|
18 |
+
if current_dir not in sys.path:
|
19 |
+
sys.path.insert(0, current_dir)
|
20 |
|
21 |
try:
|
22 |
from .configuration_sybil import SybilConfig
|
23 |
+
from .modeling_sybil import SybilForRiskPrediction
|
24 |
+
from .image_processing_sybil import SybilImageProcessor
|
25 |
except ImportError:
|
26 |
from configuration_sybil import SybilConfig
|
27 |
+
from modeling_sybil import SybilForRiskPrediction
|
28 |
+
from image_processing_sybil import SybilImageProcessor
|
29 |
|
30 |
|
31 |
@dataclass
|
32 |
class SybilOutput(BaseModelOutput):
|
33 |
"""
|
34 |
+
Output class for Sybil model predictions.
|
35 |
+
|
36 |
+
Args:
|
37 |
+
risk_scores: Risk scores for each year (1-6 years by default)
|
38 |
+
attentions: Optional attention maps if requested
|
39 |
"""
|
40 |
risk_scores: torch.FloatTensor = None
|
41 |
attentions: Optional[Dict] = None
|
42 |
|
43 |
|
44 |
+
class SybilHFWrapper:
|
45 |
"""
|
46 |
+
Hugging Face wrapper for Sybil ensemble model.
|
47 |
+
Provides a simple interface for lung cancer risk prediction from CT scans.
|
48 |
"""
|
49 |
+
|
50 |
+
def __init__(self, config: SybilConfig = None):
|
51 |
+
"""
|
52 |
+
Initialize the Sybil model ensemble.
|
53 |
+
|
54 |
+
Args:
|
55 |
+
config: Model configuration (will use default if not provided)
|
56 |
+
"""
|
57 |
+
self.config = config if config is not None else SybilConfig()
|
58 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
59 |
+
|
60 |
+
# Get the directory where this file is located
|
61 |
+
self.model_dir = os.path.dirname(os.path.abspath(__file__))
|
62 |
+
|
63 |
+
# Initialize image processor
|
64 |
+
self.image_processor = SybilImageProcessor()
|
65 |
+
|
66 |
+
# Load calibrator
|
67 |
+
self.calibrator = self._load_calibrator()
|
68 |
+
|
69 |
+
# Load ensemble models
|
70 |
+
self.models = self._load_ensemble_models()
|
71 |
+
|
72 |
+
def _load_calibrator(self) -> Dict:
|
73 |
+
"""Load ensemble calibrator data"""
|
74 |
+
calibrator_path = os.path.join(self.model_dir, "checkpoints", "sybil_ensemble_simple_calibrator.json")
|
75 |
+
|
76 |
+
if os.path.exists(calibrator_path):
|
77 |
+
with open(calibrator_path, 'r') as f:
|
78 |
+
return json.load(f)
|
79 |
+
else:
|
80 |
+
# Try alternative location
|
81 |
+
calibrator_path = os.path.join(self.model_dir, "calibrator_data.json")
|
82 |
+
if os.path.exists(calibrator_path):
|
83 |
+
with open(calibrator_path, 'r') as f:
|
84 |
+
return json.load(f)
|
85 |
+
return {}
|
86 |
+
|
87 |
+
def _load_ensemble_models(self) -> List[torch.nn.Module]:
|
88 |
+
"""Load all models in the ensemble from safetensors files"""
|
89 |
+
models = []
|
90 |
+
|
91 |
+
# Load each model in the ensemble (Sybil uses 5 models)
|
92 |
+
for i in range(1, 6):
|
93 |
+
model_subdir = os.path.join(self.model_dir, f"sybil_{i}")
|
94 |
+
weights_path = os.path.join(model_subdir, "model.safetensors")
|
95 |
+
|
96 |
+
if os.path.exists(weights_path):
|
97 |
+
# Create model instance
|
98 |
+
model = SybilForRiskPrediction(self.config)
|
99 |
+
|
100 |
+
# Load weights from safetensors
|
101 |
+
try:
|
102 |
+
state_dict = load_file(weights_path)
|
103 |
+
model.load_state_dict(state_dict, strict=False)
|
104 |
+
except Exception as e:
|
105 |
+
print(f"Warning: Could not load weights for sybil_{i}: {e}")
|
106 |
+
continue
|
107 |
+
|
108 |
+
model.to(self.device)
|
109 |
+
model.eval()
|
110 |
+
models.append(model)
|
111 |
+
else:
|
112 |
+
# Try loading from checkpoints directory
|
113 |
+
checkpoint_path = os.path.join(self.model_dir, "checkpoints", f"sybil_{i}.ckpt")
|
114 |
+
if os.path.exists(checkpoint_path):
|
115 |
+
model = SybilForRiskPrediction(self.config)
|
116 |
+
checkpoint = torch.load(checkpoint_path, map_location='cpu')
|
117 |
+
|
118 |
+
# Extract state dict
|
119 |
+
if 'state_dict' in checkpoint:
|
120 |
+
state_dict = checkpoint['state_dict']
|
121 |
+
else:
|
122 |
+
state_dict = checkpoint
|
123 |
+
|
124 |
+
# Remove 'model.' prefix if present
|
125 |
+
cleaned_state_dict = {}
|
126 |
+
for k, v in state_dict.items():
|
127 |
+
if k.startswith('model.'):
|
128 |
+
cleaned_state_dict[k[6:]] = v
|
129 |
+
else:
|
130 |
+
cleaned_state_dict[k] = v
|
131 |
+
|
132 |
+
model.load_state_dict(cleaned_state_dict, strict=False)
|
133 |
+
model.to(self.device)
|
134 |
+
model.eval()
|
135 |
+
models.append(model)
|
136 |
+
|
137 |
+
if not models:
|
138 |
+
raise ValueError("No models could be loaded from the ensemble. Please ensure model files are present.")
|
139 |
+
|
140 |
+
print(f"Loaded {len(models)} models in ensemble")
|
141 |
+
return models
|
142 |
+
|
143 |
+
def _apply_calibration(self, scores: np.ndarray) -> np.ndarray:
|
144 |
+
"""
|
145 |
+
Apply calibration to raw model outputs.
|
146 |
+
|
147 |
+
Args:
|
148 |
+
scores: Raw risk scores from the model
|
149 |
+
|
150 |
+
Returns:
|
151 |
+
Calibrated risk scores
|
152 |
+
"""
|
153 |
+
if not self.calibrator:
|
154 |
+
return scores
|
155 |
+
|
156 |
+
calibrated = np.zeros_like(scores)
|
157 |
+
|
158 |
+
for year in range(scores.shape[1]):
|
159 |
+
year_key = f"Year{year + 1}"
|
160 |
+
if year_key in self.calibrator:
|
161 |
+
cal_data = self.calibrator[year_key]
|
162 |
+
if isinstance(cal_data, list) and len(cal_data) > 0:
|
163 |
+
cal_data = cal_data[0]
|
164 |
+
|
165 |
+
# Apply linear calibration if available
|
166 |
+
if isinstance(cal_data, dict) and "coef" in cal_data and "intercept" in cal_data:
|
167 |
+
coef = cal_data["coef"][0][0] if isinstance(cal_data["coef"], list) else cal_data["coef"]
|
168 |
+
intercept = cal_data["intercept"][0] if isinstance(cal_data["intercept"], list) else cal_data["intercept"]
|
169 |
+
|
170 |
+
# Apply calibration
|
171 |
+
calibrated[:, year] = scores[:, year] * coef + intercept
|
172 |
+
calibrated[:, year] = 1 / (1 + np.exp(-calibrated[:, year])) # Sigmoid
|
173 |
+
else:
|
174 |
+
calibrated[:, year] = scores[:, year]
|
175 |
+
else:
|
176 |
+
calibrated[:, year] = scores[:, year]
|
177 |
+
|
178 |
+
return calibrated
|
179 |
+
|
180 |
+
def preprocess_dicom(self, dicom_paths: List[str]) -> torch.Tensor:
|
181 |
"""
|
182 |
+
Preprocess DICOM files for model input.
|
183 |
|
184 |
Args:
|
185 |
+
dicom_paths: List of paths to DICOM files
|
186 |
+
|
187 |
+
Returns:
|
188 |
+
Preprocessed tensor ready for model input
|
189 |
+
"""
|
190 |
+
# Use the image processor to handle DICOM files
|
191 |
+
result = self.image_processor(dicom_paths, file_type="dicom", return_tensors="pt")
|
192 |
+
pixel_values = result["pixel_values"]
|
193 |
+
|
194 |
+
# Ensure we have 5D tensor (B, C, D, H, W)
|
195 |
+
if pixel_values.ndim == 4:
|
196 |
+
pixel_values = pixel_values.unsqueeze(0) # Add batch dimension
|
197 |
+
|
198 |
+
return pixel_values.to(self.device)
|
199 |
+
|
200 |
+
def predict(self, dicom_paths: List[str], return_attentions: bool = False) -> SybilOutput:
|
201 |
+
"""
|
202 |
+
Run prediction on a CT scan series.
|
203 |
+
|
204 |
+
Args:
|
205 |
+
dicom_paths: List of paths to DICOM files for a single CT series
|
206 |
return_attentions: Whether to return attention maps
|
207 |
|
208 |
Returns:
|
209 |
+
SybilOutput with risk scores and optional attention maps
|
210 |
"""
|
211 |
+
# Preprocess the DICOM files
|
212 |
+
pixel_values = self.preprocess_dicom(dicom_paths)
|
213 |
|
214 |
+
# Run inference with ensemble
|
215 |
+
all_predictions = []
|
216 |
+
all_attentions = []
|
217 |
|
218 |
+
with torch.no_grad():
|
219 |
+
for model in self.models:
|
220 |
+
output = model(
|
221 |
+
pixel_values=pixel_values,
|
222 |
+
return_attentions=return_attentions
|
223 |
+
)
|
224 |
|
225 |
+
# Extract risk scores
|
226 |
+
if hasattr(output, 'risk_scores'):
|
227 |
+
predictions = output.risk_scores
|
228 |
+
else:
|
229 |
+
predictions = output[0] if isinstance(output, tuple) else output
|
230 |
|
231 |
+
all_predictions.append(predictions.cpu().numpy())
|
|
|
232 |
|
233 |
+
if return_attentions and hasattr(output, 'image_attention'):
|
234 |
+
all_attentions.append(output.image_attention)
|
|
|
|
|
235 |
|
236 |
+
# Average ensemble predictions
|
237 |
+
ensemble_pred = np.mean(all_predictions, axis=0)
|
238 |
+
|
239 |
+
# Apply calibration
|
240 |
+
calibrated_pred = self._apply_calibration(ensemble_pred)
|
241 |
+
|
242 |
+
# Convert back to torch tensor
|
243 |
+
risk_scores = torch.from_numpy(calibrated_pred).float()
|
244 |
+
|
245 |
+
# Average attentions if requested
|
246 |
+
attentions = None
|
247 |
+
if return_attentions and all_attentions:
|
248 |
+
attentions = {"image_attention": torch.stack(all_attentions).mean(dim=0)}
|
249 |
+
|
250 |
+
return SybilOutput(risk_scores=risk_scores, attentions=attentions)
|
251 |
+
|
252 |
+
def __call__(self, dicom_paths: List[str] = None, dicom_series: List[List[str]] = None, **kwargs) -> SybilOutput:
|
253 |
"""
|
254 |
+
Convenience method for prediction.
|
255 |
+
|
256 |
+
Args:
|
257 |
+
dicom_paths: List of DICOM file paths for a single series
|
258 |
+
dicom_series: List of lists of DICOM paths for batch processing
|
259 |
+
**kwargs: Additional arguments passed to predict()
|
260 |
+
|
261 |
+
Returns:
|
262 |
+
SybilOutput with predictions
|
263 |
"""
|
264 |
+
if dicom_series is not None:
|
265 |
+
# Batch processing
|
266 |
+
all_outputs = []
|
267 |
+
for paths in dicom_series:
|
268 |
+
output = self.predict(paths, **kwargs)
|
269 |
+
all_outputs.append(output.risk_scores)
|
270 |
|
271 |
+
risk_scores = torch.stack(all_outputs)
|
272 |
+
return SybilOutput(risk_scores=risk_scores)
|
273 |
+
elif dicom_paths is not None:
|
274 |
+
return self.predict(dicom_paths, **kwargs)
|
275 |
+
else:
|
276 |
+
raise ValueError("Either dicom_paths or dicom_series must be provided")
|
277 |
|
278 |
+
@classmethod
|
279 |
+
def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
|
280 |
"""
|
281 |
+
Load model from Hugging Face hub or local path.
|
282 |
+
|
283 |
+
Args:
|
284 |
+
pretrained_model_name_or_path: HF model ID or local path
|
285 |
+
**kwargs: Additional configuration arguments
|
286 |
+
|
287 |
+
Returns:
|
288 |
+
SybilHFWrapper instance
|
289 |
"""
|
290 |
+
# Load configuration
|
291 |
+
config = kwargs.pop("config", None)
|
292 |
+
if config is None:
|
293 |
+
try:
|
294 |
+
config = SybilConfig.from_pretrained(pretrained_model_name_or_path)
|
295 |
+
except:
|
296 |
+
config = SybilConfig()
|
297 |
+
|
298 |
+
return cls(config=config)
|
|
|
|
|
|