{{ message }}
-
Notifications
You must be signed in to change notification settings - Fork 216
Expand file tree
/
Copy pathexport_model.py
More file actions
385 lines (320 loc) · 12.8 KB
/
Copy pathexport_model.py
File metadata and controls
385 lines (320 loc) · 12.8 KB
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
"""
Export the trained ClimateVision model to ONNX for production serving.
Produces (per exported model ``<name>``):
- <dir>/<name>.onnx — standard ONNX graph
- <dir>/<name>_quantized.onnx — INT8 quantized (CPU-optimised)
- <dir>/<name>_export_info.json — metadata (opset, input shape, benchmark)
Usage:
python scripts/export_model.py \\
--checkpoint models/20240101_120000/best_model.pth
# Export every enabled analysis type to models/unet_<type>.onnx,
# reading each type's checkpoint and num_classes from config.yaml:
python scripts/export_model.py --all
# Override output path and input size:
python scripts/export_model.py \\
--checkpoint models/best_model.pth \\
--out models/production/model.onnx \\
--image-size 512
# Skip quantization (requires onnxruntime):
python scripts/export_model.py --checkpoint ... --no-quantize
"""
from __future__ import annotations
import argparse
import json
import logging
import sys
import time
from pathlib import Path
import torch
import torch.nn as nn
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s %(levelname)-8s %(message)s",
datefmt="%H:%M:%S",
)
logger = logging.getLogger(__name__)
PROJECT_ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(PROJECT_ROOT / "src"))
# ---------------------------------------------------------------------------
# Model loading
# ---------------------------------------------------------------------------
def load_model(ckpt_path: Path, n_classes: int = 2) -> tuple[nn.Module, dict]:
from climatevision.models.unet import get_model
ckpt = torch.load(ckpt_path, map_location="cpu")
cfg = ckpt.get("cfg", {})
arch = cfg.get("model", {}).get("architecture", "attention_unet")
state = ckpt.get("ema_state_dict") or ckpt.get("model_state_dict", ckpt)
# Infer in_channels from weight shape
in_ch = 4
for key, val in state.items():
if "inc" in key and "weight" in key and val.ndim == 4:
in_ch = val.shape[1]
break
model = get_model(arch, n_channels=in_ch, n_classes=n_classes)
model.load_state_dict(state, strict=False)
model.eval()
logger.info(
"Loaded %s (in_channels=%d, n_classes=%d) from epoch %d val_iou=%.4f",
arch,
in_ch,
n_classes,
ckpt.get("epoch", 0),
ckpt.get("val_iou", 0.0),
)
return model, cfg
# ---------------------------------------------------------------------------
# ONNX export
# ---------------------------------------------------------------------------
def export_onnx(
model: nn.Module,
onnx_path: Path,
image_size: int,
in_channels: int,
opset: int = 17,
) -> None:
dummy = torch.zeros(1, in_channels, image_size, image_size)
onnx_path.parent.mkdir(parents=True, exist_ok=True)
torch.onnx.export(
model,
dummy,
str(onnx_path),
export_params=True,
opset_version=opset,
do_constant_folding=True,
input_names=["image"],
output_names=["logits"],
dynamic_axes={
"image": {0: "batch", 2: "height", 3: "width"},
"logits": {0: "batch", 2: "height", 3: "width"},
},
)
size_mb = onnx_path.stat().st_size / 1e6
logger.info("ONNX model saved: %s (%.1f MB)", onnx_path, size_mb)
# ---------------------------------------------------------------------------
# ONNX validation
# ---------------------------------------------------------------------------
def validate_onnx(
onnx_path: Path, in_channels: int, image_size: int, n_classes: int = 2
) -> float:
"""
Run a forward pass with onnxruntime and return inference latency (ms).
Also verifies the output tensor signature is ``(N, n_classes, H, W)`` so a
shape regression (wrong channel count, missing dynamic axis) is caught at
export time rather than at serve time.
"""
try:
import onnxruntime as ort
import numpy as np
except ImportError:
logger.warning("onnxruntime not installed — skipping validation. "
"Run: pip install onnxruntime")
return -1.0
sess = ort.InferenceSession(
str(onnx_path),
providers=["CPUExecutionProvider"],
)
dummy = np.random.rand(1, in_channels, image_size, image_size).astype(np.float32)
# Verify output signature matches (N, n_classes, H, W)
out = sess.run(None, {"image": dummy})[0]
expected = (1, n_classes, image_size, image_size)
if tuple(out.shape) != expected:
raise ValueError(
f"ONNX output shape {tuple(out.shape)} != expected {expected} "
f"for {onnx_path}"
)
# Warm-up
for _ in range(3):
sess.run(None, {"image": dummy})
# Benchmark
N = 20
t0 = time.perf_counter()
for _ in range(N):
sess.run(None, {"image": dummy})
latency_ms = (time.perf_counter() - t0) / N * 1000
logger.info("ONNX validation OK | avg latency: %.1f ms (batch=1, %dx%d)",
latency_ms, image_size, image_size)
return latency_ms
# ---------------------------------------------------------------------------
# INT8 quantization
# ---------------------------------------------------------------------------
def quantize_onnx(onnx_path: Path, out_path: Path) -> None:
try:
from onnxruntime.quantization import quantize_dynamic, QuantType
except ImportError:
logger.warning("onnxruntime quantization not available — skipping. "
"Run: pip install onnxruntime")
return
quantize_dynamic(
str(onnx_path),
str(out_path),
weight_type=QuantType.QInt8,
)
size_mb = out_path.stat().st_size / 1e6
logger.info("INT8 quantized model: %s (%.1f MB)", out_path, size_mb)
# ---------------------------------------------------------------------------
# PyTorch benchmark helper
# ---------------------------------------------------------------------------
def benchmark_pytorch(model: nn.Module, in_channels: int, image_size: int) -> float:
device = torch.device("cpu")
dummy = torch.zeros(1, in_channels, image_size, image_size, device=device)
with torch.no_grad():
for _ in range(3):
model(dummy)
N = 20
t0 = time.perf_counter()
for _ in range(N):
model(dummy)
return (time.perf_counter() - t0) / N * 1000
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
def export_checkpoint(
ckpt_path: Path,
onnx_path: Path,
image_size: int,
opset: int,
no_quantize: bool,
n_classes: int = 2,
) -> dict:
"""Export a single ``.pth`` checkpoint to ONNX and return export metadata."""
model, cfg = load_model(ckpt_path, n_classes=n_classes)
in_channels = getattr(model, "n_channels", cfg.get("model", {}).get("in_channels", 4))
quantized_path = onnx_path.parent / f"{onnx_path.stem}_quantized.onnx"
# PyTorch baseline latency
pt_ms = benchmark_pytorch(model, in_channels, image_size)
logger.info("PyTorch (CPU) baseline: %.1f ms", pt_ms)
export_onnx(
model=model,
onnx_path=onnx_path,
image_size=image_size,
in_channels=in_channels,
opset=opset,
)
onnx_ms = validate_onnx(onnx_path, in_channels, image_size, n_classes)
q_ms = -1.0
if not no_quantize:
quantize_onnx(onnx_path, quantized_path)
if quantized_path.exists():
q_ms = validate_onnx(quantized_path, in_channels, image_size, n_classes)
ckpt = torch.load(ckpt_path, map_location="cpu")
info = {
"checkpoint": str(ckpt_path),
"architecture": cfg.get("model", {}).get("architecture", "unknown"),
"in_channels": in_channels,
"num_classes": n_classes,
"image_size": image_size,
"onnx_opset": opset,
"onnx_path": str(onnx_path),
"quantized_path": str(quantized_path) if not no_quantize else None,
"val_iou": ckpt.get("val_iou", None),
"val_f1": ckpt.get("val_f1", None),
"epoch": ckpt.get("epoch", None),
"benchmark_ms": {
"pytorch_cpu": round(pt_ms, 2),
"onnx_cpu": round(onnx_ms, 2) if onnx_ms > 0 else None,
"onnx_int8_cpu": round(q_ms, 2) if q_ms > 0 else None,
},
}
info_path = onnx_path.parent / f"{onnx_path.stem}_export_info.json"
with open(info_path, "w") as f:
json.dump(info, f, indent=2)
logger.info("Export metadata saved to %s", info_path)
# Summary
print("\n" + "=" * 55)
print(f" Export Summary — {onnx_path.name}")
print("=" * 55)
print(f" ONNX model : {onnx_path}")
if not no_quantize and quantized_path.exists():
print(f" INT8 model : {quantized_path}")
print(f" Val IoU : {info['val_iou']:.4f}" if info["val_iou"] else " Val IoU : N/A")
print(f" PyTorch (CPU): {pt_ms:.1f} ms")
if onnx_ms > 0:
print(f" ONNX (CPU) : {onnx_ms:.1f} ms ({pt_ms / onnx_ms:.1f}× speedup)")
if q_ms > 0:
print(f" INT8 (CPU) : {q_ms:.1f} ms ({pt_ms / q_ms:.1f}× speedup)")
print("=" * 55)
return info
def _discover_analysis_checkpoints() -> list[tuple[str, Path, int]]:
"""
Resolve (analysis_type, checkpoint_path, num_classes) for every enabled
analysis type that declares neural ``weights`` in config.yaml.
"""
from climatevision.data.band_mapping import (
get_model_config,
list_enabled_analysis_types,
)
found: list[tuple[str, Path, int]] = []
for analysis_type in list_enabled_analysis_types():
model_cfg = get_model_config(analysis_type)
weights = model_cfg.get("weights")
if not weights: # e.g. SAR ensemble — no neural checkpoint to export
continue
ckpt = PROJECT_ROOT / weights
n_classes = model_cfg.get("num_classes", 2)
found.append((analysis_type, ckpt, n_classes))
return found
def _export_all(args: argparse.Namespace) -> None:
"""Export every enabled analysis type's checkpoint to ``models/unet_<type>.onnx``."""
targets = _discover_analysis_checkpoints()
if not targets:
logger.error("No analysis types with neural weights found in config.yaml")
sys.exit(1)
exported, skipped = 0, 0
for analysis_type, ckpt_path, n_classes in targets:
if not ckpt_path.exists():
logger.warning("Skipping %s — checkpoint not found: %s", analysis_type, ckpt_path)
skipped += 1
continue
onnx_path = PROJECT_ROOT / "models" / f"unet_{analysis_type}.onnx"
logger.info("Exporting %s → %s", analysis_type, onnx_path)
export_checkpoint(
ckpt_path=ckpt_path,
onnx_path=onnx_path,
image_size=args.image_size,
opset=args.opset,
no_quantize=args.no_quantize,
n_classes=n_classes,
)
exported += 1
logger.info("Batch export complete — %d exported, %d skipped", exported, skipped)
if exported == 0:
sys.exit(1)
def main() -> None:
args = parse_args()
if args.all:
_export_all(args)
return
if not args.checkpoint:
logger.error("Provide --checkpoint <path> or --all")
sys.exit(1)
ckpt_path = Path(args.checkpoint)
if not ckpt_path.exists():
logger.error("Checkpoint not found: %s", ckpt_path)
sys.exit(1)
run_dir = ckpt_path.parent
onnx_path = Path(args.out) if args.out else run_dir / "model.onnx"
export_checkpoint(
ckpt_path=ckpt_path,
onnx_path=onnx_path,
image_size=args.image_size,
opset=args.opset,
no_quantize=args.no_quantize,
)
print()
print("Serve with:")
print(f" onnxruntime → sess = ort.InferenceSession('{onnx_path}')")
print()
def parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser(description="Export ClimateVision model to ONNX")
p.add_argument("--checkpoint", default=None, help="Path to best_model.pth")
p.add_argument("--all", action="store_true",
help="Export every enabled analysis type to models/unet_<type>.onnx")
p.add_argument("--out", default=None,
help="ONNX output path (default: <checkpoint_dir>/model.onnx)")
p.add_argument("--image-size", type=int, default=256,
help="Spatial size used for export (any size works at inference via dynamic axes)")
p.add_argument("--opset", type=int, default=17, help="ONNX opset version")
p.add_argument("--no-quantize", action="store_true", help="Skip INT8 quantization")
return p.parse_args()
if __name__ == "__main__":
main()
You can’t perform that action at this time.