Add export to onnx and tensorRT
Browse files- scripts/export_onnx_trt.py +130 -0
scripts/export_onnx_trt.py
ADDED
@@ -0,0 +1,130 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
import pprint
|
4 |
+
from typing import Tuple
|
5 |
+
|
6 |
+
import torch
|
7 |
+
|
8 |
+
from src.wireseghr.model import WireSegHR
|
9 |
+
|
10 |
+
|
11 |
+
class CoarseModule(torch.nn.Module):
|
12 |
+
def __init__(self, core: WireSegHR):
|
13 |
+
super().__init__()
|
14 |
+
self.core = core
|
15 |
+
|
16 |
+
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
17 |
+
logits, cond = self.core.forward_coarse(x)
|
18 |
+
return logits, cond
|
19 |
+
|
20 |
+
|
21 |
+
class FineModule(torch.nn.Module):
|
22 |
+
def __init__(self, core: WireSegHR):
|
23 |
+
super().__init__()
|
24 |
+
self.core = core
|
25 |
+
|
26 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
27 |
+
logits = self.core.forward_fine(x)
|
28 |
+
return logits
|
29 |
+
|
30 |
+
|
31 |
+
def build_model(cfg: dict, device: torch.device) -> WireSegHR:
|
32 |
+
pretrained_flag = bool(cfg.get("pretrained", False))
|
33 |
+
model = WireSegHR(backbone=cfg["backbone"], in_channels=6, pretrained=pretrained_flag)
|
34 |
+
model = model.to(device)
|
35 |
+
return model
|
36 |
+
|
37 |
+
|
38 |
+
def main():
|
39 |
+
parser = argparse.ArgumentParser(description="Export WireSegHR to ONNX and TensorRT")
|
40 |
+
parser.add_argument("--config", type=str, default="configs/default.yaml")
|
41 |
+
parser.add_argument("--ckpt", type=str, default="", help="Path to checkpoint .pt")
|
42 |
+
parser.add_argument("--out_dir", type=str, default="exports")
|
43 |
+
parser.add_argument("--coarse_size", type=int, default=1024)
|
44 |
+
parser.add_argument("--fine_patch_size", type=int, default=1024)
|
45 |
+
parser.add_argument("--opset", type=int, default=17)
|
46 |
+
parser.add_argument("--trtexec", type=str, default="", help="Optional path to trtexec to build TRT engines")
|
47 |
+
|
48 |
+
args = parser.parse_args()
|
49 |
+
|
50 |
+
import yaml
|
51 |
+
|
52 |
+
with open(args.config, "r") as f:
|
53 |
+
cfg = yaml.safe_load(f)
|
54 |
+
print("[export] Loaded config:")
|
55 |
+
pprint.pprint(cfg)
|
56 |
+
|
57 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
58 |
+
model = build_model(cfg, device)
|
59 |
+
|
60 |
+
ckpt_path = args.ckpt if args.ckpt else cfg.get("resume", "")
|
61 |
+
if ckpt_path:
|
62 |
+
assert os.path.isfile(ckpt_path), f"Checkpoint not found: {ckpt_path}"
|
63 |
+
print(f"[export] Loading checkpoint: {ckpt_path}")
|
64 |
+
state = torch.load(ckpt_path, map_location=device)
|
65 |
+
model.load_state_dict(state["model"]) # expects dict with key 'model'
|
66 |
+
model.eval()
|
67 |
+
|
68 |
+
os.makedirs(args.out_dir, exist_ok=True)
|
69 |
+
|
70 |
+
# Prepare dummy inputs (static shapes for best TRT performance)
|
71 |
+
coarse_in = torch.randn(1, 6, args.coarse_size, args.coarse_size, device=device)
|
72 |
+
fine_in = torch.randn(1, 6, args.fine_patch_size, args.fine_patch_size, device=device)
|
73 |
+
|
74 |
+
# Coarse export
|
75 |
+
coarse_wrapper = CoarseModule(model).to(device).eval()
|
76 |
+
coarse_onnx = os.path.join(args.out_dir, f"wireseghr_coarse_{args.coarse_size}.onnx")
|
77 |
+
print(f"[export] Exporting COARSE to {coarse_onnx}")
|
78 |
+
torch.onnx.export(
|
79 |
+
coarse_wrapper,
|
80 |
+
coarse_in,
|
81 |
+
coarse_onnx,
|
82 |
+
export_params=True,
|
83 |
+
opset_version=args.opset,
|
84 |
+
do_constant_folding=True,
|
85 |
+
input_names=["x_coarse"],
|
86 |
+
output_names=["logits", "cond"],
|
87 |
+
dynamic_axes=None,
|
88 |
+
)
|
89 |
+
|
90 |
+
# Fine export
|
91 |
+
fine_wrapper = FineModule(model).to(device).eval()
|
92 |
+
fine_onnx = os.path.join(args.out_dir, f"wireseghr_fine_{args.fine_patch_size}.onnx")
|
93 |
+
print(f"[export] Exporting FINE to {fine_onnx}")
|
94 |
+
torch.onnx.export(
|
95 |
+
fine_wrapper,
|
96 |
+
fine_in,
|
97 |
+
fine_onnx,
|
98 |
+
export_params=True,
|
99 |
+
opset_version=args.opset,
|
100 |
+
do_constant_folding=True,
|
101 |
+
input_names=["x_fine"],
|
102 |
+
output_names=["logits"],
|
103 |
+
dynamic_axes=None,
|
104 |
+
)
|
105 |
+
|
106 |
+
# Optional TensorRT building via trtexec
|
107 |
+
if args.trtexec:
|
108 |
+
import subprocess
|
109 |
+
|
110 |
+
def build_engine(onnx_path: str, engine_path: str):
|
111 |
+
print(f"[export] Building TRT engine: {engine_path}")
|
112 |
+
cmd = [
|
113 |
+
args.trtexec,
|
114 |
+
f"--onnx={onnx_path}",
|
115 |
+
f"--saveEngine={engine_path}",
|
116 |
+
"--explicitBatch",
|
117 |
+
"--fp16",
|
118 |
+
]
|
119 |
+
subprocess.run(cmd, check=True)
|
120 |
+
|
121 |
+
coarse_engine = os.path.join(args.out_dir, f"wireseghr_coarse_{args.coarse_size}.engine")
|
122 |
+
fine_engine = os.path.join(args.out_dir, f"wireseghr_fine_{args.fine_patch_size}.engine")
|
123 |
+
build_engine(coarse_onnx, coarse_engine)
|
124 |
+
build_engine(fine_onnx, fine_engine)
|
125 |
+
|
126 |
+
print("[export] Done.")
|
127 |
+
|
128 |
+
|
129 |
+
if __name__ == "__main__":
|
130 |
+
main()
|