Open2
PyTorchのモデルを各種解像度で一括ONNX出力（エクスポート）するコードスニペット torch.onnx.export
PINTO import onnx
from onnxsim import simplify
RESOLUTION = [
[192,320],
[192,416],
[192,640],
[192,800],
[256,320],
[256,416],
[256,448],
[256,640],
[256,800],
[256,960],
[288,480],
[288,640],
[288,800],
[288,960],
[288,1280],
[320,320],
[384,480],
[384,640],
[384,800],
[384,960],
[384,1280],
[416,416],
[480,640],
[480,800],
[480,960],
[480,1280],
[512,512],
[512,640],
[512,896],
[544,800],
[544,960],
[544,1280],
[640,640],
[736,1280],
]
MODEL = f'xxxx'
for H, W in RESOLUTION:
onnx_file = f"{MODEL}_1x3x{H}x{W}.onnx"
x = torch.randn(1, 3, H, W).cuda()
torch.onnx.export(
self.model.module,
args=(x),
f=onnx_file,
opset_version=11,
input_names=['input'],
#output_names=['output'],
)
model_onnx1 = onnx.load(onnx_file)
model_onnx1 = onnx.shape_inference.infer_shapes(model_onnx1)
onnx.save(model_onnx1, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
onnx_file = f"{MODEL}_1x3xHxW.onnx"
x = torch.randn(1, 3, 192, 320).cuda()
torch.onnx.export(
self.model.module,
args=(x),
f=onnx_file,
opset_version=11,
input_names=['input'],
#output_names=['output'],
dynamic_axes={
'input' : {2: 'height', 3: 'width'},
'output' : {2: 'height', 3: 'width'},
}
)
model_onnx1 = onnx.load(onnx_file)
model_onnx1 = onnx.shape_inference.infer_shapes(model_onnx1)
onnx.save(model_onnx1, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
onnx_file = f"{MODEL}_Nx3xHxW.onnx"
x = torch.randn(1, 3, 192, 320).cuda()
torch.onnx.export(
self.model.module,
args=(x),
f=onnx_file,
opset_version=11,
input_names=['input'],
output_names=['output'],
dynamic_axes={
'input' : {0: 'N', 2: 'height', 3: 'width'},
'output' : {0: 'N', 2: 'height', 3: 'width'},
}
)
model_onnx1 = onnx.load(onnx_file)
model_onnx1 = onnx.shape_inference.infer_shapes(model_onnx1)
onnx.save(model_onnx1, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
import sys
sys.exit(0)
import onnx
from onnxsim import simplify
RESOLUTION = [
[180,320],
[180,416],
[180,512],
[180,640],
[180,800],
[240,320],
[240,416],
[240,512],
[240,640],
[240,800],
[240,960],
[256,448],
[288,480],
[288,512],
[288,640],
[288,800],
[288,960],
[288,1280],
[320,320],
[360,480],
[360,512],
[360,640],
[360,800],
[360,960],
[360,1280],
[376,1344],
[416,416],
[480,640],
[480,800],
[480,960],
[480,1280],
[512,512],
[512,896],
[540,800],
[540,960],
[540,1280],
[640,640],
[640,960],
[720,1280],
[720,2560],
[1080,1920],
]
MODEL = f'xxxx'
for H, W in RESOLUTION:
onnx_file = f"{MODEL}_1x3x{H}x{W}.onnx"
x = torch.randn(1, 3, H, W).cuda()
torch.onnx.export(
self.model.module,
args=(x),
f=onnx_file,
opset_version=11,
input_names=['input'],
#output_names=['output'],
)
model_onnx1 = onnx.load(onnx_file)
model_onnx1 = onnx.shape_inference.infer_shapes(model_onnx1)
onnx.save(model_onnx1, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
onnx_file = f"{MODEL}_1x3xHxW.onnx"
x = torch.randn(1, 3, 180, 320).cuda()
torch.onnx.export(
self.model.module,
args=(x),
f=onnx_file,
opset_version=11,
input_names=['input'],
#output_names=['output'],
dynamic_axes={
'input' : {2: 'height', 3: 'width'},
'output' : {2: 'height', 3: 'width'},
}
)
model_onnx1 = onnx.load(onnx_file)
model_onnx1 = onnx.shape_inference.infer_shapes(model_onnx1)
onnx.save(model_onnx1, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
onnx_file = f"{MODEL}_Nx3xHxW.onnx"
x = torch.randn(1, 3, 180, 320).cuda()
torch.onnx.export(
self.model.module,
args=(x),
f=onnx_file,
opset_version=11,
input_names=['input'],
output_names=['output'],
dynamic_axes={
'input' : {0: 'N', 2: 'height', 3: 'width'},
'output' : {0: 'N', 2: 'height', 3: 'width'},
}
)
model_onnx1 = onnx.load(onnx_file)
model_onnx1 = onnx.shape_inference.infer_shapes(model_onnx1)
onnx.save(model_onnx1, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
model_onnx2 = onnx.load(onnx_file)
model_simp, check = simplify(model_onnx2)
onnx.save(model_simp, onnx_file)
import sys
sys.exit(0)