Very interesting, I have a model like this:
class YOLOv3(nn.Module):
def __init__(self, num_classes=80, ignore_thre=0.7, label_smooth=False, rfb=False, vis=False, asff=False):
super(YOLOv3, self).__init__()
self.module_list = build_yolov3_modules(
num_classes, ignore_thre, label_smooth, rfb)
self.level_0_fusion = ASFF(level=0, rfb=rfb, vis=vis)
self.level_0_header = YOLOv3Head(anch_mask=[6, 7, 8], n_classes=num_classes, stride=32, in_ch=1024,
ignore_thre=ignore_thre, label_smooth=label_smooth, rfb=rfb)
self.level_1_fusion = ASFF(level=1, rfb=rfb, vis=vis)
self.level_1_header = YOLOv3Head(anch_mask=[3, 4, 5], n_classes=num_classes, stride=16, in_ch=512,
ignore_thre=ignore_thre, label_smooth=label_smooth, rfb=rfb)
self.level_2_fusion = ASFF(level=2, rfb=rfb, vis=vis)
self.level_2_header = YOLOv3Head(anch_mask=[0, 1, 2], n_classes=num_classes, stride=8, in_ch=256,
ignore_thre=ignore_thre, label_smooth=label_smooth, rfb=rfb)
def forward(self, x, targets=None, epoch=0):
output = []
route_layers = []
for i, module in enumerate(self.module_list):
x = module(x)
if i in [6, 8, 17, 24, 32]:
route_layers.append(x)
if i == 19:
x = torch.cat((x, route_layers[1]), 1)
if i == 26:
x = torch.cat((x, route_layers[0]), 1)
print(len(route_layers))
fused_0 = self.level_0_fusion(route_layers[2], route_layers[3], route_layers[4])
x = self.level_0_header(fused_0)
output.append(x)
fused_1 = self.level_1_fusion(route_layers[2], route_layers[3], route_layers[4])
x = self.level_1_header(fused_1)
output.append(x)
fused_2 = self.level_2_fusion(route_layers[2], route_layers[3], route_layers[4])
x = self.level_2_header(fused_2)
output.append(x)
return torch.cat(output, 1)
This model when using torch.onnx.export it generate a ONNX model contains only an input and a Constant, log like this:
%level_2_header.Feature_adaption.dconv.weight : Float(256, 256, 3, 3),
%level_2_header.Feature_adaption.dconv.bias : Float(256),
%level_2_header.conv.weight : Float(340, 256, 1, 1),
%level_2_header.conv.bias : Float(340)):
%559 : Float(1, 52500, 85) = onnx::Constant[value=<Tensor>]()
return (%559)
interesting thing is that, somehow I comment out these lines:
def forward(self, x, targets=None, epoch=0):
output = []
route_layers = []
for i, module in enumerate(self.module_list):
x = module(x)
if i in [6, 8, 17, 24, 32]:
route_layers.append(x)
if i == 19:
x = torch.cat((x, route_layers[1]), 1)
if i == 26:
x = torch.cat((x, route_layers[0]), 1)
# print(len(route_layers))
# fused_0 = self.level_0_fusion(route_layers[2], route_layers[3], route_layers[4])
# x = self.level_0_header(fused_0)
# output.append(x)
# fused_1 = self.level_1_fusion(route_layers[2], route_layers[3], route_layers[4])
# x = self.level_1_header(fused_1)
# output.append(x)
# fused_2 = self.level_2_fusion(route_layers[2], route_layers[3], route_layers[4])
# x = self.level_2_header(fused_2)
# output.append(x)
# return torch.cat(output, 1)
return x
It can trace normal ONNX model.
Does anybody can help me debug this weird issue?