GOATmessi7
onnx export bug
When using this snippet export yolov3 asff model out:
# Initiate model
if args.asff:
if backbone == 'mobile':
from models.yolov3_mobilev2 import YOLOv3
print(
"For mobilenet, we currently don't support dropblock, rfb and FeatureAdaption")
else:
from models.yolov3_asff import YOLOv3
print('Training YOLOv3 with ASFF!')
model = YOLOv3(num_classes=num_class, rfb=args.rfb, asff=args.asff)
else:
if backbone == 'mobile':
from models.yolov3_mobilev2 import YOLOv3
else:
from models.yolov3_baseline import YOLOv3
print('Training YOLOv3 strong baseline!')
model = YOLOv3(num_classes=num_class, rfb=args.rfb)
if args.checkpoint:
print("loading pytorch ckpt...", args.checkpoint)
cpu_device = torch.device("cpu")
ckpt = torch.load(args.checkpoint, map_location=cpu_device)
# model.load_state_dict(ckpt,strict=False)
model.load_state_dict(ckpt)
if cuda:
print("using cuda")
torch.backends.cudnn.benchmark = True
device = torch.device("cuda")
model = model.to(device)
if args.half:
model = model.half()
model = model.eval()
print('model is ready..')
# input width is 608 and 608
dummy_input = torch.zeros((3, 800, 800)).to(device).unsqueeze(0)
onnx_f = 'weights/yolov3_asff.onnx'
torch.onnx.export(model, dummy_input, onnx_f,
verbose=True)
print('onnx exported successfully.')
The onnx model exported successfully, but inside the model, it only contains 2 constants, does anyone knows why? (you can also try using this script export yolov3 asff onnx, if you got this issue pls give me a comment)
When I export onnx, there are some troubles, it seems like your problem
Maybe the onnx doesn't support the apex operators and the dcn module. It seems that the DCN module only supports the cuda version.
https://github.com/ruinmessi/ASFF/blob/master/models/yolov3_head.py(line 134). "return refined_pred.data" => "return refined_pred" (when you convert pth to onnx)
