how to removed the Slice operation and replaced the input with a downscaled image and stacked it to match the channel number?

Question

how to removed the Slice operation and replaced the input with a downscaled image and stacked it to match the channel number?

huolianchu opened this issue 4 years ago · comments

huolianchu commented 4 years ago

Hi,

Thanks for you work

i using my own trained model file will report error
Slice operation in .pt or in .onnx?

PengZhang · Answer 1 · Sat Jun 27 2020 11:22:59 GMT+0800 (China Standard Time)

more details needed as you

Sunny Deng · Answer 2 · Sun Jun 28 2020 16:55:32 GMT+0800 (China Standard Time)

ONNX converts the operation of line 87 in common.py

return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))

into a slice operation with steps.

The easiest way to eliminate the slice operation is to change this line into:

return self.conv(torch.cat([x,x,x,x], 1))

It is also possible to change this operation into a convolution operation.

huolianchu · Answer 3 · Sun Jun 28 2020 16:59:33 GMT+0800 (China Standard Time)

ONNX converts the operation of line 87 in common.py
return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))
into a slice operation with steps.

The easiest way to eliminate the slice operation is to change this line into:
return self.conv(torch.cat([x,x,x,x], 1))
It is also possible to change this operation into a convolution operation.

Thanks a lot

tadejj · Answer 4 · Thu Jul 09 2020 16:17:02 GMT+0800 (China Standard Time)

Hi,
Can I remove slice operation for trained model? I have trained my custom dataset, and converted it to pytorchscript. But the result is afwul (all boxes are wrong). Can I use models/onnx_export.py from ultralytics for my model and remove slice operation without retraining?
Thank you so much

Sunny Deng · Answer 5 · Fri Jul 10 2020 09:33:35 GMT+0800 (China Standard Time)

Of course, you can remove the slice operation without retraining.

…

On Thu, Jul 9, 2020, 16:17 tadejj ***@***.***> wrote: Hi, Can I remove slice operation for trained model? I have trained my custom dataset, and converted it to pytorchscript. But the result is afwul (all boxes are wrong). Can I use models/onnx_export.py from ultralytics for my model and remove slice operation without retraining? Thank you so much — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#2 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABA7OF5ZBMCCCBZNXUXWFPTR2V4JDANCNFSM4OGM3S4A> .

tadejj · Answer 6 · Mon Jul 13 2020 14:21:26 GMT+0800 (China Standard Time)

Of course, you can remove the slice operation without retraining.
…
On Thu, Jul 9, 2020, 16:17 tadejj @.***> wrote: Hi, Can I remove slice operation for trained model? I have trained my custom dataset, and converted it to pytorchscript. But the result is afwul (all boxes are wrong). Can I use models/onnx_export.py from ultralytics for my model and remove slice operation without retraining? Thank you so much — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#2 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABA7OF5ZBMCCCBZNXUXWFPTR2V4JDANCNFSM4OGM3S4A .

Thanks a lot!