search

toandaominh1997 / EfficientDet.Pytorch

Implementation EfficientDet: Scalable and Efficient Object Detection in PyTorch
MIT License
1.44k stars 306 forks source link

bi fpn without detach #26

Closed rydenisbak closed 4 years ago

rydenisbak commented 4 years ago

I got better result on face detection task with this changes. Overall detach seems strange for me. We add inputs like skip connections and why we need to detach them?

pep8speaks commented 4 years ago

Hello @rydenisbak! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

Line 193:80: E501 line too long (84 > 79 characters) Line 196:80: E501 line too long (84 > 79 characters) Line 208:80: E501 line too long (115 > 79 characters)

Comment last updated at 2019-12-12 11:56:46 UTC
toandaominh1997 commented 4 years ago

@rydenisbak Thanks for your pull request, I have read your code, I feel it makes no difference. Can you explain this difference?

78978279_578935422907320_8238087675656011776_n

rydenisbak commented 4 years ago

Hello, oh excuse me. I cloned your repo two days ago and saw version with this code

for in_tensor in inputs:
    pathtd.append(in_tensor.clone().detach())

it was seemed like mistake, but then you fixed it.

Now, we have no difference