search

jiwoon-ahn / irn

Weakly Supervised Learning of Instance Segmentation with Inter-pixel Relations, CVPR 2019 (Oral)
MIT License
525 stars 100 forks source link

When training Deeplabv2, did you use any pre-trained model? #21

Closed CenIII closed 4 years ago

CenIII commented 4 years ago

When you were training Deeplabv2 using the pseudo-labels produced by your method, had the Deeplabv2 been pretrained before hand or just raw? Thanks.

jiwoon-ahn commented 4 years ago

Hi @CenIII, We used imagenet pretrained model for deeplabv2.

CenIII commented 4 years ago

Hi @CenIII, We used imagenet pretrained model for deeplabv2.

The resnet backbone of deeplabv2 seems to have different configurations (e.g. different dilation rate or stride). How did you reload the weights of the pretrained model? (My understanding is that the difference in dilation rate wont affect the weight loading?)

Second question, did you freeze the first bn layer when training?

Thanks!

jiwoon-ahn commented 4 years ago
  1. Your understanding is right. We just followed the way of the original deeplab.
  2. Yes, the first conv and bn should be freezed when finetuning.
CenIII commented 4 years ago
  1. Your understanding is right. We just followed the way of the original deeplab.
  2. Yes, the first conv and bn should be freezed when finetuning.

Thank you for replying!

Would you like to share the open source code you used to train deeplabv2? Thank you!

Ien001 commented 4 years ago

The same question as @CenIII , Thank you! @jiwoon-ahn

jiwoon-ahn commented 4 years ago

@CenIII, We made our own model from the bottom. I highly recommend to use the code of the author of DeepLab (https://bitbucket.org/aquariusjay/deeplab-public-ver2/src/master/)

CenIII commented 4 years ago

@jiwoon-ahn Thanks! What do you mean by "made from bottom"? So what's the difference of your model and training process from the original one?

jiwoon-ahn commented 4 years ago

I believe none or little gap exists between our model from the original as we did our best to minimize the difference.

CenIII commented 4 years ago

I believe none or little gap exists between our model from the original as we did our best to minimize the difference.

Got it. Thank you for your help!