search

xtudbxk / DSRG-tensorflow

a tensorflow version for DSRG (Weakly-Supervised Semantic Segmentation Network with Deep Seeded Region Growing)
75 stars 20 forks source link

Question about provided pretrained weights "vgg16_deeplab_aspp.npy" #29

Closed terenceylchow124 closed 3 years ago

terenceylchow124 commented 3 years ago

Hi, Thank you for preparing the tensorflow DSRG model. In your mark down, it was mentioned that there is the pretrained model which can be downloaded and loaded into the DSRG model. However, I am wondering is this model pretrained by imagenet (for normal classification task) or by Voc2012 (for semantic segmentation). I am not quite sure about the supervision, as the DSRG is the model designed for weakly supervised segmentation, I just want to make sure how much supervision provided by the pretrained weight. . Thanks.

xtudbxk commented 3 years ago

The model is pretrained on imagenet. And I just converts the pretrained model of deeplab from caffe to tensorflow.

terenceylchow124 commented 3 years ago

Thanks for your reply!

terenceylchow124 commented 3 years ago

Capture Sorry for interrupting again, however I am wondering if you have any experience for handling this kind of error. I am trying modify the DSRG model with building some external layers, leading the multiple outputs (two) of the model. And compute the loss like, loss = loss_s + loss_c + loss_e (external loss). Do I need to set up the anything other than building the layers, and construct the loss.

xtudbxk commented 3 years ago

It seems the 'pyfunc' cuts off the flow of backward gradients. Just remember that the 'pyfunc' cann't pass the gradients back. So I doubt that there is least one loss function whose all inputs are built upon 'pyfunc' functions.