0

I did not have any choice except asking here. I have a lot of difficulties for a long time. I have not been to observe any output from FCN32 :( I trained FCN32 on my data from scratch and always getting a black image. I added gaussian with std= 0.01 initialization for convolutional layers. But still I get black image.

I tried to add weighted loss layers. However, I was not successful to add it correctly. I am not good at python and c++.

My questions:

  1. Is there any correct PR that it can easily include this layer?
  2. My data has 5 classes that the proportion of classes differ from each other in different images. How can I create these weight matrices for each image?

I really appreciate any help. Please share if you know any resource/link/ or if I can get it from other networks' repositories.

Community
  • 1
  • 1
S.EB
  • 1,966
  • 4
  • 29
  • 54
  • What do you mean "trained FCN32 on my data from scratch"? Is your network pretrained on another dataset (like ImageNet)? – Dr. Snoopy Mar 19 '17 at 15:58
  • Thanks for your comment, No I am training FCN32 from the scratch on my data. I am not using pre-trained model, because my data is medical images and its characteristics is different. Should I use a pretrained model? If yes, could you please let me know which model should I use for the MRI images that I have? could you please let me know what can I do? Thanks a lot again. – S.EB Mar 19 '17 at 16:03
  • @MatiasValdenegro should I fine-tune from a model? – S.EB Mar 20 '17 at 05:12
  • @S.EB this should be easier – Shai Mar 20 '17 at 06:02
  • @Shai Thanks for your message, My data is medical images. Is it ok that I fine tune from other models?If yes, which model can I use? Could you please share the link? I think I should add weighted loss layer as well. It is showing errors. I do not know at which point I am doing wrong. – S.EB Mar 20 '17 at 07:57
  • @S.EB The FCN papers mention that their networks have to be pretrained on imagenet, or else training just diverges/fails. You will have to find a similar way to do the same. Maybe just pretrain in a classification dataset from your own data. – Dr. Snoopy Mar 20 '17 at 15:23
  • You should fine-tune using the fcn-32s-pascal model from https://github.com/shelhamer/fcn.berkeleyvision.org. You should also consider adding "weight_filler: { type: "bilinear" }" to the upscore deconvolution layer in your FCN32 prototxt. – adon Apr 04 '17 at 02:13

0 Answers0