search

aharley / segaware

Segmentation-Aware Convolutional Networks Using Local Attention Masks
145 stars 31 forks source link

TwoImageData #6

Open FuyuWang opened 6 years ago

FuyuWang commented 6 years ago

Could you please tell me what "TwoImageData" layer is used for, I can't find the code. Thanks!

jinde-liu commented 6 years ago

The closed session part has the same question you asked.

aharley commented 6 years ago

You can swap TwoImageData with ImageSegData in the prototxt. See the "res" example to see how to use ImageSegData. I didn't mean to close the first issue before actually fixing the prototxt. I'll leave this one open until I really fix.

FuyuWang commented 6 years ago

Thanks for your reply, but I encountered a problem: When I train vgg network using VOC2012, the loss cannot converge and output is always zero. So the network does not learn from the datasets, how to solve this problem?

jinde-liu commented 6 years ago

I have the same problem, which init caffemodel did you use? I have also found that with batch size: 1, the loss vibrate lot. It seems that the net can not convergent with batch size: 1

FuyuWang commented 6 years ago

I think it is the large network that results to hard training. I use deeplab init model.

jinde-liu commented 6 years ago

I have found the problem is that when you change TwoImageData to ImageSegData, you need to add labe_type: PIXEL/NONE to image_data_param and also top: "data_dim" to layer. I have added a silence layer to ignore information form data_dim. But adding a silence layer may not be necessary.