FCN_Tutorial.ipynb

Is the it possible to get the same functionality of the Crop layer from Caffe's TransformationParameters?

The Crop layer in Caffe's future branch does not perform cropping like the crop_size parameter from TransformationParameter. The Crop layer tracks/maps coordinates in order to align two blobs to establish a correspondence between input and output. This layer takes two bottom blobs and produces one top, which is a copy of the first bottom cropped to the size of the second so that coordinates exactly correspond. More information on the Crop layer here: BVLC/caffe#1976

For full image learning, it does not make sense to perform random cropping of the input image before feeding it to the network. This is because the input image and label need to be in lockstep during training. It is possible that the random crop of the image and the corresponding label will be out of sync due to randomness, and will cause the training loss to diverge.

I tried to follow this tutorial and used future branch, but ended up with error like Check failed: !conv_param.has_kernel_size() != !(conv_param.has_kernel_h() && conv_param.has_kernel_w()) Filter size is kernel_size OR kernel_h and kernel_w; not both
*** Check failure stack trace: ***

I then switched to the caffe-master branch and merged https://github.com/BlGene/caffe/tree/crop-nd into the master branch to get the crop layer. The error now is different,

I0222 00:01:04.905822  2873 net.cpp:106] Creating Layer score
I0222 00:01:04.905827  2873 net.cpp:454] score <- upscore
I0222 00:01:04.905833  2873 net.cpp:454] score <- data_data_0_split_1
I0222 00:01:04.905839  2873 net.cpp:411] score -> score
I0222 00:01:04.905879  2873 net.cpp:150] Setting up score
I0222 00:01:04.905897  2873 net.cpp:157] Top shape: 1 2 271 271 (146882)
I0222 00:01:04.905901  2873 net.cpp:165] Memory required for data: 35186344
I0222 00:01:04.905906  2873 layer_factory.hpp:77] Creating layer loss
I0222 00:01:04.905917  2873 net.cpp:106] Creating Layer loss
I0222 00:01:04.905923  2873 net.cpp:454] loss <- score
I0222 00:01:04.905928  2873 net.cpp:454] loss <- label
I0222 00:01:04.905936  2873 net.cpp:411] loss -> loss
I0222 00:01:04.905946  2873 layer_factory.hpp:77] Creating layer loss
F0222 00:01:04.906365  2873 softmax_loss_layer.cpp:47] Check failed: outer_num_ * inner_num_ == bottom[1]->count() (73441 vs. 65536) Number of labels must match number of predictions; e.g., if softmax axis == 1 and prediction shape is (N, C, H, W), label count (number of labels) must be N*H*W, with integer values in {0, 1, ..., C-1}.
*** Check failure stack trace: ***

The size of score should be 1 2 256 256, but no idea why it is still 1 2 271 271
Any idea? Otherwise I will try to go back to the future version and see if I can make it work.

Thanks a lot!

How did you guys get the data? I registered twice but sadly I wouldn't get an email.

@hh1985, you should continue working with the future branch for consistency purposes. I read that the implementation of the crop-nd layer may be different from the original implementation of the Crop layer. As for your error, it looks like you did not define the parameters of the deconvolution layer; it needs to be manually defined in your train/val prototxt files.

@sebastian-schlecht, the link in the tutorial should lead you to the right registration site. They could be overwhelmed with registration requests...

@vuptran, i am trying to run the tutorial, got a problem to generate _caffe.so, since the folder /caffe_FCN/include/caffe/layers does not exist, just wondering how to get it? or just copy from the master branch? thanks a lot.

I had run through the tutorial, it works. What i dont understand is "how to properly set up the network in order for the crop layer to work?" I try to modify some params, then it always shows error like " Check failed: (crop_map.coefs()[i].first) >= (1)-0.000000000000001L (0.888889 vs. 1)" . Can anyone help here? Thanks a lot.

@vuptran Hello thank you for your sharing. I'm now working on this dataset and got lost while finding the matching between images and labels. Your post is helping me a lot understand that. But my question why you only look for the contour with highest number? What about the others? There are not SAX series?
Another question: in the SAX series dictionary: what do folders and values association mean?
SAX_SERIES = { # challenge training "SC-HF-I-1": "0004", "SC-HF-I-2": "0106", ...... }
I looked in SC-HF-I-1 folder in contour label and didn't find any contour named: IM-0001-0004-......txt but in the corresponding image, I've found it so it's like I've an datum without label. Same goes for the SC-HF-I-2 folder, it has instead 0107 as a contour number.
Please explain me a bit?
Thank you so much.

@vuptran The link in which data of sunnybrook is given is not working so if any of you have download the data then Please share with me.

he link in which data of sunnybrook is given is not working so if any of you have download the data then Please share with me.