The Caffe network does not train (despite several attempts)

[a.b...@unicas.it]

Hi,

I am post-doc at University of Cassino (Italy), and our research is focused on detecting abnormalities on biomedical images. We have published a number of works in this field [1][2][3], based on different kinds of machine learning techniques (e.g. boosting, cascade, markov chains, SVM). 

We now want to move on to Deep Learning, and we are testing various frameworks before we decide which one we will use for our long-term research. As a test case, we are considering two different classification tasks: (A) detecting microcalcifications on digital mammograms; and (B) vessel segmentation in retina fundus images. We have a long-term experience on these problems and their related data sets as well.

First we tested Torch, but we were not 100% satisfied of the user experience / LUA language, so then we tested Caffe. It's now 7 months since we are using it. We have two PhD students who are working on the tasks (A) and (B), respectively, but none of them was able to train the network with Caffe, as you can see from their previous posts (A and B).

For instance, for task (A), the network structure is composed by 11 layers (2x convolutional, max pooling, ReLu, 2x convolutional, ReLu, 3x fully connected), mini-batch size is 512, image patch size is 12x12. For more details, please see the attached protoxt.

Our training data are {0, 1}-labeled, 1:1 balanced (50,000 positives and 50,000 negatives), normalized (we also tested w/o normalization), and shuffled.

The problem is: after several attempts with different learning parameters (learning rate, momentum, etc.), we cannot understand why our network does not train (training loss = 0.69, testing accuracy = 0.5). We are at a crucial point now, as we have to decide whether to go on with Caffe or to consider another framework.

We thank you in advance for your helpfulness. We think Caffe has a great potential, and perhaps there is some small (but crucial) detail we are missing.