Unknown Issue with the FCN. Python kernel dies when creating test-net

[Mateo]

Hello,

I'm trying to finetune the FCN-AlexNet PASCAL with my own datset. First of all , I tried the model with the original dataset (VOC-2011 and sbdd) ant it works well, with the scripts solve.py and net.py.

When I tried the network with my owndataset,While I execute the line "solver.step(1)"(I just want to check if it can do at least one iteration) the python kernel dies when it is creating the test-net. What's weird, is that it doesn't show any error message.(see the log.txt for the kernel messages)

The differences between the orignal dataset and my own are :

1. My data set are Ultra-sound medical images , in gray scale, the VOC-Pascal are more general images in color. 

2. The Segmentation ground-truth of the original data-set, for the training phase is stored in a .mat file, and for the validation phase, they are stored in .png . In my case , I store the both in png.

3. The original work is done for classifying 20 classes with the ground ( 21 classes) . In my case, I only want to detect 1 class and the ground. So, in the original prototxt, I changed the  num_output from 21 to 2 .

-Regarding the images size, they are from different sizes , but they are between the range (300,512) ->width and a fixed 512 height.

-In order to use my own dataset, I use the same folders of the original data set, I only change the original files for my owns, and I modify the .txt with the lists.

I attach the files : net.py, solve.py, voc_layers.py  surgery.py and solver.prototxt. I also attach one example of my image.

Personally I have no idea could be happening .I'd be very grateful if someone finds the error.

Mateo

Mode: CPU

Operating System : Windows 7 Professional

Python Version : 2.7

[Mateo]

In the solver.prototxt, the momentum is suposed to be 0.9 ( I tried with 0.1 , but i got the same issue, I forgot to put the original value again)