I want to finetune vgg face on Eyediap Dataset.
I have used the attached network model(vgg-eyediap.prototxt) and solver.prototxt.
In this experiment I have used small part of Eyediap gaze, 26k for train and 4.5k frames for test.
The finetuning was done without error, for 1000 iteration.
sudo ./build/tools/extract_to_txt test /home/vgg-eyediap_snapshot_iter_1000.caffemodel /home/vgg-eyediap.prototxt /home/feature.txt fc8_new 250 txt GPU 0
To find out whether I have extracted good feature or not, I have extracted the last fully connected features, which should be the same as labels(headpose.txt),right?
5000 lines of Features are extracted without error.
The problem is features extracted from fc8_new are zeros except the first column.
Please help me to find the solution.
Thanks