Max Lotz
Sep 19, 2017, 12:49:48 PM9/19/17
to Caffe Users
I am attempting to use the nyud-fcn32s-color model but seem to be having several problems.
Firstly I attempted to train it on the nyud rgb images. I was having problems understanding and using the python layer so instead I used an ImageData layer. However when I inputted my 480x640 single channel .png label images caffe seemed to always automatically read it as a 3 channel image as I would get a dimension mismatch (it was 3x the size). Instead I created an lmdb for the images and one for the labels and this seemed to solve the problem.
However when I attempted to train the model using the weights at http://dl.caffe.berkeleyvision.org/nyud-fcn32s-color-heavy.caffemodel as a starting point, the loss would not decrease. I then evaluated the output of some of the NYUV2 rgb images on the aforementioned pretrained weights and the output was the same for each image. I visualized the conv1_1 layer filters of these weights and they certainly don't seem random.
I have attached the files I have used to what I have described above.
Firstly I attempted to train it on the nyud rgb images. I was having problems understanding and using the python layer so instead I used an ImageData layer. However when I inputted my 480x640 single channel .png label images caffe seemed to always automatically read it as a 3 channel image as I would get a dimension mismatch (it was 3x the size). Instead I created an lmdb for the images and one for the labels and this seemed to solve the problem.
However when I attempted to train the model using the weights at http://dl.caffe.berkeleyvision.org/nyud-fcn32s-color-heavy.caffemodel as a starting point, the loss would not decrease. I then evaluated the output of some of the NYUV2 rgb images on the aforementioned pretrained weights and the output was the same for each image. I visualized the conv1_1 layer filters of these weights and they certainly don't seem random.
I have attached the files I have used to what I have described above.
Reply all
Reply to author
Forward
0 new messages