Skip to first unread message
Rishabh Shah
Apr 11, 2017, 9:09:15 AM4/11/17
to Caffe Users
Hi all,
I have been working on semantic segmentation where I feed the data using Python layers created here . I want to train my own network in caffe using Python. My attempts at this have been futile. What I am currently doing is shown in the following Python code.
However after each solver iteration there is no update of the weights as my weights have all been initialized to zero. Even though, in my prototxt I have specified weight_filler and bias_filler parameters in train.prototxt, all my weights have been initialized to zero. Why aren't my weight initialized to parameters mentioned in weight_filler? Also, is there any approach to solve the issue?
I have been working on semantic segmentation where I feed the data using Python layers created here . I want to train my own network in caffe using Python. My attempts at this have been futile. What I am currently doing is shown in the following Python code.
caffe.set_mode_cpu()
net = caffe.Net('/home/stylabs/projects/codes/imageSegmentation/semanticSegmentation/fcn.berkeleyvision.org-master/voc-fcn8s-myModel/train.prototxt', caffe.TRAIN)
solver = caffe.SGDSolver('solver.prototxt')
# solver.net.copy_from(weights)
# surgeries
interp_layers = [k for k in solver.net.params.keys() if 'up' in k]
surgery.interp(solver.net, interp_layers)
# scoring
val = np.loadtxt('/home/stylabs/projects/data/imageSegmentation/semanticSegmentation/104_108_63_122_300x300/test.txt', dtype=str)
for _ in range(50):
solver.step(100)
However after each solver iteration there is no update of the weights as my weights have all been initialized to zero. Even though, in my prototxt I have specified weight_filler and bias_filler parameters in train.prototxt, all my weights have been initialized to zero. Why aren't my weight initialized to parameters mentioned in weight_filler? Also, is there any approach to solve the issue?
Reply all
Reply to author
Forward
0 new messages