john1...@gmail.com
Dec 16, 2016, 8:17:35 AM12/16/16
to Caffe Users
Hello all, I am using the newest version of FCN in the http://fcn.berkeleyvision.org
In the file /voc-fcn8s/solve.py line 10 and line 17, the author used a pre-weight as
weights = '../voc-fcn16s/voc-fcn16s.caffemodel'
...
solver.net.copy_from(weights)Is it possible to ignore these lines? I do not want to use pre-weight for training, because I want to apply it in my own database, instead of pascal. Thanks
Przemek D
Dec 19, 2016, 2:13:11 AM12/19/16
to Caffe Users
You will most likely fail. According to the FAQ on the very page you linked:
To (...) train your own FCNs, it is crucial to transplant the weights from the corresponding ILSVRC net such as VGG16.
Why do you not want to use pretrained weights? Transfer learning almost always works better than learning from scratch, no matter what your dataset is. You can skip the last (classification) layer (the one with 21 outputs for PASCAL) and train it to your own data, but the convolutions on the earlier layers will be very difficult for you to get from scratch.
To (...) train your own FCNs, it is crucial to transplant the weights from the corresponding ILSVRC net such as VGG16.
Why do you not want to use pretrained weights? Transfer learning almost always works better than learning from scratch, no matter what your dataset is. You can skip the last (classification) layer (the one with 21 outputs for PASCAL) and train it to your own data, but the convolutions on the earlier layers will be very difficult for you to get from scratch.
Reply all
Reply to author
Forward
0 new messages