- put training data and labels in ./data/dataset/train_data/
- generate train.txt like the already existed one
google for vgg16.npy weights file and put it in ./data/weights/initial_weights/
p.s. if you don't use vgg16 weights just comment 'hed_class.assign_init_weights(sess)' in train.py
cd to the root directory './hed-tf'
python train.py -gpu '0' # default gpu 0
p.s. it seems that if you do not have gpu ,tf will run it in cpu in this program
change the learning rate in train.py
this repository is not exactly identical to hed paper.
I use dilate conv at conv1_1, you can change it easily.
for other changes , to see the paper and official code carefully.