General approach to execution is a set of scripts since there are a lot of parameters for running. There are 2 main scripts:
extra/scripts/run_pascal.sh
extra/scripts/run_densecrf.sh
So, if you want to execute Deeplab, you need to set up these scripts respectively to your configuration.
-
Set variable DEST_DIR in extra/scripts/deploy.sh as you need to and run that script.
-
To run system you need to configure variables in extra/scripts/run_pascal.sh and extra/scripts/run_densecrf.sh properly.
RUN_TRAIN, RUN_TEST, RUN_TRAIN2, RUN_TEST2 stands for train, val, trainval, test subsets respectively, NET_ID — for net name, DATA_ROOT — for dataset files directory, EXP — for 'exper' directory, CAFFE_BIN — for install/tools/caffe binary.
- If you want to learn model from a scratch, you need to download initial model weights:
cd $DEST_DIR/model/$MODEL
wget http://ccvl.stat.ucla.edu/ccvl/init_models/vgg16_20M.caffemodel
cp -s $PWD/vgg16_20M.caffemodel init.caffemodel
The next time you run run_pascal.sh script, it automatically sets initial weights of your model with these values on the TRAIN step.