chainer implementation of pspnet
install chainer
pip install chainer=='1.24.0'
download the cityscape dataset
your data directory consists of
your_path|----gtFine
|----leftImg8bit
run the code for training
the use of multiple gpus is strongly recommended. pspnet consists of very deep layers.
if you have 3 gpus, run
python main.py --data_dir your_path --gpus 0 1 2
you can also set the input size and batchsize by the --size and --batchsize option. large input size lead to high performance, please set to larger as possible (depending on your gpu memory).
if you have only single gpu
python main.py --data_dir your_path --gpus 0 0 0
in this case, due to the limitation of gpu memory, the input size must be to smaller.
but even if input size is smaller than 100, pspnet shows good result.
after training 100 epochs, run the code for testing
python main.py --mode test --data_dir your_path --gpus 0 0 0 --load pspnet 100
test is on single gpu