- FCN with VGG (pretrain model from pytorch)
- optimizer : option:['Adam','rmrprop','SGD']
- criterion : torch.nn.BCEWithLogitsLoss
- hyper params: reference config.py
- about the class num reduce part, please reference label_changer.py, where show how some label class mix to one class(such as car,bike,truck to vehicle)
- if need to change the class num distribution, please rewrite code in cityscape_loader.py(Cityscapes_Loader. init ,transform_classes)
- requirement list
CUDA
scipy = 0.19.1
pytorch
matplotlib
tensorboardX
PIL - change params in config.py (including the CityScapes datasets path)
- train with GPU(recommend)
single GPU runCUDA_VISIBLE_DEVICES=0 python train.py - run demo for your images
move your images to path ./demo/
runpython demo.py
- batch_size not too big, or else lead to GPU out of memory.
- the input images need to resize into (512,1024), or else cause out of memory.
- learn rate = 1 e-4 ,by using Adam
- Result image predict
- The fig4 is the trend of Loss and mIoU within 100 epochs. We can see that after 42epochs(got best mIoU) the train_loss still decrease but val_loss begin increase, this would lead to overfitting, need to early stopping in this epoch.
- The different class group may result in different accuracy, for example, person class and sky class fuse together may lead to stop reducing of Loss, and mIoU will stay in a level no matter how epochs train.Just like the fig5 show.
- Use Adam optimizer could significant increase regression speed. As fig6 show.
