The program performs semantic segmentation on KITTI dataset.
Data loader, model, multi-GPU training, evaluation and visualization are all involved in the implementation. The data loader and model part are flexible.
Install the requirements first.
pip install -r requirements.txtTo run the code, read the help first. Or, you can directly run train.py and see what will happen.
python train.py -h
usage: train.py [-h] [--batch-size N] [--split Split] [--resize-ratio Resize]
[--numloader Nl] [--epochs E] [--lr Lr] [--lr-pretrain LR-P]
[--momentum M] [--decay D] [--step Step] [--gamma Gamma]
[--logdir Log] [--vgg Vgg] [--fcn FCN] [--mode Mode]
[--model Model]
pytorch FCN on Kitti data set
optional arguments:
-h, --help show this help message and exit
--batch-size N the training batch size(default: 8)
--split Split the split ratio indicates how much percentage of train
set, the rest will for validation (default: 0.8)
--resize-ratio Resize
how much ratio to resize(shrink) the training
image(default: 0.6)
--numloader Nl the num of CPU for data loading. 0 means only use one
CPU. (default: 8)
--epochs E the required total training epochs(default: 100)
--lr Lr the learning rate for decoder part(default: 1e-3)
--lr-pretrain LR-P the learning rate for encoder(pre-trained)
part(default: 1e-4)
--momentum M the momentum of optimizer(default: 0)
--decay D the weight for L2 regularization(default: 1e-5)
--step Step learning rate will decay after the step(default: 10)
--gamma Gamma learning rate will decay gamma percent after few
steps(default: 0.5)
--logdir Log the folder to store the tensorboard logs(default: log)
--vgg Vgg the configuration of vgg(default: 11)
--fcn FCN the configuration of FCN(default: 1)
--mode Mode train or test(default: train)
--model Model the pre-trained model or the model for test. if not
specify, it will initial a new model (default: )
The training model of each epoch will be stored on your drive (in ./store folder).
Run tensorboard to tracing the training state:
tensorboard --logdir log