This is a distributed training framework for supervised depth prediction based on Pytorch 1.0 (Pytorch version >= 1.2 is best). Now it provide the implementation of DORN(state of the art in KITTI depth prediction benchmark), and you can implement your model in your customed dataset with a little modification.
- Distributed & Single GPU Flexible selection between distributed training with multi gpus and a single gpu.
- Flexible Visulization Implementation You can implement your visulizers for network comprehensive analysis.
- Suport Various Optimizers and Learning-rate Policy Provide all the optimizers and learning-rate schedulers in pytorch. And support poly lr_scheduler and warmup, which are widely used in segmentation and detection.
- Support Grad Clip Provide grad clip to avoid gradient exploding.
- Mixed Precision Training Support mixed precision training with NVIDIA apex lib.
- Sync BN Support Sync BN when training with multi gps.
- Break-point Restoration Support continue to train from a break-point.