TRAIN_README.md

Training for TDNet

1. Prepare Data

Download and save datasets.

• Cityscapes: put leftImg8bit_sequence and gtFine in the same folder.
• Camvid
• NYUDepthV2 (~14GB)

2. Download Models

Download the pretained models for subnetworks, and put them into ./pretrained.

• td2_fa18: takes two FANet-18 as subnetworks.
• td2_psp50: takes two PSPNet-50 as subnetworks.
• td4_psp18: takes four PSPNet-18 as subnetworks.

3. Modify Codes

Modify *.yml files in ./config

• ''data:path'': path to dataset
• ''training:batch_size'': batch_size
• ''training:train_augmentations:rcrop'': input size for training

4. Training

Run

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train.py --config configs/*.yml

**Notice

Still working on this implementation.