Skip to content

Latest commit

 

History

History
74 lines (57 loc) · 2.77 KB

before_start.md

File metadata and controls

74 lines (57 loc) · 2.77 KB
Raw

Getting Started

we visualize our training details via wandb (https://wandb.ai/site).

visualization

  1. you'll need to login

    $ wandb login

    you can find you API key in (https://wandb.ai/authorize), and copy & paste it in terminal.

  2. you can (optionally) add the key to the main.py for the server use, with

    import os
os.environ['WANDB_API_KEY'] = "you key"

checkpoints

  1. for the deeplabv3+ experiments, we utilize exactly same checkpoints as provided by the CPS in here.
  2. for the pspnet experiments, we follow the CCT as provided in here.

training

The code is trained under 4xV100(32Gb) for the voc12 dataset, and 2xV100 (32Gb) for the cityscapes.

Our approach performs robust under different hardware's test, please see the training logs for more details.

VOC12 Setting

(global) we utilize batch_size=64 (32 labelled, 32 unlabelled data) for the training, with learning rate 1e-2, under 4 GPUs.

(local) in each GPU, we utilize batch_size=16 (8 labelled, 8 unlabelled data) under the learning rate 2.5e-3.

  1. augset experiment

    hyper-param 1/16 (662) 1/8 (1323) 1/4 (2646) 1/2 (5291)
    epoch 80 80 200 300
    weight 1.5 1.5 1.5 1.5

    run the scripts with

    # -l -> labelled_num; -g -> gpus; -b -> resnet backbone;
./scripts/train_voc_aug.sh -l 1323 -g 2 -b 101

  2. high-quality experiment

    hyper-param 1/16 (92) 1/8 (183) 1/4 (366) 1/2 (732)
    epoch 80 80 80 80
    weight 0.06 0.6 0.6 0.6

    run the scripts with

    # -l -> labelled_num; -g -> gpus; -b -> resnet backbone;
./scripts/train_voc_hq.sh -l 732 -g 2 -b 101

P.S., for 1464 high quality setting, our experiments show that, the training under half of the batch_size (i.e., GPU=2xV100) are likely to perform higher than the paper reported result.

Cityscapes Setting

we utilize batch_size=16 (8 labelled, 8 unlabelled data) under the learning rate 4.5e-3. (I have to reduce the batch size, as our dgx faced some issues in 2021 Nov. )

hyper-param 1/8 (372) 1/4 (744) 1/2 (1488)
epoch 320 450 550
weight 3.0 3.0 3.0

run the scripts with

# -l -> labelled_num; -g -> gpus; -b -> resnet backbone;
./scripts/train_city.sh -l 372 -g 2 -b 50