Tutorial 2: Finetuning Models

Flow estimators pre-trained on the FlyingChairs and FlyingThings3d can serve as a good pre-trained model for other datasets. This tutorial provides instruction for users to use the models provided in the Model Zoo for other datasets to obtain better performance. MMFlow also provides out-of-the-box tools for training models. This section will show how to train predefined models on standard datasets.

Modify training schedule

The fine-tuning hyper-parameters vary from the default schedule. It usually requires smaller learning rate and less training iterations.

# optimizer
optimizer = dict(type='Adam', lr=1e-5, weight_decay=0.0004, betas=(0.9, 0.999))
optimizer_config = dict(grad_clip=None)
# learning policy
lr_config = dict(
        45000, 65000, 85000, 95000, 97500, 100000, 110000, 120000, 130000,
runner = dict(type='IterBasedRunner', max_iters=150000)
checkpoint_config = dict(by_epoch=False, interval=10000)
evaluation = dict(interval=10000, metric='EPE')

Use pre-trained model

Users can load a pre-trained model by setting the load_from field of the config to the model’s path or link. The users might need to download the model weights before training to avoid the download time during training.

# use the pre-trained model for the whole PWC-Net
load_from = ''  # model path can be found in model zoo

Training on a single GPU

We provide tools/ to launch training jobs on a single GPU. The basic usage is as follows.

python tools/ \
    ${CONFIG_FILE} \
    [optional arguments]

During training, log files and checkpoints will be saved to the working directory, which is specified by work_dir in the config file or via CLI argument --work-dir.

This tool accepts several optional arguments, including:

  • --work-dir ${WORK_DIR}: Override the working directory.

  • --resume-from ${CHECKPOINT_FILE}: Resume from a previous checkpoint file.

  • --cfg-option: Override some settings in the used config, the key-value pair in xxx=yyy format will be merged into config file. For example, ‘–cfg-option model.encoder.in_channels=6’.


Difference between resume-from and load-from:

resume-from loads both the model weights and optimizer status, and the iteration is also inherited from the specified checkpoint. It is usually used for resuming the training process that is interrupted accidentally. load-from only loads the model weights and the training iteration starts from 0. It is usually used for finetuning.

Training on multiple GPUs

MMFlow implements distributed training with MMDistributedDataParallel.

We provide tools/ to launch training on multiple GPUs. The basic usage is as follows.

bash ./tools/ \
    ${CONFIG_FILE} \
    ${GPU_NUM} \
    [optional arguments]

Optional arguments remain the same as stated above and has additional arguments to specify the number of GPUs.

Launch multiple jobs simultaneously

If you would like to launch multiple jobs on a single machine, e.g., 2 jobs of 4-GPU training on a machine with 8 GPUs, you need to specify different ports (29500 by default) for each job to avoid communication conflict.

If you use to launch training jobs, you can set the port in commands.

CUDA_VISIBLE_DEVICES=0,1,2,3 PORT=29500 ./tools/ ${CONFIG_FILE} 4
CUDA_VISIBLE_DEVICES=4,5,6,7 PORT=29501 ./tools/ ${CONFIG_FILE} 4

Training on multiple nodes

MMFlow relies on torch.distributed package for distributed training. Thus, as a basic usage, one can launch distributed training via PyTorch’s launch utility.

Manage jobs with Slurm

Slurm is a good job scheduling system for computing clusters. On a cluster managed by Slurm, you can use to spawn training jobs. It supports both single-node and multi-node training.

The basic usage is as follows.


Below is an example of using 8 GPUs to train PWC-Net on a Slurm partition named dev, and set the work-dir to some shared file systems.

GPUS=8 ./tools/ dev pwc_chairs configs/pwcnet/ work_dir/pwc_chairs

You can check the source code to review full arguments and environment variables.

When using Slurm, the port option need to be set in one of the following ways:

  1. Set the port through --cfg-options. This is more recommended since it does not change the original configs.

    GPUS=4 GPUS_PER_NODE=4 ./tools/ ${PARTITION} ${JOB_NAME} ${WORK_DIR} --cfg-options 'dist_params.port=29500'
    GPUS=4 GPUS_PER_NODE=4 ./tools/ ${PARTITION} ${JOB_NAME} ${WORK_DIR} --cfg-options 'dist_params.port=29501'
  2. Modify the config files to set different communication ports.

    In, set

    dist_params = dict(backend='nccl', port=29500)

    In, set

    dist_params = dict(backend='nccl', port=29501)

    Then you can launch two jobs with and

Read the Docs v: latest
On Read the Docs
Project Home

Free document hosting provided by Read the Docs.