Getting Started

Here we provide the basic tutorials about the usage SimREC. Make sure you’ve already install the environments for SimREC, please refer to Installation.

Training

In SimREC, we provide tools/train_engine.py and tools/eval_engine.py for training and evaluation.

The following script will start training simrec model on refcoco dataset on a single GPU:

$ bash tools/train.sh configs/simrec_refcoco_scratch.py 1

All of the checkpoints, logs and tensorboard logs will be saved to cfg.train.output_dir, you can modify them in the config file, we highly recommend the users to put their own config file under /configs to easily reuse the default config files:

# config.py
from .common.train import train

train.output_dir = "/your/own/path"

For distributed data parallel training, you can modify the training script as follows:

$ bash tools/train.sh configs/simrec_refcoco_scratch.py 4

To run ddp training mode by simply modifing the last number of the training scripts to 4.

Override the config in command line

You can override the config file in command line. For example, you can enable the SyncBatchNorm in scripts for ddp training like:

$ bash tools/train.sh configs/simrec_refcoco_scratch.py 4 train.sync_bn.enabled=True

which may give you a better result.

Resume training

In SimREC, we support two resume training ways adopted from Swin-Transformer:

• Automatically resume training:

SimREC automatically saves last_checkpoint.pth during training time to cfg.train.output_dir, if set cfg.train.auto_resume.enabled=True, before training, SimREC will find if there is last_checkpoint.pth in cfg.train.output_dir and automatically resume from it.

• Resume training from specific checkpoint:

Firstly, you should disable auto-resume function which will override the cfg.train.resume_path by setting cfg.train.auto_resume.enabled=False, and you should update cfg.train.resume_path to the specific checkpoint you want to resume from as follows:

# config.py
from .common.train import train

train.auto_resume.enabled=False
train.resume_path = "path/to/specific/checkpoint.pth"

Evaluation

Run bash tools/eval.sh under to evaluate the saved checkpoint.

bash tools/eval.sh config/simrec_refcoco_scratch.py 4 /path/to/checkpoint