This repository provides a training pipeline that can be used for both image and video datasets. The dataset and training parameters can be easily modified through the provided arguments.

By changing the --dataset_name argument in train.sh, you can select between the following datasets:

• "HuggingFaceM4/COCO" for COCO images.
• "webvid" for WebVid videos.

You can use any video dataset as long as they follow the path format:

• --num_train_data and --num_val_data: These arguments are only relevant for WebVid datasets. They specify the number of data samples to use for training and validation, respectively.
• --num_result: Specifies the number of results to visualize during each validation step.

In the train.sh file, adjust the dataset argument like this to use the COCO dataset:

--dataset_name "HuggingFaceM4/COCO"

Training Options:
--num_train_data: Number of training samples (for WebVid datasets).
--num_val_data: Number of validation samples (for WebVid datasets).
--num_result: Number of validation results to visualize.
--exp_name: Name of the experiment, useful for organizing multiple runs.
--lr_mult: Learning rate multiplier for affine layers.
--train_batch_size: Batch size for training.

The rest of the arguments remain consistent with the Wouaf framework.

• Run train.sh