datapipelines

Iterable datapipelines for pytorch training.

The functions sdata.create_dataset() and sdata.create_loader() provide interfaces for your pytorch training code, where the former returns a dataset and the latter a wrapper around a pytorch dataloader.

A dataset as returned by sdata.create_dataset() consists of 5 main modules should be defined in a yaml-config:

1. A base datapipeline, which reads data as tar files from local fs and assembles them to samples. Each sample comes as a python-dict.
2. A list of preprocessors which can be either used to transform the entries of a sample or to filter out unsuitable samples. The former kinds are called mappers, the latter filters. This repository provides a basic set of mappers and filters which provide basic (not too application specific) data transforms and filters.
3. A list of decoders whose elements can be either defined as a string matching one of the predefined webdataset image decoders decoders or some custom decoder (in the config-style) for handling more specific needs. Note that decoding will be skipped alltogether when setting decoders=None (or in config-style yaml decoders: null).
4. A list of postprocessors which are used to filter or transform the data after it has been decoded and should again be either mappers or filters.
5. error_handler: A webdataset-style function for handling any errors which occur in the datapipeline, preprocessors, decoders or postprocessors.

A wrapper around a pytorch dataloader, which can be plugged in to your training, is returned by sdata.create_loader(). You can pass the dataset either as an IterableDataset as returned by sdata.create_dataset() or via the config which would instantiate this dataset. Apart from the known batch_size, num_workers, partial and collation_fn parameteters for pytorch dataloaders, the function can be configured via the following arguments.

1. batched_transforms of batched mappers and filters which transform an entire training batch before being passed to the dataloader defined in the same style than the preprocessors and postprocessors from above.
2. loader_kwargs defining additional keyword arguments for the dataloader (such as prefetch_factor, ...)
3. error_handler: A webdataset-style function for handling any errors which occur in the batched_transforms.

Examples

Here, it is most effective to look at the configs in examples/configs/ for the following examples. These will show you how this works.

For a simple example, see examples/image_simple.py, find config here.

NOTE: You have to add your dataset in tar-form which should follow the webdataset-format. To find the parts which have to be adapted, search for comments conaining USER: in the respective config.

Installation

Pytorch 2 and later

python3 -m venv .pt2
source .pt2/bin/activate
pip3 install wheel
pip3 install -r requirements_pt2.txt

Pytorch 1.13

python3 -m venv .pt1
source .pt1/bin/activate
pip3 install wheel
pip3 install -r requirements_pt1.txt