BatchFlow

BatchFlow helps you conveniently work with random or sequential batches of your data and define data processing and machine learning workflows even for datasets that do not fit into memory.

For more details see the documentation and tutorials.

Main features:

flexible batch generaton
deterministic and stochastic pipelines
datasets and pipelines joins and merges
data processing actions
flexible model configuration
within batch parallelism
batch prefetching
ready to use ML models and proven NN architectures
convenient layers and helper functions to build custom models
a powerful research engine with parallel model training and extended experiment logging.

Basic usage

my_workflow = my_dataset.pipeline()
              .load('/some/path')
              .do_something()
              .do_something_else()
              .some_additional_action()
              .save('/to/other/path')

The trick here is that all the processing actions are lazy. They are not executed until their results are needed, e.g. when you request a preprocessed batch:

my_workflow.run(BATCH_SIZE, shuffle=True, n_epochs=5)

for batch in my_workflow.gen_batch(BATCH_SIZE, shuffle=True, n_epochs=5):
    # only now the actions are fired and data is being changed with the workflow defined earlier
    # actions are executed one by one and here you get a fully processed batch

NUM_ITERS = 1000
for i in range(NUM_ITERS):
    processed_batch = my_workflow.next_batch(BATCH_SIZE, shuffle=True, n_epochs=None)
    # only now the actions are fired and data is changed with the workflow defined earlier
    # actions are executed one by one and here you get a fully processed batch

Train a neural network

BatchFlow includes ready-to-use proven architectures like VGG, Inception, ResNet and many others. To apply them to your data just choose a model, specify the inputs (like the number of classes or images shape) and call train_model. Of course, you can also choose a loss function, an optimizer and many other parameters, if you want.

from batchflow.models.torch import ResNet34

my_workflow = my_dataset.pipeline()
              .init_model('model', ResNet34, config={'loss': 'ce', 'classes': 10})
              .load('/some/path')
              .some_transform()
              .another_transform()
              .train_model('ResNet34', inputs=B.images, targets=B.labels)
              .run(BATCH_SIZE, shuffle=True)

For more advanced cases and detailed API see the documentation.

Installation

BatchFlow module is in the beta stage. Your suggestions and improvements are very welcome.

BatchFlow supports python 3.6 or higher.

Stable python package

With poetry

poetry add batchflow

With old-fashioned pip

pip3 install batchflow

Development version

With poetry

poetry add --editable git+https://github.com/analysiscenter/batchflow

With old-fashioned pip

pip install --editable git+https://github.com/analysiscenter/batchflow

Extras

Some batchflow functions and classed require additional dependencies. In order to use that functionality you might need to install batchflow with extras (e.g. batchflow[nn]):

image - working with image datasets and plotting
nn - for neural networks (includes torch, torchvision, ...)
datasets - loading standard datasets (MNIST, CIFAR, ...)
profile - performance profiling
jupyter - utility functions for notebooks
research - multiprocess research
telegram - for monitoring pipelines via a telegram bot
dev - batchflow development (pylint, pytest, ...)

You can install several extras at once, like batchflow[image,nn,research].

Projects based on BatchFlow

SeismiQB - ML for seismic interpretation
SeismicPro - ML for seismic processing
PyDEns - DL Solver for ODE and PDE
RadIO - ML for CT imaging
CardIO - ML for heart signals

Citing BatchFlow

Please cite BatchFlow in your publications if it helps your research.

Roman Khudorozhkov et al. BatchFlow library for fast ML workflows. 2017. doi:10.5281/zenodo.1041203

@misc{roman_kh_2017_1041203,
  author       = {Khudorozhkov, Roman and others},
  title        = {BatchFlow library for fast ML workflows},
  year         = 2017,
  doi          = {10.5281/zenodo.1041203},
  url          = {https://doi.org/10.5281/zenodo.1041203}
}