GulpIO Save Abandoned
Binary storage format for deep learning on videos.
====== GulpIO
About
Binary storage format for deep learning on videos.
Status
.. image:: https://travis-ci.org/TwentyBN/GulpIO.svg?branch=master :target: https://travis-ci.org/TwentyBN/GulpIO
Install
There are many ways to setup a Python environment for installation. Here we outline an approach using virtualenvironment. Note: this package does not support legacy Python and will only work with Python 3.x.
The following will setup a virtualenvironment and activate it:
.. code::
python -m venv gulpio
source gulpio/bin/activate
Then, install the package using:
.. code::
pip install gulpio
Usage
'Gulp' a Dataset
The gulpio package has been designed to be infinitely hackable and to support
arbitrary datasets. To this end, we are providing a so-called adapter
pattern. Specifically there exists an abstract class in gulpio.adapters:
the AbstractDatasetAdapter. In order to ingest your dataset, you basically
need to implement your own custom adapter that inherits from this.
You should be able to get going quickly by looking at the following examples, that we use internally to gulp our video datasets.
- The class:
gulpio.adapters.Custom20BNJsonAdapteradapters.py <src/main/python/gulpio/adapters.py>_ - The script:
gulp_20bn_json_videoscommand line script <src/main/scripts/gulp_20bn_json_videos>_
And an example invocation would be:
.. code::
gulp_20bn_json_videos videos.json input_dir output_dir
...
Additionally, if you would like to ingest your dataset from the command line,
the register_adapter script can be used to generate the command line interface
for the new adapter. Write your adapter that inherits from the AbstractDatasetAdapter
in the adapter.py file, then simply call:
.. code::
gulp_register_adapter gulpio.adapters <NewAdapterClassName>
The script that provides the command line interface will be in the main directory of the repository. To use it, execute ./new_adapter_class_name.
Sanity check the 'Gulped' Files
A very basic test to check the correctness of the gulped files is provided by the gulp_sanity_check script.
For execution run:
.. code::
gulp_sanity_check <folder containing the gulped files>
It tests:
- The presence of any content in the
.gulpand.gmeta-files - The file size of the
.gulpfile corresponds to the required file size that is given in the.gmetafile - Duplicate appearances of any video-ids
The file names of the files where any test fails will be printed. Currently no script to fix possible errors is provided, 'regulping' is the only solution.
Read a 'Gulped' Dataset
In order to read from the gulps, you can let yourself be inspired by the following snippet:
.. code:: python
# import the main interface for reading
from gulpio import GulpDirectory
# instantiate the GulpDirectory
gulp_directory = GulpDirectory('/tmp/something_something_gulps')
# iterate over all chunks
for chunk in gulp_directory:
# for each 'video' get the metadata and all frames
for frames, meta in chunk:
# do something with the metadata
for i, f in enumerate(frames):
# do something with the frames
pass
Alternatively, a video with a specific id can be directly accessed via:
.. code:: python
# import the main interface for reading
from gulpio import GulpDirectory
#instantiate the GulpDirectory
gulp_directory = GulpDirectory('/tmp/something_something_gulps')
frames, meta = gulp_directory[<id>]
For down-sampling or loading only a part of a video, a python slice can be passed as well:
.. code:: python
frames, meta = gulp_directory[<id>, slice(1,10,2)]
or:
.. code:: python
frames, meta = gulp_directory[<id>, 1:10:2]
Loading Data
You can use GulpIO data iterator and augmentation functions to load GulpIO dataset into memory.
For a working example given on different deep learning libraries please refer to examples/GulpIOTrainingExample.ipynb.
We provide archetypical dataset wrappers that work for general supervised cases of image and video datasets. If you need a particular use,
you might need to create your own dataset by inheriting dataset.py and overwriting __getitem__ and __len__.
Below is an example loading an image dataset with GulpIO loader and defining augmentation pipeline. Transformations are applied to each instance on the fly. Some transformations have separate video and image versions since some of the augmentations need to be aligned video-wise.
.. code:: python
from gulpio.dataset import GulpImageDataset
from gulpio.loader import DataLoader
from gulpio.transforms import Scale, CenterCrop, Compose, UnitNorm
# define data augmentations. Notice that there are different functions for videos and images
transforms = Compose([
Scale(28), # resize image by the shortest edge
CenterCrop(28),
UnitNorm(), # instance wise mean and std norm
])
# define dataset wrapper and pick this up by the data loader interface.
dataset = GulpImageDataset('/path/to/train_data', transform=transforms)
loader = DataLoader(dataset, batch_size=256, shuffle=True, num_workers=0, drop_last=True)
dataset_val = GulpImageDataset('/path/to/validation_data/', transform=transforms)
loader_val = DataLoader(dataset_val, batch_size=256, shuffle=True, num_workers=0, drop_last=True)
Here we iterate through the dataset we loaded. Iterator returns data and label as numpy arrays. You might need to cast these into the format of you deep learning library.
.. code:: python
for data, label in loader:
# train your model here
# ...
Format Description
When gulping a dataset, two different files are created for every chunk: a
*.gulp data file that contains the actual data and a *.gmeta meta file
that contains the metadata.
The layout of the *.gulp file is as follows:
.. code::
|-jpeg-|-pad-|-jpeg-|-pad-|...
Essentially, the data file is simply a series of concatenated JPEG images, i.e. the frames of the video. Each frame is padded to be divisible by four bytes, since this makes it easier to read JPEGs from disk.
Here is a more visual example:
.. image:: docs/data_file_layout.png
As you can see there are 6 records in the example. They have the following paddings and lengths:
===== ===== ===== FRAME LEN PAD ===== ===== ===== 0 4 1 1 4 2 2 4 0 3 4 1 4 4 3 5 8 1 ===== ===== =====
The layout of the meta file is a mapping, where each id representing a
video is mapped to two further mappings, meta_data, which contains
arbitrary, user-defined meta-data. And a triplet, frame_info, which
contains the offset (index) into the data file, the number of bytes used for
padding and the total length of the frame (including padding). ([<offset>, <padding>, <total_length>].) The frame_info is required to recover the
frames from the data file.
.. code::
'id'
|
|-> meta_data: [{}]
|
|-> frame_info: [[], [], ...]
.
.
.
By default, the meta file is serialized in JSON format.
For example, here is a meta file snippet:
.. code::
{"702766": {"frame_info": [[0, 3, 7260],
[7260, 3, 7252],
[14512, 2, 7256],
[21768, 2, 7260],
[29028, 1, 7308],
[36336, 1, 7344],
[43680, 0, 7352],
[51032, 1, 7364],
[58396, 0, 7348],
[65744, 1, 7352],
[73096, 1, 7352],
[80448, 1, 7408],
[87856, 1, 7400],
[95256, 0, 7376],
[102632, 1, 7384],
[110016, 2, 7404],
[117420, 0, 7396],
[124816, 1, 7400],
[132216, 2, 7428],
[139644, 1, 7420],
[147064, 0, 7428],
[154492, 2, 7472],
[161964, 3, 7456],
[169420, 2, 7444],
[176864, 2, 7436]],
"meta_data": [{"label": "something something",
"id": 702766}]},
"803959": {"frame_info": [[184300, 1, 9256],
[193556, 3, 9232],
[202788, 2, 9340],
[212128, 2, 9184],
[221312, 1, 9112],
[230424, 3, 9100],
[239524, 0, 9144],
[248668, 1, 9120],
[257788, 0, 9104],
[266892, 0, 9220],
[276112, 1, 9140],
[285252, 1, 9076],
[294328, 2, 9100],
[303428, 0, 9224],
[312652, 3, 9200],
[321852, 3, 9136],
[330988, 2, 9136],
[340124, 1, 9152],
[349276, 0, 8984],
[358260, 1, 9048],
[367308, 0, 9116],
[376424, 1, 9136],
[385560, 1, 9108],
[394668, 2, 9084],
[403752, 1, 9112],
[412864, 2, 9108]],
"meta_data": [{"label": "something something",
"id": 803959}]},
"803957": {"frame_info": [[421972, 2, 8592],
[430564, 1, 8608],
[439172, 2, 8872],
[448044, 3, 8852],
[456896, 2, 8860],
[465756, 0, 8908],
[474664, 2, 8912],
[483576, 1, 8884],
[492460, 1, 8752],
[501212, 3, 8692],
[509904, 0, 8612],
[518516, 0, 8816],
[527332, 2, 8784],
[536116, 1, 8840],
[544956, 1, 8844],
[553800, 1, 8988],
[562788, 0, 8992],
[571780, 0, 8972],
[580752, 3, 9044],
[589796, 2, 9012],
[598808, 3, 9060],
[607868, 2, 9032],
[616900, 1, 9052],
[625952, 2, 9056],
[635008, 0, 9084],
[644092, 2, 9100]],
"meta_data": [{"label": "something something",
"id": 803957}]},
"773430": {"frame_info": [[653192, 1, 7964],
[661156, 2, 7996],
[669152, 1, 7960],
[677112, 0, 8024],
[685136, 0, 8008],
[693144, 1, 7972],
[701116, 0, 7980],
[709096, 0, 8036],
[717132, 0, 8016],
[725148, 0, 8016],
[733164, 1, 8004],
[741168, 1, 8008],
[749176, 1, 7996],
[757172, 1, 8016],
[765188, 1, 8032],
[773220, 0, 8040],
[781260, 2, 8044],
[789304, 2, 8004],
[797308, 1, 8008],
[805316, 0, 8056],
[813372, 3, 8088],
[821460, 0, 8044]],
"meta_data": [{"label": "something something",
"id": 773430}]},
"803963": {"frame_info": [[829504, 2, 8952],
[838456, 1, 8928],
[847384, 0, 8972],
[856356, 1, 8992],
[865348, 1, 8936],
[874284, 1, 8992],
[883276, 3, 8988],
[892264, 1, 9008],
[901272, 2, 8996],
[910268, 2, 8976],
[919244, 0, 9180],
[928424, 0, 9128],
[937552, 2, 9100],
[946652, 2, 9096],
[955748, 3, 9044],
[964792, 0, 9096],
[973888, 2, 9068],
[982956, 1, 8996],
[991952, 3, 8928],
[1000880, 1, 9040],
[1009920, 0, 9084],
[1019004, 0, 9076],
[1028080, 2, 9056],
[1037136, 2, 9040],
[1046176, 2, 9052],
[1055228, 3, 9096]],
"meta_data": [{"label": "something something",
"id": 803963}]}
}
Benchmarks
- Benchmarks are available in a seperate repo: https://github.com/TwentyBN/GulpIO-benchmarks
Prior Art
- Inspired by: MXNet based RecordIO: http://mxnet.io/architecture/note_data_loading.html
- GulpIO data loader is branched from great
PyTorch <http://pytorch.org>_ implementation.
License
All code except gulpio.loader is Copyright (c) Twenty Billion Neurons and
licensed under the MIT License, see the file LICENSE.txt for details.
The code in gulpio.loader which came from the PyTorch project is licensed
under a 3-clause BSD License, see the file LICENSE_PYTORCH for details.
Open Source Agenda Badge
