Keras Sequence Video Generators

This package offers classes that generate sequences of frames from video files using Keras (officially included in TensorFlow as of the 2.0 release). The resulting frame sequences work with the Time Distributed, GRU, LSTM, and other recurrent layers.

See articles:

• The basics of Video frame as input
• Usage of these generators here

An provided example of usage can be displayed in nbviewer here.

Requirements are:

• Python >= 3.6 (Python 2 will never been supported)
• OpenCV
• numpy
• Keras >= 2
• TensorFlow >= 1.15 (or other backend, not tested, TensorFlow is needed by Keras)

TensorFlow 2 works as well. This requirements is not integrated in the setup.py to let you choose the version, or to let you try with other backend. We mean that you will need to install a backend yourself (e.g. pip install tensorflow)

If you want to compile the package, you need:

• sphinx to compile doc (work in progress)
• setuptools

Installation

You can install the package via pip:

pip install keras-video-generators

If you want to build from source, clone the repository then:

python setup.py build

Usage

The module name (keras_video) is different from the installation package name (keras-video-generators). Import the entire module with

import keras_video

or load a single generator:

from keras_video import VideoFrameGenerator

The package contains three generators that inherit the Sequence interface and may be used with model.fit_generator():

• VideoFrameGenerator provides a pre-determined number of frames from the entire video
• SlidingFrameGenerator provides frames with decay for the entire video or with a sequence time
• OpticalFlowGenerator provides an optical flow sequence from frames with different methods (experimental)

Each generator accepts a standard set of parameters:

• glob_pattern; must contain {classname}, e.g. './videos/{classname}/*.avi' - the "classname" in string is used to detect classes
• nb_frames; the number of frames in the sequence
• batch_size; the number of sequences in one batch
• transformation; can be None or ImageDataGenerator (Keras) for data augmentation
• use_frame_cache; use with caution, if set to True, the class will keep frames in memory (without augmentation). You need a lot of memory

See the class documentation for all parameters.

Changelog

v1.1.0

• TensorFlow 2.5 compatibility
• Code cleanup
• Fixup some unbound variables

v1.0.14

• Changes to get first and last frames in sequence

v1.0.13

• try to fix SageMaker problem by avoiding usage of internal keras from tensorflow

v1.0.12

• fix transformation error with SlidingFrameGenerator

v1.0.11

• set generator to be Iterable
• frame cache was disabled by error, it's back now
• fixup import Sequence from tensorflow.keras
• fix frame count problems for video with bad headers

v1.0.10

• fix Windows problems with path using backslashes
• add auto discovery for classes if "None" is sent
• add travis tests

v1.0.9

• fix frame counter in SlidingFrameGenerator

v1.0.8

• fix tiny video frame count
• refactorisation
• pep8
• fix problems for video without headers

v1.0.7

• fix name check in classes to get them from filename
• add split_test and split_val

v1.0.5

• fix package generation

v1.0.4

• fix assertion