ELEKTRONN2 a is highly configurable toolkit for training 3d/2d CNNs and general Neural Networks
ELEKTRONN2 is a flexible and extensible Python toolkit that facilitates design, training and application of neural networks.
It can be used for general machine learning tasks, but its main focus is on convolutional neural networks (CNNs) for high-throughput 3D and 2D image analysis.
ELEKTRONN2 is especially useful for efficiently assessing experimental neural network architectures thanks to its powerful interactive shell that can be entered at any time during training, temporarily pausing all calculations.
The shell interface provides shortcuts and autocompletions for frequently used operations (e.g. adjusting the learning rate) and also provides a complete python shell with full read/write access to the network model, the plotting subsystem and all training parameters and hyperparameters. Changes made in the shell take effect immediately, so you can monitor, analyse and manipulate your training sessions directly during their run time, without losing any training progress.
Computationally expensive calculations are automatically compiled and transparently executed as highly-optimized CUDA binaries on your GPU if a CUDA-compatible graphics card is available .
ELEKTRONN2 is written in Python (2.7 / and 3.4+) and is a complete rewrite of the previously published ELEKTRONN library. The largest improvement is the development of a functional interface that allows easy creation of complex data-flow graphs with loops between arbitrary points in contrast to simple “chain”-like models. Currently, the only supported platform is Linux (x86_64).
|||You can find out if your graphics card is compatible here. Usage on systems without CUDA is possible but generally not recommended because it is very slow.|
ELEKTRONN2 is being superceded by the more flexible, PyTorch-based elektronn3 library. elektronn3 is actively developed and supported, so we encourage you to use it instead of ELEKTRONN2 (if elektronn3’s more experimental status and currently less extensive documentation are acceptable for you).
Although other high-level libraries are available (Keras, Lasagne), they all lacked desired features and flexibility for our work, mostly in terms of an intuitive method to specify complicated computational graphs and utilities for training and data handling, especially in the domain of specialised large-scale 3-dimensional image data analysis for connectomics research (e.g. tracing, mapping and segmenting neurons in in SBEM  data sets).
Although the mentioned use cases are ELEKTRONN2’s specialty, it can be used and extended for a wide range of other tasks thanks to its modular object-oriented API design (for example, new operations can be implemented as subclasses of the Node class and seamlessly integrated into neural network models).
|||Serial Block-Face Scanning Electron Microscopy, a method to generate high resolution 3D images from small samples, such as brain tissue.|
Example visualisation of ELEKTRONN2’s usage on a 3D SBEM data set (blending input to output from left to right).
|Left (input: raw data)||Right (output: predictions by ELEKTRONN2, color-coded)|
|3D electron microscopy images of a zebra finch brain (area X dataset j0126 by Jörgen Kornfeld).||Probability of barriers (union of cell boundaries and extracellular space, marked in white) and mitochondria (marked in blue) predicted by ELEKTRONN2.|
See the installation instructions at https://elektronn2.readthedocs.io/en/latest/installation.html.
ELEKTRONN2 adds another abstraction layer to Theano. To create a model, the user has to connect different types of node objects and thereby builds a graph as with Theano. But the creation of the raw Theano graph, composed of symbolic variables and trainable model parameters, is hidden and managed through usage of sensible default values and bundling of stereotypical Theano operations into a single ELEKTRONN2 node. For example, creating a convolution layer consists of initialising weights, performing the convolution, adding the bias, applying the activation function and optional operations such as dropout or batch normalisation. Involved parameters might be trainable (e.g. convolution weights) or non-trainable but changeable during training (e.g. dropout rates).
- Perceptron / fully-connected / dot-product layer, works for arbitrary dimensional input
- Convolution, 1-,2-,3-dimensional
- Max/Average Pooling, 1,2,3-dimensional
- UpConv, 1,2,3-dimensional
- Max Fragment Pooling (MFP), 1,2,3-dimensional
- Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM) unit
- Recurrence / Scan over arbitrary sub-graph: support of multiple inputs multiple outputs and feed-back of multiple values per iteration
- Batch normalisation with automatic accumulation of whole data set statistics during training
- Gaussian noise layer (for Variational Auto Encoders)
- Activation functions: tanh, sigmoid, relu, prelu, abs, softplus, maxout, softmax-layer
- Local Response Normalisation (LRN), feature-wise or spatially
- Basic operations such as concatenation, slicing, cropping, or element-wise functions
- Bernoulli / Multinoulli negative log likelihood
- Gaussian negative log likelihood
- Squared Deviation Loss, (margin optional)
- Absolute Deviation Loss, (margin optional)
- Weighted sum of losses for multi-task training
- Stochastic Gradient Descent (SGD)
- Automatic creation of training directory to which all files (parameters, log files, previews etc.) will be saved
- Frequent printing and logging of current state, iteration speed etc.
- Frequent plotting of monitored states (error samples on training and validation data, classification errors and custom monitoring targets)
- Frequent saving of intermediate parameter states and history of monitored variables
- Frequent preview prediction images for CNN training
- Customisable schedules for non-trainable meta-parameters (e.g. dropout rates, learning rate, momentum)
- Fully functional python command line during training, usable for debugging/inspection (e.g. of inputs, gradient statistics) or for changing meta-parameters
- Randomised patch extraction from a list of of input/target image pairs
- Data augmentation trough histogram distortions, rotation, shear, stretch, reflection and perspective distortion
- Real-time data augmentation through a queue with background threads.
- Array interface for KNOSSOS data sets with caching, pre-fetching and support for multiple data sets as channel axis.
- Viewer for multichannel 3-dimensional image arrays within the Python runtime
- Function to convert ID images to boundary images
- Utilities needed for skeltonisation agent training and application
- Visualisation of the computational graph
- Class for profiling within loops
- KD Tree that supports append (realised through mixture of KD-Tree and brute-force search and amortised rebuilds)
- Daemon script for the synchronised start of experiments on several hosts, based on resource occupation.
The documentation is hosted at https://elektronn2.readthedocs.io/ (built automatically from the sources in the docs/ subdirectory of the code repository).
ELEKTRONN2 was funded by Winfried Denk’s lab at the Max Planck Institute of Neurobiology.
Jörgen Kornfeld was academic advisor to this project.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size elektronn2-0.6.0-py2.py3-none-any.whl (325.2 kB)||File type Wheel||Python version py2.py3||Upload date||Hashes View|
|Filename, size elektronn2-v0.6.0.tar.gz (325.0 kB)||File type Source||Python version None||Upload date||Hashes View|
Hashes for elektronn2-0.6.0-py2.py3-none-any.whl