I just started playing with Chainer which is a deep learning package in Python (core numerical computation is written in C++/CUDA) developed by Preferred Networks. It seems to be useful for fast prototyping of neural networks research, especially for recurrent things, due to "define-by-run" strategy and dynamically allocating memory. Maybe I am going to put my stuff for extending Chainer to this repo.

• Chainer >= 1.1.1
• chainer-cuda-deps
• future
• futures

This is defined as an inverted operation of Convolution which is effectively used in Fully Convnets for (learned) upsampling strided output maps. I referred to an implementation in Caffe to implement this.

Maximum/Minimum takes two arrays with the same shape and behaves like numpy.maximum and cuda.gpuarray.maximum. Since these are not parameterized, you can use the shortcut functions maximum and minimum in the layers module.

Utility functions/classes are in the utils module.

A generator function which yields a tuple of mini-batches infinitely. Each generator call create threads for prefetching data in the background. See doc.

This transforms an array with shape of (b, c, h, w) into a gray image consists of tiled b * n images except for c=3 case, which outputs a color image with tiled b images. This is useful for visualizing inputs, outputs, feature maps and filters in vision tasks.

Go to tests folder, then run:

PYTHONPATH=.. nodetests