Deepflow

A node based deep neural network library on top of CuDNN

Status

A work in progress. Not mature enough for a release but cool enough for poking around.

Features

• Node base like Tensorflow
• Multiple execution phases per graph
• Custom solver per variable
• Live weight/data display
• Create C++ code from Deepflow model

Requirments

• NVIDIA Graphics Card
• GPU Only

Dependencies

• Visual Studio 2015
• CUDA 9.0
• OpenCV 3.0
• cuDNN v7.1
• Protocol Buffers
• glog
• gflags

Working Examples

• face_dcgan: A least-square generative adversarial network.

• face_ac: Deep auto-encoder on 128x128 CelebA dataset.

• face_vae: Deep variational auto-encoder on 128x128 CelebA dataset.

• mnist_ac: Deep auto-encoder on MNIST dataset.

• mnist_dcgan: Least-square gan (generative adversarial) on MNIST dataset.

• mnist_lenet

Current Nodes

Nodes	Nodes	Nodes	Nodes
data_generator	variable	place_holder	conv2d
image_batch_generator	pooling	convolution_2d	transposed_conv2d
image_reader	add	square	matmult
mnist_reader	subtract	bias_add	dropout
argmax	argmin	reduce_max	reduce_min
reduce_mean	reduce_sum	reduce_absmax	reduce_norm1
reduce_norm2	leaky_relu	sigmoid	relu
tanh	clipped_relu	elu	phaseplexer
random_selector	softmax_loss	euclidean_loss	print
display	psnr	softmax	equal
cast_float	accumulator	batch_normalization	logger
negate	multiplexer	random_selector	restructure
softmax	...

Current Solvers

• AdaDelta
• Adam
• RMSProp
• SGD