LifangHe / TGAN

Code for the paper "Tensorizing Generative Adversarial Nets"

TGAN

Code for the paper "Tensorizing Generative Adversarial Nets"

Tensor Layer

Given an input tensor X, we apply multilinear transformation to it, then we perform element-wise activations to form a tensor layer. For details, refer to the paper [1]. The function 'tensor_layer(tensor, matrices, bias, activation_function)' in 'tnsr.py' defines the presented tensor layer which forms TGAN. Refer hyperparameter.png for hyperparameters used for the paper.

def tensor_layer(tensor, matrices, bias, activation_function):
	"""
	INPUT: tensor, matrices, bias and activation function
		   tensor: tensorflow obejct 
		   matrices: list of matrix (tf object again)
		   bias: you know
		   activation function: function such as tf.nn.relu
	OUTPUT: tensor-layer
			mode-dot operation is applied and it changes
			the dimensions of the tensor by contraction operations
	"""

Example

To run an example of tensorized GAN with MNIST dataset, run the following

python tgan.py

Libraries

• Python 2.7.14
• Tensorflow 1.1.0
• Numpy 1.12.1
• Matplotlib 2.1.0

Reference

[1] Cao, Xingwei, and Qibin Zhao. "Tensorizing Generative Adversarial Nets." arXiv preprint arXiv:1710.10772 (2017).