cGAN

Conditional GANs implemented in PyTorch

The paper: A Study of Smoothing Methods for Language Models Applied to Information Retrieval

This work tries to reproduce cGAN (Conditional GAN) refer only to the paper, without any later tricks but an alternative heuristic loss function. See model1.py

Discrinator	Generator
X Input ∈ R^784; y input ∈ $r^10$	z Input ∈ $r^100$; y input ∈ $r^10$
240,5 Maxout; OneHot. 50,5 Maxout.	FC. 1000 RELU; OneHot. FC. 200 RELU
Dropout(0.5). 290,4 Maxout.	Dropout(0.5). FC. 784 RELU. Sigmoid
Dropout(0.5). FC. Sigmoid.

Hype-parameters	learning rate	decay facotr	momentum	optimizer
Values	0.1->0.000001	1.00004	>0.7	SGD

model2.py is used as a contrast model which has additional Batch Normlization layer in Generator.

Discrinator	Generator
X Input ∈ R^784; y input ∈ $r^10$	z Input ∈ $r^100$; y input ∈ $r^10$
240,5 Maxout; OneHot. 50,5 Maxout.	FC. BatchNorm. 1000 RELU; OneHot. FC. BatchNorm. 200 RELU
Dropout(0.5). 290,4 Maxout.	Dropout(0.5). FC. 784 RELU. Sigmoid
Dropout(0.5). FC. Sigmoid.

Result

Name	epoch1	epoch10	epoch50	gif	remarks
Model1					It still need time to converge.
Model2					It shows that BN really herlps accelerating training.