DLP GAN Implementation
What
Final project for CS5173_DeepLearning. The objective was to implement and evaluate Draw Landscape Pictures (DLP) GAN model proposed for generating Chinese landscape paintings.
Why
Adversarial learning has been my area of interest and after learning about Generative Adversarial Network (GAN) in this course, i wanted to explore how we can leverage this technology to re-imagine art under a different lens.
We’ve seen an extensive amount of GAN models trained on prominent Western art culture (Van Gogh, Claude Monet, …) but research and application on Asian art style are minimal in the current landscape. This is why the DLP paper stood out to me as an interesting topic to dive deeper. The novelty of this model came from its Dual-Consistency Loss, which aims to balance realism and abstraction by enforcing consistency at the semantic and feature levels.
How
Since the paper provides no official code implementation, i proceeded with a full re-implentation in PyTorch based on the methodology discussed. After some 80+ hours of researching, coding, training and result synthesising later. The re-implmented model show results that were comparable to the orignial paper in terms of structural similiarity score but with fewer training sample and less training time.
Results
Trained on 1 RTX ADA6000 for 20 hours with 2192 high quality Chinese landscape painting and 2000 high quality landscape photos.