CartoonGAN on the Web | aiarts.week.07 | Midterm Process Blog – Casey & Eric

Group Member  –  Casey, Eric

Github Repo  –  CartoonGAN-Application

Previous Report  –   Project Proposal

cartoonGAN_wa_intro_img

Methodology

Model Training

Web Application Interface

Experiments

Model Training

model_training_stage_1
stage snapshot #1
model_training_stage_2
stage snapshot #2

***Issues***

#1 CPU/GPU

During our training experiments, the first few attempts we made on training a model on Aku no Hana (Japan 2012) was surprisingly slow. The training speed was unreasonably low and we struggled a while to figure out why.

It turned out that we have been training our Aku no Hana model on the server’s CPU instead of GPU by accident. And as a result of the speed difference up to 50 times between CPU and GPU, our model were only trained over 40 batches in a day. After fixing this problem and switching to the server’s GPUs, the process went back to a normal speed, and we were able to see the training results within hours.

#2 Resource-specific data collection

As stated earlier, we collected our target cartoon dataset by making 256*256 screenshots every 5 seconds with ffmpeg. One of the problems with the dataset collected this way is that there existed a lot of unusable screenshots, that were either of a single color, or repeating the same content for multiple sequential screenshots. The reason for this issue is that the target anime work is 

model_training_stage_3
stage snapshot #3

&nbsp

Web Application

screensshot.01

User-Case Flow

user_case_flow.01

Leave a Reply