search

hzwer / ICCV2019-LearningToPaint

ICCV2019 - Learning to Paint With Model-based Deep Reinforcement Learning
MIT License
2.25k stars 312 forks source link

Renderer input features #14

Closed tmquan closed 5 years ago

tmquan commented 5 years ago

Hi @hzwer , Could you clarify the input feature of the neural renderer as it is 10-value vector or 13-value vector (+RGB). If training with 10-value vector, how the painter can generate color pictures?

Bests,

hzwer commented 5 years ago

The renderer uses a 10-dimentional vector to generate grayscale strokes. The agent multiplies the other 3-dimentional action by the grayscale image to generate colorful images.

AhmedAkram96 commented 4 years ago

hi @hzwer Thank you for the great work in the paper and the code ! I have a question that is a little bit related to this issue.

I'm trying to render the output actions of the agent (actor) in the learning to paint project using the fluid renderer of spiral ( https://github.com/deepmind/spiral ), So the issue I'm currently facing is understanding the specific meaning of the agent output actions parameters, so I can map it to to the input actions of the other renderer. I understand that the last 3 parameters are for the RGB; However, I'm not sure I understand how the first 10 parameters represent the shape, position and transparency?

please note that the other renderer is non-differentiable

hzwer commented 4 years ago

Hi @AhmedAkram96, please check https://github.com/megvii-research/ICCV2019-LearningToPaint/blob/master/baseline/Renderer/stroke_gen.py to know the meaning of actions parameters.