Closed OrNot closed 6 years ago
Most of the data pixels in the image are empty so small feature map should not hinder the accuracy or speed. I personally believe stripping away some data was necessary as we are processing images real-time as we perform actions. I have in fact got satisfactory results with an image input of mere 20x40 as well. That being said, please try out and let us know if you find a better configuration. The current training configuration requires at least 1 million steps, which is quite OK in case of reinforcement learning but if we can reduce that somehow that's certainly an achievement.
Thanks for your reply. I am following your codes and try to re-implement with pytorch. I modify the model to align the architecture as proposed in the paper double-dqn. I have run the loops nearly 20000 but the highest score is still under 100. Is it a normal training or should I cease the process to double check my code ?
20000 is a small count to judge. I would suggest keep it for another 20000. If you see no progress, check your code.
The 3 layers of conv with proposed values of stride , pooling .ect. leads the feature map to be small too fast ( supposed 80*80 input). is there any other settings I missed ?
Thanks a lot