Avoiding unnecessary casting

[VCasecnikovs]

No description provided.

[VCasecnikovs]

Q - value is overwritten

[toshikwa]

Have you compared the performance?

[VCasecnikovs]

Good morning,
I haven't checked the performance changes yet, I just started to train sacd on PacMan yesterday and for now, I can not get the same results...
After successful training, I'll try to do some benchmarks.

[VCasecnikovs]

Typo fix just changes cateorical to categorical : )

[VCasecnikovs]

Now it is possible to set render in your config.

[toshikwa]

I haven't checked the performance changes yet

I'm sorry, but I meant about the speed test.

I have compared the speed performance of "cast into uint8 -> transfer to GPU -> cast into float32"(master branch) with "cast into float32 -> transfer"(yours), but I also included here.
As a result, the former was much faster, so I don't feel like merging this commit.

I haven't had time to check the issues so far, but I really appreciate your contribution ;)
Thanks ;)

[VCasecnikovs]

Thank You for Your fast answer. The problem is that with such casting it could ruin floating points for some non-image data. So, maybe it could be better to add a flag, so the user would be able to choose from these two options.

[VCasecnikovs]

I have tested the performance, it is the same, results are in issue #22.

[toshikwa]

@VCasecnikovs
I'm sorry but I would not merge it because I concern about the training speed.
This repository only supports Atari env now, however, I am planning to make it compatible with other envs (after I finish my thesis).
At that time, I will be willing to merge this functionality as an optional argument.

I really appreciate your supports.
Thanks ;)

[VCasecnikovs]

@ku2482,
thank you for the answer.
The speed is also the same, I have added PrintScreen with relative to #22.
Best regards,
Vadims.