I installed the starter kit as instructed in the README on two different machines, and the track colour is a different pixel value on each in obs.top_down_rbg. This means my vision modules don't work across platforms. Do you know why this is? Can you confirm how I can get this setup to match competition specs? Thanks
Posted by: Team12 @ Nov. 9, 2019, 7:56 p.m.Do you mind sharing some technical details of the two machines?
relevant information would be:
- OS
- GPU models
- graphics drivers
- when you run the simulator, you should see a log lines that looks like this:
```
Known pipe types:
<list of graphics pipes, e.g. glxGraphicsPipe>
(all display modules loaded.)
```
Please share the output from your two machines.
Also screenshots from the two machines showing the differently colored roads would be very helpful. You can upload them to an image sharing website like imgur and share the links here.
Posted by: HuaweiUK @ Nov. 9, 2019, 10:39 p.m.Here are the two different images in env_obs.top_down_rgb:
https://ibb.co/w64s8ry
https://ibb.co/LzWnQkm
I know they are different tracks here but this occurs for the same scenario file on each machine.
One is Nvidia GTX 1060 GPU and one has no GPU (laptop).
Both say this pipeline
```
Warning: Deprecated vehicle classes 'rail_fast' in input network.
Known pipe types:
glxGraphicsPipe
(all display modules loaded.)
```
@HuwaeiTeam
Can you please confirm the track colour in the test env?
Posted by: Team12 @ Nov. 14, 2019, 12:38 p.m.My apologies for losing track of this thread,
The images your agent will receive during evaluation will look like the second image you linked (https://ibb.co/LzWnQkm).
If the track color differs from your more powerful machine which you would like to do training on, a temporary workaround would be to perform some image processing to discourage your models from over fitting to the track color, this can be done either through randomization of colors, or by inspecting the rgb values and normalizing them to a canonical track and background color.
We are stilling looking into the cause of the differences between machines. Good news is that we can replicate the differences on a few machines we have here to test with and so we're optimistic we can get a fix out soon.
Posted by: HuaweiUK @ Nov. 15, 2019, 3:21 p.m.Thanks for getting back to me! That's a great help
Posted by: Team12 @ Nov. 15, 2019, 9:36 p.m.