Ray Version?

[jhpenger]

What version of ray should I use for this?

[robertnishihara]

I believe 0.4.0. But you could also try the latest version of Ray.

[jhpenger]

Works for 0.4.0. But:
I'm not able to train it under current ray version: 0.5.3 after I updated the old config parameters to their appropriate names.
I am getting an error in tensorflow 2Dconv:

ValueError: Negative dimension size caused by subtracting 11 from 10 for 'default/vision_net/fc1/Conv2D' (op: 'Conv2D') with input shapes: [?,10,10,32], [11,11,32,256].

[robertnishihara]

Ok, good to know! It was initially tested with 0.4.0.

[jhpenger]

@robertnishihara Can you please check visionnet.py on the main repo? line 54 & 59
Were changed from [256, [10, 10], 1] to [256, [11, 11], 1] on ray-project/ray#2679
Those 2 default conv filters with 11s instead of 10s was the cause of the conv2D error.
Was able to get sonic-on-ray working when I changed the kernel from [11, 11] back to [10, 10].
Any thoughts on what went wrong there?
and what conv filters did you use for this originally?