Unable to shuffle if the fields are not of same length

[michael20at]

Hm, tried ml-agents 0.7 with my new 2080ti (runs great on normal Tensorflow),
updated Unity to the newest version 2018.3.7f1, training of 3D Balls example via cmd starts fine, but after a few episode (sometimes only a few, sometimes after 9) I get

"Unable to shuffle if the fields are not of same length"!

What could be the problem?

[michael20at]

Hm, it kinda works if I increase the buffer size by ten in the config file, but I still don't know why?

[xiaomaogy]

This is kind of weird and unexpected, I don't see other people raising this issue, so I guess maybe delete and rerun and try to follow the basic guide step by step might help.

[Ina299]

I also experienced this issue.
I use ml-agents 0.6 on 1080ti.
batch_size: 1536 beta: 0.006 buffer_size: 15360 epsilon: 0.17 gamma: 0.995 hidden_units: 512 lambd: 0.9 learning_rate: 0.00012 max_steps: 5.0e15 normalize: True num_epoch: 4 num_layers: 4 time_horizon: 768 summary_freq: 2000 use_recurrent: False use_curiosity: False
I saw this phenomenon when I took agent' s life time too long.
So I guess agent should be Done() before buffer_size.

[ervteng]

Are you using a Numpy version greater than 1.14.1 by any chance? Newer versions of Numpy are known to have similar issues.

[rainysolar]

I have the same issue.
I use ml-agents 0.8.1 on 2070 card which only support CUDA 10.
CUDA 10 needs TensorFlow >=1.13.0.
And Numpy 1.14.1 is not support TensorFlow>=1.13.0

[yttiktak]

Same issue. Using Ubuntu Bionic, numpy v 1.16.2
Trying to track it down, seems a numpy array (advantages) changes shape from (N,) to (1,N) when the ppy trainer.py subtracts the mean from it, about line 325.
I hacked out a thing to bypass policy updates when that happens. Walker seems to be training now. My 5 yr old approves of the walker. It got up to a reward of 300 so far.

[JadenTravnik]

This doesnt seem to happen when using Numpy 1.14.5 and tensorflow 1.7 as mentioned in setup.py.
I'm still not sure what the issue is but at least one doesn't get the bug.

[EfveZombie]

It should be a compatibility problem with higher version of Tensorflow or Numpy.

In my situation (py 3.7 && tf 1.13 && np 1.16), simply adding parameter dtype = float to all the calls of numpy.array() in ml-agents/mlagents/trainers/buffer.py could fix this problem.

But I dont know if that would cause any potential problems.

[ervteng]

The fields in the buffer should be all numerical, so I don't see an issue with adding the dtype. We'll definitely track this as we upgrade to newer versions of Numpy.

[ervteng]

Hi @EfveZombie, we found the root cause of this issue and fixed it on the latest develop branch. Let us know if it fixes your issue. Thanks!

[awjuliani]

Closing this issue as it was addressed in a recent release of ML-Agents.

[github-actions]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.