error with batch size > 1

[YandanYang]

I got the following error when running with batch_size=4.
It seems like the chunk number for each sample is different.
Does the code support batch_size>1? Will this error affect the training results?

[collate debug] stack failed for key='state' err=ValueError: all input arrays must have the same shape
  sample[0]: shape=(4, 64), dtype=float64, ndim=2
  sample[1]: shape=(3, 64), dtype=float64, ndim=2
  sample[2]: shape=(4, 64), dtype=float64, ndim=2
  sample[3]: shape=(4, 64), dtype=float64, ndim=2
  lengths axis0=[4, 3, 4, 4]
[collate debug] stack failed for key='state_mask' err=ValueError: all input arrays must have the same shape
  sample[0]: shape=(4, 64), dtype=bool, ndim=2
  sample[1]: shape=(3, 64), dtype=bool, ndim=2
  sample[2]: shape=(4, 64), dtype=bool, ndim=2
  sample[3]: shape=(4, 64), dtype=bool, ndim=2
  lengths axis0=[4, 3, 4, 4]
[collate debug] stack failed for key='action' err=ValueError: all input arrays must have the same shape
  sample[0]: shape=(96, 32), dtype=float64, ndim=2
  sample[1]: shape=(72, 32), dtype=float64, ndim=2
  sample[2]: shape=(96, 32), dtype=float64, ndim=2
  sample[3]: shape=(96, 32), dtype=float64, ndim=2
  lengths axis0=[96, 72, 96, 96]
[collate debug] stack failed for key='action_mask' err=ValueError: all input arrays must have the same shape
  sample[0]: shape=(96, 32), dtype=bool, ndim=2
  sample[1]: shape=(72, 32), dtype=bool, ndim=2
  sample[2]: shape=(96, 32), dtype=bool, ndim=2
  sample[3]: shape=(96, 32), dtype=bool, ndim=2
  lengths axis0=[96, 72, 96, 96]
[collate debug] stack failed for key='lapa_action' err=ValueError: all input arrays must have the same shape
  sample[0]: shape=(96, 32), dtype=float64, ndim=2
  sample[1]: shape=(72, 32), dtype=float64, ndim=2
  sample[2]: shape=(96, 32), dtype=float64, ndim=2
  sample[3]: shape=(96, 32), dtype=float64, ndim=2
  lengths axis0=[96, 72, 96, 96]
这个问题严重吗？影响训练效果吗

[NVshawnhwang]

Hi,
Default training configurations are batch_size = 1.
You can read more about the default configurations here: https://github.com/dreamzero0/dreamzero/blob/main/README.md#training-configuration

[zrporz]

I follow the setting in https://github.com/dreamzero0/dreamzero/blob/main/docs/DATASET_TO_GEAR_AND_TRAIN.md which default batch_size is 4 and meet the similar problem.

[NVshawnhwang]

Hi,
Default training configurations are batch_size =1. Increasing the batch size is not officially supported, however, if you would like to increase the batch size to b > 1, it is possible to do with modifications to the code.