- InPlaceAccumulatorV2#

InPlaceAccumulatorV2 - 1 (


This version of the operator has been available since version 1 of domain


In-place accumulator for tensors. Differs from older op by adding overwrite_flag for reset, and making output buffer as optional. Set the overwrite_flag to false for gradient accumulation and to True for overwriting the accumulation buffer during gradient computation (equivalent to reset grad + train step)


Between 2 and 3 inputs.

  • accumulation_buffer (heterogeneous) - T: historical result of accumulator

  • value (heterogeneous) - T_GRAD: the value that will be added to the accumulator

  • overwrite_flag (optional, heterogeneous) - T_BOOL: Indicates if tensor should be overwritten. Default is accumulation


Between 1 and 2 outputs.

  • updated_flag (heterogeneous) - T_BOOL: Whether the update was completed

  • accumulation_buffer_out (optional, heterogeneous) - T: updated result of accumulator