Following the original DPSGD algorithm (Abadi et al, 2016)

Hi. First off, I presume that the code, particularly "dp_optimizer.py" implements the original algorithm proposed by Abadai et al, 2016 (https://arxiv.org/abs/1607.00133). If this is not the case, correct me please. If so, it is meant to clip gradient corresponding to each individual input sample. However, in the code, "def process_microbatch", right in the very beginning it takes mean over individual gradients in the microbatch each of which is corresponding to one input sample so that the impact of individual input samples is canceled out. So, it seems the Abadi algorithm is not followed. Is this true?

Read dp_optimizer.py and inspect process_microbatch first, then compare its gradient aggregation and clipping order with the Abadi et al. 2016 algorithm linked in the issue. Done means confirming whether the implementation matches the paper and documenting the conclusion or identifying the specific correction needed.