Skip to content

Why do the three frameworks give slightly different numbers for the same algorithm? #1

Description

@Terriao

Several reasons combine: low-level operators (sparse convolution, padding, pooling) are not bit-identical across frameworks; floating-point accumulation order differs; default weight-initialization and BatchNorm semantics vary; and graph-mode auto-tuning makes different choices. We document these in detail in the accompanying paper. We use fixed environment and a strict consistency protocol to keep the differences as small as possible.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions