Skip to content

Windows Training Speed Issue - Solution #33

Description

@ssw03270

Windows Training Speed Issue - Solution

Environment:

  • Windows 11
  • Python 3.11
  • PyTorch 2.5.1 + CUDA 12.4
  • RTX 3090 Ti
  • Dataset: CMU only (1867 train / 207 test files)

Problem:
Training was extremely slow on Windows (~68 seconds per epoch) compared to the paper's reported ~2 hours.

Root Cause:
Windows multiprocessing overhead with dataloader_num_workers causes significant slowdown.

Solution:
Set dataloader_num_workers: 0 in [options/train_avatarposer.json]

Results:

num_workers Speed
16 (default) ~68s/epoch
4 ~16s/epoch
0 ~2.3s/epoch ✓

With num_workers=0, training speed matches the paper's benchmark (~2.4s/epoch on Linux).

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

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions