Skip to content

Fix/tokenizer sampling#106

Closed
fenneishi wants to merge 15 commits into
NVIDIA:mainfrom
fenneishi:fix/tokenizer-sampling
Closed

Fix/tokenizer sampling#106
fenneishi wants to merge 15 commits into
NVIDIA:mainfrom
fenneishi:fix/tokenizer-sampling

Conversation

@fenneishi
Copy link
Copy Markdown

@fenneishi fenneishi commented Feb 16, 2025

Description

This PR addresses issue #102

Comprehensive Testing

Tested all video tokenizer configurations:

  • Cosmos-Tokenizer-0.1: CV4x8x8, CV8x8x8, CV8x16x16, DV4x8x8, DV8x8x8, DV8x16x16
  • Cosmos-Tokenizer-1.0: CV8x8x8, DV8x16x16

Validation includes:

  1. Automated shape verification
    • Latent space dimensions
    • Reconstruction output dimensions
  2. Interactive quality assessment
    • Side-by-side visual comparison
    • Human evaluation interface
    • Manual quality verification

image

pytest -s -v .\cosmos1\models\tokenizer\tests\test_tokenizer.py

Tokenizer Test Results.pdf

ielh1 and others added 15 commits January 6, 2025 22:18
- Only update curr_res when spatial down-sampling occurs
- Fix assertion limits for num_spatial/temporal_downs to num_resolutions-1
Model tested configs:
- Cosmos-0.1-Tokenizer-CV4x8x8
- Cosmos-0.1-Tokenizer-CV8x8x8
- Cosmos-0.1-Tokenizer-CV8x16x16
- Cosmos-0.1-Tokenizer-DV4x8x8
- Cosmos-0.1-Tokenizer-DV8x8x8
- Cosmos-0.1-Tokenizer-DV8x16x16
- Cosmos-1.0-Tokenizer-CV8x8x8
- Cosmos-1.0-Tokenizer-DV8x16x16

Automated checks:
- Validate latent space dimensions for each config.
- Verify reconstruction output shape match input for each config.

Quality assurance:
- Add interactive human evaluation interface.
- Visual comparison between input and reconstruction.
- Manual quality verification with accept/reject options.
@fenneishi fenneishi marked this pull request as ready for review February 16, 2025 09:48
@sophiahhuang sophiahhuang added the enhancement New feature or request label Feb 20, 2025
@lfengad lfengad closed this Jun 1, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

enhancement New feature or request

Projects

None yet

Development

Successfully merging this pull request may close these issues.

5 participants