Skip to content

171 upgrade tf 2190 #172

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 100 commits into from
Apr 12, 2025
Merged

171 upgrade tf 2190 #172

merged 100 commits into from
Apr 12, 2025

Conversation

david-thrower
Copy link
Owner

@david-thrower david-thrower commented Apr 11, 2025

Summary of changes:

  • Replaced the text embedding base model with an interleaved Rotaty Positional Embedding iRoPE in the Phishing detection NLP proof of concept
  • Proof of concept that the entire model can scale at O(n) timing as sequence length increases infinitely.

Benchmarks:

seq_len     val_bianry_accuracy       min/model        total_min         timing_relative_to_1024                      Commit_SHA
3072          0.955                    65.942           329.715                    2.817                     4bc217b36d1baf8b9a81fe53000e9e13e6e87801
1536          0.96                     37.27             186.36                    1.591                     286ba81a1e51493d748ded727bd602a4398248a8
1024          0.952                    23.42             117.08                    1.0                       9893bfc55d4f7b753eff79e0c5c70e4992c61085
  • Upgraded tf to 2.19.0
  • Upgraded Jax
  • Upgraded model architecture in both phishing CICD example and CIFAR10 example to accommodate tf 2.19.0.
  • Removed obsolete BERT embedding CICD test.

Comment temporarily disable time-consuming workflows. Comment out BERT based text classification workflow possibly permanently, as this is obsolete.
Add branch to workflow.
Added a baseline fine tuning of the full GPT2 to compare against Cerebros text classifier.
Amendments to Cerebros model.
Reduce seq length to accelerate job completion.
Up timeout to 300 min.
Correct history indexing error.
Temporary test to fast forward to cerebros model.
Comment out an artifact of GPT test so we this can lint and run.
Fix errors from trying to work too fast ...
Re-corrected the metrics BinaryAccuracy to correct AI introduced error.
Correct metric to rank by (binary accuracy) ...
Uncomment out GPT test ...
Upped number of trials to 5.
Make seq len 750, fix typo.
Added branch to the workflow...
Added a positional embedding and a LayerNorm to the text embedding.
Missed position embedding in copy and paste ...
Synchronize embedding dim across embeddings.
Corrected import of PositionEmbedding.
Remove layernorm, concat instead of add.
Try addition to merge embeddings without LayerNorm
Restore optimal run with position embedding. Reduce max levels to fit the optimal run and reduce overhead. Test this to see if it works. if successful, add back the commented out comparison and PR. Then open an issue to optimize the params around this new model. We may need to run this on Katib to optimize the hyperparameters, as the model is fundamentally different than the original and can probably be optimized considerably.
Hard set levels to the known optimum.
Corrected hard set on levels to correct optima.
Restore the best model yet.
Add back the CICD test for image CLS. Prepare for PR.
Added back baseline workflow in best trial thus far.
Added all CICD tests to be used back to best NLP configuration.
Upgrade tensorflow-text to v 2.19.0
Add branch to workflows.
Typo on requirements.txt
Test to fast forward to Cerebros NLP test and check for compatibility.
Attempt to correct issue with tf v 2.19.0 graph scope.
Another attempt to resolve tf v 2.19.0 graph scope compatibility...
Run a full CICD run.
AI suggested tf 2.15.0 -> 2.19.0 compat fix.
Add back the baseline GPT2 task.
Fix a typo in string termination ...
@david-thrower david-thrower linked an issue Apr 11, 2025 that may be closed by this pull request
@david-thrower david-thrower marked this pull request as draft April 11, 2025 22:38
Uncommented out CICD test that was left commented out by error.
@david-thrower
Copy link
Owner Author

All checks have passed on this version of the code:

@david-thrower david-thrower marked this pull request as ready for review April 11, 2025 23:00
@Aidyn-Lopez
Copy link
Collaborator

Looks great, and I approve.

Copy link
Collaborator

@Aidyn-Lopez Aidyn-Lopez left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks great, and I approve.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

upgrade-tf-2.19.0
2 participants