Use `tf.matmul` instead of `np.dot` in `test_retrieval_transformer` - Githubissues

NVIDIA-Merlin / models

Merlin Models is a collection of deep learning recommender system model reference implementations

https://nvidia-merlin.github.io/models/main/index.html

Apache License 2.0

262 stars 50 forks source link

Use `tf.matmul` instead of `np.dot` in `test_retrieval_transformer` #1122

Closed oliverholworthy closed 1 year ago

oliverholworthy commented 1 year ago

Fixes issue with test_retrieval_transformer failing in Blossom tests

Implementation Details :construction:

The test test_retrieval_transformer is currently failing in Blossom with a tolerance error. This seems to be something to do with small differences in numerical results when running the computation on CPU vs GPU.

Using tf.matmul to run this operation on the same GPU device (when available) as the operation in the model runs.

Testing Details :mag:

Tested this by submitting a custom blossom job and running the test multiple times with pytest-repeat

github-actions[bot] commented 1 year ago

Documentation preview

https://nvidia-merlin.github.io/models/review/pr-1122