search

facebookresearch / FBTT-Embedding

This is a Tensor Train based compression library to compress sparse embedding tables used in large-scale machine learning models such as recommendation and natural language processing. We showed this library can reduce the total model size by up to 100x in Facebook’s open sourced DLRM model while achieving same model quality. Our implementation is faster than the state-of-the-art implementations. Existing the state-of-the-art library also decompresses the whole embedding tables on the fly therefore they do not provide memory reduction during runtime of the training. Our library decompresses only the requested rows therefore can provide 10,000 times memory footprint reduction per embedding table. The library also includes a software cache to store a portion of the entries in the table in decompressed format for faster lookup and process.
MIT License
192 stars 27 forks source link

Fix approx-uniform scaling and make it default initializer #11

Closed bilgeacun closed 3 years ago

bilgeacun commented 3 years ago

Summary: approx-uniform performs the best so setting it as default. Without scaling, full tensor values generated by approx-uniform is betwen (-1, 1). They should be multiplied by (1/sqrt(n))^1/3 to be in the range (-sqrt(1/n), sqrt(1/n)) -- changed the division to multiplication.

Differential Revision: D26744953

facebook-github-bot commented 3 years ago

This pull request was exported from Phabricator. Differential Revision: D26744953

facebook-github-bot commented 3 years ago

This pull request has been merged in facebookresearch/FBTT-Embedding@e16249aa5bde3f470a412b4a080ff3fa7e9e0a45.