models

models copied to clipboard

viswa-nvidia

[WIP] inference benchmarking using transformers

2

this is a PR just for sharing code / progress there is no intention to merge it

radekosmulski

Save and load SOK model embeddings

9

Fixes # (issue) ### Goals :soccer: Provide an interfact for SOKEmbedding load/dump and a sample code for SOKEmbedding model load/dump ### Implementation Details :construction: ### Testing Details :mag: This ticket...

wenjing-nv

This task is about improving the feedback loop on Merlin Models Pull Requests through the reorganization and configuration of the tests. Refactoring the implementation of the tests themselves is beyond...

oliverholworthy

[BUG] Training TwoTowerModel when loading dataset using `part_size=128MB` crashes

11

**Describe the bug** I am trying to train a TwoTowerModel. I load the datasets (train and val) using NVTabular.Dataset and later pass them to `model.fit`. When I add the `part_size="128MB"`...

mats-claassen

### Describe the documentation you'd like Inline Code Coverage of Merlin Models is around 40%. We should aim for X%. Merlin Models: ============================ Coverage for /workspace/01_MerlinDev/62_DocStrings/models/merlin/ ============================ --------------------------------------------------------- Summary ---------------------------------------------------------...

bschifferer

### Description We want to add a CI test, that collects metrics to understand if performance changes between releases. First step: - Add regression tests for session based transformer model...

bschifferer

### Description REES46 eCoommerce with XLNet + MLM Masking + Item Features (best configuration from paper) Create benchmark similar to https://github.com/NVIDIA-Merlin/Transformers4Rec/issues/610 . Provide results in similar tables/presentation, that we can...

bschifferer

Investigate if embeddings lookup with `tf.RaggedTensor` is slower than with `tf.Tensor` with latter versions of TF

1

In previous experiments from @vysarge (June 2022) it was found that `tf.RaggedTensor` representation is slower than using fixed-length dense `tf.Tensor` for embedding lookup, as shown in this [spreadsheet](https://docs.google.com/spreadsheets/d/1jlKDVeoMvpQfyCF9RFmR3VxckbPbrvBg2POwbF2p7RM/edit#gid=135622185). This tasks...

gabrielspmoreira

Benchmark, improve and document mixed precision (AMP) support in Models

2

- [ ] Train different models with mixed precision and inspect if it doesn't break the API, leads to better performance at similar accuracy without leading to numeric instabilities /...

gabrielspmoreira