models

models copied to clipboard

### Description The `UniformNegativeSampling` class name might be misleading since it is not really doing uniform negative sampling. It'd be useful to find another name that'd be more accurate and...

rnyak

### Description We have been using the `int_domain.max` property of the `ColumnSchema` to compute embedding cardinalities. This task proposes to replace the usage of `int_domain.max` to infer the cardinality with...

oliverholworthy

# 🚀 Feature request Currently, we don’t support variable-length continuous columns out of the box, therefore, when we train Two-Tower model with variable length multi-hot continuous columns we get an...

rnyak

# 🚀 Feature request ## Motivation When building latent factor models (for example with LightFM). Evaluation requires that the validation or test dataset has no user-item interaction overlap. LightFM raises...

oliverholworthy

Right now we are converting all features to dense when we store them in the feature-context ([source](https://github.com/NVIDIA-Merlin/models/blob/main/merlin/models/tf/models/base.py#L293)). It would be better to store them as `tf.RaggedTensors`, when retrieving them the...

marcromeyn

[FEA] cannot properly save and load the TF Retrieval model

4

### Bug description We want to be able to save the entire Two-Tower model and load back to be able to do `model.evaluate()` and `model.predict()`. However, we get the following...

rnyak

First example of youtube-dnn

3

### Goals :soccer: This PR introduces a first draft of the Youtube-DNN retrieval model example. It also fixes some related issues raised when defining the notebook. - [x] First draft...

sararb

# 🚀 Feature request As a user, I want to visualize the architecture and way to debug/understand the architecture. NVTabular + Merlin Models are a powerful tool to provide easy...

bschifferer

Address TODO comments in retrieval.py and ranking.py.

3

viswa-nvidia

Enable dynamic memory growth

2

Currently Tensorflow will use all memory it has available on startup. In newer versions dynamic memory-growth can be enabled. This should be added as an option in `configure_tensorflow`.

marcromeyn