models

models copied to clipboard

Tf/dep prediction tasks

1

Fixes # (issue) ### Goals :soccer: ### Implementation Details :construction: ### Testing Details :mag:

marcromeyn

### Description The pairwise losses were originally tested with 2D prediction tensors (batch size x # items/classes). For sequential/session-based recommendation, the prediction tensors are 3D (batch size x seq length...

gabrielspmoreira

Make it easier to sub-class SequentialBlock & ParallelBlock

3

### Goals :soccer: Right now blocks like `InputBlockV2` are implemented as functions and not as classes. This because, as of now, creating it in a sub-classes way adds extra overhead...

marcromeyn

Change `BatchedDataset` from Keras Sequence to Generator

5

Relates to #722 (issue) Alternative to #724 ### Goals :soccer: - Support Tensorflow/Keras 2.10 - Improve performance of dataloader by removing redundant calls to dataloader `stop` method ### Implementation Details...

oliverholworthy

Create an example (notebook or python script) including guidelines on how to train transformer-based models with techniques supported in the HuggingFace trainer class, such as Early stopping, fp16, lr scheduler,...

sararb

[Task] Ensure that the data augmentation negative sampling layer can't be used as a sampler for prediction

1

### Description We'd like to make sure it's clear that the negative sampling data augmentation layer is not confused with the samplers used in the prediction blocks. The data augmentation...

oliverholworthy

Add OutputBlock

3

### Goals :soccer: This PR introduces a `OutputBlock` which is the equivalent to the `InputBlock`. It loops over all columns in the schema and based on the tags it adds...

marcromeyn

### Description In the datasets API, we support the item-level preprocessing of three variants of [movielens dataset](https://github.com/NVIDIA-Merlin/models/blob/main/merlin/datasets/entertainment/movielens/dataset.py). The goal of this task is to add a pre-processing function to generate...

sararb

### Goals :soccer: Create integration tests for ranking models (based on our research scripts) to check for - **Stability of our API (unit tests)** - by using advanced options (hparams)...

gabrielspmoreira

[FEA] Implement model interpretation/explanation

2

# 🚀 Feature request Would like to understand if there is any plan to implement/integrate model interpretation/explanation techniques like [DeepLIFT](https://arxiv.org/abs/1704.02685) or GradientExplainer from [SHAP](https://github.com/slundberg/shap) into Merlin? SHAP natively supports model...

zhiruiwang