models

models copied to clipboard

[BUG] Dataloader doesnt release memory and memory growth

3

### Bug description I run a training script and reinitialize nvtabular dataloaders. After each initialization, the available GPU memory decreases (fmem = pynvml_mem_size(kind="free", index=0) ). That is unexpected. The available...

bschifferer

[WIP] add an example notebook explaining hyperparameter optimization using Optuna and showing retrieval on h&m data

28

We have had a couple of discussions recently on exploring using H&M data. Also, some time ago we discussed creating an example around using HPO (it can be a nice...

radekosmulski

[WIP] Add initial draft of example notebook using horovod

2

A draft PR that shows the workflow. Depends on #783. Currently uses a workaround that re-partitions the dataset, i.e., `ddf = train.to_ddf().repartition(npartitions=hvd.size())`. After some preprocessing with nvtabular, the training code...

edknv

[Task] Add RNN-based model apis for session-based recommendation

1

### Description Currently there is no out-of-the-box api for creating an LSTM, BiLSTM and GRU architectures for a session-based (or sequential) task. - For LSTM , I assume we'd use...

rnyak

Reproduce selected results from Transformers4Rec paper with Merlin Models API

1

### Description For Transformers4Rec, we have created a training/eval [script](https://github.com/NVIDIA-Merlin/Transformers4Rec/tree/main/examples/t4rec_paper_experiments) for reproducing the [paper experiments](https://dl.acm.org/doi/10.1145/3460231.3474255), that takes a set of hparams as command line arguments and a preprocessed dataset. This...

gabrielspmoreira

Update the retrieval integration tests on CI to use the refactored MF and TwoTower

1

viswa-nvidia

[FEA] Mixed Precision Support for Merlin Models TensorFlow

1

# 🚀 Feature request

bschifferer

Reduce running time of unit tests

1

### Goals :soccer: - Improve iteration speed by reducing the running time of unit tests ### Implementation Details :construction: ### Testing Details :mag:

oliverholworthy

### Description Currently we do not have any unit test for using LSTM or BiLSTM for session-based recommendation tasks. it'd be useful to add one. For LSTM we'd use `tf.keras.layers.LSTM`...

rnyak

[BUG] Data parallel training freezes due to different number of batches

7

### Bug description In data parallel training, we start multiple workers with different initialization of the dataloader and train with horovod. After each batch update, the parameters are synced. Merlin...

bschifferer