Memformers utilize a external dynamic memory to store history information. This repo contains implementation of the pre-trained model MemBART and its training code. Our implementation is based on huggingface transformers. Currently, we provide two checkpoints «qywu/membart-large» (checkpooint) and «qywu/membart-base»(checkpooint). You can directly load the checkpoint with: Note that due to BART denosing pre-training, it needs to further fine-tune the model on the downstream tasks to get better performance. Then, you can refer to the code in examples/finetunedialog` for details about finetuning or further pre-training MemBart on your tasks. Memformer: A Memory-Augmented Transformer for Sequence Modeling Stateful Memory-Augmented Transformers for Dialogue Modeling
Модальности:
Генерация текста
Задача: Генерация текста
Автор: qywu
Теги: endpoints_compatible
Лайков: 3 | Загрузок: 11
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
