search

alex-berard / seq2seq

Attention-based sequence to sequence learning
Apache License 2.0
388 stars 122 forks source link

Multi-task training #23

Closed menamine closed 5 years ago

menamine commented 5 years ago

Hi all,

For the multi-task training, what kind of architecture did you use to accomplish this task? Did you use a separate encoder/decoder for each task, or did you use shared parameters between the encoder and/or the decoder? Could you make available an article about the implemented architecture.

Than you in advance.

alex-berard commented 5 years ago

Hi, Actually, it depends on your configuration. If you take a look at LibriSpeech/Multi-Task.yaml, you can see that 3 tasks are defined: MT, ASR and AST. AST and ASR share the same encoder (speech.en), and AST and MT share the same decoder (char.fr) You can find more info about this model in https://arxiv.org/abs/1802.04200. This sequence to sequence multi-task training approach was proposed in https://arxiv.org/abs/1511.06114

menamine commented 5 years ago

Thank you for the answer ^^, it is clear. Best.