Added tutorial about domain adaptation, including LHUC #409
Conversation
|
This is a nice use case for the recently introduced config files. :-) |
tutorials/adapt/README.md
Outdated
| this tutorial, we show two methods on how to perform domain adaptation of a | ||
| general translation system using Sockeye. | ||
|
|
||
| We assume you already have an already trained Sockeye model, for example the |
There was a problem hiding this comment.
'we assume you already have a trained Sockeye model'
tutorials/adapt/README.md
Outdated
| general translation system using Sockeye. | ||
|
|
||
| We assume you already have an already trained Sockeye model, for example the | ||
| one trained in the [second tutorial](../wmt/README.md). We also assume that you |
There was a problem hiding this comment.
'for example from the second tutorial.'
tutorials/adapt/README.md
Outdated
|
|
||
| We assume you already have an already trained Sockeye model, for example the | ||
| one trained in the [second tutorial](../wmt/README.md). We also assume that you | ||
| have two training sets, one composed of general or out-of-domain data, and one |
There was a problem hiding this comment.
if would introduce the abbreviations you use below here ('id' and 'ood')
tutorials/adapt/README.md
Outdated
|
|
||
| ## Continuation of training | ||
|
|
||
| This method consists in taking the parameters of an already trained system and |
There was a problem hiding this comment.
maybe just: 'This method fine-tunes a trained model and starts a second training run on in-domain data, initialized with the parameters obtained from the out-domain data.'
tutorials/adapt/README.md
Outdated
| This method consists in taking the parameters of an already trained system and | ||
| use them as initialization of a new training run only on the in-domain data. | ||
| Thus you "continue training" on the data you are more interested in. Freitag | ||
| and Al-Onaizan (2016) showed that, while being quite a straightforward |
There was a problem hiding this comment.
maybe add a "# references" section at the bottom with the full citation.
tutorials/adapt/README.md
Outdated
| use them as initialization of a new training run only on the in-domain data. | ||
| Thus you "continue training" on the data you are more interested in. Freitag | ||
| and Al-Onaizan (2016) showed that, while being quite a straightforward | ||
| technique, this method can achieve good results. |
There was a problem hiding this comment.
'showed that this straightforward technique can achieve good results.'
tutorials/adapt/README.md
Outdated
|
|
||
| ## Leaning Hidden Unit Contribution | ||
|
|
||
| Leaning Hidden Unit Contribution (LHUC) is a method proposed by Vilar (2018), |
There was a problem hiding this comment.
full citation at the bottom maybe
Included tutorial about adapting NMT models, including LHUC.
Pull Request Checklist
until you can check this box.
By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.