translation.rst

Translation

The Task

Translation is the task of translating text from a source language to another, such as English to Romanian. This task is a subset of Sequence to Sequence tasks, which requires the model to generate a variable length sequence given an input sequence. In our case, the task will take an English sequence as input, and output the same sequence in Romanian.

---

Example

Let's look at an example. We'll use WMT16 English/Romanian, a dataset of English to Romanian samples, based on the Europarl corpora. The data set contains a train.csv and valid.csv. Each CSV file looks like this:

input,target
"Written statements and oral questions (tabling): see Minutes","Declaraţii scrise şi întrebări orale (depunere): consultaţi procesul-verbal"
"Closure of sitting","Ridicarea şedinţei"
...

In the above the input/target columns represent the English and Romanian translation respectively. Once we've downloaded the data using ~flash.core.data.download_data, we create the ~flash.text.seq2seq.translation.data.TranslationData. We select a pre-trained backbone to use for our ~flash.text.seq2seq.translation.model.TranslationTask and finetune on the WMT16 data. The backbone can be any Seq2Seq translation model from HuggingFace/transformers.

Note

When changing the backbone, make sure you pass in the same backbone to the ~flash.text.seq2seq.translation.data.TranslationData and the ~flash.text.seq2seq.translation.model.TranslationTask!

Next, we use the trained ~flash.text.seq2seq.translation.model.TranslationTask for inference. Finally, we save the model. Here's the full example:

../../../flash_examples/translation.py

To learn how to view the available backbones / heads for this task, see backbones_heads.

---

Flash Zero

The translation task can be used directly from the command line with zero code using flash_zero. You can run the above example with:

flash translation

To view configuration options and options for running the translation task with your own data, use:

flash translation --help

---

Serving

The ~flash.text.seq2seq.translation.model.TranslationTask is servable. This means you can call .serve to serve your ~flash.core.model.Task. Here's an example:

../../../flash_examples/serve/translation/inference_server.py

You can now perform inference from your client like this:

../../../flash_examples/serve/translation/client.py

---

Accelerate Training & Inference with Torch ORT

Torch ORT converts your model into an optimized ONNX graph, speeding up training & inference when using NVIDIA or AMD GPUs. Enabling Torch ORT requires a single flag passed to the TranslationTask once installed. See installation instructions here.

Note

Not all Transformer models are supported. See this table for supported models + branches containing fixes for certain models.

...

model = TranslationTask(backbone="t5-large", num_classes=datamodule.num_classes, enable_ort=True)