Skip to content

Latest commit

 

History

History
241 lines (173 loc) · 11.7 KB

README.md

File metadata and controls

241 lines (173 loc) · 11.7 KB
Raw



Twitter License Transformers Accelerate Author

Easy-Translate is a script for translating large text files with a 💥SINGLE COMMAND💥. Easy-Translate is designed to be as easy as possible for beginners and as seamless and customizable as possible for advanced users. We support almost any model, including M2M100, NLLB200, LLaMA, Bloom and more 🥳. We also provide a script for Easy-Evaluation of your translations 📋

Easy-Translate is built on top of 🤗HuggingFace's Transformers and 🤗HuggingFace's Accelerate library.

We currently support:

  • CPU / multi-CPU / GPU / multi-GPU / TPU acceleration
  • BF16 / FP16 / FP32 / 8 Bits / 4 Bits precision.
  • Automatic batch size finder: Forget CUDA OOM errors. Set an initial batch size, if it doesn't fit, we will automatically adjust it.
  • Multiple decoding strategies: Greedy Search, Beam Search, Top-K Sampling, Top-p (nucleus) sampling, etc. See Decoding Strategies for more information.
  • 🆕 Load huge models in a single GPU with 8-bits / 4-bits quantization and support for splitting the model between GPU and CPU. See Loading Huge Models for more information.
  • 🆕 LoRA models support
  • 🆕 Support for any Seq2SeqLM or CausalLM model from HuggingFace's Hub.
  • 🆕 Prompt support! See Prompting for more information.

Test the 🔌 Online Demo here: https://huggingface.co/spaces/Iker/Translate-100-languages

Supported languages

See the Supported languages table for a table of the supported languages and their ids.

Supported Models

💥 EasyTranslate now supports any Seq2SeqLM (m2m100, nllb200, small100, mbart, MarianMT, T5, FlanT5, etc.) and any CausalLM (GPT2, LLaMA, Vicuna, Falcon) model from 🤗 Hugging Face's Hub!! We still recommend you to use M2M100 or NLLB200 for the best results, but you can experiment with any other MT model, as well as prompting LLMs to generate translations (See Prompting Section for more details). You can also see the examples folder for examples of how to use EasyTranslate with different models.

M2M100

M2M100 is a multilingual encoder-decoder (seq-to-seq) model trained for Many-to-Many multilingual translation introduced in this paper and first released in this repository.

M2M100 can directly translate between 9,900 directions of 100 languages.

NLLB200

No Language Left Behind (NLLB) open-sources models capable of delivering high-quality translations directly between any pair of 200+ languages — including low-resource languages like Asturian, Luganda, Urdu and more. It aims to help people communicate with anyone, anywhere, regardless of their language preferences. It was introduced in this paper and first released in this repository.

NLLB can directly translate between +40,000 of +200 languages.

Other MT Models supported

We support every MT model in the 🤗 Hugging Face's Hub. If you find a model that doesn't work, please open an issue for us to fix it or a PR with the fix. This includes, among many others:

Citation

If you use this software please cite

@inproceedings{garcia-ferrero-etal-2022-model,
    title = "Model and Data Transfer for Cross-Lingual Sequence Labelling in Zero-Resource Settings",
    author = "Garc{\'\i}a-Ferrero, Iker  and
      Agerri, Rodrigo  and
      Rigau, German",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2022",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.findings-emnlp.478",
    pages = "6403--6416",
}

Requirements

Pytorch >= 1.10.0 
See: https://pytorch.org/get-started/locally/

Accelerate >= 0.12.0
pip install accelerate

HuggingFace Transformers 
If you plan to use NLLB200, please use >= 4.28.0, as an important bug was fixed in this version. 
pip install --upgrade transformers

BitsAndBytes (Optional, required for 8-bits / 4-bits quantization)
pip install bitsandbytes

PEFT (Optional, required for loading LoRA models)
pip install peft

Translate a file

Run python translate.py -h for more info.
See the examples folder for examples of how to run different models.

Using a single CPU / GPU

python3 translate.py \
--sentences_path sample_text/en.txt \
--output_path sample_text/en2es.translation.m2m100_1.2B.txt \
--source_lang en \
--target_lang es \
--model_name facebook/m2m100_1.2B

Multi-GPU

See Accelerate documentation for more information (multi-node, TPU, Sharded model...): https://huggingface.co/docs/accelerate/index
You can use the Accelerate CLI to configure the Accelerate environment (Run accelerate config in your terminal) instead of using the --multi_gpu and --num_processes flags.

# Use 2 GPUs
accelerate launch --multi_gpu --num_processes 2 --num_machines 1 translate.py \
--sentences_path sample_text/en.txt \
--output_path sample_text/en2es.translation.m2m100_1.2B.txt \
--source_lang en \
--target_lang es \
--model_name facebook/m2m100_1.2B

Automatic batch size finder

We will automatically find a batch size that fits in your GPU memory. The default initial batch size is 128 (You can set it with the --starting_batch_size 128 flag). If we find an Out Of Memory error, we will automatically decrease the batch size until we find a working one.

Loading Huge Models

Huge models such as LLaMA 65B or nllb-moe-54b can be loaded in a single GPU with 8 bits and 4 bits quantification with minimal performance degradation. See BitsAndBytes. Set precision to 8 or 4 with the --precision flag.

pip install bitsandbytes

python3 translate.py \
--sentences_path sample_text/en.txt \
--output_path sample_text/en2es.translation.nllb200-moe-54B.txt \
--source_lang eng_Latn \
--target_lang spa_Latn \
--model_name facebook/nllb-moe-54b \
--precision 8 \
--force_auto_device_map \
--starting_batch_size 8

If even the quantified model does not fit in your GPU memory, you can set the --force_auto_device_map flag. The model will be split across GPUs and CPU to fit it in memory. CPU offloading is slow, but will allow you to run huge models that do not fit in your GPU memory.

Prompting

You can use LLMs such as LLaMA, Vicuna, GPT2, FlanT5, etc, instead of a translation model. These models require a prompt to define the task. You can either have the prompt already in the input file (each sentence includes the prompt) or you can use the --prompt flag to add the prompt to each sentence. In this case, you need to include the token %%SENTENCE%% in the prompt. This token will be replaced by the sentence to translate. You do not need to specify the --source_lang and --target_lang flags in this case.

python3 translate.py \
--sentences_path sample_text/en.txt \
--output_path sample_text/en2es.FlanT5.translation.txt \
--model_name google/flan-t5-large \
--prompt "Translate English to Spanish: %%SENTENCE%%"

Decoding/Sampling strategies

You can choose the decoding/sampling strategy to use and the number of candidate translations to output for each input sentence. By default, we will use beam-search with num_beams set to 5, and we will output the most likely candidate translation. This should be the best configuration for most use cases. You can change this behaviour with the following flags:

--num_beams: Number of beams to use for beam-search decoding (default: 5)
--do_sample: Whether to use sampling instead of beam-search decoding (default: False)
--temperature: Sampling temperature (default: 0.8)
--top_k: Top k sampling (default: 100)
--top_p: Top p sampling (default: 0.75)
--repetition_penalty: Repetition penalty (default: 1.0)
--keep_special_tokens: Whether to keep special tokens (default: False)
--keep_tokenization_spaces: Whether to keep tokenization spaces (default: False)
--num_return_sequences: Number of candidate translations to output for each input sentence (default: 1)

Please, note that running --do_sample with --num_beams > 1 and 8 bits or 4 bits quantification may be numerically unstable and produce an error.

Evaluate translations

To run the evaluation script you need to install bert_score: pip install bert_score and 🤗HuggingFace's Evaluate model: pip install evaluate.

The evaluation script will calculate the following metrics:

Run the following command to evaluate the translations:

python3 eval.py \
--pred_path sample_text/en2es.translation.m2m100_1.2B.txt \
--gold_path sample_text/es.txt

If you want to save the results to a file use the --output_path flag.

See sample_text/en2es.m2m100_1.2B.json for a sample output.