SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[ported model] FSMT (FairSeq MachineTranslation) (#6940)

Browse Source 
* ready for PR

* cleanup

* correct FSMT_PRETRAINED_MODEL_ARCHIVE_LIST

* fix

* perfectionism

* revert change from another PR

* odd, already committed this one

* non-interactive upload workaround

* backup the failed experiment

* store langs in config

* workaround for localizing model path

* doc clean up as in https://github.com/huggingface/transformers/pull/6956

* style

* back out debug mode

* document: run_eval.py --num_beams 10

* remove unneeded constant

* typo

* re-use bart's Attention

* re-use EncoderLayer, DecoderLayer from bart

* refactor

* send to cuda and fp16

* cleanup

* revert (moved to another PR)

* better error message

* document run_eval --num_beams

* solve the problem of tokenizer finding the right files when model is local

* polish, remove hardcoded config

* add a note that the file is autogenerated to avoid losing changes

* prep for org change, remove unneeded code

* switch to model4.pt, update scores

* s/python/bash/

* missing init (but doesn't impact the finetuned model)

* cleanup

* major refactor (reuse-bart)

* new model, new expected weights

* cleanup

* cleanup

* full link

* fix model type

* merge porting notes

* style

* cleanup

* have to create a DecoderConfig object to handle vocab_size properly

* doc fix

* add note (not a public class)

* parametrize

* - add bleu scores integration tests

* skip test if sacrebleu is not installed

* cache heavy models/tokenizers

* some tweaks

* remove tokens that aren't used

* more purging

* simplify code

* switch to using decoder_start_token_id

* add doc

* Revert "major refactor (reuse-bart)"

This reverts commit 226dad15ca6a9ef4e26178526e878e8fc5c85874.

* decouple from bart

* remove unused code #1

* remove unused code #2

* remove unused code #3

* update instructions

* clean up

* move bleu eval to examples

* check import only once

* move data+gen script into files

* reuse via import

* take less space

* add prepare_seq2seq_batch (auto-tested)

* cleanup

* recode test to use json instead of yaml

* ignore keys not needed

* use the new -y in transformers-cli upload -y

* [xlm tok] config dict: fix str into int to match definition (#7034)

* [s2s] --eval_max_generate_length (#7018)

* Fix CI with change of name of nlp (#7054)

* nlp -> datasets

* More nlp -> datasets

* Woopsie

* More nlp -> datasets

* One last

* extending to support allen_nlp wmt models

- allow a specific checkpoint file to be passed
- more arg settings
- scripts for allen_nlp models

* sync with changes

* s/fsmt-wmt/wmt/ in model names

* s/fsmt-wmt/wmt/ in model names (p2)

* s/fsmt-wmt/wmt/ in model names (p3)

* switch to a better checkpoint

* typo

* make non-optional args such - adjust tests where possible or skip when there is no other choice

* consistency

* style

* adjust header

* cards moved (model rename)

* use best custom hparams

* update info

* remove old cards

* cleanup

* s/stas/facebook/

* update scores

* s/allen_nlp/allenai/

* url maps aren't needed

* typo

* move all the doc / build /eval generators to their own scripts

* cleanup

* Apply suggestions from code review

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Apply suggestions from code review

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* fix indent

* duplicated line

* style

* use the correct add_start_docstrings

* oops

* resizing can't be done with the core approach, due to 2 dicts

* check that the arg is a list

* style

* style

Co-authored-by: Sam Shleifer <sshleifer@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
This commit is contained in:
Stas Bekman
2020-09-17 08:31:29 -07:00
committed by GitHub
parent 492bb6aa48
commit 1eeb206bef
19 changed files with 3534 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

94
model_cards/facebook/wmt19-de-en/README.md Normal file
View File

@@ -0,0 +1,94 @@
---
<!-- This file has been auto-generated by src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py - DO NOT EDIT or your changes will be lost -->
language: de, en
thumbnail:
tags:
- translation
- wmt19
license: Apache 2.0
datasets:
- http://www.statmt.org/wmt19/ ([test-set](http://matrix.statmt.org/test_sets/newstest2019.tgz?1556572561))
metrics:
- http://www.statmt.org/wmt19/metrics-task.html
---
# FSMT
## Model description
This is a ported version of [fairseq wmt19 transformer](https://github.com/pytorch/fairseq/blob/master/examples/wmt19/README.md) for de-en.
For more details, please see, [Facebook FAIR's WMT19 News Translation Task Submission](https://arxiv.org/abs/1907.06616).
The abbreviation FSMT stands for FairSeqMachineTranslation
All four models are available:
* [wmt19-en-ru](https://huggingface.co/facebook/wmt19-en-ru)
* [wmt19-ru-en](https://huggingface.co/facebook/wmt19-ru-en)
* [wmt19-en-de](https://huggingface.co/facebook/wmt19-en-de)
* [wmt19-de-en](https://huggingface.co/facebook/wmt19-de-en)
## Intended uses & limitations
#### How to use
```python
from transformers.tokenization_fsmt import FSMTTokenizer
from transformers.modeling_fsmt import FSMTForConditionalGeneration
mname = "facebook/wmt19-de-en"
tokenizer = FSMTTokenizer.from_pretrained(mname)
model = FSMTForConditionalGeneration.from_pretrained(mname)
input = "Maschinelles Lernen ist großartig, oder?"
input_ids = tokenizer.encode(input, return_tensors="pt")
outputs = model.generate(input_ids)
decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(decoded) # Machine learning is great, isn't it?
```
#### Limitations and bias
- The original (and this ported model) doesn't seem to handle well inputs with repeated sub-phrases, [content gets truncated](https://discuss.huggingface.co/t/issues-with-translating-inputs-containing-repeated-phrases/981)
## Training data
Pretrained weights were left identical to the original model released by fairseq. For more details, please, see the [paper](https://arxiv.org/abs/1907.06616).
## Eval results
pair | fairseq | transformers
-------|---------|----------
de-en | [42.3](http://matrix.statmt.org/matrix/output/1902?run_id=6750) | 41.35
The score is slightly below the score reported by `fairseq`, since `transformers`` currently doesn't support:
- model ensemble, therefore the best performing checkpoint was ported (``model4.pt``).
- re-ranking
The score was calculated using this code:
```bash
git clone https://github.com/huggingface/transformers
cd transformers
export PAIR=de-en
export DATA_DIR=data/$PAIR
export SAVE_DIR=data/$PAIR
export BS=8
export NUM_BEAMS=15
mkdir -p $DATA_DIR
sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
echo $PAIR
PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
```
note: fairseq reports using a beam of 50, so you should get a slightly higher score if re-run with `--num_beams 50`.
## TODO
- port model ensemble (fairseq uses 4 model checkpoints)

94
model_cards/facebook/wmt19-en-de/README.md Normal file
View File

@@ -0,0 +1,94 @@
---
<!-- This file has been auto-generated by src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py - DO NOT EDIT or your changes will be lost -->
language: en, de
thumbnail:
tags:
- translation
- wmt19
license: Apache 2.0
datasets:
- http://www.statmt.org/wmt19/ ([test-set](http://matrix.statmt.org/test_sets/newstest2019.tgz?1556572561))
metrics:
- http://www.statmt.org/wmt19/metrics-task.html
---
# FSMT
## Model description
This is a ported version of [fairseq wmt19 transformer](https://github.com/pytorch/fairseq/blob/master/examples/wmt19/README.md) for en-de.
For more details, please see, [Facebook FAIR's WMT19 News Translation Task Submission](https://arxiv.org/abs/1907.06616).
The abbreviation FSMT stands for FairSeqMachineTranslation
All four models are available:
* [wmt19-en-ru](https://huggingface.co/facebook/wmt19-en-ru)
* [wmt19-ru-en](https://huggingface.co/facebook/wmt19-ru-en)
* [wmt19-en-de](https://huggingface.co/facebook/wmt19-en-de)
* [wmt19-de-en](https://huggingface.co/facebook/wmt19-de-en)
## Intended uses & limitations
#### How to use
```python
from transformers.tokenization_fsmt import FSMTTokenizer
from transformers.modeling_fsmt import FSMTForConditionalGeneration
mname = "facebook/wmt19-en-de"
tokenizer = FSMTTokenizer.from_pretrained(mname)
model = FSMTForConditionalGeneration.from_pretrained(mname)
input = "Machine learning is great, isn't it?"
input_ids = tokenizer.encode(input, return_tensors="pt")
outputs = model.generate(input_ids)
decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(decoded) # Maschinelles Lernen ist großartig, oder?
```
#### Limitations and bias
- The original (and this ported model) doesn't seem to handle well inputs with repeated sub-phrases, [content gets truncated](https://discuss.huggingface.co/t/issues-with-translating-inputs-containing-repeated-phrases/981)
## Training data
Pretrained weights were left identical to the original model released by fairseq. For more details, please, see the [paper](https://arxiv.org/abs/1907.06616).
## Eval results
pair | fairseq | transformers
-------|---------|----------
en-de | [43.1](http://matrix.statmt.org/matrix/output/1909?run_id=6862) | 42.83
The score is slightly below the score reported by `fairseq`, since `transformers`` currently doesn't support:
- model ensemble, therefore the best performing checkpoint was ported (``model4.pt``).
- re-ranking
The score was calculated using this code:
```bash
git clone https://github.com/huggingface/transformers
cd transformers
export PAIR=en-de
export DATA_DIR=data/$PAIR
export SAVE_DIR=data/$PAIR
export BS=8
export NUM_BEAMS=15
mkdir -p $DATA_DIR
sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
echo $PAIR
PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
```
note: fairseq reports using a beam of 50, so you should get a slightly higher score if re-run with `--num_beams 50`.
## TODO
- port model ensemble (fairseq uses 4 model checkpoints)

94
model_cards/facebook/wmt19-en-ru/README.md Normal file
View File

@@ -0,0 +1,94 @@
---
<!-- This file has been auto-generated by src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py - DO NOT EDIT or your changes will be lost -->
language: en, ru
thumbnail:
tags:
- translation
- wmt19
license: Apache 2.0
datasets:
- http://www.statmt.org/wmt19/ ([test-set](http://matrix.statmt.org/test_sets/newstest2019.tgz?1556572561))
metrics:
- http://www.statmt.org/wmt19/metrics-task.html
---
# FSMT
## Model description
This is a ported version of [fairseq wmt19 transformer](https://github.com/pytorch/fairseq/blob/master/examples/wmt19/README.md) for en-ru.
For more details, please see, [Facebook FAIR's WMT19 News Translation Task Submission](https://arxiv.org/abs/1907.06616).
The abbreviation FSMT stands for FairSeqMachineTranslation
All four models are available:
* [wmt19-en-ru](https://huggingface.co/facebook/wmt19-en-ru)
* [wmt19-ru-en](https://huggingface.co/facebook/wmt19-ru-en)
* [wmt19-en-de](https://huggingface.co/facebook/wmt19-en-de)
* [wmt19-de-en](https://huggingface.co/facebook/wmt19-de-en)
## Intended uses & limitations
#### How to use
```python
from transformers.tokenization_fsmt import FSMTTokenizer
from transformers.modeling_fsmt import FSMTForConditionalGeneration
mname = "facebook/wmt19-en-ru"
tokenizer = FSMTTokenizer.from_pretrained(mname)
model = FSMTForConditionalGeneration.from_pretrained(mname)
input = "Machine learning is great, isn't it?"
input_ids = tokenizer.encode(input, return_tensors="pt")
outputs = model.generate(input_ids)
decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(decoded) # Машинное обучение - это здорово, не так ли?
```
#### Limitations and bias
- The original (and this ported model) doesn't seem to handle well inputs with repeated sub-phrases, [content gets truncated](https://discuss.huggingface.co/t/issues-with-translating-inputs-containing-repeated-phrases/981)
## Training data
Pretrained weights were left identical to the original model released by fairseq. For more details, please, see the [paper](https://arxiv.org/abs/1907.06616).
## Eval results
pair | fairseq | transformers
-------|---------|----------
en-ru | [36.4](http://matrix.statmt.org/matrix/output/1914?run_id=6724) | 33.47
The score is slightly below the score reported by `fairseq`, since `transformers`` currently doesn't support:
- model ensemble, therefore the best performing checkpoint was ported (``model4.pt``).
- re-ranking
The score was calculated using this code:
```bash
git clone https://github.com/huggingface/transformers
cd transformers
export PAIR=en-ru
export DATA_DIR=data/$PAIR
export SAVE_DIR=data/$PAIR
export BS=8
export NUM_BEAMS=15
mkdir -p $DATA_DIR
sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
echo $PAIR
PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
```
note: fairseq reports using a beam of 50, so you should get a slightly higher score if re-run with `--num_beams 50`.
## TODO
- port model ensemble (fairseq uses 4 model checkpoints)

94
model_cards/facebook/wmt19-ru-en/README.md Normal file
View File

@@ -0,0 +1,94 @@
---
<!-- This file has been auto-generated by src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py - DO NOT EDIT or your changes will be lost -->
language: ru, en
thumbnail:
tags:
- translation
- wmt19
license: Apache 2.0
datasets:
- http://www.statmt.org/wmt19/ ([test-set](http://matrix.statmt.org/test_sets/newstest2019.tgz?1556572561))
metrics:
- http://www.statmt.org/wmt19/metrics-task.html
---
# FSMT
## Model description
This is a ported version of [fairseq wmt19 transformer](https://github.com/pytorch/fairseq/blob/master/examples/wmt19/README.md) for ru-en.
For more details, please see, [Facebook FAIR's WMT19 News Translation Task Submission](https://arxiv.org/abs/1907.06616).
The abbreviation FSMT stands for FairSeqMachineTranslation
All four models are available:
* [wmt19-en-ru](https://huggingface.co/facebook/wmt19-en-ru)
* [wmt19-ru-en](https://huggingface.co/facebook/wmt19-ru-en)
* [wmt19-en-de](https://huggingface.co/facebook/wmt19-en-de)
* [wmt19-de-en](https://huggingface.co/facebook/wmt19-de-en)
## Intended uses & limitations
#### How to use
```python
from transformers.tokenization_fsmt import FSMTTokenizer
from transformers.modeling_fsmt import FSMTForConditionalGeneration
mname = "facebook/wmt19-ru-en"
tokenizer = FSMTTokenizer.from_pretrained(mname)
model = FSMTForConditionalGeneration.from_pretrained(mname)
input = "Машинное обучение - это здорово, не так ли?"
input_ids = tokenizer.encode(input, return_tensors="pt")
outputs = model.generate(input_ids)
decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(decoded) # Machine learning is great, isn't it?
```
#### Limitations and bias
- The original (and this ported model) doesn't seem to handle well inputs with repeated sub-phrases, [content gets truncated](https://discuss.huggingface.co/t/issues-with-translating-inputs-containing-repeated-phrases/981)
## Training data
Pretrained weights were left identical to the original model released by fairseq. For more details, please, see the [paper](https://arxiv.org/abs/1907.06616).
## Eval results
pair | fairseq | transformers
-------|---------|----------
ru-en | [41.3](http://matrix.statmt.org/matrix/output/1907?run_id=6937) | 39.20
The score is slightly below the score reported by `fairseq`, since `transformers`` currently doesn't support:
- model ensemble, therefore the best performing checkpoint was ported (``model4.pt``).
- re-ranking
The score was calculated using this code:
```bash
git clone https://github.com/huggingface/transformers
cd transformers
export PAIR=ru-en
export DATA_DIR=data/$PAIR
export SAVE_DIR=data/$PAIR
export BS=8
export NUM_BEAMS=15
mkdir -p $DATA_DIR
sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
echo $PAIR
PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
```
note: fairseq reports using a beam of 50, so you should get a slightly higher score if re-run with `--num_beams 50`.
## TODO
- port model ensemble (fairseq uses 4 model checkpoints)