The purpose of the trainer is to provide the user with a flexible way of scheduling various sources of input data, as well as augment the training data with tittle casing, all caps, etc. This is particularly useful when you have multiple data sources and you want to pretrain the model first on backtranslated data, gradually add other sources of data, and finally fine tune, all in one go.
Alternatively, this tool is particularly suited to training multilingual models, as it provides an easy way to define the desired mixture of datasets from different language sources.
You've got two options: Install directly from PyPI:
pip install opustrainer
or clone this repository, and install it in editable mode so you can change the source, but still use all the commands:
git clone git@github.com:hplt-project/opustrainer.git
cd opustrainer
pip install -e .
% opustrainer-train --help
usage: opustrainer-train [-h] --config CONFIG [--state STATE] [--sync] [--temporary-directory TEMPORARY_DIRECTORY] [--do-not-resume] [--no-shuffle] [--log-level LOG_LEVEL] [--log-file LOG_FILE] ...
Feeds marian tsv data for training.
positional arguments:
trainer Trainer program that gets fed the input. If empty it is read from config.
options:
-h, --help show this help message and exit
--config CONFIG, -c CONFIG
YML configuration input.
--state STATE, -s STATE
YML state file, defaults to ${CONFIG}.state.
--sync Do not shuffle async
--temporary-directory TEMPORARY_DIRECTORY, -T TEMPORARY_DIRECTORY
Temporary dir, used for shuffling and tracking state
--do-not-resume, -d Do not resume from the previous training state
--no-shuffle, -n Do not shuffle, for debugging
--log-level LOG_LEVEL
Set log level. Available levels: DEBUG, INFO, WARNING, ERROR, CRITICAL. Default is INFO
--log-file LOG_FILE, -l LOG_FILE
Target location for logging. Always logs to stderr and optionally to a file.
Once you fix the paths in the configuration file, train_config.yml
you can run a test case by doing:
opustrainer-train -c train_config.yml /path/to/marian -c marian_config --any --other --flags
You can check resulting mixed file in /tmp/test
. If your neural network trainer doesn't support training from stdin
, you can use this tool to generate a training dataset and then disable data reordering or shuffling at your trainer implementation, as your training input should be balanced.
At the start of the training all datasets are shuffled. Each time a dataset's end is reached, it is re-shuffled. Shuffling in the system temp directory but can be repositioned using --temporary-directory
or the TMPDIR
environment variable. By default, the training state is kept in the same place as the configuration file. If training is interrupted, re-running the trainer should resume from where it was (depending on how much your neural network trainer has buffered, that part will be skipped).
Define your training process via a configuration file. You define the datasets on top, the stages and then for each stage a mixing criteria and a stage termination criteria. An example configuration file is provided below. The path to the trainer
is a path to any neural network trainer that supports having stdin as training input format.
# Datasets are already TSV files. We support reading gzip'd files, as well as multiple dataset file per name
datasets:
clean: test/data/clean
medium: test/data/medium
dirty: test/data/dirty
stages:
- start
- mid
- end
start:
- clean 0.8
- medium 0.2
- dirty 0
- until clean 2 # Until two epochs of clean
mid:
- clean 0.6
- medium 0.3
- dirty 0.1
- until medium 1
end:
- clean 0.4
- medium 0.3
- dirty 0.3
- until dirty 5 # use `inf` to mean until forever
modifiers:
- UpperCase: 0.05 # Apply uppercase randomly to 5% of sentences. See below
- TitleCase: 0.05
seed: 1111
num_fields: 2 # Assures every line has exactly 2 fields.
trainer: /path/to/trainer/run.py
If num_fields
is provided, at read time, the trainer will strip any extra TSV fields that the dataset contains (such as optinal alignment field that you are not going to use). Furthermore, any line that doesn't have enough fields gets filtered (eg lines missing alignment info when you do actually care about alignment).
If you want to change which modifiers are used for a specific stage, you can the extended stage configuration format. If a modifiers
is mentioned here, it will override the curriculum-wide defined modifiers
for just this stage.
In the extended format, the list of datasets is defined in the mix
key. You can optionally add a modifiers
key. For example:
start:
mix:
- clean 0.8
- medium 0.2
- dirty 0
- until clean 2 # Until two epochs of clean
modifiers:
- UpperCase: 0.05
- TitleCase: 0.05
Note that you can use YAML references if you wish to extensively combine global and local modifiers.
Modifiers are randomly applied to the sentences that go into the trainer. Each modifier has a probability associated with it that is the chance that a sentence is modified by the modifier. E.g. a modifier with a probability of 0.05 will affect about 1 in every 20 sentences.
Modifiers are applied one after another, in the order you defined them, all with their own probability regardless of the modifiers that got applied before it. E.g. if you have the following configuration:
modifiers:
- UpperCase: 0.05
- TitleCase: 0.05
This means that 1 in 20 sentences will be uppercased, and 1 in 20 will be titlecased. And effectively 0.05 * 0.05
so 1 in 400 will first be uppercased and then titlecased.
Turns the entire source and target sentence to upper case, e.g. 'heLLo' becomes 'HELLO'.
modifiers:
- UpperCase: 0.05
Makes the first letter of every word uppercase, and the rest lowercase. Words are split by spaces. E.g. 'heLLo' becomes 'Hello'.
modifiers:
- TitleCase: 0.05
Introduce typos in the source side of the sentence pair.
The probability of the modifier itself is the chance a sentence is affected. The probabilities of each of the types of typos describes the chance a word is affected. Each type of typo occurs at most once in a sentence.
You can specify a probability for each modifier individually. If any of the typo classes is omitted, it has a probability of 0. Alternatively, you can omit all typo classes. Then all of them will have a default 10% probability.
modifiers:
- Typos: 0.05
char_swap: 0.1 # Swaps two random consecutive word characters in the string.
missing_char: 0.1 # Skips a random word character in the string.
extra_char: 0.1 # Adds an extra, keyboard-neighbor, letter next to a random word character.
nearby_char: 0.1 # Replaces a random word character with keyboard-neighbor letter.
similar_char: 0.1 # Replaces a random word character with another visually similar character.
skipped_space: 0.1 # Skips a random space from the string.
random_space: 0.1 # Adds a random space in the string.
repeated_char: 0.1 # Repeats a random word character.
unichar: 0.1 # Replaces a random consecutive repeated letter with a single letter.
Adds a modifier that merges up to n
lines lines together. The idea is that sometimes we want to see longer sequences so that we are more robust.
modifiers:
- Merge: 0.01
min_lines: 2 # Minimum lines to merge together
max_lines: 4 # Maximum lines to merge together
Adds a noise modifier that inserts sentence pair containing identical random unicode noise on the source and target side. This is useful to teach the model to copy things it doesn't understand (IE notranslate).
modifiers:
- Noise: 0.01
min_word_length: 2 # Minimum word length for each word in the noisy sentence
max_word_length: 5 # Maximum word length for each word in the noisy sentence
max_words: 6 # Maximum number of words in each noisy sentence
Adds a placeholder tag to the source sentence that can be used by the model to hint how it should translate that word. The word to hint is chosen at random from the target sentence. Only words with a 1-to-1 mapping between source and target are considered.
This modifier needs a third column in the training data with per-word (technically: space separated token) alignment information.
- Tags: 0.05
custom_detok_src: null
custom_detok_trg: zh
spm_vocab: path/to/vocab.enzh.spm
template: "__source__ {src} __target__ {trg} __done__"
All options are optional.
You can specify custom detokenizer languages using custom_detok_src
and custom_detok_trg
if the dataset you're reading from has been tokenized by the Moses tokenizer. This can be helpful to do for languages that do not use spaces to delimit words. The default tokenisation strategy is splitting/joining by spaces.
The spm_vocab
option can be used to recompute the alignment info to match the tokenisation from the sentencepiece vocabulary. This is mostly useful for Marian, which takes untokenised input but expects the alignment info to match the sentencepiece tokenisation it performs. Note that at the moment alignment info is only produced when spm_vocab
is given.
The format for telling the translation model the intention to translate a word in a certain way can be controlled by template
. Here {src}
and {trg}
are replaced by the selected words from the source and target side of the sentence pair.
Sometimes we want to just replace the source token with the target token directly, so during terminology inference the model doesn't try to think too hard what to do, but always places the hinted token on the target side. See contrib/test_enzh_noise_config.yml
for example usage.
modifiers:
- Tags: 0.1
custom_detok_src: null # Null value for the src detokenizer
custom_detok_trg: zh
replace: 0.4 # 0.4 out of the time tags is triggered, instead replace the target token with random noise, and use that random noise to tag a corresponding source word.
If alignment information is present, we can augment the training data with inline unicode noise that appears at the appropriate location on both the source and the target. This is useful to teach the model to copy things it doesn't understand (IE notranslate). See contrib/test_enzh_noise_config.yml
for example usage.
modifiers:
- Tags: 0.1
custom_detok_src: null # Null value for the src detokenizer
custom_detok_trg: zh
augment: 0.4 # 0.4 out of the time tags is triggered, instead augment the source and the target with random noise. If you want 100% only noise without tag functionality use augment: 1
Note: Due to how most modifiers are implemented, they will have a normalising effect on spaces. Sequences of spaces will be collapsed into a single space. This is also true for the Tags modifier.
Note: Even if the probability of the Tags modifier is set to 0, it will apply detokenisation and optionally re-computation of the alignment on every sentence pair, regardless whether it was picked out to be modified or not.
Prepends a random subsection of the target sentence before the source sentence.
This is useful for teaching the model to force decode a specific string if the user is absolutely certain it has to appear in the output. For example I like pie. Me gustan los pasteles.
becomes __start__ los pasteles __end__ I like pie. Me gustan los pasteles.
Note: The Prefix modifier must always be used as the last modifier, but ideally never used together with "Tags".
modifiers:
- Prefix: 0.5
min_words: 2
max_words: 5
template: "__start__ {trg} __end__ "
In order to train models supporting vocabulary "hints" we provide a tags system where we laverage word alignment information to provide "hints" for the translation model to what it should produce on the Target side. The work is very similar to Dinu et al. 2019. An example. Given an alignment augmented tsv training line:
I like pies! \t Ich mag Kuchen! \t 0-0 1-1 2-2
The machine translation traing system would see something like:
I like <tag0> mag <tag0> pies! \t Ich mag Kuchen!
Where the numeric ID of the tag is drawn from a random distribution from the total number of tags. The probability asigned to the Tags
modifier determines how likely is it for a tag augmentation to appear on any given word.
Finally, if your corpus is augmented with alignment info, but you don't want to use any tags, just set the probability of this modifier to 0.
In the future, this will be handled by a training Pipeline, but until then here's the basic scripts used
For producing alignment augmented corpus use this script:
#!/bin/bash -v
# Usage: ./align_corpus.sh source_corpus target_corpus src trg
# install fast align
mkdir -p bin
# download and compile fast_align
if [ ! -e bin/fast_align ]; then
git clone https://github.com/clab/fast_align
mkdir -p fast_align/build
cd fast_align/build
cmake ..
make -j4
cp fast_align atools ../../bin
cd ../../
fi
# Prepare the corpus for fast align
test -s $2/corpus.tmp.${3}-${4}.falign || cat $1 | sed 's/\t/ ||| /' > $2/corpus.tmp.${3}-${4}.falign
# Align it
test -s $2/align.${3}-${4}.s2t || bin/fast_align -vod -i $2/corpus.tmp.${3}-${4}.falign > $2/align.${3}-${4}.s2t
test -s $2/align.${3}-${4}.t2s || bin/fast_align -vodr -i $2/corpus.tmp.${3}-${4}.falign > $2/align.${3}-${4}.t2s
test -s $2/corpus.${3}-${4}.aln || bin/atools -i $2/align.${3}-${4}.s2t -j $2/align.${3}-${4}.t2s -c grow-diag-final-and > $2/corpus.${3}-${4}.aln
For creating vocabulary with tags support, use this script:
#!/usr/bin/env bash
#Usage ./vocab.sh en de path-to-corpora char-cov vocab_size
char_cov=${4:-'0.9995'} # Default char coverage
vocab_size=${5:-'32000'} # Default vocab size
# Set up some constants
# Language pairs
src=$1
trg=$2
prefix="--model_prefix=model.${src}-${trg}"
# Placeholders array
placeholders="--user_defined_symbols=__source__,__target__,__done__,__start__,__end__"
# Character coverage. CJK is recommended to have 0.9995, vocab languages proabbly you want 1.
char_cov="--character_coverage=${char_cov}"
# First clone and compile SPM
spm_exec="sentencepiece/build/src/spm_train"
if [ ! -e ${spm_exec} ]; then
git clone https://github.com/google/sentencepiece.git
cd sentencepiece
mkdir build
cd build
cmake ..
make -j4
cd ..
cd ..
if [ ! -e ${spm_exec} ]; then
echo "Failed to compile sentencepiece"
exit 1
fi
fi
$spm_exec --bos_id=-1 --eos_id=0 --unk_id=1 ${placeholders} ${char_cov} ${prefix} --vocab_size=${vocab_size} --input=${3} --input_sentence_size=20000000 --byte_fallback #--input_format=tsv seems broken
- Terminology support (using a dictionary), where augmentation happens not by using alignment scores but by taking values from a dictionary.
- A one click run training
This project has received funding from the European Union’s Horizon Europe research and innovation programme under grant agreement No 101070350 and from UK Research and Innovation (UKRI) under the UK government’s Horizon Europe funding guarantee [grant number 10052546]