test_discut / reading.py
poyum's picture
gradio space init
f709e5e
#!/usr/bin/python
# -*- coding: utf-8 -*-
import os, sys
import datasets
import transformers
import disrpt_io
import utils
# TODO to rm when dealt with this issue of loading languages
##from ersatz import utils
##LANGUAGES = utils.MODELS.keys()
LANGUAGES = []
def read_dataset( input_path, output_path, config, add_lang_token=True,add_frame_token=True,lang_token="",frame_token="" ):
'''
- Read the file in input_path
- Return a Dataset corresponding to the file
Parameters
----------
input_path : str
Path to the dataset
output_path : str
Path to an output directory that can be used to write new split files
tokenizer : AutoTokenizer
Tokenizer corresponding the checkpoint model
add_lang_token : bool
If True, add a special language token at the beginning of each sequence
Returns
-------
Dataset
Contain Dataset built from train_path and dev_path for train mode,
only dev / test pasth else
Tokenizer
The tokenizer used for the dataset
'''
model_checkpoint = config["model_checkpoint"]
# -- Init tokenizer
tokenizer = transformers.AutoTokenizer.from_pretrained( model_checkpoint )
# -- Read and tokenize
dataset = DatasetSeq( input_path, output_path, config, tokenizer, add_lang_token=add_lang_token,add_frame_token=add_frame_token,lang_token=lang_token,frame_token=frame_token )
dataset.read_and_tokenize()
# TODO move in class? or do elsewhere
LABEL_NAMES_BIO = retrieve_bio_labels( dataset ) # TODO should do it only once for all
dataset.set_label_names_bio(LABEL_NAMES_BIO)
return dataset, tokenizer
# --------------------------------------------------------------------------
# DatasetDict
class DatasetDisc( ):
def __init__(self, annotations_file, output_path, config, tokenizer, dset=None ):
"""
Here we save the location of our input file,
load the data, i.e. retrieve the list of texts and associated labels,
build the vocabulary if none is given,
and define the pipelines used to prepare the data
"""
self.annotations_file = annotations_file
if isinstance(annotations_file, str) and not os.path.isfile(annotations_file):
print("this is a string dataset")
self.basename = "input"
else:
self.basename = os.path.basename( self.annotations_file )
self.dset = self.basename.split(".")[2].split('_')[1]
self.corpus_name = self.basename.split('_')[0]
self.tokenizer = tokenizer
self.config = config
# If a sentence splitter is used, the files with the new segmentation will be saved here
self.output_path = output_path
# Retriev info from config: TODO check against info from dir name?
self.mode = config["type"]
self.task = config["task"]
self.trace = config["trace"]
self.tok_config = config["tok_config"]
self.sent_spliter = config["sent_spliter"]
# Additional fields
self.id2label, self.label2id = {}, {}
# -- Use disrpt_io to read the file and retrieve annotated data
self.corpus = init_corpus( self.task ) # initialize a Corpus instance, depending on the task
def read_and_tokenize( self ):
print("\n-- READ FROM FILE:", self.annotations_file )
try:
self.read_annotations( )
except Exception as err:
print(f"Unexpected {err=}, {type(err)=}", file=sys.stderr)
raise
# print( "Problem when reading", self.annotations_file )
#print("\n-- SET LABELS")
self.set_labels( )
print( "self.label2id", self.label2id )
#print("\n-- TOKENIZE DATASET")
self.tokenize_dataset()
if self.trace:
if self.dset:
print( "\n-- FINISHED READING", self.dset, "PRINTING TRACE --")
self.print_trace()
def tokenize_datasets( self ):
# Specific to subclasses
raise NotImplementedError
def set_labels( self ):
# Specific to subclasses
raise NotImplementedError
# outside the class?
# TODO use **kwags instead?
def read_annotations( self ):
'''
Generate a Corpus object based on the input_file.
Since .tok files are not segmented into sentences, a sentence splitter
is used (here, ersatz)
'''
if os.path.isfile(self.annotations_file):
self.corpus.from_file(self.annotations_file)
lang = os.path.basename(self.annotations_file).split(".")[0]
frame = os.path.basename(self.annotations_file).split(".")[1]
base = os.path.basename(self.annotations_file)
else:
# on suppose que c’est du texte brut déjà au format attendu
src = self.mode if self.mode in ["tok", "conllu", "split"] else "conllu"
self.corpus.from_string(self.annotations_file,src=src)
lang = self.lang_token
frame = self.frame_token
base = "input.text"
#print(f"[DEBUG] lang? {lang}")
for doc in self.corpus.docs:
doc.lang = lang
doc.frame = frame
# print(corpus)
# Split corpus into sentences using Ersatz
if self.mode == 'tok':
kwargs={}
from wtpsplit import SaT
sat_version="sat-3l"
if "sat_model" in self.config:
sat_version=self.config["sat_model"]
sat_model = SaT(sat_version)
kwargs["sat_model"] = sat_model
self.corpus.sentence_split(model = self.sent_spliter, lang="default-multilingual",sat_model=sat_model)
# Writing files with the split sentences
parts = base.split(".")[:-1]
split_filename = ".".join(parts) + ".split"
split_file = os.path.join(self.output_path, split_filename)
self.corpus.format(file=split_file)
# no need for sentence splitting if mode = conllu or split, no need to write files
def print_trace( self ):
print( "\n| Annotation_file: ", self.annotations_file )
print( '| Output_path:', self.output_path )
print( '| Nb_of_instances:', len(self.dataset), "(", len(self.dataset['labels']), ")" )
# "(", len(self.dataset['tokens']), len(self.dataset['labels']), ")" )
def print_stats( self ):
print( "| Annotation_file: ", self.annotations_file )
if self.dset: print( "| Data_split: ", self.dset )
print( "| Task: ", self.task )
print( "| Lang: ", self.lang )
print( "| Mode: ", self.mode )
print( "| Label_names: ", self.LABEL_NAMES)
#print( "---Number_of_documents", len( self.corpus.docs ) )
print( "| Number_of_instances: ", len(self.dataset) )
# TODO : add number of docs: not computed for .rels for now
# -------------------------------------------------------------------------------------------------
class DatasetSeq(DatasetDisc):
def __init__( self, annotations_file, output_path, config, tokenizer, add_lang_token=True, add_frame_token=True,
dset=None,lang_token="",frame_token="" ):
"""
Class for tasks corresponding to sequence labeling problem
(seg, conn).
Here we save the location of our input file,
load the data, i.e. retrieve the list of texts and associated
labels,
build the vocabulary if none is given,
and define the pipelines used to prepare the data """
DatasetDisc.__init__( self, annotations_file, output_path, config,
tokenizer )
self.add_lang_token = add_lang_token
self.add_frame_token=add_frame_token
self.lang_token = lang_token
self.frame_token=frame_token
if self.mode == 'tok' and self.output_path == None:
self.output_path = os.path.dirname( self.annotations_file )
self.output_path = os.path.join( self.output_path,
self.basename.replace("."+self.mode, ".split") )
self.sent_spliter = None
if "sent_spliter" in self.config:
self.sent_spliter = self.config["sent_spliter"] #only for seg
self.LABEL_NAMES_BIO = None
# # TODO not used, really a good idea?
# self.data_collator = transformers.DataCollatorForTokenClassification(tokenizer=self.tokenizer,
# padding=self.tok_config["padding"] )
def tokenize_dataset( self ):
# -- Create a HuggingFace Dataset object
if self.trace:
print(f"\n-- Creating dataset from generator (add_lang_token={self.add_lang_token})")
self.dataset = datasets.Dataset.from_generator(
gen,
gen_kwargs={"corpus": self.corpus, "label2id": self.label2id, "mode": self.mode, "add_lang_token": self.add_lang_token,"add_frame_token":self.add_frame_token},
)
if self.trace:
print( self.dataset[0])
# Keep track of the alignement between words ans subtokens, even if not ##
# BERT* add a tokenisation based on punctuation even if given with a list of words
self.all_word_ids = []
# Align labels according to tokenized subwords
if self.trace:
print( "\n-- Mapping dataset labels and subwords ")
self.tokenized_datasets = self.dataset.map(
tokenize_and_align_labels,
fn_kwargs = {"tokenizer":self.tokenizer,
"id2label":self.id2label,
"label2id":self.label2id,
"all_word_ids":self.all_word_ids,
"config":self.config},
batched=True,
remove_columns=self.dataset.column_names,
)
if self.trace:
print( self.tokenized_datasets[0])
def set_labels(self):
self.LABEL_NAMES = self.corpus.LABELS
self.id2label = {i: label for i, label in enumerate( self.LABEL_NAMES )}
self.label2id = {v: k for k,v in self.id2label.items()}
def set_label_names_bio( self, LABEL_NAMES_BIO ):
self.LABEL_NAMES_BIO = LABEL_NAMES_BIO
def print_trace( self ):
super().print_trace()
print( '\n--First sentence: original tokens and labels.\n')
print( self.dataset[0]['tokens'] )
print( self.dataset[0]['labels'] )
print( "\n---First sentence: tokenized version:\n")
print( self.tokenized_datasets[0] )
# print( '\nSource word ids:', len(self.all_word_ids) )
# # TODO prepaper a compute_stats before printing, to allow partial printing without trace mode
# def print_stats( self ):
# super().print_stats()
# print( "| Number_of_documents", len( self.corpus.docs ) )
def init_corpus( task ):
if task.strip().lower() == 'conn':
return disrpt_io.ConnectiveCorpus()
elif task == 'seg':
return disrpt_io.SegmentCorpus()
else:
raise NotImplementedError
def gen( corpus, label2id, mode, add_lang_token=True,add_frame_token=True ):
# Ajout d'un token spécial langue au début de chaque séquence
source = "split"
if mode == 'conllu':
source = "conllu"
for doc in corpus.docs:
lang = getattr(doc, 'lang', 'xx') if hasattr(doc, 'lang') else 'xx'
lang_token = f"[LANG={lang}]"
frame = getattr(doc, 'frame', 'xx') if hasattr(doc, 'lang') else 'xx'
frame_token = f"[FRAME={frame}]"
sent_list = doc.sentences[source] if source in doc.sentences else doc.sentences
for sentence in sent_list:
labels = []
tokens = []
if add_lang_token:
tokens.append(lang_token)
labels.append(-100)
if add_frame_token:
tokens.append(frame_token)
labels.append(-100)
#print(f"[DEBUG] Ajout du token frame {frame_token} pour la phrase: {' '.join([t.form for t in sentence.toks])}")
for t in sentence.toks:
tokens.append(t.form)
if t.label == '_':
if 'O' in label2id:
labels.append(label2id['O'])
else:
labels.append(list(label2id.values())[0])
else:
labels.append(label2id[t.label])
yield {
"tokens": tokens,
"labels": labels
}
def get_tokenizer( model_checkpoint ):
return transformers.AutoTokenizer.from_pretrained(model_checkpoint)
def tokenize_and_align_labels( dataset, tokenizer, id2label, label2id, all_word_ids, config ):
'''
(Done in batches)
To preprocess our whole dataset, we need to tokenize all the inputs and
apply align_labels_with_tokens() on all the labels.
(with HG, we could use Dataset.map to process batches)
The word_ids() function needs to get the index of the example we want
the word IDs of when the inputs to the tokenizer are lists of texts
(or in our case, list of lists of words), so we add that too:
"tok_config"
'''
tokenized_inputs = tokenizer(
dataset["tokens"],
truncation=config["tok_config"]['truncation'],
padding=config["tok_config"]['padding'],
max_length=config["tok_config"]['max_length'],
is_split_into_words=True
)
# tokenized_inputs = tokenizer(
# dataset["tokens"], truncation=True, padding=True, is_split_into_words=True
# )
all_labels = dataset["labels"]
new_labels = []
#print( "tokenized_inputs.word_ids()", tokenized_inputs.word_ids() )
#print( [tokenizer.decode(tok) for tok in tokenized_inputs['input_ids']])
##with progressbar.ProgressBar(max_value=len(all_labels)) as bar:
##for i in tqdm(range(len(all_labels))):
for i, labels in enumerate(all_labels):
word_ids = tokenized_inputs.word_ids(i)
new_labels.append(align_labels_with_tokens(labels, word_ids, id2label, label2id, tokenizer, tokenized_inputs ))
# Used to fill the self.word_ids field of the Dataset object, but should probably be done some<here else
all_word_ids.append( word_ids )
##bar.update(i)
tokenized_inputs["labels"] = new_labels
return tokenized_inputs
def align_labels_with_tokens(labels, word_ids, id2label, label2id, tokenizer, tokenized_inputs):
'''
BERT like tokenization will create new tokens, we need to align labels.
Special tokens get a label of -100. This is because by default -100 is an
index that is ignored in the loss function we will use (cross entropy).
Then, each token gets the same label as the token that started the word
it’s inside, since they are part of the same entity. For tokens inside a
word but not at the beginning, we replace the B- with I- (since the token
does not begin the entity). [Taken from HF website course on NER]
'''
count = 0
new_labels = []
current_word = None
for word_id in word_ids:
count += 1
if word_id==0: # ou 1 peut etre
#TODO
#add lang token -100
pass
if word_id != current_word:
# Start of a new word!
current_word = word_id
label = -100 if word_id is None else labels[word_id]
new_labels.append(label)
elif word_id is None:
# Special token
new_labels.append(-100)
else:
# Same word as previous token
label = labels[word_id]
# On ne cherche 'B-' que si label != -100
if label != -100 and 'B-' in id2label[label]:
label = -100
new_labels.append(label)
return new_labels
def retrieve_bio_labels( dataset ):
'''
Needed for compute_metrics, I think? It seems to be using a classic metrics for BIO
scheme, thus we create a mapping to BIO labels, i.e.:
'_' --> 'O'
'Seg=B-Conn' --> 'B'
'Seg=I-Conn' --> 'I'
Should also work for segmentation TODO: check
datasets: dict: DatasetSeq instances for train/dev/test
Return: list: original label names
list: label names mapped to BIO
'''
# need a Dataset instance to retrieve the original label sets
task = dataset.task
LABEL_NAMES_BIO = []
LABEL_NAMES = dataset.LABEL_NAMES
label2idx, idx2newl = {}, {}
if task in ["conn", "seg"]:
for i,l in enumerate( LABEL_NAMES ):
label2idx[l] = i
for l in label2idx:
nl = ''
if 'B' in l:
nl = 'B'
elif 'I' in l:
nl = 'I'
else:
nl = 'O'
idx2newl[label2idx[l]] = nl
for i in sorted(idx2newl):
LABEL_NAMES_BIO.append(idx2newl[i])
#label_names = ['O', 'B', 'I']
return LABEL_NAMES_BIO
# def _compute_distrib( dataset, id2label ):
# distrib = {}
# multi = []
# for inst in dataset:
# label = id2label[inst['label']]
# if label in distrib:
# distrib[label] += 1
# else:
# distrib[label] = 1
# len_labels = len( inst["all_labels"])
# if len_labels > 1:
# #count_multi += 1
# multi.append( len_labels )
# return distrib, multi
# Defines the language code for the sentence spliter, should be done in disrpt_io?
def set_language( lang ):
#lang = "default-multilingual" #default value
# patch
if lang=="sp": lang="es"
if lang not in LANGUAGES:
lang = "default-multilingual"
return lang
# ------------------------------------------------------------------
if __name__=="__main__":
import argparse, os
parser = argparse.ArgumentParser(
description='DISCUT: reading data from disrpt_io and converting to HuggingFace'
)
# TRAIN AND DEV are (list of) FILES or DIRECTORIES
parser.add_argument("-t", "--train",
help="Training file. Default: data_test/eng.sample.rstdt/eng.sample.rstdt_train.conllu",
default="data_test/eng.sample.rstdt/eng.sample.rstdt_train.conllu")
parser.add_argument("-d", "--dev",
help="Dev file. Default: data/eng.sample.rstdt/eng.sample.rstdt_dev.conllu",
default="data_test/eng.sample.rstdt/eng.sample.rstdt_dev.conllu")
# OUTPUT DIRECTORY
parser.add_argument("-o", "--output",
help="Directory where models and pred will be saved. Default: /home/cbraud/experiments/expe_discut_2025/",
default="")
# CONFIG FILE
parser.add_argument("-c", "--config",
help="Config file. Default: ./config_seg.json",
default="./config_seg.json")
# TRACE / VERBOSITY
parser.add_argument( '-v', '--trace',
action='store_true',
default=False,
help="Whether to print full messages. If used, it will override the value in config file.")
args = parser.parse_args()
train_path = args.train
dev_path = args.dev
print(dev_path)
if not os.path.isfile(dev_path[0]):
print( "ERROR with dev file:", dev_path)
output_path = args.output
config_file = args.config
#eval = args.eval
trace = args.trace
print( '\n-[JEDIS]--PROGRAM (reader) ARGUMENTS')
print( '| Train_path', train_path )
print( '| Dev_path', dev_path )
print( "| Output_path", output_path )
print( '| Config', config_file )
print( '\n-[JEDIS]--CONFIG INFO')
config = utils.read_config( config_file )
utils.print_config(config)
# WE override the config file if the user says no trace in arguments
# easier than modifying the config files each time
if not trace:
config['trace'] = False
print( "\n-[JEDIS]--READING DATASETS" )
# dictionnary containing train (if model=='train') and/or dev (test) Dataset instance
datasets, tokenizer = read_dataset( train_path, dev_path, config, add_lang_token=True )