def test_integration(self):
level_origin = logging.get_verbosity()
logger = logging.get_logger(
"transformers.models.bart.tokenization_bart")
msg = "Testing 1, 2, 3"
# should be able to log warnings (if default settings weren't overridden by `pytest --log-level-all`)
if level_origin <= logging.WARNING:
with CaptureLogger(logger) as cl:
logger.warning(msg)
self.assertEqual(cl.out, msg + "\n")
# this is setting the level for all of `transformers.*` loggers
logging.set_verbosity_error()
# should not be able to log warnings
with CaptureLogger(logger) as cl:
logger.warning(msg)
self.assertEqual(cl.out, "")
# should be able to log warnings again
logging.set_verbosity_warning()
with CaptureLogger(logger) as cl:
logger.warning(msg)
self.assertEqual(cl.out, msg + "\n")
# restore to the original level
logging.set_verbosity(level_origin)
def LoggingLevel(level):
"""
This is a context manager to temporarily change transformers modules logging level to the desired value and have it
restored to the original setting at the end of the scope.
For example ::
with LoggingLevel(logging.INFO):
AutoModel.from_pretrained("gpt2") # calls logger.info() several times
"""
orig_level = transformers_logging.get_verbosity()
try:
transformers_logging.set_verbosity(level)
yield
finally:
transformers_logging.set_verbosity(orig_level)
def instantiate_trainer(config):
verbosity = config.get("verbosity", logging.INFO)
t_logging.set_verbosity(verbosity)
logger.setLevel(verbosity)
# debug (see torch.autograd.detect_anomaly)
set_detect_anomaly(bool(config.get("debug", False)))
# model
model_args = dict(name="ViT-B/32",
jit=False,
training=True,
Class="CLIPDecoder")
model_args.update(config.get("model", {}))
model_args["Class"] = getattr(clip.model, model_args["Class"])
logger.info(f"loading model from pre-trained CLIP {model_args}...")
model, image_preprocess = load(**model_args)
# data
train_dataset, eval_dataset = get_datasets(
image_preprocess=image_preprocess, **config.get("dataset", {}))
# training
criterion_args = config.get("criterion", {})
# get criterion class (e.g. nn.NLLLoss) by name
CriterionClass = getattr(nn, criterion_args.pop("Class", "NLLLoss"))
criterion = CriterionClass(**criterion_args)
learner_args = config.get("learner", {})
LearnerClass = getattr(sys.modules[__name__],
learner_args.pop("Class", "LanguageModel"))
learner = LearnerClass(model, criterion)
training_args = Seq2SeqTrainingArguments(**config.get("training", {}))
trainer = CLIPTrainer(model=learner,
args=training_args,
data_collator=collate_batch,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics)
# training callbacks
for callback in config.get("callbacks", []):
CallbackClass = getattr(trainer_callback, callback.pop("Class"))
trainer.add_callback(CallbackClass(**callback))
return trainer, training_args, config
def test_set_level(self):
logger = logging.get_logger()
# the current default level is logging.WARNING
level_origin = logging.get_verbosity()
logging.set_verbosity_error()
self.assertEqual(logger.getEffectiveLevel(), logging.get_verbosity())
logging.set_verbosity_warning()
self.assertEqual(logger.getEffectiveLevel(), logging.get_verbosity())
logging.set_verbosity_info()
self.assertEqual(logger.getEffectiveLevel(), logging.get_verbosity())
logging.set_verbosity_debug()
self.assertEqual(logger.getEffectiveLevel(), logging.get_verbosity())
# restore to the original level
logging.set_verbosity(level_origin)
help='directory with .wav files for queries')
parser.add_argument('--references_dir',
default='references',
help='directory with .wav files for references')
parser.add_argument('--model', default='wav2vec2-large-xlsr-53')
parser.add_argument(
'--hft_logging',
default=40,
help=
'HuggingFace Transformers verbosity level (40 = errors, 30 = warnings, 20 = info, 10 = debug)'
)
args = parser.parse_args()
logging.set_verbosity(args.hft_logging)
def load_wav2vec2_featurizer(model, layer=None):
"""
Loads Wav2Vec2 featurization pipeline and returns it as a function.
Featurizer returns a list with all hidden layer representations if "layer" argument is None.
Otherwise, only returns the specified layer representations.
"""
model_spec = KNOWN_MODELS.get(model, model)
model_kwargs = {}
if layer is not None:
model_kwargs["num_hidden_layers"] = layer if layer > 0 else 0
if type(model_spec) is dict:
from transformers import BertForMaskedLM, BertTokenizer, logging
from gensim.models import KeyedVectors
from nltk.stem import PorterStemmer
import nltk
import torch
import math
from . import config
# Word embeddings.
embeddings = None
# Surpress warnings.
logging.set_verbosity(logging.ERROR)
# Load BERT model and tokenizer.
print("\nLoading BERT model...")
tokenizer = BertTokenizer.from_pretrained("bert-base-multilingual-uncased")
model = BertForMaskedLM.from_pretrained("bert-base-multilingual-uncased")
print("\nBERT model loaded!")
# Create stemmer object.
stemmer = PorterStemmer()
def load_embeddings(language):
"""Load word embeddings for selected language."""
try:
print("\nAttempting to load embeddings...")
wv_path = "simplifier/embeddings/" + config.lang + ".kv"
wv_model = KeyedVectors.load(wv_path, mmap='r')
import os
import torch
import soundfile as sf
import pandas as pd
import numpy as np
from tqdm import tqdm
from argparse import ArgumentParser
from transformers import logging
from transformers.models.wav2vec2 import Wav2Vec2Model
logging.set_verbosity(40)
parser = ArgumentParser(
prog='Wav2Vec2 Featurizer',
description='Extract features aggregated across intervals specified in input csv file',
)
parser.add_argument('--model', default='wav2vec2-large-xlsr-53')
parser.add_argument('--intervals_csv', help = 'CSV file of intervals')
parser.add_argument('--features_csv', help = 'Name of output CSV file')
args = parser.parse_args()
KNOWN_MODELS = {
# Pre-trained
'wav2vec2-base': 'facebook/wav2vec2-base',
'wav2vec2-large': {'name' : 'facebook/wav2vec2-large', 'revision' : '85c73b1a7c1ee154fd7b06634ca7f42321db94db' },
# March 11, 2021 version: https://huggingface.co/facebook/wav2vec2-large/commit/85c73b1a7c1ee154fd7b06634ca7f42321db94db
'wav2vec2-large-lv60': 'facebook/wav2vec2-large-lv60',
