Secure your code as it's written. Use Snyk Code to scan source code in minutes - no build needed - and fix issues immediately.
import collections
import os
import shutil
from absl import logging
from absl.testing import absltest
import numpy as np
import six
from t5.data import sentencepiece_vocabulary
from t5.data import utils as dataset_utils
import tensorflow.compat.v1 as tf
import tensorflow_datasets as tfds
TaskRegistry = dataset_utils.TaskRegistry
MixtureRegistry = dataset_utils.MixtureRegistry
mock = absltest.mock
TEST_DATA_DIR = os.path.join(
os.path.dirname(os.path.abspath(__file__)), "test_data")
# _ProxyTest is required because py2 does not allow instantiating
# absltest.TestCase directly.
class _ProxyTest(absltest.TestCase):
"""Instance of TestCase to reuse methods for testing."""
maxDiff = None
def runTest(self):
pass
# Tasks is not the best description. For example, glue_v002_equal refers to a
# mixture. Calling it "finetune tasks" because we consider all glue tasks as
# a single dataset to train on.
_finetune_tasks = [
"glue_v002_proportional", # mixture
"super_glue_v102_proportional", # mixture
"cnn_dailymail_v002",
"squad_v010_allanswers",
"wmt_t2t_ende_v003",
"wmt15_enfr_v003",
"wmt16_enro_v003"
]
# ========================== GLUE and SuperGLUE ================================
MixtureRegistry.add(
"glue_v002_proportional",
_glue_tasks, default_rate=rate_num_examples)
MixtureRegistry.add(
"super_glue_v102_proportional",
_super_glue_tasks,
default_rate=rate_num_examples)
# mnli and its associated dev sets: mnli_matched and mnli_mismatched
MixtureRegistry.add(
"glue_mnli_and_dev_v002",
[t for t in _glue_tasks if "mnli" in t],
default_rate=1.0)
# ============================== Co-training ===================================
# C4, glue, squad, superglue
# The supervised tasks here are all small datasets
# Mix them proportionally to their dataset sizes.
# TODO(noam): This should be called "small_mix" or something, but we will
# keep it as en_mix to avoid restarting experiments.
# TODO(noam): some rates should be reduced - but not now to avoid restarting
# experiments. They are:
# - Tasks duplicated between glue and superglue (see _dedupe)
# - squad and glue_qnli are duplicates
# - glue_sst2 may contain overlapping phrases (related examples with itself)
# - we seem to overtrain on super_glue_record - don't know why
MixtureRegistry.add(
"en_mix",
[("c4_v020_unsupervised", rate_unsupervised)] +
_glue_tasks + _super_glue_tasks +
["squad_v010_allanswers"],
default_rate=rate_num_examples)
MixtureRegistry.add(
"all_equal",
_supervised_tasks + ["c4_v020_unsupervised"],
default_rate=1.,
)
def _dedupe(name):
if "glue" in name and "rte" in name:
return functools.partial(rate_num_examples, scale=0.5)
def add(cls, name, tasks, default_rate=None):
super(MixtureRegistry, cls).add(name, Mixture, tasks, default_rate)
def get_mixture_or_task(task_or_mixture_name):
"""Return the Task or Mixture from the appropriate registry."""
mixtures = MixtureRegistry.names()
tasks = TaskRegistry.names()
if task_or_mixture_name in mixtures:
if task_or_mixture_name in tasks:
logging.warning("%s is both a Task and a Mixture, returning Mixture",
task_or_mixture_name)
return MixtureRegistry.get(task_or_mixture_name)
if task_or_mixture_name in tasks:
return TaskRegistry.get(task_or_mixture_name)
else:
raise ValueError("No Task or Mixture found with name: %s" %
task_or_mixture_name)
def get_mixture_or_task(task_or_mixture_name):
"""Return the Task or Mixture from the appropriate registry."""
mixtures = MixtureRegistry.names()
tasks = TaskRegistry.names()
if task_or_mixture_name in mixtures:
if task_or_mixture_name in tasks:
logging.warning("%s is both a Task and a Mixture, returning Mixture",
task_or_mixture_name)
return MixtureRegistry.get(task_or_mixture_name)
if task_or_mixture_name in tasks:
return TaskRegistry.get(task_or_mixture_name)
else:
raise ValueError("No Task or Mixture found with name: %s" %
task_or_mixture_name)