How to use the t5.data.utils.MixtureRegistry.add function in t5

# Special case to treat all GLUE tasks as one task.
  if task_name == "glue_v002_proportional":
    task_names -= set(_glue_tasks)
    # No de-duping needed
    tasks = [(t, rate_num_examples) for t in task_names]
  # Special case to treat all Super GLUE tasks as one task.
  elif task_name == "super_glue_v102_proportional":
    task_names -= set(_super_glue_tasks)
    # No de-duping needed
    tasks = [(t, rate_num_examples) for t in task_names]
  else:
    task_names -= {task_name}
    # Use de-duping since we have GLUE and SuperGLUE
    tasks = [(t, _dedupe(t)) for t in task_names]

  MixtureRegistry.add("leave_one_out_{}".format(task_name), tasks)

# ================= Pre-train on supervised tasks ==============================

_large_translation_tasks = ["wmt_t2t_ende_v003",
                            "wmt15_enfr_v003"]

_large_supervised_tasks = _large_translation_tasks + ["cnn_dailymail_v002"]

MixtureRegistry.add(
    "large_supervised_equal",
    _large_supervised_tasks,
    default_rate=1.0)

MixtureRegistry.add(
    "large_supervised_proportional",
    _large_supervised_tasks,

# ========================== 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

_large_supervised_tasks,
    default_rate=rate_num_examples)

MixtureRegistry.add(
    "large_translation_equal",
    _large_translation_tasks,
    default_rate=1.0)

# =========================== Squad + Trivia QA ================================
MixtureRegistry.add(
    "squad_trivia_qa_equal",
    ["squad_v010_allanswers", "trivia_qa_v010"],
    default_rate=1.0)

# ================================= WSC + DPR ==================================
MixtureRegistry.add(
    "wsc_dpr_simple_proportional",
    _wsc_dpr_tasks,
    default_rate=rate_num_examples)

MixtureRegistry.add("leave_one_out_{}".format(task_name), tasks)

# ================= Pre-train on supervised tasks ==============================

_large_translation_tasks = ["wmt_t2t_ende_v003",
                            "wmt15_enfr_v003"]

_large_supervised_tasks = _large_translation_tasks + ["cnn_dailymail_v002"]

MixtureRegistry.add(
    "large_supervised_equal",
    _large_supervised_tasks,
    default_rate=1.0)

MixtureRegistry.add(
    "large_supervised_proportional",
    _large_supervised_tasks,
    default_rate=rate_num_examples)

MixtureRegistry.add(
    "large_translation_equal",
    _large_translation_tasks,
    default_rate=1.0)

# =========================== Squad + Trivia QA ================================
MixtureRegistry.add(
    "squad_trivia_qa_equal",
    ["squad_v010_allanswers", "trivia_qa_v010"],
    default_rate=1.0)

# ================================= WSC + DPR ==================================

# 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)
  return rate_num_examples

MixtureRegistry.add(
    "all_proportional",
    [(t, _dedupe(t)) for t in _supervised_tasks + ["c4_v020_unsupervised"]],
)

["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)
  return rate_num_examples

MixtureRegistry.add(
    "all_proportional",
    [(t, _dedupe(t)) for t in _supervised_tasks + ["c4_v020_unsupervised"]],
)

# all_mix is the same as all_proportional except it uses rate_unsupervised
# for c4_v020_unsupervised. This is useful if you want to specify a specific
# rate for the unsupervised task which is different from the global value for
# rate_num_examples.maximum
# If you use this task, you should set a maximum rate value via gin e.g.
# --gin_param="dataset_utils.rate_num_examples.maximum = 1e6"
MixtureRegistry.add(
    "all_mix",
    ([("c4_v020_unsupervised", rate_unsupervised)] +
     [(t, _dedupe(t)) for t in _supervised_tasks]),
)

"large_supervised_equal",
    _large_supervised_tasks,
    default_rate=1.0)

MixtureRegistry.add(
    "large_supervised_proportional",
    _large_supervised_tasks,
    default_rate=rate_num_examples)

MixtureRegistry.add(
    "large_translation_equal",
    _large_translation_tasks,
    default_rate=1.0)

# =========================== Squad + Trivia QA ================================
MixtureRegistry.add(
    "squad_trivia_qa_equal",
    ["squad_v010_allanswers", "trivia_qa_v010"],
    default_rate=1.0)

# ================================= WSC + DPR ==================================
MixtureRegistry.add(
    "wsc_dpr_simple_proportional",
    _wsc_dpr_tasks,
    default_rate=rate_num_examples)

if "glue" in name and "rte" in name:
    return functools.partial(rate_num_examples, scale=0.5)
  return rate_num_examples

MixtureRegistry.add(
    "all_proportional",
    [(t, _dedupe(t)) for t in _supervised_tasks + ["c4_v020_unsupervised"]],
)

# all_mix is the same as all_proportional except it uses rate_unsupervised
# for c4_v020_unsupervised. This is useful if you want to specify a specific
# rate for the unsupervised task which is different from the global value for
# rate_num_examples.maximum
# If you use this task, you should set a maximum rate value via gin e.g.
# --gin_param="dataset_utils.rate_num_examples.maximum = 1e6"
MixtureRegistry.add(
    "all_mix",
    ([("c4_v020_unsupervised", rate_unsupervised)] +
     [(t, _dedupe(t)) for t in _supervised_tasks]),
)

# ================== Leave-one-out cotrain then finetune =======================

for task_name in _finetune_tasks:
  task_names = set(_supervised_tasks + ["c4_v020_unsupervised"])

  # Special case to treat all GLUE tasks as one task.
  if task_name == "glue_v002_proportional":
    task_names -= set(_glue_tasks)
    # No de-duping needed
    tasks = [(t, rate_num_examples) for t in task_names]
  # Special case to treat all Super GLUE tasks as one task.