def main(argv):
if len(argv) > 1:
raise RuntimeError(argv)
absl_logging.use_python_logging()
retriever_config = tf_utils.REALMSave(
**utils.from_json_file(_FLAG_RETRIEVER_CONFIG_PATH.value))
extra = "_FROM_SUBSET" if _FLAG_USE_SUBSET.value else ""
time_stamp = time.strftime("%Y%m%d-%H%M%S")
target_path = os.path.join(_FLAG_OUTPUT_PATH.value,
time_stamp + extra).strip()
if target_path[-1] != "/":
target_path += "/"
##############################################################################
# Setup devices and strategy
##############################################################################
with utils.log_duration(LOGGER, "main", "Initializing devices"):
tpu_config = tf_utils.init_tpus()
device_type = tf_utils.current_accelerator_type()
LOGGER.debug("Devices: %s", str(tf_utils.devices_to_use()))
if device_type == "TPU":
if tpu_config is None:
raise RuntimeError("We should have a tpu_config.")
strategy = tf.distribute.TPUStrategy(tpu_config.resolver)
batch_size = len(
tf_utils.devices_to_use()) * _FLAG_BATCH_SIZE.value
elif device_type == "GPU" or device_type == "CPU":
strategy = tf.distribute.MirroredStrategy()
batch_size = len(
tf_utils.devices_to_use()) * _FLAG_BATCH_SIZE.value
else:
raise RuntimeError(device_type)
##############################################################################
# Load the dataset.
##############################################################################
eli5 = {}
keys = ["train", "eval", "test"]
gpt2_tokenizer = transformers.GPT2TokenizerFast.from_pretrained("gpt2-xl")
gpt2_tokenizer.pad_token = gpt2_tokenizer.eos_token
with utils.log_duration(LOGGER, "main", "Loading the ELI5 datasets."):
for split in tqdm.tqdm(keys):
load_path = os.path.join(_FLAGS_DATASET_ROOT.value,
"HuggingfaceDatasets",
f"{split}_kilt_eli5.hf")
with tf.device("/job:localhost"):
eli5[split] = datasets.load_from_disk(load_path)
if _FLAG_USE_SUBSET.value:
_warn_subset()
##############################################################################
#
##############################################################################
with utils.log_duration(LOGGER, "Main", "Load the textual dataset"):
# Extract the appropriate text
# The buffer_size is taken from the original ORQA code.
blocks_dataset = tf.data.TFRecordDataset(retriever_config.text_records,
buffer_size=512 * 1024 * 1024)
blocks_dataset = blocks_dataset.batch(
retriever_config.num_block_records, drop_remainder=True)
blocks = tf.data.experimental.get_single_element(blocks_dataset)
with tempfile.TemporaryDirectory() as tmp_dir:
############################################################################
# Prepare the output file.
############################################################################
tmp_dir = pathlib.Path(tmp_dir)
h5_output_path = tmp_dir / "codes.h5"
output_file = h5py.File(h5_output_path, "w")
flags_dict = {
flag.name: flag.value
for flag in flags.FLAGS.flags_by_module_dict()[argv[0]]
}
utils.to_json_file(tmp_dir / "params.json", flags_dict)
for split in keys:
with utils.log_duration(
LOGGER, "main",
"Creating the output hdf5 file, embeddings."):
num_entries = len(eli5[split]["id"])
if _FLAG_USE_SUBSET.value:
num_entries = min(num_entries, _FLAG_SUBSET_AMOUNT.value)
split_group = output_file.create_group(split)
with utils.log_duration(
LOGGER, "main",
"Creating the output hdf5 file, retrieval."):
split_group.create_dataset(
constants.CTH5Fields.distances,
shape=(num_entries, _FLAG_NUM_RETRIEVALS.value),
dtype=np.float32,
)
split_group.create_dataset(
constants.CTH5Fields.gpt2_question_ids_inputs,
shape=(num_entries, _FLAG_CONTEXT_SIZE.value),
dtype=np.int32)
if split != "test":
split_group.create_dataset(
constants.CTH5Fields.gpt2_answer_ids_inputs,
shape=(num_entries, _FLAG_CONTEXT_SIZE.value),
dtype=np.int32)
split_group.create_dataset(
constants.CTH5Fields.gpt2_retrieved_ids,
shape=(
num_entries,
_FLAG_NUM_RETRIEVALS.value,
_FLAG_MAX_LENGTH_RETRIEVALS.value,
),
dtype=np.int32)
with utils.log_duration(LOGGER, "main",
"Loading the reference db."):
checkpoint_path = os.path.join(
retriever_config.query_embedder_path, "encoded",
"encoded.ckpt")
reference_db_device = tf_utils.device_mapping().CPUs[0].name
with tf.device(reference_db_device):
reference_db = tf_utils.load_reference_db(
checkpoint_path,
variable_name="block_emb",
)
############################################################################
# Prep the encoder and the tokenizer
############################################################################
with utils.log_duration(
LOGGER, "main",
"Loading the encoder model and the tokenizer."):
with strategy.scope():
query_encoder = hub.load(retriever_config.query_embedder_path,
tags={})
encode_fn = _make_encode_fn(query_encoder)
encode_fn_strategy_run = _make_encode_fn_strategy_run_fn(
strategy=strategy,
encode_fn=encode_fn,
)
vocab_file = os.path.join(retriever_config.query_embedder_path,
"assets", "vocab.txt")
utils.check_exists(vocab_file)
do_lower_case = query_encoder.signatures["tokenization_info"](
)["do_lower_case"]
tokenization_info = dict(vocab_file=vocab_file,
do_lower_case=do_lower_case)
tokenizer, vocab_lookup_table = bert_utils.get_tf_tokenizer(
query_encoder, tokenization_info)
############################################################################
# Preprocess the dataset
############################################################################
cls_token_id = tf.cast(vocab_lookup_table.lookup(tf.constant("[CLS]")),
tf.int32)
sep_token_id = tf.cast(vocab_lookup_table.lookup(tf.constant("[SEP]")),
tf.int32)
transform = _make_transform_fn(
bert_tokenizer=tokenizer,
bert_cls_token_id=cls_token_id,
bert_sep_token_id=sep_token_id,
)
with utils.log_duration(LOGGER, "main", "generating codes"):
tqdm_splits = tqdm.tqdm(keys)
for split in tqdm_splits:
tqdm_splits.set_description(f"Split `{split}`")
eli5: Dict[str, datasets.Dataset]
write_start = 0
if _FLAG_USE_SUBSET.value:
_warn_subset(tqdm_splits)
eli5[split] = eli5[split][:_FLAG_SUBSET_AMOUNT.value]
utils.check_operator(operator.le, len(eli5[split]["id"]),
_FLAG_SUBSET_AMOUNT.value)
utils.check_operator(operator.le,
len(eli5[split]["input"]),
_FLAG_SUBSET_AMOUNT.value)
else:
utils.check_equal(len(eli5[split]), len(eli5[split]["id"]))
utils.check_equal(len(eli5[split]),
len(eli5[split]["input"]))
if split != "test":
for_slices = dict(sample_id=eli5[split]["id"],
question=eli5[split]["input"],
answer=[
sample["answer"][0]
for sample in eli5[split]["output"]
])
else:
for_slices = dict(
sample_id=eli5[split]["id"],
question=eli5[split]["input"],
)
ds = tf.data.Dataset.from_tensor_slices(for_slices)
ds = ds.map(transform,
num_parallel_calls=tf.data.experimental.AUTOTUNE)
ds = ds.apply(
tf.data.experimental.dense_to_ragged_batch(batch_size))
ds = ds.map(_squeeze,
num_parallel_calls=tf.data.experimental.AUTOTUNE)
tqdm_inner = tqdm.tqdm(enumerate(ds),
total=len(eli5[split]["id"]) //
_FLAG_BATCH_SIZE.value,
desc=f"Split `{split}`: Batches")
for i, batch in tqdm_inner:
######################################################################
# Enforce the current real batch size
######################################################################
current_batch_size = batch["sample_id"].shape[0]
for k, v in batch.items():
utils.check_equal(v.shape[0], current_batch_size)
######################################################################
gpt2_question_ids_inputs = _prep_field(
batch["question"], gpt2_tokenizer)
utils.check_equal(gpt2_question_ids_inputs.dtype, np.int32)
utils.check_equal(gpt2_question_ids_inputs.shape[0],
current_batch_size)
if split != "test":
gpt2_answer_ids_inputs = _prep_field(
batch["answer"], gpt2_tokenizer)
utils.check_equal(gpt2_answer_ids_inputs.dtype,
np.int32)
utils.check_equal(gpt2_answer_ids_inputs.shape[0],
current_batch_size)
assert len(gpt2_answer_ids_inputs.shape) == 2, (
gpt2_answer_ids_inputs.shape)
######################################################################
# Save the gpt2 tokenized question and answer
######################################################################
end = write_start + current_batch_size
utils.check_equal(
output_file[split][
constants.CTH5Fields.gpt2_question_ids_inputs]
[write_start:end].shape[0], current_batch_size)
output_file[split][
constants.CTH5Fields.gpt2_question_ids_inputs][
write_start:end] = gpt2_question_ids_inputs
if split != "test":
output_file[split][
constants.CTH5Fields.gpt2_answer_ids_inputs][
write_start:end] = gpt2_answer_ids_inputs
######################################################################
# Encode the samples.
######################################################################
batch = strategy.experimental_distribute_values_from_function(
tf_utils.make_dict_distribute_fn(batch))
embeddings = encode_fn_strategy_run(batch)
embeddings = tf_utils.process_strat_output(
embeddings, "embeddings", strategy, current_batch_size)
utils.check_isinstance(embeddings, ops.EagerTensor)
utils.check_equal(embeddings.shape[0], current_batch_size)
# pytype doesn't seem to see that we check the type
utils.check_equal(embeddings.shape[1],
_FLAG_EMBEDDING_DEPTH.value) # pytype: disable=attribute-error
######################################################################
# Retrieve.
######################################################################
with tf.device(reference_db_device):
top_k, inner_prods = tf_utils.mips_exact_search(
embeddings, _FLAG_NUM_RETRIEVALS.value,
reference_db)
top_k = tf_utils.process_strat_output(
top_k, "top_k", strategy, current_batch_size)
utils.check_equal(
inner_prods.shape,
(current_batch_size, _FLAG_NUM_RETRIEVALS.value))
utils.check_equal(
top_k.shape,
(current_batch_size, _FLAG_NUM_RETRIEVALS.value))
output_file[split]["distances"][
write_start:end] = inner_prods
gathered = tf.gather(blocks, top_k).numpy()
utils.check_equal(gathered.shape[0], current_batch_size)
utils.check_equal(write_start + gathered.shape[0], end)
for j in range(gathered.shape[0]):
local_gathered = gathered[j].tolist()
utils.check_equal(len(local_gathered),
_FLAG_NUM_RETRIEVALS.value)
local_gathered = [
sample.decode() for sample in local_gathered
]
token_ids = np.array(
gpt2_tokenizer.batch_encode_plus(
local_gathered,
padding="max_length",
truncation=True,
).input_ids)
for line in token_ids:
assert not np.all(line == 0), line
token_ids[token_ids ==
gpt2_tokenizer.eos_token_id] = -1
output_file[split][
constants.CTH5Fields.gpt2_retrieved_ids][
write_start +
j] = token_ids[:, :_FLAG_MAX_LENGTH_RETRIEVALS.
value]
write_start += current_batch_size
############################################################################
# Upload the results to GCS
############################################################################
LOGGER.debug("DONE WITH THE PRODUCTION")
output_file.close()
with utils.log_duration(LOGGER, "main", "gsutil transfer"):
command = [
"/root/google-cloud-sdk/bin/gsutil", "-m", "cp", "-r",
str(tmp_dir / "*"), target_path
]
LOGGER.debug("Command: %s", " ".join(command))
subprocess.check_call(command)
LOGGER.debug("ALL DONE")
def main(argv):
if len(argv) > 1:
raise RuntimeError(argv)
absl_logging.use_python_logging()
utils.log_module_args(LOGGER, argv[0])
retriever_config = tf_utils.REALMSave(
**utils.from_json_file(_FLAG_RETRIEVER_CONFIG_PATH.value))
assert not _FLAG_USE_SUBSET.value
time_stamp = time.strftime("%Y%m%d-%H%M%S")
target_path = os.path.join(_FLAG_OUTPUT_PATH.value, time_stamp.strip())
if target_path[-1] != "/":
target_path += "/"
##############################################################################
# Setup devices and strategy
##############################################################################
with utils.log_duration(LOGGER, "main", "Initializing devices"):
tpu_config = tf_utils.init_tpus()
device_type = tf_utils.current_accelerator_type()
LOGGER.debug("Devices: %s", str(tf_utils.devices_to_use()))
if device_type == "TPU":
if tpu_config is None:
raise RuntimeError("We should have a tpu_config.")
strategy = tf.distribute.TPUStrategy(tpu_config.resolver)
batch_size = len(
tf_utils.devices_to_use()) * _FLAG_BATCH_SIZE.value
elif device_type == "GPU" or device_type == "CPU":
strategy = tf.distribute.MirroredStrategy()
batch_size = len(
tf_utils.devices_to_use()) * _FLAG_BATCH_SIZE.value
else:
raise RuntimeError(device_type)
##############################################################################
# Load the dataset.
##############################################################################
eli5 = {}
keys = ["train", "eval", "test"]
gpt2_tokenizer = transformers.GPT2TokenizerFast.from_pretrained("gpt2-xl")
gpt2_tokenizer.pad_token = gpt2_tokenizer.eos_token
with utils.log_duration(LOGGER, "main", "Loading the ELI5 datasets."):
for split in tqdm.tqdm(keys):
load_path = os.path.join(_FLAG_DATASET_ROOT.value,
"HuggingfaceDatasets",
f"{split}_kilt_eli5.hf")
with tf.device("/job:localhost"):
eli5[split] = datasets.load_from_disk(load_path)
##############################################################################
#
##############################################################################
with utils.log_duration(LOGGER, "Main", "Load the textual dataset"):
# Extract the appropriate text
# The buffer_size is taken from the original ORQA code.
blocks_dataset = tf.data.TFRecordDataset(retriever_config.text_records,
buffer_size=512 * 1024 * 1024)
blocks_dataset = blocks_dataset.batch(
retriever_config.num_block_records, drop_remainder=True)
blocks = tf.data.experimental.get_single_element(blocks_dataset)
############################################################################
# Prepare the output file.
############################################################################
writers = {}
all_paths = {}
for split in keys:
maybe_subset = "_subset" if _FLAG_USE_SUBSET.value else ""
paths = [
os.path.join(target_path + maybe_subset, f"{split}_{i}.tfr")
for i in range(_FLAG_NUM_SHARDS.value)
]
all_paths[split] = paths
writers[split] = [tf.io.TFRecordWriter(filename) for filename in paths]
with utils.log_duration(LOGGER, "main", "Loading the reference db."):
checkpoint_path = os.path.join(
retriever_config.query_embedder_path, "encoded",
"encoded.ckpt")
reference_db_device = tf_utils.device_mapping().CPUs[0].name
with tf.device(reference_db_device):
reference_db = tf_utils.load_reference_db(
checkpoint_path,
variable_name="block_emb",
)
############################################################################
# Prep the encoder and the tokenizer
############################################################################
with utils.log_duration(LOGGER, "main",
"Loading the encoder model and the tokenizer."):
with strategy.scope():
query_encoder = hub.load(retriever_config.query_embedder_path,
tags={})
encode_fn = _make_encode_fn(query_encoder)
encode_fn_strategy_run = make_encode_fn_strategy_run_fn(
strategy=strategy,
encode_fn=encode_fn,
)
vocab_file = os.path.join(retriever_config.query_embedder_path,
"assets", "vocab.txt")
utils.check_exists(vocab_file)
do_lower_case = query_encoder.signatures["tokenization_info"](
)["do_lower_case"]
tokenization_info = dict(vocab_file=vocab_file,
do_lower_case=do_lower_case)
tokenizer, vocab_lookup_table = bert_utils.get_tf_tokenizer(
query_encoder, tokenization_info)
############################################################################
# Preprocess the dataset
############################################################################
cls_token_id = tf.cast(vocab_lookup_table.lookup(tf.constant("[CLS]")),
tf.int32)
sep_token_id = tf.cast(vocab_lookup_table.lookup(tf.constant("[SEP]")),
tf.int32)
transform = _make_transform_fn(
bert_tokenizer=tokenizer,
bert_cls_token_id=cls_token_id,
bert_sep_token_id=sep_token_id,
)
feature_dtypes = {
constants.CTH5Fields.distances: tf.float32,
constants.CTH5Fields.gpt2_retrieved_ids: tf.int32,
constants.CTH5Fields.gpt2_answer_ids_inputs: tf.int32,
constants.CTH5Fields.gpt2_question_ids_inputs: tf.int32,
}
with utils.log_duration(LOGGER, "main", "generating codes"):
for split in keys:
sample_count = 0
eli5: Dict[str, datasets.Dataset]
if split != "test":
for_slices = dict(sample_id=eli5[split]["id"],
question=eli5[split]["input"],
answer=[
sample["answer"][0]
for sample in eli5[split]["output"]
])
else:
for_slices = dict(
sample_id=eli5[split]["id"],
question=eli5[split]["input"],
)
ds = tf.data.Dataset.from_tensor_slices(for_slices)
ds = ds.map(transform,
num_parallel_calls=tf.data.experimental.AUTOTUNE)
ds = ds.apply(
tf.data.experimental.dense_to_ragged_batch(batch_size))
ds = ds.map(_squeeze,
num_parallel_calls=tf.data.experimental.AUTOTUNE)
tqdm_inner = tqdm.tqdm(enumerate(ds),
total=len(eli5[split]["id"]) //
_FLAG_BATCH_SIZE.value,
desc=f"Split `{split}`: Batches")
for i, batch in tqdm_inner:
features = collections.defaultdict(list)
######################################################################
# Enforce the current real batch size
######################################################################
current_batch_size = batch["sample_id"].shape[0]
for k, v in batch.items():
utils.check_equal(v.shape[0], current_batch_size)
######################################################################
gpt2_question_ids_inputs = _prep_field(batch["question"],
gpt2_tokenizer)
utils.check_equal(gpt2_question_ids_inputs.dtype, np.int32)
utils.check_equal(gpt2_question_ids_inputs.shape[0],
current_batch_size)
if split != "test":
gpt2_answer_ids_inputs = _prep_field(
batch["answer"], gpt2_tokenizer)
utils.check_equal(gpt2_answer_ids_inputs.dtype, np.int32)
utils.check_equal(gpt2_answer_ids_inputs.shape[0],
current_batch_size)
assert len(gpt2_answer_ids_inputs.shape) == 2, (
gpt2_answer_ids_inputs.shape)
######################################################################
# Save the gpt2 tokenized question and answer
######################################################################
features[constants.CTH5Fields.gpt2_question_ids_inputs].extend(
gpt2_question_ids_inputs)
if split != "test":
features[
constants.CTH5Fields.gpt2_answer_ids_inputs].extend(
gpt2_answer_ids_inputs)
######################################################################
# Encode the samples.
######################################################################
batch = strategy.experimental_distribute_values_from_function(
tf_utils.make_dict_distribute_fn(batch))
embeddings = encode_fn_strategy_run(batch)
embeddings = tf_utils.process_strat_output(
embeddings, "embeddings", strategy, current_batch_size)
utils.check_isinstance(embeddings, ops.EagerTensor)
utils.check_equal(embeddings.shape[0], current_batch_size)
# pytype doesn't seem to see that we check the type
utils.check_equal(embeddings.shape[1],
_FLAG_EMBEDDING_DEPTH.value) # pytype: disable=attribute-error
######################################################################
# Retrieve.
######################################################################
with tf.device(reference_db_device):
top_k, inner_prods = tf_utils.mips_exact_search(
embeddings, _FLAG_NUM_RETRIEVALS.value, reference_db)
top_k = tf_utils.process_strat_output(top_k, "top_k", strategy,
current_batch_size)
utils.check_equal(
inner_prods.shape,
(current_batch_size, _FLAG_NUM_RETRIEVALS.value))
utils.check_equal(
top_k.shape,
(current_batch_size, _FLAG_NUM_RETRIEVALS.value))
features[constants.CTH5Fields.distances].extend(inner_prods)
gathered = tf.gather(blocks, top_k).numpy()
utils.check_equal(gathered.shape[0], current_batch_size)
retrievals = []
for j in range(gathered.shape[0]):
local_gathered = gathered[j].tolist()
utils.check_equal(len(local_gathered),
_FLAG_NUM_RETRIEVALS.value)
local_gathered = [
sample.decode() for sample in local_gathered
]
token_ids = np.array(
gpt2_tokenizer.batch_encode_plus(
local_gathered,
padding="max_length",
truncation=True,
).input_ids)
for line in token_ids:
assert not np.all(line == 0), line
token_ids[token_ids == gpt2_tokenizer.eos_token_id] = -1
retrievals.append(token_ids)
features[constants.CTH5Fields.gpt2_retrieved_ids] = retrievals
utils.check_equal(
retrievals[0].shape,
(_FLAG_NUM_RETRIEVALS.value, _FLAG_CONTEXT_SIZE.value))
for k, v in features.items():
utils.check_equal(len(v), current_batch_size)
for k in range(current_batch_size):
feature = tf.train.Features(
feature={
k: _bytes_feature(
tf.io.serialize_tensor(
tf.cast(v[k], feature_dtypes[k])))
for k, v in features.items()
})
writers[split][
sample_count % _FLAG_NUM_SHARDS.value].write(
tf.train.Example(
features=feature).SerializeToString())
sample_count += 1
if sample_count % 1000 == 0:
LOGGER.debug("Paths: %s", str(all_paths[split][0]))
LOGGER.debug("Done.")
def main(argv):
#######################################################################
# Initial Setup. Logging, Flags, Random seeds.
#######################################################################
if len(argv) > 1:
raise app.UsageError("Too many command-line arguments.")
absl_logging.use_python_logging()
flags_dict = {
flag.name: flag.value
for flag in FLAGS.flags_by_module_dict()[argv[0]]
}
if FLAGS.use_subset:
message = (f"{colorama.Back.RED}{colorama.Fore.WHITE}"
f"{colorama.Style.BRIGHT}USING A SUBSET OF THE DATASET"
f"{colorama.Style.RESET_ALL}")
LOGGER.warning(
message
)
utils.log_module_args(LOGGER, argv[0])
if not FLAGS.output_dir.startswith("gs://"):
utils.check_exists(FLAG_OUTPUT_DIR.value)
if not tf.io.gfile.isdir(FLAG_OUTPUT_DIR.value):
raise RuntimeError("Output dir needs to be a directory.")
tf.random.set_seed(FLAG_RANDOM_SEED.value)
np.random.seed(FLAG_RANDOM_SEED.value)
# Prepare the instance output directory path and save the config there
folder_name = time.strftime(
f"{FLAG_RUN_NAME.value}_{FLAG_APPROACH_TYPE.value}_%Y%m%d-%H%M%S"
)
instance_output_dir = os.path.join(FLAG_OUTPUT_DIR.value, folder_name).strip()
if not instance_output_dir.endswith("/"):
instance_output_dir += "/"
json_target = os.path.join(instance_output_dir, "training_params.json")
if not json_target.strip().startswith("gs://"):
subprocess.check_call(["mkdir", "-p", instance_output_dir])
utils.to_json_file(json_target, instance_output_dir)
##############################################################################
# Initialization and Configuration of the Devices.
##############################################################################
tpu_setup = None
# current_acelerator_type is always "CPU" in the beginning with TPUs
if tf_utils.current_accelerator_type() == "CPU":
tpu_setup = tf_utils.init_tpus()
LOGGER.debug("Devices we are computing on:\n%s",
utils.wrap_iterable(map(str, tf_utils.devices_to_use())))
LOGGER.debug("All devices:")
LOGGER.debug(tf_utils.device_mapping())
if tf_utils.current_accelerator_type() == "GPU":
tf.config.set_soft_device_placement(True)
if tf_utils.current_accelerator_type() != "TPU":
tf.debugging.set_log_device_placement(True)
if FLAG_DISTRIBUTE_MODE.value in constants.PURE_DATA_PARALLEL_STRATEGIES:
actual_num_replicas = len(tf_utils.devices_to_use())
elif FLAG_DISTRIBUTE_MODE.value in constants.DATA_PARALLEL_DMC:
actual_num_replicas = FLAG_NUM_REPLICAS.value
else:
actual_num_replicas = 1
##############################################################################
# We load the retriever model if it is needed.
##############################################################################
# Not currently used.
retriever = None
if (FLAG_APPROACH_TYPE.value ==
constants.ApproachTypeChoices.lm_and_realm):
config_path = FLAG_RETRIEVER_CONFIG_PATH.value
realm_save = tf_utils.REALMSave(**utils.from_json_file(config_path))
# Approx 15 min when not in dev mode, on CPU
with utils.log_duration(LOGGER, "main",
"whole of BERTScaNNRetriever.__init__",
logging.INFO):
scann_config = retrievers.ScannConfig(
**utils.from_json_file(FLAG_SCANN_CONFIG_PATH.value))
retriever = retrievers.BERTScaNNRetriever(
retriever_module_path=realm_save.query_embedder_path,
block_records_path=realm_save.text_records,
num_block_records=realm_save.num_block_records,
mode=tf.estimator.ModeKeys.EVAL,
scann_config=scann_config)
elif (FLAG_APPROACH_TYPE.value ==
constants.ApproachTypeChoices.cached_realm):
config_path = FLAG_RETRIEVER_CONFIG_PATH.value
realm_save = tf_utils.REALMSave(**utils.from_json_file(config_path))
# Approx 15 min when not in dev mode, on CPU
with utils.log_duration(LOGGER, "main",
"whole of FullyCachedRetriever.__init__",
logging.INFO):
retriever = retrievers.FullyCachedRetriever(
db_path=FLAG_FULLYCACHED_H5_PATH.value,
block_records_path=realm_save.text_records,
num_block_records=realm_save.num_block_records,
)
##############################################################################
# Distributed training task
##############################################################################
if FLAG_TASK.value == constants.TaskChoices.train:
with utils.log_duration(LOGGER, "main", "Load model"):
utils.print_mem("before loading model", LOGGER)
model_specific = task_specific.load_model(FLAG_MODEL_LOAD_PATH.value,
FLAG_MODEL_KEY.value,
FLAG_DISTRIBUTE_MODE.value,
tpu_setup,
FLAG_NUM_REPLICAS.value)
utils.print_mem("after loading model", LOGGER)
model_or_replicas = model_specific.model
if isinstance(model_or_replicas, list):
model_or_replicas: List[transformers.TFGPT2LMHeadModel]
else:
model_or_replicas: transformers.TFGPT2LMHeadModel
tokenizer = model_specific.tokenizer
def make_optimizer():
return tensor2tensor.utils.adafactor.AdafactorOptimizer(
learning_rate=FLAG_LEARNING_RATE.value)
if model_specific.strategy:
with model_specific.strategy.scope():
optimizer = make_optimizer()
else:
optimizer = make_optimizer()
############################################################################
# Prepare the dataset functions
############################################################################
rg = np.random.default_rng(FLAG_RANDOM_SEED.value)
def call_lm_preproc(
repeat,
split,
random_seed
):
"""Using functools.partial prevents the linter from doing its job."""
if FLAG_DATASET_NAME.value == constants.DatasetNameChoices.kilt_eli5:
return task_specific.create_lm_ds_kilt_eli5(
tokenizer=tokenizer,
context_window_size=(
model_or_replicas[0].config.n_positions
if isinstance(model_or_replicas, list)
else model_or_replicas.config.n_positions
),
dataset_name=FLAG_DATASET_NAME.value,
# Batches are split over the replicas:
batch_size=FLAG_BATCH_SIZE.value * actual_num_replicas,
db_path=FLAG_DB_PATH.value,
random_seed=random_seed,
use_subset=FLAG_USE_SUBSET.value,
subset_size=FLAG_SUBSET_SIZE.value,
use_helper_words=FLAG_USE_HELPER_WORDS.value,
approach_type=FLAG_APPROACH_TYPE.value,
num_retrievals=FLAG_NUM_RETRIEVALS.value,
retrieval_temperature=FLAG_RETRIEVAL_TEMPERATURE.value,
retriever=retriever,
repeat=repeat,
split=split,
enable_debug_checks=FLAG_DATASET_DEBUG.value,
retrieval_bank_size=FLAG_RETRIEVAL_BANK_SIZE.value,
dataset_type=FLAG_DATASET_TYPE.value,
qty_shuffle=FLAG_QTY_SHUFFLE.value,
tfr_prefix=FLAG_TFR_PREFIX.value,
max_length_generation=FLAG_MAX_LENGTH_GENERATION.value,
)
else:
raise NotImplementedError(
f"FLAG_DATASET_NAME.value unsupported: `{FLAG_DATASET_NAME.value}`"
)
make_training_dataset: Callable[Ellipsis, tf.data.Dataset] = functools.partial(
call_lm_preproc,
split="train",
repeat=False,
)
make_eval_dataset: Callable[Ellipsis, tf.data.Dataset] = functools.partial(
call_lm_preproc,
split="eval",
repeat=True,
)
############################################################################
# Prepare the step functions
############################################################################
utils.check_contained(
FLAG_DISTRIBUTE_MODE.value, constants.DistributeModeChoices.choices()
)
tf_function_flags = dict(
experimental_compile=FLAG_EXPERIMENTAL_COMPILE.value,
experimental_relax_shapes=not FLAG_INPUT_FIXED_SIZE.value
)
if (FLAG_DISTRIBUTE_MODE.value ==
constants.DistributeModeChoices.split_and_data_parallel):
if not isinstance(model_or_replicas, list):
raise RuntimeError(type(model_or_replicas))
training_step = build_manual_data_parallel_training_step(
model_or_replicas, optimizer, tf_function_flags
)
else:
training_step = build_regular_training_step(
model_or_replicas,
optimizer,
strategy=model_specific.strategy,
tf_function_kwargs=tf_function_flags
)
evaluation_step = build_evaluation_step(
model_or_replicas, tf_function_flags
)
secs_since_last_ckpt = time.time()
# Model checkpoints are saved to the tmp_directory and then rsynced to GCS
##########################################################################
# Prepare the different logging facilities
##########################################################################
train_log_dir = os.path.join(instance_output_dir, "tensorboard", "train")
eval_log_dir = os.path.join(instance_output_dir, "tensorboard", "eval")
flags_log_dir = os.path.join(instance_output_dir, "tensorboard", "params")
writers = dict(
train=tf.summary.create_file_writer(train_log_dir),
eval=tf.summary.create_file_writer(eval_log_dir),
flags=tf.summary.create_file_writer(flags_log_dir)
)
with writers["flags"].as_default():
tf.summary.text(
"Flags",
# Tensorboard takes Markdown:
json.dumps(flags_dict, indent=4).replace("\n", "\n\n"),
step=0
)
ma_loss = dict(
train=utils.MovingAverage(0.9),
eval=utils.MovingAverage(0.9)
)
step_counters = dict(train=0, eval=0)
batch_counters = dict(train=0, eval=0)
prev_batch_end = time.time()
# The eval ds has no real concept of epoch, repeats forever, shuffling
# each time it reaches its end
with utils.log_duration(LOGGER, "main", "All of make_eval_dataset"):
eval_ds_instance = make_eval_dataset(
random_seed=rg.integers(-2**63, 2**63 - 1),
)
LOGGER.debug("Distributing the eval dataset to the replicas.")
if FLAG_DATASET_TYPE.value == "tfr":
eval_ds_instance = (
model_specific.strategy.experimental_distribute_dataset(
eval_ds_instance
)
)
LOGGER.debug("Done distributing the eval dataset to the replcias.")
eval_ds_instance = iter(eval_ds_instance)
##########################################################################
# Training Loop
##########################################################################
for epoch in itertools.count():
####################################################################
# Epoch Setup
####################################################################
LOGGER.debug("EPOCH %d START", epoch)
# Shuffle differently every epoch
with utils.log_duration(
LOGGER, "main", "All of make_training_dataset"
):
train_ds_instance = make_training_dataset(
random_seed=rg.integers(-2**63, 2**63 - 1),
)
LOGGER.debug(
"Attempting to distribute the training dataset to the replicas."
)
if FLAG_DATASET_TYPE.value == "tfr":
train_ds_instance = (
model_specific.strategy.experimental_distribute_dataset(
train_ds_instance
)
)
LOGGER.debug(
"Done distributing the training dataset to the replicas."
)
train_ds_instance = iter(train_ds_instance)
# This allows us to see if we reached the end of the training iterator,
# in which case "did_at_least_one_training_batch == False".
# We could also test that it did all the batches, to similar results.
did_at_least_one_training_batch = True
split = "eval"
while did_at_least_one_training_batch:
# Invert split
if split == "train":
split = "eval"
else:
split = "train"
# Prepare to test if we did at least one training batch
if split == "train":
did_at_least_one_training_batch = False
if split == "train":
dataset_iterator = itertools.islice(
train_ds_instance, FLAG_BATCHES_BETWEEN_EVALS.value
)
else:
# The evaluation DS is tiny, so we reshuffle and take a random
dataset_iterator = itertools.islice(
eval_ds_instance, FLAG_NUMBER_EVAL_BATCHES.value
)
LOGGER.debug("Batching")
for batch in dataset_iterator:
# LOGGER.debug("Input sentence:\n\"%s\"",
# tokenizer.decode([x for x in batch["input_ids"][0]
# if x != tokenizer.eos_token_id]))
# LOGGER.debug("Label:\n\"%s\"",
# tokenizer.decode([(x if x != -100 else 0)
# for x in batch["label_ids"][0]]))
if FLAG_DATASET_TYPE.value != "tfr":
batch = (
model_specific.strategy
.experimental_distribute_values_from_function(
tf_utils.make_dict_distribute_fn(batch)
))
# We only care about training epochs as, obviously, we don't train
# over eval samples; the number of eval samples seen only
# contributes to lowering the variance in the evaluation of when to
# do early stopping.
if split == "train":
did_at_least_one_training_batch = True
input_ids = batch["input_ids"]
label_ids = batch["label_ids"]
####################################################################
# Training Step
####################################################################
step_counters[split] += (
FLAG_BATCH_SIZE.value * actual_num_replicas
)
if split == "train":
batch_counters[split] += 1
training_kwargs = dict(
input_ids=input_ids,
label_ids=label_ids,
)
if model_specific.strategy:
utils.print_mem("before running", LOGGER)
LOGGER.debug("Training, Calling strategy.run")
loss = model_specific.strategy.run(
training_step,
kwargs=training_kwargs
)
LOGGER.debug("Training, Done with strategy.run")
utils.print_mem("after running", LOGGER)
else:
loss = training_step(**training_kwargs) # pytype: disable=wrong-arg-count
# If we are in the strategy-free data parallel mode, we need
# to change the weights of all replicas to those of the model at
# index 0
if (
FLAG_DISTRIBUTE_MODE.value ==
constants.DistributeModeChoices.split_and_data_parallel
):
for replica in model_or_replicas[1:]:
replica.set_weights(model_or_replicas[0].get_weights())
####################################################################
# Evaluation Step
####################################################################
elif split == "eval":
evaluation_kwargs = dict(
input_ids=input_ids,
label_ids=label_ids,
)
if model_specific.strategy:
loss = model_specific.strategy.run(
evaluation_step,
kwargs=evaluation_kwargs
)
else:
loss = evaluation_step(**evaluation_kwargs)
else:
raise ValueError(f"Unexpected value for split: {split}")
####################################################################
# Logging
####################################################################
if (FLAG_DISTRIBUTE_MODE.value in
constants.PURE_DATA_PARALLEL_STRATEGIES):
utils.check_equal(len(loss.values), actual_num_replicas)
LOGGER.debug("Split: %s", split)
LOGGER.debug("Real num replicas: %s", actual_num_replicas)
LOGGER.debug("Loss: %s", loss)
LOGGER.debug("Loss values: %s", loss.values)
average_loss = float(tf.math.reduce_mean(loss.values).numpy())
else:
average_loss = float(loss.numpy())
# tf.debugging.check_numerics(loss)
now = time.time()
batch_duration = now - prev_batch_end
prev_batch_end = now
ma_loss[split].update(average_loss)
# Actual logging
LOGGER.info("Epoch: # %d", epoch)
LOGGER.info("Tensorboard_dir: %s", instance_output_dir)
LOGGER.info("Batch: %s # %d", split, batch_counters[split])
LOGGER.info("Step: %s # %d", split, step_counters[split])
if FLAG_USE_SUBSET.value:
LOGGER.warning(">> USING A SUBSET OF THE DATASET <<")
LOGGER.info(
"%(split)s Batch loss: %(metric)f",
dict(split=split, metric=average_loss)
)
LOGGER.info(
"%(split)s Moving average loss: %(metric)f",
dict(split=split, metric=ma_loss[split].average)
)
LOGGER.info(
"%(split)s Moving average ppl: %(metric)f",
dict(split=split, metric=np.exp(ma_loss[split].average))
)
LOGGER.info(
"%(split)s Batch duration: %(duration)s",
dict(
split=split,
duration=utils.TimeStamp.from_seconds(
batch_duration).format()
)
)
if FLAG_DISTRIBUTE_MODE.value in constants.DATA_PARALLEL_DMC:
LOGGER.info(
"%(split)s Duration per sample: %(duration)s",
dict(
split=split,
duration=utils.TimeStamp.from_seconds(
batch_duration / (
FLAG_BATCH_SIZE.value * actual_num_replicas
)
)
)
)
# Write to Tensorboard
with writers[split].as_default():
tf.summary.scalar(
f"Loss/{split}", average_loss, step_counters[split]
)
tf.summary.scalar(
f"PPL/{split}", np.exp(average_loss), step_counters[split]
)
writers[split].flush()
# Save every 5 min
if (time.time() - secs_since_last_ckpt) / (60 * 20) >= 1:
secs_since_last_ckpt = time.time()
save_model(
train_steps=step_counters["train"],
model_or_replicas=model_or_replicas,
instance_output_dir=instance_output_dir
)
secs_since_last_ckpt = time.time()
save_model(
train_steps=step_counters["train"],
model_or_replicas=model_or_replicas,
instance_output_dir=instance_output_dir
)
#############################################################
# Post Training Cleanup
#######################################################################
for writer in writers.values():
writer.close()