def test_past_stop_threshold_none_false(self):
"""Tests that check None returns false."""
self.assertFalse(model_helpers.past_stop_threshold(None, -1.5))
self.assertFalse(model_helpers.past_stop_threshold(None, None))
self.assertFalse(model_helpers.past_stop_threshold(None, 1.5))
# Zero should be okay, though.
self.assertTrue(model_helpers.past_stop_threshold(0, 1.5))
def test_past_stop_threshold_none_false(self): """Tests that check None returns false.""" self.assertFalse(model_helpers.past_stop_threshold(None, -1.5)) self.assertFalse(model_helpers.past_stop_threshold(None, None)) self.assertFalse(model_helpers.past_stop_threshold(None, 1.5)) # Zero should be okay, though. self.assertTrue(model_helpers.past_stop_threshold(0, 1.5))
def run_wide_deep(flags_obj):
"""Run Wide-Deep training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
# Clean up the model directory if present
shutil.rmtree(flags_obj.model_dir, ignore_errors=True)
model = build_estimator(flags_obj.model_dir, flags_obj.model_type)
train_file = os.path.join(flags_obj.data_dir, 'adult.data')
test_file = os.path.join(flags_obj.data_dir, 'adult.test')
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
def train_input_fn():
return input_fn(
train_file, flags_obj.epochs_between_evals, True, flags_obj.batch_size)
def eval_input_fn():
return input_fn(test_file, 1, False, flags_obj.batch_size)
run_params = {
'batch_size': flags_obj.batch_size,
'train_epochs': flags_obj.train_epochs,
'model_type': flags_obj.model_type,
}
benchmark_logger = logger.config_benchmark_logger(flags_obj)
benchmark_logger.log_run_info('wide_deep', 'Census Income', run_params)
loss_prefix = LOSS_PREFIX.get(flags_obj.model_type, '')
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks, batch_size=flags_obj.batch_size,
tensors_to_log={'average_loss': loss_prefix + 'head/truediv',
'loss': loss_prefix + 'head/weighted_loss/Sum'})
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
tf.logging.info('Results at epoch %d / %d',
(n + 1) * flags_obj.epochs_between_evals,
flags_obj.train_epochs)
tf.logging.info('-' * 60)
for key in sorted(results):
tf.logging.info('%s: %s' % (key, results[key]))
benchmark_logger.log_evaluation_result(results)
if model_helpers.past_stop_threshold(
flags_obj.stop_threshold, results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
export_model(model, flags_obj.model_type, flags_obj.export_dir)
def run_loop(name, train_input_fn, eval_input_fn, model_column_fn,
build_estimator_fn, flags_obj, tensors_to_log, early_stop=False):
"""Define training loop."""
model_helpers.apply_clean(flags.FLAGS)
print('+' * 50)
print('mode type: ' + flags_obj.model_type)
print('batch size: ' + str(flags_obj.batch_size))
model = build_estimator_fn(
model_dir=flags_obj.model_dir, model_type=flags_obj.model_type,
model_column_fn=model_column_fn,
inter_op=flags_obj.inter_op_parallelism_threads,
intra_op=flags_obj.intra_op_parallelism_threads)
run_params = {
'batch_size': flags_obj.batch_size,
'train_epochs': flags_obj.train_epochs,
'model_type': flags_obj.model_type,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('wide_deep', name, run_params,
test_id=flags_obj.benchmark_test_id)
loss_prefix = LOSS_PREFIX.get(flags_obj.model_type, '')
tensors_to_log = {k: v.format(loss_prefix=loss_prefix)
for k, v in tensors_to_log.items()}
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks, model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size, tensors_to_log=tensors_to_log)
train_hooks = []
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
tf.logging.info('Results at epoch %d / %d',
(n + 1) * flags_obj.epochs_between_evals,
flags_obj.train_epochs)
tf.logging.info('-' * 60)
for key in sorted(results):
tf.logging.info('%s: %s' % (key, results[key]))
benchmark_logger.log_evaluation_result(results)
if early_stop and model_helpers.past_stop_threshold(
flags_obj.stop_threshold, results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
export_model(model, flags_obj.model_type, flags_obj.export_dir,
model_column_fn)
def test_past_stop_threshold(self):
"""Tests for normal operating conditions."""
self.assertTrue(model_helpers.past_stop_threshold(0.54, 1))
self.assertTrue(model_helpers.past_stop_threshold(54, 100))
self.assertFalse(model_helpers.past_stop_threshold(0.54, 0.1))
self.assertFalse(model_helpers.past_stop_threshold(-0.54, -1.5))
self.assertTrue(model_helpers.past_stop_threshold(-0.54, 0))
self.assertTrue(model_helpers.past_stop_threshold(0, 0))
self.assertTrue(model_helpers.past_stop_threshold(0.54, 0.54))
def test_past_stop_threshold(self): """Tests for normal operating conditions.""" self.assertTrue(model_helpers.past_stop_threshold(0.54, 1)) self.assertTrue(model_helpers.past_stop_threshold(54, 100)) self.assertFalse(model_helpers.past_stop_threshold(0.54, 0.1)) self.assertFalse(model_helpers.past_stop_threshold(-0.54, -1.5)) self.assertTrue(model_helpers.past_stop_threshold(-0.54, 0)) self.assertTrue(model_helpers.past_stop_threshold(0, 0)) self.assertTrue(model_helpers.past_stop_threshold(0.54, 0.54))
def run_loop(name, train_input_fn, eval_input_fn, model_column_fn,
build_estimator_fn, flags_obj, tensors_to_log, early_stop=False):
"""Define training loop."""
model_helpers.apply_clean(flags.FLAGS)
model = build_estimator_fn(
model_dir=flags_obj.model_dir, model_type=flags_obj.model_type,
model_column_fn=model_column_fn,
inter_op=flags_obj.inter_op_parallelism_threads,
intra_op=flags_obj.intra_op_parallelism_threads)
run_params = {
'batch_size': flags_obj.batch_size,
'train_epochs': flags_obj.train_epochs,
'model_type': flags_obj.model_type,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('wide_deep', name, run_params,
test_id=flags_obj.benchmark_test_id)
loss_prefix = LOSS_PREFIX.get(flags_obj.model_type, '')
tensors_to_log = {k: v.format(loss_prefix=loss_prefix)
for k, v in tensors_to_log.items()}
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks, model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size, tensors_to_log=tensors_to_log)
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
tf.logging.info('Results at epoch %d / %d',
(n + 1) * flags_obj.epochs_between_evals,
flags_obj.train_epochs)
tf.logging.info('-' * 60)
for key in sorted(results):
tf.logging.info('%s: %s' % (key, results[key]))
benchmark_logger.log_evaluation_result(results)
if early_stop and model_helpers.past_stop_threshold(
flags_obj.stop_threshold, results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
export_model(model, flags_obj.model_type, flags_obj.export_dir,
model_column_fn)
def train(self):
self.createDatasets()
if self.model is None:
self._getModelEstimator()
estimator = self.model.getEstimator()
run_params = {
'batch_size': self.flags.batch_size,
'train_epochs': self.flags.train_epochs,
'model_type': 'deep',
}
benchmark_logger = logger.config_benchmark_logger(self.flags)
benchmark_logger.log_run_info('deep', 'Readmission Patient',
run_params)
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(self.flags.train_epochs //
self.flags.epochs_between_evals):
# Break from loop if privacy budget is exceedeed and differential privacy is enabled
if self.flags.enable_dp and self.model.is_privacy_budget_exceeded(
):
break
print('n: ' + str(n))
estimator.train(input_fn=self._input_fn_train)
results = estimator.evaluate(input_fn=self._input_fn_eval)
# Display evaluation metrics
tf.logging.info('Results at epoch %d / %d',
(n + 1) * self.flags.epochs_between_evals,
self.flags.train_epochs)
tf.logging.info('-' * 60)
for key in sorted(results):
tf.logging.info('%s: %s' % (key, results[key]))
benchmark_logger.log_evaluation_result(results)
if model_helpers.past_stop_threshold(self.flags.stop_threshold,
results['accuracy']):
break
# Export the model
print('export the model?')
if n % 10 == 0 and self.flags.export_dir is not None:
self.export_model()
def main(argv):
parser = WideDeepArgParser()
flags = parser.parse_args(args=argv[1:])
# Clean up the model directory if present
shutil.rmtree(flags.model_dir, ignore_errors=True)
model = build_estimator(flags.model_dir, flags.model_type)
train_file = os.path.join(flags.data_dir, 'adult.data')
test_file = os.path.join(flags.data_dir, 'adult.test')
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
def train_input_fn():
return input_fn(train_file, flags.epochs_between_evals, True,
flags.batch_size)
def eval_input_fn():
return input_fn(test_file, 1, False, flags.batch_size)
loss_prefix = LOSS_PREFIX.get(flags.model_type, '')
train_hooks = hooks_helper.get_train_hooks(
flags.hooks,
batch_size=flags.batch_size,
tensors_to_log={
'average_loss': loss_prefix + 'head/truediv',
'loss': loss_prefix + 'head/weighted_loss/Sum'
})
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags.train_epochs // flags.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
print('Results at epoch', (n + 1) * flags.epochs_between_evals)
print('-' * 60)
for key in sorted(results):
print('%s: %s' % (key, results[key]))
if model_helpers.past_stop_threshold(flags.stop_threshold,
results['accuracy']):
break
# Export the model
if flags.export_dir is not None:
export_model(model, flags.model_type, flags.export_dir)
def main(argv):
parser = WideDeepArgParser()
flags = parser.parse_args(args=argv[1:])
# Clean up the model directory if present
shutil.rmtree(flags.model_dir, ignore_errors=True)
model = build_estimator(flags.model_dir, flags.model_type)
train_file = os.path.join(flags.data_dir, 'adult.data')
test_file = os.path.join(flags.data_dir, 'adult.test')
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
def train_input_fn():
return input_fn(
train_file, flags.epochs_between_evals, True, flags.batch_size)
def eval_input_fn():
return input_fn(test_file, 1, False, flags.batch_size)
loss_prefix = LOSS_PREFIX.get(flags.model_type, '')
train_hooks = hooks_helper.get_train_hooks(
flags.hooks, batch_size=flags.batch_size,
tensors_to_log={'average_loss': loss_prefix + 'head/truediv',
'loss': loss_prefix + 'head/weighted_loss/Sum'})
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags.train_epochs // flags.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=train_hooks)
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
print('Results at epoch', (n + 1) * flags.epochs_between_evals)
print('-' * 60)
for key in sorted(results):
print('%s: %s' % (key, results[key]))
if model_helpers.past_stop_threshold(
flags.stop_threshold, results['accuracy']):
break
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
if FLAGS.seed is not None:
np.random.seed(FLAGS.seed)
num_gpus = flags_core.get_num_gpus(FLAGS)
batch_size = distribution_utils.per_device_batch_size(
int(FLAGS.batch_size), num_gpus)
eval_per_user = rconst.NUM_EVAL_NEGATIVES + 1
eval_batch_size = int(FLAGS.eval_batch_size
or max([FLAGS.batch_size, eval_per_user]))
if eval_batch_size % eval_per_user:
eval_batch_size = eval_batch_size // eval_per_user * eval_per_user
tf.logging.warning(
"eval examples per user does not evenly divide eval_batch_size. "
"Overriding to {}".format(eval_batch_size))
ncf_dataset, cleanup_fn = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset,
data_dir=FLAGS.data_dir,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_neg=FLAGS.num_neg,
epochs_per_cycle=FLAGS.epochs_between_evals,
match_mlperf=FLAGS.ml_perf,
deterministic=FLAGS.seed is not None)
model_helpers.apply_clean(flags.FLAGS)
train_estimator, eval_estimator = construct_estimator(
num_gpus=num_gpus,
model_dir=FLAGS.model_dir,
params={
"use_seed": FLAGS.seed is not None,
"hash_pipeline": FLAGS.hash_pipeline,
"batch_size": batch_size,
"learning_rate": FLAGS.learning_rate,
"num_users": ncf_dataset.num_users,
"num_items": ncf_dataset.num_items,
"mf_dim": FLAGS.num_factors,
"model_layers": [int(layer) for layer in FLAGS.layers],
"mf_regularization": FLAGS.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in FLAGS.mlp_regularization],
"num_neg": FLAGS.num_neg,
"use_tpu": FLAGS.tpu is not None,
"tpu": FLAGS.tpu,
"tpu_zone": FLAGS.tpu_zone,
"tpu_gcp_project": FLAGS.tpu_gcp_project,
"beta1": FLAGS.beta1,
"beta2": FLAGS.beta2,
"epsilon": FLAGS.epsilon,
"match_mlperf": FLAGS.ml_perf,
},
batch_size=flags.FLAGS.batch_size,
eval_batch_size=eval_batch_size)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size, # for ExamplesPerSecondHook
tensors_to_log={"cross_entropy": "cross_entropy"})
run_params = {
"batch_size": FLAGS.batch_size,
"eval_batch_size": eval_batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
approx_train_steps = int(ncf_dataset.num_train_positives *
(1 + FLAGS.num_neg) // FLAGS.batch_size)
pred_input_fn = data_preprocessing.make_pred_input_fn(
ncf_dataset=ncf_dataset)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
train_input_fn, train_record_dir, batch_count = \
data_preprocessing.make_train_input_fn(ncf_dataset=ncf_dataset)
if np.abs(approx_train_steps - batch_count) > 1:
tf.logging.warning(
"Estimated ({}) and reported ({}) number of batches differ by more "
"than one".format(approx_train_steps, batch_count))
train_estimator.train(input_fn=train_input_fn,
hooks=train_hooks,
steps=batch_count)
tf.gfile.DeleteRecursively(train_record_dir)
tf.logging.info("Beginning evaluation.")
eval_results = eval_estimator.evaluate(pred_input_fn)
tf.logging.info("Evaluation complete.")
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[rconst.HR_KEY]
ndcg = eval_results[rconst.NDCG_KEY]
tf.logging.info("Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
cleanup_fn() # Cleanup data construction artifacts and subprocess.
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def run_loop(estimator,
schedule_manager,
train_hooks=None,
benchmark_logger=None,
bleu_source=None,
bleu_ref=None,
bleu_threshold=None,
vocab_file=None):
"""Train and evaluate model, and optionally compute model's BLEU score.
**Step vs. Epoch vs. Iteration**
Steps and epochs are canonical terms used in TensorFlow and general machine
learning. They are used to describe running a single process (train/eval):
- Step refers to running the process through a single or batch of examples.
- Epoch refers to running the process through an entire dataset.
E.g. training a dataset with 100 examples. The dataset is
divided into 20 batches with 5 examples per batch. A single training step
trains the model on one batch. After 20 training steps, the model will have
trained on every batch in the dataset, or, in other words, one epoch.
Meanwhile, iteration is used in this implementation to describe running
multiple processes (training and eval).
- A single iteration:
1. trains the model for a specific number of steps or epochs.
2. evaluates the model.
3. (if source and ref files are provided) compute BLEU score.
This function runs through multiple train+eval+bleu iterations.
Args:
estimator: tf.Estimator containing model to train.
schedule_manager: A schedule.Manager object to guide the run loop.
train_hooks: List of hooks to pass to the estimator during training.
benchmark_logger: a BenchmarkLogger object that logs evaluation data
bleu_source: File containing text to be translated for BLEU calculation.
bleu_ref: File containing reference translations for BLEU calculation.
bleu_threshold: minimum BLEU score before training is stopped.
vocab_file: Path to vocab file that will be used to subtokenize bleu_source.
Returns:
Dict of results of the run. Contains the keys `eval_results`,
`train_hooks`, `bleu_cased`, and `bleu_uncased`. `train_hooks` is a list the
instances of hooks used during training.
Raises:
ValueError: if both or none of single_iteration_train_steps and
single_iteration_train_epochs were defined.
NotFoundError: if the vocab file or bleu files don't exist.
"""
if bleu_source:
_validate_file(bleu_source)
if bleu_ref:
_validate_file(bleu_ref)
if vocab_file:
_validate_file(vocab_file)
evaluate_bleu = bleu_source is not None and bleu_ref is not None
if evaluate_bleu and schedule_manager.use_tpu:
raise ValueError(
"BLEU score can not be computed when training with a TPU, "
"as it requires estimator.predict which is not yet "
"supported.")
# Print details of training schedule.
tf.logging.info("Training schedule:")
tf.logging.info("\t1. Train for {}".format(
schedule_manager.train_increment_str))
tf.logging.info("\t2. Evaluate model.")
if evaluate_bleu:
tf.logging.info("\t3. Compute BLEU score.")
if bleu_threshold is not None:
tf.logging.info(
"Repeat above steps until the BLEU score reaches %f" %
bleu_threshold)
if not evaluate_bleu or bleu_threshold is None:
tf.logging.info("Repeat above steps %d times." %
schedule_manager.train_eval_iterations)
if evaluate_bleu:
# Create summary writer to log bleu score (values can be displayed in
# Tensorboard).
bleu_writer = tf.summary.FileWriter(
os.path.join(estimator.model_dir, BLEU_DIR))
if bleu_threshold is not None:
# Change loop stopping condition if bleu_threshold is defined.
schedule_manager.train_eval_iterations = INF
# Loop training/evaluation/bleu cycles
stats = {}
for i in xrange(schedule_manager.train_eval_iterations):
tf.logging.info("Starting iteration %d" % (i + 1))
# Train the model for single_iteration_train_steps or until the input fn
# runs out of examples (if single_iteration_train_steps is None).
estimator.train(dataset.train_input_fn,
steps=schedule_manager.single_iteration_train_steps,
hooks=train_hooks)
eval_results = estimator.evaluate(
input_fn=dataset.eval_input_fn,
steps=schedule_manager.single_iteration_eval_steps)
tf.logging.info("Evaluation results (iter %d/%d):" %
(i + 1, schedule_manager.train_eval_iterations))
tf.logging.info(eval_results)
benchmark_logger.log_evaluation_result(eval_results)
# The results from estimator.evaluate() are measured on an approximate
# translation, which utilize the target golden values provided. The actual
# bleu score must be computed using the estimator.predict() path, which
# outputs translations that are not based on golden values. The translations
# are compared to reference file to get the actual bleu score.
if evaluate_bleu:
uncased_score, cased_score = evaluate_and_log_bleu(
estimator, bleu_source, bleu_ref, vocab_file)
stats["bleu_uncased"] = uncased_score
stats["bleu_cased"] = cased_score
# Write actual bleu scores using summary writer and benchmark logger
global_step = get_global_step(estimator)
summary = tf.Summary(value=[
tf.Summary.Value(tag="bleu/uncased",
simple_value=uncased_score),
tf.Summary.Value(tag="bleu/cased", simple_value=cased_score),
])
bleu_writer.add_summary(summary, global_step)
bleu_writer.flush()
benchmark_logger.log_metric("bleu_uncased",
uncased_score,
global_step=global_step)
benchmark_logger.log_metric("bleu_cased",
cased_score,
global_step=global_step)
# Stop training if bleu stopping threshold is met.
if model_helpers.past_stop_threshold(bleu_threshold,
uncased_score):
bleu_writer.close()
break
stats["eval_results"] = eval_results
stats["train_hooks"] = train_hooks
return stats
def run_loop(estimator,
schedule_manager,
train_hooks=None,
benchmark_logger=None,
bleu_source=None,
bleu_ref=None,
bleu_threshold=None,
vocab_file=None):
if bleu_source:
_validate_file(bleu_source)
if bleu_ref:
_validate_file(bleu_ref)
if vocab_file:
_validate_file(vocab_file)
evaluate_bleu = bleu_source is not None and bleu_ref is not None
if evaluate_bleu and schedule_manager.use_tpu:
raise ValueError(
"BLEU score can not be computed when training with a TPU, "
"as it requires estimator.predict which is not yet "
"supported.")
# Print details of training schedule.
tf.logging.info("Training schedule:")
tf.logging.info("\t1. Train for {}".format(
schedule_manager.train_increment_str))
tf.logging.info("\t2. Evaluate model.")
if evaluate_bleu:
tf.logging.info("\t3. Compute BLEU score.")
if bleu_threshold is not None:
tf.logging.info(
"Repeat above steps until the BLEU score reaches %f" %
bleu_threshold)
if not evaluate_bleu or bleu_threshold is None:
tf.logging.info("Repeat above steps %d times." %
schedule_manager.train_eval_iterations)
if evaluate_bleu:
# Create summary writer to log bleu score (values can be displayed in
# Tensorboard).
bleu_writer = tf.summary.FileWriter(
os.path.join(estimator.model_dir, BLEU_DIR))
if bleu_threshold is not None:
# Change loop stopping condition if bleu_threshold is defined.
schedule_manager.train_eval_iterations = INF
# Loop training/evaluation/bleu cycles
stats = {}
for i in xrange(schedule_manager.train_eval_iterations):
tf.logging.info("Starting iteration %d" % (i + 1))
# Train the model for single_iteration_train_steps or until the input fn
# runs out of examples (if single_iteration_train_steps is None).
estimator.train(dataset.train_input_fn,
steps=schedule_manager.single_iteration_train_steps,
hooks=train_hooks)
eval_results = estimator.evaluate(
input_fn=dataset.eval_input_fn,
steps=schedule_manager.single_iteration_eval_steps)
tf.logging.info("Evaluation results (iter %d/%d):" %
(i + 1, schedule_manager.train_eval_iterations))
tf.logging.info(eval_results)
benchmark_logger.log_evaluation_result(eval_results)
# The results from estimator.evaluate() are measured on an approximate
# translation, which utilize the target golden values provided. The actual
# bleu score must be computed using the estimator.predict() path, which
# outputs translations that are not based on golden values. The translations
# are compared to reference file to get the actual bleu score.
if evaluate_bleu:
uncased_score, cased_score = evaluate_and_log_bleu(
estimator, bleu_source, bleu_ref, vocab_file)
stats["bleu_uncased"] = uncased_score
stats["bleu_cased"] = cased_score
# Write actual bleu scores using summary writer and benchmark logger
global_step = get_global_step(estimator)
summary = tf.Summary(value=[
tf.Summary.Value(tag="bleu/uncased",
simple_value=uncased_score),
tf.Summary.Value(tag="bleu/cased", simple_value=cased_score),
])
bleu_writer.add_summary(summary, global_step)
bleu_writer.flush()
benchmark_logger.log_metric("bleu_uncased",
uncased_score,
global_step=global_step)
benchmark_logger.log_metric("bleu_cased",
cased_score,
global_step=global_step)
# Stop training if bleu stopping threshold is met.
if model_helpers.past_stop_threshold(bleu_threshold,
uncased_score):
bleu_writer.close()
break
stats["eval_results"] = eval_results
stats["train_hooks"] = train_hooks
return stats
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing and not FLAGS.use_synthetic_data:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
if FLAGS.seed is not None:
np.random.seed(FLAGS.seed)
params = parse_flags(FLAGS)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
if FLAGS.use_synthetic_data:
producer = data_pipeline.DummyConstructor()
num_users, num_items = data_preprocessing.DATASET_TO_NUM_USERS_AND_ITEMS[
FLAGS.dataset]
num_train_steps = rconst.SYNTHETIC_BATCHES_PER_EPOCH
num_eval_steps = rconst.SYNTHETIC_BATCHES_PER_EPOCH
else:
num_users, num_items, producer = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset, data_dir=FLAGS.data_dir, params=params,
constructor_type=FLAGS.constructor_type,
deterministic=FLAGS.seed is not None)
num_train_steps = (producer.train_batches_per_epoch //
params["batches_per_step"])
num_eval_steps = (producer.eval_batches_per_epoch //
params["batches_per_step"])
assert not producer.train_batches_per_epoch % params["batches_per_step"]
assert not producer.eval_batches_per_epoch % params["batches_per_step"]
producer.start()
params["num_users"], params["num_items"] = num_users, num_items
model_helpers.apply_clean(flags.FLAGS)
estimator = construct_estimator(model_dir=FLAGS.model_dir, params=params)
benchmark_logger, train_hooks = log_and_get_hooks(params["eval_batch_size"])
target_reached = False
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_LOOP)
for cycle_index in range(total_training_cycle):
assert FLAGS.epochs_between_evals == 1 or not mlperf_helper.LOGGER.enabled
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_EPOCH,
value=cycle_index)
train_input_fn = producer.make_input_fn(is_training=True)
estimator.train(input_fn=train_input_fn, hooks=train_hooks,
steps=num_train_steps)
tf.logging.info("Beginning evaluation.")
eval_input_fn = producer.make_input_fn(is_training=False)
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_START,
value=cycle_index)
eval_results = estimator.evaluate(eval_input_fn, steps=num_eval_steps)
tf.logging.info("Evaluation complete.")
hr = float(eval_results[rconst.HR_KEY])
ndcg = float(eval_results[rconst.NDCG_KEY])
loss = float(eval_results["loss"])
mlperf_helper.ncf_print(
key=mlperf_helper.TAGS.EVAL_TARGET,
value={"epoch": cycle_index, "value": FLAGS.hr_threshold})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_ACCURACY,
value={"epoch": cycle_index, "value": hr})
mlperf_helper.ncf_print(
key=mlperf_helper.TAGS.EVAL_HP_NUM_NEG,
value={"epoch": cycle_index, "value": rconst.NUM_EVAL_NEGATIVES})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_STOP, value=cycle_index)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}, Loss = {:.4f}".format(
cycle_index + 1, hr, ndcg, loss))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
target_reached = True
break
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_STOP,
value={"success": target_reached})
producer.stop_loop()
producer.join()
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_FINAL)
def run_ncf(_):
"""Run NCF training and eval loop."""
params = ncf_common.parse_flags(FLAGS)
num_users, num_items, num_train_steps, num_eval_steps, producer = (
ncf_common.get_inputs(params))
params["num_users"], params["num_items"] = num_users, num_items
producer.start()
model_helpers.apply_clean(flags.FLAGS)
estimator = construct_estimator(model_dir=FLAGS.model_dir, params=params)
benchmark_logger, train_hooks = log_and_get_hooks(params["eval_batch_size"])
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
target_reached = False
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_LOOP)
for cycle_index in range(total_training_cycle):
assert FLAGS.epochs_between_evals == 1 or not mlperf_helper.LOGGER.enabled
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_EPOCH,
value=cycle_index)
train_input_fn = producer.make_input_fn(is_training=True)
estimator.train(input_fn=train_input_fn, hooks=train_hooks,
steps=num_train_steps)
tf.logging.info("Beginning evaluation.")
eval_input_fn = producer.make_input_fn(is_training=False)
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_START,
value=cycle_index)
eval_results = estimator.evaluate(eval_input_fn, steps=num_eval_steps)
tf.logging.info("Evaluation complete.")
hr = float(eval_results[rconst.HR_KEY])
ndcg = float(eval_results[rconst.NDCG_KEY])
loss = float(eval_results["loss"])
mlperf_helper.ncf_print(
key=mlperf_helper.TAGS.EVAL_TARGET,
value={"epoch": cycle_index, "value": FLAGS.hr_threshold})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_ACCURACY,
value={"epoch": cycle_index, "value": hr})
mlperf_helper.ncf_print(
key=mlperf_helper.TAGS.EVAL_HP_NUM_NEG,
value={"epoch": cycle_index, "value": rconst.NUM_EVAL_NEGATIVES})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_STOP, value=cycle_index)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}, Loss = {:.4f}".format(
cycle_index + 1, hr, ndcg, loss))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
target_reached = True
break
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_STOP,
value={"success": target_reached})
producer.stop_loop()
producer.join()
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_FINAL)
def run_ncf(_):
"""Run NCF training and eval loop."""
# Data preprocessing
# The file name of training and test dataset
train_fname = os.path.join(
FLAGS.data_dir, FLAGS.dataset + "-" + constants.TRAIN_RATINGS_FILENAME)
test_fname = os.path.join(
FLAGS.data_dir, FLAGS.dataset + "-" + constants.TEST_RATINGS_FILENAME)
neg_fname = os.path.join(
FLAGS.data_dir, FLAGS.dataset + "-" + constants.TEST_NEG_FILENAME)
assert os.path.exists(train_fname), (
"Run data_download.py first to download and extract {} dataset".format(
FLAGS.dataset))
tf.logging.info("Data preprocessing...")
ncf_dataset = dataset.data_preprocessing(
train_fname, test_fname, neg_fname, FLAGS.num_neg)
# Create NeuMF model and convert it to Estimator
tf.logging.info("Creating Estimator from Keras model...")
layers = [int(layer) for layer in FLAGS.layers]
mlp_regularization = [float(reg) for reg in FLAGS.mlp_regularization]
keras_model = neumf_model.NeuMF(
ncf_dataset.num_users, ncf_dataset.num_items, FLAGS.num_factors,
layers, FLAGS.batch_size, FLAGS.mf_regularization,
mlp_regularization)
num_gpus = flags_core.get_num_gpus(FLAGS)
estimator = convert_keras_to_estimator(keras_model, num_gpus, FLAGS.model_dir)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
batch_size=FLAGS.batch_size # for ExamplesPerSecondHook
)
run_params = {
"batch_size": FLAGS.batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(
model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
# Training and evaluation cycle
def train_input_fn():
return dataset.input_fn(
True,
distribution_utils.per_device_batch_size(FLAGS.batch_size, num_gpus),
ncf_dataset, FLAGS.epochs_between_evals)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
estimator.train(input_fn=train_input_fn, hooks=train_hooks)
# Evaluate the model
eval_results = evaluate_model(
estimator, FLAGS.batch_size, num_gpus, ncf_dataset)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
eval_speech_dataset.speech_labels,
eval_speech_dataset.entries,
input_fn_eval,
)
# Log the WER and CER results.
benchmark_logger.log_evaluation_result(eval_results)
tf.logging.info(
"Iteration {}: WER = {:.2f}, CER = {:.2f}".format(
cycle_index + 1, eval_results[_WER_KEY], eval_results[_CER_KEY]
)
)
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(
flags_obj.wer_threshold, eval_results[_WER_KEY]
):
break
def define_deep_speech_flags():
"""Add flags for run_deep_speech."""
# Add common flags
flags_core.define_base(
data_dir=False # we use train_data_dir and eval_data_dir instead
)
flags_core.define_performance(
num_parallel_calls=False,
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
"""
model_helpers.apply_clean(flags.FLAGS)
# Ensures flag override logic is only executed if explicitly triggered.
if flags_obj.tf_gpu_thread_mode:
override_flags_and_set_envars_for_gpu_thread_pool(flags_obj)
# Creates session config. allow_soft_placement = True, is required for
# multi-GPU and is not harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
# Creates a `RunConfig` that checkpoints every 24 hours which essentially
# results in checkpoints determined only by `epochs_between_evals`.
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy,
session_config=session_config,
save_checkpoints_secs=60 * 60 * 24)
# Initializes model with all but the dense layer from pretrained ResNet.
if flags_obj.pretrained_model_checkpoint_path is not None:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start='^(?!.*dense)')
else:
warm_start_settings = None
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
warm_start_from=warm_start_settings,
params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj),
'fine_tune': flags_obj.fine_tune
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs):
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=num_epochs,
dtype=flags_core.get_tf_dtype(flags_obj),
datasets_num_private_threads=flags_obj.
datasets_num_private_threads,
num_parallel_batches=flags_obj.datasets_num_parallel_batches)
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1,
dtype=flags_core.get_tf_dtype(flags_obj))
if flags_obj.eval_only or not flags_obj.train_epochs:
# If --eval_only is set, perform a single loop with zero train epochs.
schedule, n_loops = [0], 1
else:
# Compute the number of times to loop while training. All but the last
# pass will train for `epochs_between_evals` epochs, while the last will
# train for the number needed to reach `training_epochs`. For instance if
# train_epochs = 25 and epochs_between_evals = 10
# schedule will be set to [10, 10, 5]. That is to say, the loop will:
# Train for 10 epochs and then evaluate.
# Train for another 10 epochs and then evaluate.
# Train for a final 5 epochs (to reach 25 epochs) and then evaluate.
n_loops = math.ceil(flags_obj.train_epochs /
flags_obj.epochs_between_evals)
schedule = [
flags_obj.epochs_between_evals for _ in range(int(n_loops))
]
schedule[-1] = flags_obj.train_epochs - sum(
schedule[:-1]) # over counting.
with tf.Session() as sess:
run_metadata = tf.RunMetadata()
for cycle_index, num_train_epochs in enumerate(schedule):
tf.logging.info('Starting cycle: %d/%d', cycle_index, int(n_loops))
if num_train_epochs:
sess.run(classifier.train(
input_fn=lambda: input_fn_train(num_train_epochs),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps),
run_metadata=run_metadata)
tf.logging.info('Starting to evaluate.')
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data, which
# will iterate forever. Passing steps=flags_obj.max_train_steps allows the
# eval (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = sess.run(classifier.evaluate(
input_fn=input_fn_eval, steps=flags_obj.max_train_steps),
run_metadata=run_metadata)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
fetched_timeline = timeline.Timeline(run_metadata.step_stats)
chrome_trace = fetched_timeline.generate_chrome_trace_format()
with open('timeline_run_loop.json', 'w') as f:
f.write(chrome_trace)
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
export_dtype = flags_core.get_tf_dtype(flags_obj)
if flags_obj.image_bytes_as_serving_input:
input_receiver_fn = functools.partial(image_bytes_serving_input_fn,
shape,
dtype=export_dtype)
else:
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size, dtype=export_dtype)
classifier.export_savedmodel(flags_obj.export_dir,
input_receiver_fn,
strip_default_attrs=True)
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
num_gpus = flags_core.get_num_gpus(FLAGS)
batch_size = distribution_utils.per_device_batch_size(
int(FLAGS.batch_size), num_gpus)
eval_batch_size = int(FLAGS.eval_batch_size or FLAGS.batch_size)
ncf_dataset = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset,
data_dir=FLAGS.data_dir,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_neg=FLAGS.num_neg,
epochs_per_cycle=FLAGS.epochs_between_evals,
match_mlperf=FLAGS.ml_perf)
model_helpers.apply_clean(flags.FLAGS)
train_estimator, eval_estimator = construct_estimator(
num_gpus=num_gpus,
model_dir=FLAGS.model_dir,
params={
"batch_size": batch_size,
"learning_rate": FLAGS.learning_rate,
"num_users": ncf_dataset.num_users,
"num_items": ncf_dataset.num_items,
"mf_dim": FLAGS.num_factors,
"model_layers": [int(layer) for layer in FLAGS.layers],
"mf_regularization": FLAGS.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in FLAGS.mlp_regularization],
"use_tpu": FLAGS.tpu is not None,
"tpu": FLAGS.tpu,
"tpu_zone": FLAGS.tpu_zone,
"tpu_gcp_project": FLAGS.tpu_gcp_project,
},
batch_size=flags.FLAGS.batch_size,
eval_batch_size=eval_batch_size)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size # for ExamplesPerSecondHook
)
run_params = {
"batch_size": FLAGS.batch_size,
"eval_batch_size": eval_batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
approx_train_steps = int(ncf_dataset.num_train_positives *
(1 + FLAGS.num_neg) // FLAGS.batch_size)
pred_input_fn = data_preprocessing.make_pred_input_fn(
ncf_dataset=ncf_dataset)
total_training_cycle = 1 if FLAGS.inference_only else FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
if not FLAGS.inference_only:
# Train the model
train_input_fn, train_record_dir, batch_count = \
data_preprocessing.make_train_input_fn(ncf_dataset=ncf_dataset)
if np.abs(approx_train_steps - batch_count) > 1:
tf.logging.warning(
"Estimated ({}) and reported ({}) number of batches differ by more "
"than one".format(approx_train_steps, batch_count))
train_estimator.train(input_fn=train_input_fn,
hooks=train_hooks,
steps=batch_count)
tf.gfile.DeleteRecursively(train_record_dir)
# Evaluate the model
eval_results = evaluate_model(eval_estimator, ncf_dataset,
pred_input_fn)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.fatal("Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# Export SavedModel
if FLAGS.export_savedmodel:
eval_estimator.export_savedmodel(FLAGS.model_dir,
serving_input_receiver_fn)
print("SavedModel successfully exported to: {}/<timestamp>".format(
FLAGS.model_dir))
# Some of the NumPy vector math can be quite large and likes to stay in
# memory for a while.
gc.collect()
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
movielens_dataset.construct_train_eval_csv(data_dir=FLAGS.data_dir,
dataset=FLAGS.dataset)
tf.logging.info("Data preprocessing...")
ncf_dataset = movielens_dataset.data_preprocessing(FLAGS.data_dir,
FLAGS.dataset,
FLAGS.num_neg)
model_helpers.apply_clean(flags.FLAGS)
# Create NeuMF model and convert it to Estimator
tf.logging.info("Creating Estimator from Keras model...")
layers = [int(layer) for layer in FLAGS.layers]
mlp_regularization = [float(reg) for reg in FLAGS.mlp_regularization]
keras_model = neumf_model.NeuMF(ncf_dataset.num_users,
ncf_dataset.num_items, FLAGS.num_factors,
layers, FLAGS.batch_size,
FLAGS.mf_regularization,
mlp_regularization)
num_gpus = flags_core.get_num_gpus(FLAGS)
estimator = convert_keras_to_estimator(keras_model, num_gpus,
FLAGS.model_dir)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size # for ExamplesPerSecondHook
)
run_params = {
"batch_size": FLAGS.batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
# Training and evaluation cycle
def get_train_input_fn():
return movielens_dataset.get_input_fn(
True,
distribution_utils.per_device_batch_size(FLAGS.batch_size,
num_gpus), ncf_dataset,
FLAGS.data_dir, FLAGS.dataset, FLAGS.epochs_between_evals)
def get_pred_input_fn():
return movielens_dataset.get_input_fn(
False,
distribution_utils.per_device_batch_size(FLAGS.batch_size,
num_gpus), ncf_dataset,
FLAGS.data_dir, FLAGS.dataset, 1)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
estimator.train(input_fn=get_train_input_fn(), hooks=train_hooks)
# Evaluate the model
eval_results = evaluate_model(estimator, FLAGS.batch_size, num_gpus,
ncf_dataset, get_pred_input_fn())
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.info("Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def run_deep_speech(_):
"""Run deep speech training and eval loop."""
tf.set_random_seed(flags_obj.seed)
# Data preprocessing
tf.logging.info("Data preprocessing...")
train_speech_dataset = generate_dataset(flags_obj.train_data_dir)
eval_speech_dataset = generate_dataset(flags_obj.eval_data_dir)
# Number of label classes. Label string is "[a-z]' -"
num_classes = len(train_speech_dataset.speech_labels)
# Use distribution strategy for multi-gpu training
num_gpus = flags_core.get_num_gpus(flags_obj)
distribution_strategy = distribution_utils.get_distribution_strategy(num_gpus)
run_config = tf.estimator.RunConfig(
train_distribute=distribution_strategy)
estimator = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=flags_obj.model_dir,
config=run_config,
params={
"num_classes": num_classes,
}
)
# Benchmark logging
run_params = {
"batch_size": flags_obj.batch_size,
"train_epochs": flags_obj.train_epochs,
"rnn_hidden_size": flags_obj.rnn_hidden_size,
"rnn_hidden_layers": flags_obj.rnn_hidden_layers,
"rnn_type": flags_obj.rnn_type,
"is_bidirectional": flags_obj.is_bidirectional,
"use_bias": flags_obj.use_bias
}
dataset_name = "LibriSpeech"
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info("deep_speech", dataset_name, run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
per_device_batch_size = distribution_utils.per_device_batch_size(
flags_obj.batch_size, num_gpus)
def input_fn_train():
return dataset.input_fn(
per_device_batch_size, train_speech_dataset)
def input_fn_eval():
return dataset.input_fn(
per_device_batch_size, eval_speech_dataset)
total_training_cycle = (flags_obj.train_epochs //
flags_obj.epochs_between_evals)
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: %d/%d",
cycle_index + 1, total_training_cycle)
# Perform batch_wise dataset shuffling
train_speech_dataset.entries = dataset.batch_wise_dataset_shuffle(
train_speech_dataset.entries, cycle_index, flags_obj.sortagrad,
flags_obj.batch_size)
estimator.train(input_fn=input_fn_train, hooks=train_hooks)
# Evaluation
tf.logging.info("Starting to evaluate...")
eval_results = evaluate_model(
estimator, eval_speech_dataset.speech_labels,
eval_speech_dataset.entries, input_fn_eval)
# Log the WER and CER results.
benchmark_logger.log_evaluation_result(eval_results)
tf.logging.info(
"Iteration {}: WER = {:.2f}, CER = {:.2f}".format(
cycle_index + 1, eval_results[_WER_KEY], eval_results[_CER_KEY]))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(
flags_obj.wer_threshold, eval_results[_WER_KEY]):
break
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None,
num_images=None,
zeroshot_eval=False):
model_helpers.apply_clean(flags.FLAGS)
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
session_config = config_utils.get_session_config(flags_obj)
run_config = config_utils.get_run_config(flags_obj, flags_core,
session_config,
num_images['train'])
tf.logging.info("ERR1!!!!")
def gen_estimator(period=None):
resnet_size = int(flags_obj.resnet_size)
data_format = flags_obj.data_format
batch_size = flags_obj.batch_size
resnet_version = int(flags_obj.resnet_version)
loss_scale = flags_core.get_loss_scale(flags_obj)
dtype_tf = flags_core.get_tf_dtype(flags_obj)
num_epochs_per_decay = flags_obj.num_epochs_per_decay
learning_rate_decay_factor = flags_obj.learning_rate_decay_factor
end_learning_rate = flags_obj.end_learning_rate
learning_rate_decay_type = flags_obj.learning_rate_decay_type
weight_decay = flags_obj.weight_decay
zero_gamma = flags_obj.zero_gamma
lr_warmup_epochs = flags_obj.lr_warmup_epochs
base_learning_rate = flags_obj.base_learning_rate
use_resnet_d = flags_obj.use_resnet_d
use_dropblock = flags_obj.use_dropblock
dropblock_kp = [float(be) for be in flags_obj.dropblock_kp]
label_smoothing = flags_obj.label_smoothing
momentum = flags_obj.momentum
bn_momentum = flags_obj.bn_momentum
train_epochs = flags_obj.train_epochs
piecewise_lr_boundary_epochs = [
int(be) for be in flags_obj.piecewise_lr_boundary_epochs
]
piecewise_lr_decay_rates = [
float(dr) for dr in flags_obj.piecewise_lr_decay_rates
]
use_ranking_loss = flags_obj.use_ranking_loss
use_se_block = flags_obj.use_se_block
use_sk_block = flags_obj.use_sk_block
mixup_type = flags_obj.mixup_type
dataset_name = flags_obj.dataset_name
kd_temp = flags_obj.kd_temp
no_downsample = flags_obj.no_downsample
anti_alias_filter_size = flags_obj.anti_alias_filter_size
anti_alias_type = flags_obj.anti_alias_type
cls_loss_type = flags_obj.cls_loss_type
logit_type = flags_obj.logit_type
embedding_size = flags_obj.embedding_size
pool_type = flags_obj.pool_type
arc_s = flags_obj.arc_s
arc_m = flags_obj.arc_m
bl_alpha = flags_obj.bl_alpha
bl_beta = flags_obj.bl_beta
exp = None
if install_hyperdash and flags_obj.use_hyperdash:
exp = Experiment(flags_obj.model_dir.split("/")[-1])
resnet_size = exp.param("resnet_size", int(flags_obj.resnet_size))
batch_size = exp.param("batch_size", flags_obj.batch_size)
exp.param("dtype", flags_obj.dtype)
learning_rate_decay_type = exp.param(
"learning_rate_decay_type", flags_obj.learning_rate_decay_type)
weight_decay = exp.param("weight_decay", flags_obj.weight_decay)
zero_gamma = exp.param("zero_gamma", flags_obj.zero_gamma)
lr_warmup_epochs = exp.param("lr_warmup_epochs",
flags_obj.lr_warmup_epochs)
base_learning_rate = exp.param("base_learning_rate",
flags_obj.base_learning_rate)
use_dropblock = exp.param("use_dropblock", flags_obj.use_dropblock)
dropblock_kp = exp.param(
"dropblock_kp", [float(be) for be in flags_obj.dropblock_kp])
piecewise_lr_boundary_epochs = exp.param(
"piecewise_lr_boundary_epochs",
[int(be) for be in flags_obj.piecewise_lr_boundary_epochs])
piecewise_lr_decay_rates = exp.param(
"piecewise_lr_decay_rates",
[float(dr) for dr in flags_obj.piecewise_lr_decay_rates])
mixup_type = exp.param("mixup_type", flags_obj.mixup_type)
dataset_name = exp.param("dataset_name", flags_obj.dataset_name)
exp.param("autoaugment_type", flags_obj.autoaugment_type)
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'resnet_size': resnet_size,
'data_format': data_format,
'batch_size': batch_size,
'resnet_version': resnet_version,
'loss_scale': loss_scale,
'dtype': dtype_tf,
'num_epochs_per_decay': num_epochs_per_decay,
'learning_rate_decay_factor': learning_rate_decay_factor,
'end_learning_rate': end_learning_rate,
'learning_rate_decay_type': learning_rate_decay_type,
'weight_decay': weight_decay,
'zero_gamma': zero_gamma,
'lr_warmup_epochs': lr_warmup_epochs,
'base_learning_rate': base_learning_rate,
'use_resnet_d': use_resnet_d,
'use_dropblock': use_dropblock,
'dropblock_kp': dropblock_kp,
'label_smoothing': label_smoothing,
'momentum': momentum,
'bn_momentum': bn_momentum,
'embedding_size': embedding_size,
'train_epochs': train_epochs,
'piecewise_lr_boundary_epochs': piecewise_lr_boundary_epochs,
'piecewise_lr_decay_rates': piecewise_lr_decay_rates,
'with_drawing_bbox': flags_obj.with_drawing_bbox,
'use_ranking_loss': use_ranking_loss,
'use_se_block': use_se_block,
'use_sk_block': use_sk_block,
'mixup_type': mixup_type,
'kd_temp': kd_temp,
'no_downsample': no_downsample,
'dataset_name': dataset_name,
'anti_alias_filter_size': anti_alias_filter_size,
'anti_alias_type': anti_alias_type,
'cls_loss_type': cls_loss_type,
'logit_type': logit_type,
'arc_s': arc_s,
'arc_m': arc_m,
'pool_type': pool_type,
'bl_alpha': bl_alpha,
'bl_beta': bl_beta,
'train_steps': total_train_steps,
})
return classifier, exp
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs):
return input_function(is_training=True,
use_random_crop=flags_obj.training_random_crop,
num_epochs=num_epochs,
flags_obj=flags_obj)
def input_fn_eval():
return input_function(is_training=False,
use_random_crop=False,
num_epochs=1,
flags_obj=flags_obj)
ckpt_keeper = checkpoint_utils.CheckpointKeeper(
save_dir=flags_obj.model_dir,
num_to_keep=flags_obj.num_best_ckpt_to_keep,
keep_epoch=flags_obj.keep_ckpt_every_eval,
maximize=True)
if zeroshot_eval:
dataset = data_config.get_config(dataset_name)
model = model_fns_predict.Model(
int(flags_obj.resnet_size),
flags_obj.data_format,
resnet_version=int(flags_obj.resnet_version),
num_classes=dataset.num_classes,
zero_gamma=flags_obj.zero_gamma,
use_se_block=flags_obj.use_se_block,
use_sk_block=flags_obj.use_sk_block,
no_downsample=flags_obj.no_downsample,
anti_alias_filter_size=flags_obj.anti_alias_filter_size,
anti_alias_type=flags_obj.anti_alias_type,
bn_momentum=flags_obj.bn_momentum,
embedding_size=flags_obj.embedding_size,
pool_type=flags_obj.pool_type,
bl_alpha=flags_obj.bl_alpha,
bl_beta=flags_obj.bl_beta,
dtype=flags_core.get_tf_dtype(flags_obj),
loss_type=flags_obj.cls_loss_type)
def train_and_evaluate(hooks):
tf.logging.info('Starting cycle: %d/%d', cycle_index, int(n_loops))
if num_train_epochs:
classifier.train(input_fn=lambda: input_fn_train(num_train_epochs),
hooks=hooks,
steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
if zeroshot_eval:
tf.reset_default_graph()
eval_results = recall_metric.recall_at_k(
flags_obj,
flags_core,
input_fns.input_fn_ir_eval,
model,
num_images['validation'],
eval_similarity=flags_obj.eval_similarity,
return_embedding=True)
else:
eval_results = classifier.predict(input_fn=input_fn_eval)
return eval_results
total_train_steps = flags_obj.train_epochs * int(
num_images['train'] / flags_obj.batch_size)
if flags_obj.eval_only or not flags_obj.train_epochs:
schedule, n_loops = [0], 1
elif flags_obj.export_only:
schedule, n_loops = [], 0
else:
n_loops = math.ceil(flags_obj.train_epochs /
flags_obj.epochs_between_evals)
schedule = [
flags_obj.epochs_between_evals for _ in range(int(n_loops))
]
schedule[-1] = flags_obj.train_epochs - sum(
schedule[:-1]) # over counting.
schedule = config_utils.get_epoch_schedule(flags_obj, schedule,
num_images)
tf.logging.info('epoch schedule:')
tf.logging.info(schedule)
classifier, exp = gen_estimator()
if flags_obj.pretrained_model_checkpoint_path:
warm_start_hook = WarmStartHook(
flags_obj.pretrained_model_checkpoint_path)
train_hooks.append(warm_start_hook)
for cycle_index, num_train_epochs in enumerate(schedule):
tf.logging.info("ERR123!!!!")
eval_results = train_and_evaluate(train_hooks)
return eval_results
if zeroshot_eval:
metric = eval_results['recall_at_1']
else:
metric = eval_results['accuracy']
tf.logging.info("ERR1234!!!!")
ckpt_keeper.save(metric, flags_obj.model_dir)
if exp:
exp.metric("accuracy", metric)
benchmark_logger.log_evaluation_result(eval_results)
tf.logging.info("ERR12345!!!!")
if model_helpers.past_stop_threshold(flags_obj.stop_threshold, metric):
break
if model_helpers.past_stop_threshold(total_train_steps,
eval_results['global_step']):
break
if exp:
exp.end()
if flags_obj.export_dir is not None:
export_utils.export_pb(flags_core, flags_obj, shape, classifier)
def resnet_main(
flags_obj, model_function, input_function, dataset_name, shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
Returns:
Dict of results of the run.
"""
model_helpers.apply_clean(flags.FLAGS)
# Ensures flag override logic is only executed if explicitly triggered.
if flags_obj.tf_gpu_thread_mode:
override_flags_and_set_envars_for_gpu_thread_pool(flags_obj)
# Creates session config. allow_soft_placement = True, is required for
# multi-GPU and is not harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
# Creates a `RunConfig` that checkpoints every 24 hours which essentially
# results in checkpoints determined only by `epochs_between_evals`.
run_config = tf.estimator.RunConfig(
train_distribute=distribution_strategy,
session_config=session_config,
save_checkpoints_secs=60*60*24)
# Initializes model with all but the dense layer from pretrained ResNet.
if flags_obj.pretrained_model_checkpoint_path is not None:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start='^(?!.*dense)')
else:
warm_start_settings = None
classifier = tf.estimator.Estimator(
model_fn=model_function, model_dir=flags_obj.model_dir, config=run_config,
warm_start_from=warm_start_settings, params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj),
'fine_tune': flags_obj.fine_tune
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet', dataset_name, run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs):
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=num_epochs,
dtype=flags_core.get_tf_dtype(flags_obj),
datasets_num_private_threads=flags_obj.datasets_num_private_threads,
num_parallel_batches=flags_obj.datasets_num_parallel_batches)
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1,
dtype=flags_core.get_tf_dtype(flags_obj))
if flags_obj.eval_only or not flags_obj.train_epochs:
# If --eval_only is set, perform a single loop with zero train epochs.
schedule, n_loops = [0], 1
else:
# Compute the number of times to loop while training. All but the last
# pass will train for `epochs_between_evals` epochs, while the last will
# train for the number needed to reach `training_epochs`. For instance if
# train_epochs = 25 and epochs_between_evals = 10
# schedule will be set to [10, 10, 5]. That is to say, the loop will:
# Train for 10 epochs and then evaluate.
# Train for another 10 epochs and then evaluate.
# Train for a final 5 epochs (to reach 25 epochs) and then evaluate.
n_loops = math.ceil(flags_obj.train_epochs / flags_obj.epochs_between_evals)
schedule = [flags_obj.epochs_between_evals for _ in range(int(n_loops))]
schedule[-1] = flags_obj.train_epochs - sum(schedule[:-1]) # over counting.
for cycle_index, num_train_epochs in enumerate(schedule):
tf.logging.info('Starting cycle: %d/%d', cycle_index, int(n_loops))
if num_train_epochs:
classifier.train(input_fn=lambda: input_fn_train(num_train_epochs),
hooks=train_hooks, max_steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data, which
# will iterate forever. Passing steps=flags_obj.max_train_steps allows the
# eval (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(
flags_obj.stop_threshold, eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
export_dtype = flags_core.get_tf_dtype(flags_obj)
if flags_obj.image_bytes_as_serving_input:
input_receiver_fn = functools.partial(
image_bytes_serving_input_fn, shape, dtype=export_dtype)
else:
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size, dtype=export_dtype)
classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn,
strip_default_attrs=True)
return eval_results
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing and not FLAGS.use_synthetic_data:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
if FLAGS.seed is not None:
np.random.seed(FLAGS.seed)
params = parse_flags(FLAGS)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
if FLAGS.use_synthetic_data:
producer = data_pipeline.DummyConstructor()
num_users, num_items = data_preprocessing.DATASET_TO_NUM_USERS_AND_ITEMS[
FLAGS.dataset]
num_train_steps = rconst.SYNTHETIC_BATCHES_PER_EPOCH
num_eval_steps = rconst.SYNTHETIC_BATCHES_PER_EPOCH
else:
num_users, num_items, producer = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset,
data_dir=FLAGS.data_dir,
params=params,
constructor_type=FLAGS.constructor_type,
deterministic=FLAGS.seed is not None)
num_train_steps = (producer.train_batches_per_epoch //
params["batches_per_step"])
num_eval_steps = (producer.eval_batches_per_epoch //
params["batches_per_step"])
assert not producer.train_batches_per_epoch % params["batches_per_step"]
assert not producer.eval_batches_per_epoch % params["batches_per_step"]
producer.start()
params["num_users"], params["num_items"] = num_users, num_items
model_helpers.apply_clean(flags.FLAGS)
estimator = construct_estimator(model_dir=FLAGS.model_dir, params=params)
benchmark_logger, train_hooks = log_and_get_hooks(
params["eval_batch_size"])
target_reached = False
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_LOOP)
for cycle_index in range(total_training_cycle):
assert FLAGS.epochs_between_evals == 1 or not mlperf_helper.LOGGER.enabled
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_EPOCH,
value=cycle_index)
train_input_fn = producer.make_input_fn(is_training=True)
estimator.train(input_fn=train_input_fn,
hooks=train_hooks,
steps=num_train_steps)
tf.logging.info("Beginning evaluation.")
eval_input_fn = producer.make_input_fn(is_training=False)
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_START,
value=cycle_index)
eval_results = estimator.evaluate(eval_input_fn, steps=num_eval_steps)
tf.logging.info("Evaluation complete.")
hr = float(eval_results[rconst.HR_KEY])
ndcg = float(eval_results[rconst.NDCG_KEY])
loss = float(eval_results["loss"])
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_TARGET,
value={
"epoch": cycle_index,
"value": FLAGS.hr_threshold
})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_ACCURACY,
value={
"epoch": cycle_index,
"value": hr
})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_HP_NUM_NEG,
value={
"epoch": cycle_index,
"value": rconst.NUM_EVAL_NEGATIVES
})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_STOP,
value=cycle_index)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}, Loss = {:.4f}".format(
cycle_index + 1, hr, ndcg, loss))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
target_reached = True
break
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_STOP,
value={"success": target_reached})
producer.stop_loop()
producer.join()
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_FINAL)
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
"""
model_helpers.apply_clean(flags.FLAGS)
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy,
session_config=session_config)
# initialize our model with all but the dense layer from pretrained resnet
if flags_obj.pretrained_model_checkpoint_path is not None:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start='^(?!.*dense)')
else:
warm_start_settings = None
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
warm_start_from=warm_start_settings,
params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj),
'fine_tune': flags_obj.fine_tune
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs):
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=num_epochs,
num_gpus=flags_core.get_num_gpus(flags_obj),
dtype=flags_core.get_tf_dtype(flags_obj))
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1,
dtype=flags_core.get_tf_dtype(flags_obj))
if flags_obj.eval_only or not flags_obj.train_epochs:
# If --eval_only is set, perform a single loop with zero train epochs.
schedule, n_loops = [0], 1
else:
# Compute the number of times to loop while training. All but the last
# pass will train for `epochs_between_evals` epochs, while the last will
# train for the number needed to reach `training_epochs`. For instance if
# train_epochs = 25 and epochs_between_evals = 10
# schedule will be set to [10, 10, 5]. That is to say, the loop will:
# Train for 10 epochs and then evaluate.
# Train for another 10 epochs and then evaluate.
# Train for a final 5 epochs (to reach 25 epochs) and then evaluate.
n_loops = math.ceil(flags_obj.train_epochs /
flags_obj.epochs_between_evals)
schedule = [
flags_obj.epochs_between_evals for _ in range(int(n_loops))
]
schedule[-1] = flags_obj.train_epochs - sum(
schedule[:-1]) # over counting.
for cycle_index, num_train_epochs in enumerate(schedule):
tf.logging.info('Starting cycle: %d/%d', cycle_index, int(n_loops))
if num_train_epochs:
classifier.train(input_fn=lambda: input_fn_train(num_train_epochs),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data, which
# will iterate forever. Passing steps=flags_obj.max_train_steps allows the
# eval (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size)
classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn)
def resnet_main(
flags_obj, model_function, input_function, dataset_name, shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
"""
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
if flags_core.get_num_gpus(flags_obj) == 0:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:CPU:0')
elif flags_core.get_num_gpus(flags_obj) == 1:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:GPU:0')
else:
distribution = tf.contrib.distribute.MirroredStrategy(
num_gpus=flags_core.get_num_gpus(flags_obj)
)
run_config = tf.estimator.RunConfig(train_distribute=distribution,
session_config=session_config)
classifier = tf.estimator.Estimator(
model_fn=model_function, model_dir=flags_obj.model_dir, config=run_config,
params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj)
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
benchmark_logger = logger.config_benchmark_logger(flags_obj)
benchmark_logger.log_run_info('resnet', dataset_name, run_params)
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
batch_size=flags_obj.batch_size)
def input_fn_train():
return input_function(
is_training=True, data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=flags_obj.epochs_between_evals)
def input_fn_eval():
return input_function(
is_training=False, data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1)
total_training_cycle = (flags_obj.train_epochs //
flags_obj.epochs_between_evals)
for cycle_index in range(total_training_cycle):
tf.logging.info('Starting a training cycle: %d/%d',
cycle_index, total_training_cycle)
classifier.train(input_fn=input_fn_train, hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data, which
# will iterate forever. Passing steps=flags_obj.max_train_steps allows the
# eval (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(
flags_obj.stop_threshold, eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size)
classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn)
def run_loop(name, train_input_fn, eval_input_fn, model_column_fn,
build_estimator_fn, flags_obj, tensors_to_log, early_stop=False):
"""Define training loop."""
model_helpers.apply_clean(flags.FLAGS)
model = build_estimator_fn(
model_dir=flags_obj.model_dir, model_type=flags_obj.model_type,
model_column_fn=model_column_fn)
run_params = {
'batch_size': flags_obj.batch_size,
'train_epochs': flags_obj.train_epochs,
'model_type': flags_obj.model_type,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('wide_deep', name, run_params,
test_id=flags_obj.benchmark_test_id)
loss_prefix = LOSS_PREFIX.get(flags_obj.model_type, '')
tensors_to_log = {k: v.format(loss_prefix=loss_prefix)
for k, v in tensors_to_log.items()}
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks, model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size, tensors_to_log=tensors_to_log)
profiler_hook = tf.train.ProfilerHook(save_steps= 100, save_secs= None, output_dir="profs", show_memory=True, show_dataflow=True)
#DOGA DEBUG GRAPH
gdef = gpb.GraphDef()
with open('/tmp/census_model/graph.pbtxt', 'r') as fh:
graph_str = fh.read()
pbtf.Parse(graph_str, gdef)
with tf.Graph().as_default() as graph:
tf.import_graph_def(gdef)
operations_tensors = {}
operations_names = tf.get_default_graph().get_operations()
count1 = 0
count2 = 0
for operation in operations_names:
operation_name = operation.name
operations_info = tf.get_default_graph().get_operation_by_name(operation_name).values()
if len(operations_info) > 0:
if not (operations_info[0].shape.ndims is None):
operation_shape = operations_info[0].shape.as_list()
operation_dtype_size = operations_info[0].dtype.size
if not (operation_dtype_size is None):
operation_no_of_elements = 1
for dim in operation_shape:
if not(dim is None):
operation_no_of_elements = operation_no_of_elements * dim
total_size = operation_no_of_elements * operation_dtype_size
operations_tensors[operation_name] = total_size
else:
count1 = count1 + 1
else:
count1 = count1 + 1
operations_tensors[operation_name] = -1
else:
count2 = count2 + 1
operations_tensors[operation_name] = -1
print(count1)
print(count2)
with open('tensors_sz.json', 'w') as f:
json.dump(operations_tensors, f)
# Train and evaluate the model every `flags.epochs_between_evals` epochs.
for n in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
model.train(input_fn=train_input_fn, hooks=[profiler_hook])
results = model.evaluate(input_fn=eval_input_fn)
# Display evaluation metrics
tf.logging.info('Results at epoch %d / %d',
(n + 1) * flags_obj.epochs_between_evals,
flags_obj.train_epochs)
tf.logging.info('-' * 60)
for key in sorted(results):
tf.logging.info('%s: %s' % (key, results[key]))
benchmark_logger.log_evaluation_result(results)
if early_stop and model_helpers.past_stop_threshold(
flags_obj.stop_threshold, results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
export_model(model, flags_obj.model_type, flags_obj.export_dir,
model_column_fn)
def run_mnist(flags_obj):
"""Run MNIST training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
model_function = model_fn
if flags_obj.multi_gpu:
validate_batch_size_for_multi_gpu(flags_obj.batch_size)
# There are two steps required if using multi-GPU: (1) wrap the model_fn,
# and (2) wrap the optimizer. The first happens here, and (2) happens
# in the model_fn itself when the optimizer is defined.
model_function = tf.contrib.estimator.replicate_model_fn(
model_fn, loss_reduction=tf.losses.Reduction.MEAN)
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
mnist_classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
params={
'data_format': data_format,
'multi_gpu': flags_obj.multi_gpu
})
# Set up training and evaluation input functions.
def train_input_fn():
"""Prepare data for training."""
# When choosing shuffle buffer sizes, larger sizes result in better
# randomness, while smaller sizes use less memory. MNIST is a small
# enough dataset that we can easily shuffle the full epoch.
ds = dataset.train(flags_obj.data_dir)
ds = ds.cache().shuffle(buffer_size=50000).batch(flags_obj.batch_size)
# Iterate through the dataset a set number (`epochs_between_evals`) of times
# during each training session.
ds = ds.repeat(flags_obj.epochs_between_evals)
return ds
def eval_input_fn():
return dataset.test(flags_obj.data_dir).batch(
flags_obj.batch_size).make_one_shot_iterator().get_next()
# Set up hook that outputs training logs every 100 steps.
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks, batch_size=flags_obj.batch_size)
# Train and evaluate model.
for _ in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
mnist_classifier.train(input_fn=train_input_fn, hooks=train_hooks)
eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
print('\nEvaluation results:\n\t%s\n' % eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
image = tf.placeholder(tf.float32, [None, 28, 28])
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'image': image,
})
mnist_classifier.export_savedmodel(flags_obj.export_dir, input_fn)
def resnet_main(flags, model_function, input_function, shape=None):
"""Shared main loop for ResNet Models.
Args:
flags: FLAGS object that contains the params for running. See
ResnetArgParser for created flags.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
shape: list of ints representing the shape of the images used for training.
This is only used if flags.export_dir is passed.
"""
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
if flags.multi_gpu:
validate_batch_size_for_multi_gpu(flags.batch_size)
# There are two steps required if using multi-GPU: (1) wrap the model_fn,
# and (2) wrap the optimizer. The first happens here, and (2) happens
# in the model_fn itself when the optimizer is defined.
model_function = tf.contrib.estimator.replicate_model_fn(
model_function, loss_reduction=tf.losses.Reduction.MEAN)
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags.intra_op_parallelism_threads,
allow_soft_placement=True)
# Set up a RunConfig to save checkpoint and set session config.
run_config = tf.estimator.RunConfig().replace(
save_checkpoints_secs=1e9, session_config=session_config)
classifier = tf.estimator.Estimator(model_fn=model_function,
model_dir=flags.model_dir,
config=run_config,
params={
'resnet_size': flags.resnet_size,
'data_format': flags.data_format,
'batch_size': flags.batch_size,
'multi_gpu': flags.multi_gpu,
'version': flags.version,
'loss_scale': flags.loss_scale,
'dtype': flags.dtype
})
if flags.benchmark_log_dir is not None:
benchmark_logger = logger.BenchmarkLogger(flags.benchmark_log_dir)
benchmark_logger.log_run_info('resnet')
else:
benchmark_logger = None
for _ in range(flags.train_epochs // flags.epochs_between_evals):
train_hooks = hooks_helper.get_train_hooks(
flags.hooks,
batch_size=flags.batch_size,
benchmark_log_dir=flags.benchmark_log_dir)
print('Starting a training cycle.')
def input_fn_train():
return input_function(True, flags.data_dir, flags.batch_size,
flags.epochs_between_evals,
flags.num_parallel_calls, flags.multi_gpu)
classifier.train(input_fn=input_fn_train,
hooks=train_hooks,
max_steps=flags.max_train_steps)
print('Starting to evaluate.')
# Evaluate the model and print results
def input_fn_eval():
return input_function(False, flags.data_dir, flags.batch_size, 1,
flags.num_parallel_calls, flags.multi_gpu)
# flags.max_train_steps is generally associated with testing and profiling.
# As a result it is frequently called with synthetic data, which will
# iterate forever. Passing steps=flags.max_train_steps allows the eval
# (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags.max_train_steps)
print(eval_results)
if benchmark_logger:
benchmark_logger.log_estimator_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags.stop_threshold,
eval_results['accuracy']):
break
if flags.export_dir is not None:
warn_on_multi_gpu_export(flags.multi_gpu)
# Exports a saved model for the given classifier.
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags.batch_size)
classifier.export_savedmodel(flags.export_dir, input_receiver_fn)
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
movielens_dataset.construct_train_eval_csv(
data_dir=FLAGS.data_dir, dataset=FLAGS.dataset)
tf.logging.info("Data preprocessing...")
ncf_dataset = movielens_dataset.data_preprocessing(
FLAGS.data_dir, FLAGS.dataset, FLAGS.num_neg)
model_helpers.apply_clean(flags.FLAGS)
# Create NeuMF model and convert it to Estimator
tf.logging.info("Creating Estimator from Keras model...")
layers = [int(layer) for layer in FLAGS.layers]
mlp_regularization = [float(reg) for reg in FLAGS.mlp_regularization]
keras_model = neumf_model.NeuMF(
ncf_dataset.num_users, ncf_dataset.num_items, FLAGS.num_factors,
layers, FLAGS.batch_size, FLAGS.mf_regularization,
mlp_regularization)
num_gpus = flags_core.get_num_gpus(FLAGS)
estimator = convert_keras_to_estimator(keras_model, num_gpus, FLAGS.model_dir)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size # for ExamplesPerSecondHook
)
run_params = {
"batch_size": FLAGS.batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(
model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
# Training and evaluation cycle
def get_train_input_fn():
return movielens_dataset.get_input_fn(
True,
distribution_utils.per_device_batch_size(FLAGS.batch_size, num_gpus),
ncf_dataset, FLAGS.data_dir, FLAGS.dataset, FLAGS.epochs_between_evals)
def get_pred_input_fn():
return movielens_dataset.get_input_fn(
False,
distribution_utils.per_device_batch_size(FLAGS.batch_size, num_gpus),
ncf_dataset, FLAGS.data_dir, FLAGS.dataset, 1)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
estimator.train(input_fn=get_train_input_fn(), hooks=train_hooks)
# Evaluate the model
eval_results = evaluate_model(
estimator, FLAGS.batch_size, num_gpus, ncf_dataset, get_pred_input_fn())
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def densenet_main(
flags_obj, model_function, input_function, dataset_name, shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_densenet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
"""
model_helpers.apply_clean(flags.FLAGS)
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
'''
session_config = tf.ConfigProto(
inter_op_parallelism_threads=1,
intra_op_parallelism_threads=1,
allow_soft_placement=True)
'''
session_config = tf.ConfigProto(allow_soft_placement=True)
# sirius:
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
run_config = tf.estimator.RunConfig(
train_distribute=distribution_strategy, session_config=session_config,
save_summary_steps=500)
# print all flags inside model main
# for k,v in tf.flags.FLAGS.__flags.items():
# print('=================================')
# for k,v in flags_obj.items():
# print('***',v.__dict__['name'],v.__dict__['_value'])
# Note: 这里的flags_obj定义了多种类型的flags
# print(flags_obj)
train_dir = r'E:\denseNet\resnet_cifar10\train_dir'
export_dir_all = r'E:\denseNet\resnet_cifar10\export_dir'
model_name = 'd_{}_k_{}'.format(flags.FLAGS.d, flags.FLAGS.k)
model_dir = os.path.join(train_dir, model_name)
export_dir = os.path.join(export_dir_all, model_name)
# Note flags
# parameters that will be passed into model fn
classifier = tf.estimator.Estimator(
model_fn=model_function, model_dir=model_dir, config=run_config,
params={
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj),
# network parameters
'd': flags_obj.d,
'k':flags_obj.k,
'compressionRate':flags_obj.compressionRate,
'expansion':flags_obj.expansion,
'bottleneck':flags_obj.bottleneck
})
# Note flags
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('densenet', dataset_name, run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
model_dir=model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train():
return input_function(
is_training=True, data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=flags_obj.epochs_between_evals,
num_gpus=flags_core.get_num_gpus(flags_obj))
def input_fn_eval():
return input_function(
is_training=False, data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_device_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1)
total_training_cycle = (flags_obj.train_epochs //
flags_obj.epochs_between_evals)
# print('*** total_training_cycle',total_training_cycle)
for cycle_index in range(total_training_cycle):
tf.logging.info('Starting a training cycle: %d/%d',
cycle_index, total_training_cycle)
classifier.train(input_fn=input_fn_train, hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data, which
# will iterate forever. Passing steps=flags_obj.max_train_steps allows the
# eval (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(
flags_obj.stop_threshold, eval_results['accuracy']):
break
# export model at last
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size)
classifier.export_savedmodel(export_dir, input_receiver_fn)
def main(_):
# Data preprocessing
# The file name of training and test dataset
train_fname = os.path.join(
FLAGS.data_dir, FLAGS.dataset + "-" + constants.TRAIN_RATINGS_FILENAME)
test_fname = os.path.join(
FLAGS.data_dir, FLAGS.dataset + "-" + constants.TEST_RATINGS_FILENAME)
neg_fname = os.path.join(FLAGS.data_dir,
FLAGS.dataset + "-" + constants.TEST_NEG_FILENAME)
assert os.path.exists(train_fname), (
"Run data_download.py first to download and extract {} dataset".format(
FLAGS.dataset))
tf.logging.info("Data preprocessing...")
ncf_dataset = dataset.data_preprocessing(train_fname, test_fname,
neg_fname, FLAGS.num_neg)
# Create NeuMF model and convert it to Estimator
tf.logging.info("Creating Estimator from Keras model...")
layers = [int(layer) for layer in FLAGS.layers]
mlp_regularization = [float(reg) for reg in FLAGS.mlp_regularization]
keras_model = neumf_model.NeuMF(ncf_dataset.num_users,
ncf_dataset.num_items, FLAGS.num_factors,
layers, FLAGS.batch_size,
FLAGS.mf_regularization,
mlp_regularization)
num_gpus = flags_core.get_num_gpus(FLAGS)
estimator = convert_keras_to_estimator(keras_model, num_gpus,
FLAGS.model_dir)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
batch_size=FLAGS.batch_size # for ExamplesPerSecondHook
)
run_params = {
"batch_size": FLAGS.batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.config_benchmark_logger(FLAGS)
benchmark_logger.log_run_info(model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params)
# Training and evaluation cycle
def train_input_fn():
return dataset.input_fn(
True, per_device_batch_size(FLAGS.batch_size, num_gpus),
ncf_dataset, FLAGS.epochs_between_evals)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
estimator.train(input_fn=train_input_fn, hooks=train_hooks)
# Evaluate the model
eval_results = evaluate_model(estimator, FLAGS.batch_size, num_gpus,
ncf_dataset)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.info("Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def test_past_stop_threshold_not_number(self):
"""Tests for error conditions."""
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold("str", 1)
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold("str", tf.constant(5))
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold("str", "another")
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold(0, None)
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold(0.7, "str")
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold(tf.constant(4), None)
def run_deep_speech(_):
"""Run deep speech training and eval loop."""
tf.set_random_seed(flags_obj.seed)
# Data preprocessing
tf.logging.info("Data preprocessing...")
train_speech_dataset = generate_dataset(flags_obj.train_data_dir)
eval_speech_dataset = generate_dataset(flags_obj.eval_data_dir)
# Number of label classes. Label string is "[a-z]' -"
num_classes = len(train_speech_dataset.speech_labels)
# Use distribution strategy for multi-gpu training
num_gpus = flags_core.get_num_gpus(flags_obj)
distribution_strategy = distribution_utils.get_distribution_strategy(
num_gpus=num_gpus)
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy)
estimator = tf.estimator.Estimator(model_fn=model_fn,
model_dir=flags_obj.model_dir,
config=run_config,
params={
"num_classes": num_classes,
})
# Benchmark logging
run_params = {
"batch_size": flags_obj.batch_size,
"train_epochs": flags_obj.train_epochs,
"rnn_hidden_size": flags_obj.rnn_hidden_size,
"rnn_hidden_layers": flags_obj.rnn_hidden_layers,
"rnn_type": flags_obj.rnn_type,
"is_bidirectional": flags_obj.is_bidirectional,
"use_bias": flags_obj.use_bias
}
dataset_name = "LibriSpeech"
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info("deep_speech",
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
per_replica_batch_size = distribution_utils.per_replica_batch_size(
flags_obj.batch_size, num_gpus)
def input_fn_train():
return dataset.input_fn(per_replica_batch_size, train_speech_dataset)
def input_fn_eval():
return dataset.input_fn(per_replica_batch_size, eval_speech_dataset)
total_training_cycle = (flags_obj.train_epochs //
flags_obj.epochs_between_evals)
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: %d/%d", cycle_index + 1,
total_training_cycle)
# Perform batch_wise dataset shuffling
train_speech_dataset.entries = dataset.batch_wise_dataset_shuffle(
train_speech_dataset.entries, cycle_index, flags_obj.sortagrad,
flags_obj.batch_size)
estimator.train(input_fn=input_fn_train,
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
if flags_obj.skip_eval:
break
# Evaluation
tf.logging.info("Starting to evaluate...")
eval_results = evaluate_model(estimator,
eval_speech_dataset.speech_labels,
eval_speech_dataset.entries,
input_fn_eval)
# Log the WER and CER results.
benchmark_logger.log_evaluation_result(eval_results)
tf.logging.info("Iteration {}: WER = {:.2f}, CER = {:.2f}".format(
cycle_index + 1, eval_results[_WER_KEY], eval_results[_CER_KEY]))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(flags_obj.wer_threshold,
eval_results[_WER_KEY]):
break
def run_loop(estimator,
schedule_manager,
train_hooks=None,
benchmark_logger=None,
bleu_source=None,
bleu_ref=None,
bleu_threshold=None,
vocab_file=None):
"""Train and evaluate model, and optionally compute model's BLEU score. # 此函数用来训练和验证模型，可选的还有，计算bleu分数
**Step vs. Epoch vs. Iteration** # 下面一段介绍step epoch iteration的概念和区别
Steps and epochs are canonical terms used in TensorFlow and general machine
learning. They are used to describe running a single process (train/eval):
- Step refers to running the process through a single or batch of examples.
- Epoch refers to running the process through an entire dataset.
E.g. training a dataset with 100 examples. The dataset is
divided into 20 batches with 5 examples per batch. A single training step
trains the model on one batch. After 20 training steps, the model will have
trained on every batch in the dataset, or, in other words, one epoch.
Meanwhile, iteration is used in this implementation to describe running
multiple processes (training and eval).
- A single iteration:
1. trains the model for a specific number of steps or epochs.
2. evaluates the model.
3. (if source and ref files are provided) compute BLEU score.
This function runs through multiple train+eval+bleu iterations.
Args:
estimator: tf.Estimator containing model to train. # 估计器中包含模型结构
schedule_manager: A schedule.Manager object to guide the run loop. # 控制训练进程的
train_hooks: List of hooks to pass to the estimator during training. # 不懂
benchmark_logger: a BenchmarkLogger object that logs evaluation data # 专门用来打benchmark的日志的
bleu_source: File containing text to be translated for BLEU calculation. # 计算bleu 分用到的待翻译文件
bleu_ref: File containing reference translations for BLEU calculation. # 计算bleu 分用到的标准答案文件
bleu_threshold: minimum BLEU score before training is stopped. # 在训练结束前所需要达到的bleu分数
vocab_file: Path to vocab file that will be used to subtokenize bleu_source. # 指定的词表文件仅仅用于计算bleu分数？？不用于训练？？？
Raises:
ValueError: if both or none of single_iteration_train_steps and
single_iteration_train_epochs were defined.
NotFoundError: if the vocab file or bleu files don't exist.
"""
if bleu_source: # 如果指定了用于计算bleu的文件，则需要检查文件是否存在
_validate_file(bleu_source)
if bleu_ref:
_validate_file(bleu_ref)
if vocab_file:
_validate_file(vocab_file)
evaluate_bleu = bleu_source is not None and bleu_ref is not None
if evaluate_bleu and schedule_manager.use_tpu: # 如果用TPU就不能计算bleu了
raise ValueError(
"BLEU score can not be computed when training with a TPU, "
"as it requires estimator.predict which is not yet "
"supported.")
# Print details of training schedule. # 打印一下训练控制进度所用到的信息，例如是按照bleu分数来指定结束，还是按照指定的steps数量来结束
tf.logging.info("Training schedule:")
tf.logging.info("\t1. Train for {}".format(
schedule_manager.train_increment_str))
tf.logging.info("\t2. Evaluate model.")
if evaluate_bleu:
tf.logging.info("\t3. Compute BLEU score.")
if bleu_threshold is not None:
tf.logging.info(
"Repeat above steps until the BLEU score reaches %f" %
bleu_threshold)
if not evaluate_bleu or bleu_threshold is None:
tf.logging.info("Repeat above steps %d times." %
schedule_manager.train_eval_iterations)
if evaluate_bleu: # 用来打印bleu分日志的
# Create summary writer to log bleu score (values can be displayed in
# Tensorboard).
bleu_writer = tf.summary.FileWriter(
os.path.join(estimator.model_dir, BLEU_DIR))
if bleu_threshold is not None:
# Change loop stopping condition if bleu_threshold is defined.
schedule_manager.train_eval_iterations = INF
# Loop training/evaluation/bleu cycles
for i in xrange(schedule_manager.train_eval_iterations): # 这个循环控制迭代
tf.logging.info("Starting iteration %d" % (i + 1))
# Train the model for single_iteration_train_steps or until the input fn
# runs out of examples (if single_iteration_train_steps is None).
estimator.train( # 训练一个迭代周期
dataset.train_input_fn, # dataset什么时候加载的数据
steps=schedule_manager.single_iteration_train_steps,
hooks=train_hooks)
eval_results = estimator.evaluate( # 验证一次
input_fn=dataset.eval_input_fn,
steps=schedule_manager.single_iteration_eval_steps)
tf.logging.info("Evaluation results (iter %d/%d):" % # 打印验证结果
(i + 1, schedule_manager.train_eval_iterations))
tf.logging.info(eval_results)
benchmark_logger.log_evaluation_result(eval_results)
# The results from estimator.evaluate() are measured on an approximate
# translation, which utilize the target golden values provided. The actual
# bleu score must be computed using the estimator.predict() path, which
# outputs translations that are not based on golden values. The translations
# are compared to reference file to get the actual bleu score.
if evaluate_bleu: # 以下是关于bleu的东西
uncased_score, cased_score = evaluate_and_log_bleu(
estimator, bleu_source, bleu_ref, vocab_file)
# Write actual bleu scores using summary writer and benchmark logger
global_step = get_global_step(estimator)
summary = tf.Summary(value=[
tf.Summary.Value(tag="bleu/uncased",
simple_value=uncased_score),
tf.Summary.Value(tag="bleu/cased", simple_value=cased_score),
])
bleu_writer.add_summary(summary, global_step)
bleu_writer.flush()
benchmark_logger.log_metric("bleu_uncased",
uncased_score,
global_step=global_step)
benchmark_logger.log_metric("bleu_cased",
cased_score,
global_step=global_step)
# Stop training if bleu stopping threshold is met.
if model_helpers.past_stop_threshold(bleu_threshold,
uncased_score):
bleu_writer.close()
break
def run_mnist(flags_obj):
"""Run MNIST training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
model_helpers.apply_clean(flags_obj)
model_function = model_fn
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
run_config = tf.estimator.RunConfig(
train_distribute=distribution_strategy, session_config=session_config)
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
mnist_classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'data_format': data_format,
})
# Set up training and evaluation input functions.
def train_input_fn():
"""Prepare data for training."""
# When choosing shuffle buffer sizes, larger sizes result in better
# randomness, while smaller sizes use less memory. MNIST is a small
# enough dataset that we can easily shuffle the full epoch.
ds = dataset.train(flags_obj.data_dir)
ds = ds.cache().shuffle(buffer_size=50000).batch(flags_obj.batch_size)
# Iterate through the dataset a set number (`epochs_between_evals`) of times
# during each training session.
ds = ds.repeat(flags_obj.epochs_between_evals)
return ds
def eval_input_fn():
return dataset.test(flags_obj.data_dir).batch(
flags_obj.batch_size).make_one_shot_iterator().get_next()
# Set up hook that outputs training logs every 100 steps.
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks, model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
# Train and evaluate model.
for _ in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
mnist_classifier.train(input_fn=train_input_fn, hooks=train_hooks)
eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
print('\nEvaluation results:\n\t%s\n' % eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
image = tf.placeholder(tf.float32, [None, 28, 28])
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'image': image,
})
mnist_classifier.export_savedmodel(flags_obj.export_dir, input_fn,
strip_default_attrs=True)
def test_past_stop_threshold_not_number(self):
"""Tests for error conditions."""
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold('str', 1)
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold('str', tf.constant(5))
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold('str', 'another')
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold(0, None)
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold(0.7, 'str')
with self.assertRaises(ValueError):
model_helpers.past_stop_threshold(tf.constant(4), None)
def run_mnist(flags_obj):
"""Run MNIST training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
model_helpers.apply_clean(flags_obj)
model_function = model_fn
# Get number of GPUs as defined by the --num_gpus flags and the number of
# GPUs available on the machine.
num_gpus = flags_core.get_num_gpus(flags_obj)
multi_gpu = num_gpus > 1
if multi_gpu:
# Validate that the batch size can be split into devices.
distribution_utils.per_device_batch_size(flags_obj.batch_size, num_gpus)
# There are two steps required if using multi-GPU: (1) wrap the model_fn,
# and (2) wrap the optimizer. The first happens here, and (2) happens
# in the model_fn itself when the optimizer is defined.
model_function = tf.contrib.estimator.replicate_model_fn(
model_fn, loss_reduction=tf.losses.Reduction.MEAN,
devices=["/device:GPU:%d" % d for d in range(num_gpus)])
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
mnist_classifier = tf.estimator.Estimator(
model_fn=model_function,
params={
'data_format': data_format,
'multi_gpu': multi_gpu
})
# Set up training and evaluation input functions.
def train_input_fn():
"""Prepare data for training."""
# When choosing shuffle buffer sizes, larger sizes result in better
# randomness, while smaller sizes use less memory. MNIST is a small
# enough dataset that we can easily shuffle the full epoch.
ds = dataset.train(flags_obj.data_dir)
def invert(image, label):
return (image * -1.0) + 1.0, label
def brightness(image, label):
return tf.image.random_brightness(image, max_delta=0.2), label
if INVERT:
inverted = ds.map(invert)
ds = ds.concatenate(inverted)
if BRIGHTNESS:
ds = ds.concatenate(ds.map(brightness))
ds = ds.cache().shuffle(buffer_size=50000).batch(flags_obj.batch_size)
# Iterate through the dataset a set number (`epochs_between_evals`) of times
# during each training session.
ds = ds.repeat(flags_obj.epochs_between_evals)
return ds
def eval_input_fn():
return dataset.test(flags_obj.data_dir).batch(
flags_obj.batch_size).make_one_shot_iterator().get_next()
# Set up hook that outputs training logs every 100 steps.
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks,
batch_size=flags_obj.batch_size)
# Train and evaluate model.
for _ in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
mnist_classifier.train(input_fn=train_input_fn, hooks=train_hooks)
eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
print('\nEvaluation results:\n\t%s\n' % eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
image = tf.placeholder(tf.float32, [None, 28, 28])
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'image': image,
})
mnist_classifier.export_savedmodel(flags_obj.export_dir, input_fn)
def our_test_fn():
images = []
for i in list(range(1,10)) + ['dog']:
images.append(np.array(imageio.imread('{}.png'.format(i)).ravel()/255.0, dtype='float32'))
images = np.array(images)
return tf.convert_to_tensor(images)
# Check our own examples
predictions = mnist_classifier.predict(input_fn=our_test_fn)
table = []
for i in list(range(1, 10)) + ['dog']:
prediction = next(predictions)
if i == 'dog':
print("{}. CNN thinks it's a {} ({:.1f}%)".format(i, prediction['classes'], prediction['probabilities'][prediction['classes']]*100))
else:
print("{} at {:.1f}. CNN thinks it's a {} ({:.1f}%)".format(i, prediction['probabilities'][i]*100, prediction['classes'], prediction['probabilities'][prediction['classes']]*100))
table.append((i, prediction['probabilities']))
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
"""
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
if flags_core.get_num_gpus(flags_obj) == 0:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:CPU:0')
elif flags_core.get_num_gpus(flags_obj) == 1:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:GPU:0')
else:
distribution = tf.contrib.distribute.MirroredStrategy(
num_gpus=flags_core.get_num_gpus(flags_obj))
run_config = tf.estimator.RunConfig(train_distribute=distribution,
session_config=session_config)
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj)
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + "-synthetic"
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
batch_size=flags_obj.batch_size)
def input_fn_train():
return input_function(is_training=True,
data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size(
flags_obj.batch_size,
flags_core.get_num_gpus(flags_obj)),
num_epochs=flags_obj.epochs_between_evals,
num_gpus=flags_core.get_num_gpus(flags_obj))
def input_fn_eval():
return input_function(is_training=False,
data_dir=flags_obj.data_dir,
batch_size=per_device_batch_size(
flags_obj.batch_size,
flags_core.get_num_gpus(flags_obj)),
num_epochs=1)
total_training_cycle = (flags_obj.train_epochs //
flags_obj.epochs_between_evals)
for cycle_index in range(total_training_cycle):
tf.logging.info('Starting a training cycle: %d/%d', cycle_index,
total_training_cycle)
classifier.train(input_fn=input_fn_train,
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data, which
# will iterate forever. Passing steps=flags_obj.max_train_steps allows the
# eval (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size)
classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn)
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
num_gpus = flags_core.get_num_gpus(FLAGS)
batch_size = distribution_utils.per_device_batch_size(
int(FLAGS.batch_size), num_gpus)
eval_batch_size = int(FLAGS.eval_batch_size or FLAGS.batch_size)
ncf_dataset = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset, data_dir=FLAGS.data_dir,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_neg=FLAGS.num_neg,
epochs_per_cycle=FLAGS.epochs_between_evals,
match_mlperf=FLAGS.ml_perf)
model_helpers.apply_clean(flags.FLAGS)
train_estimator, eval_estimator = construct_estimator(
num_gpus=num_gpus, model_dir=FLAGS.model_dir, params={
"batch_size": batch_size,
"learning_rate": FLAGS.learning_rate,
"num_users": ncf_dataset.num_users,
"num_items": ncf_dataset.num_items,
"mf_dim": FLAGS.num_factors,
"model_layers": [int(layer) for layer in FLAGS.layers],
"mf_regularization": FLAGS.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in FLAGS.mlp_regularization],
"use_tpu": FLAGS.tpu is not None,
"tpu": FLAGS.tpu,
"tpu_zone": FLAGS.tpu_zone,
"tpu_gcp_project": FLAGS.tpu_gcp_project,
}, batch_size=flags.FLAGS.batch_size, eval_batch_size=eval_batch_size)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size # for ExamplesPerSecondHook
)
run_params = {
"batch_size": FLAGS.batch_size,
"eval_batch_size": eval_batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(
model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
approx_train_steps = int(ncf_dataset.num_train_positives
* (1 + FLAGS.num_neg) // FLAGS.batch_size)
pred_input_fn = data_preprocessing.make_pred_input_fn(ncf_dataset=ncf_dataset)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
for cycle_index in range(total_training_cycle):
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
# Train the model
train_input_fn, train_record_dir, batch_count = \
data_preprocessing.make_train_input_fn(ncf_dataset=ncf_dataset)
if np.abs(approx_train_steps - batch_count) > 1:
tf.logging.warning(
"Estimated ({}) and reported ({}) number of batches differ by more "
"than one".format(approx_train_steps, batch_count))
train_estimator.train(input_fn=train_input_fn, hooks=train_hooks,
steps=batch_count)
tf.gfile.DeleteRecursively(train_record_dir)
# Evaluate the model
eval_results = evaluate_model(
eval_estimator, ncf_dataset, pred_input_fn)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
hr = eval_results[_HR_KEY]
ndcg = eval_results[_NDCG_KEY]
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# Some of the NumPy vector math can be quite large and likes to stay in
# memory for a while.
gc.collect()
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
break
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing and not FLAGS.use_synthetic_data:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
if FLAGS.seed is not None:
np.random.seed(FLAGS.seed)
num_gpus = flags_core.get_num_gpus(FLAGS)
batch_size = distribution_utils.per_device_batch_size(
int(FLAGS.batch_size), num_gpus)
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
eval_per_user = rconst.NUM_EVAL_NEGATIVES + 1
eval_batch_size = int(FLAGS.eval_batch_size or
max([FLAGS.batch_size, eval_per_user]))
if eval_batch_size % eval_per_user:
eval_batch_size = eval_batch_size // eval_per_user * eval_per_user
tf.logging.warning(
"eval examples per user does not evenly divide eval_batch_size. "
"Overriding to {}".format(eval_batch_size))
if FLAGS.use_synthetic_data:
ncf_dataset = None
cleanup_fn = lambda: None
num_users, num_items = data_preprocessing.DATASET_TO_NUM_USERS_AND_ITEMS[
FLAGS.dataset]
num_train_steps = data_preprocessing.SYNTHETIC_BATCHES_PER_EPOCH
num_eval_steps = data_preprocessing.SYNTHETIC_BATCHES_PER_EPOCH
else:
ncf_dataset, cleanup_fn = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset, data_dir=FLAGS.data_dir,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_neg=FLAGS.num_neg,
epochs_per_cycle=FLAGS.epochs_between_evals,
num_cycles=total_training_cycle,
match_mlperf=FLAGS.ml_perf,
deterministic=FLAGS.seed is not None,
use_subprocess=FLAGS.use_subprocess,
cache_id=FLAGS.cache_id)
num_users = ncf_dataset.num_users
num_items = ncf_dataset.num_items
num_train_steps = int(np.ceil(
FLAGS.epochs_between_evals * ncf_dataset.num_train_positives *
(1 + FLAGS.num_neg) / FLAGS.batch_size))
num_eval_steps = int(np.ceil((1 + rconst.NUM_EVAL_NEGATIVES) *
ncf_dataset.num_users / eval_batch_size))
model_helpers.apply_clean(flags.FLAGS)
params = {
"use_seed": FLAGS.seed is not None,
"hash_pipeline": FLAGS.hash_pipeline,
"batch_size": batch_size,
"eval_batch_size": eval_batch_size,
"learning_rate": FLAGS.learning_rate,
"num_users": num_users,
"num_items": num_items,
"mf_dim": FLAGS.num_factors,
"model_layers": [int(layer) for layer in FLAGS.layers],
"mf_regularization": FLAGS.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in FLAGS.mlp_regularization],
"num_neg": FLAGS.num_neg,
"use_tpu": FLAGS.tpu is not None,
"tpu": FLAGS.tpu,
"tpu_zone": FLAGS.tpu_zone,
"tpu_gcp_project": FLAGS.tpu_gcp_project,
"beta1": FLAGS.beta1,
"beta2": FLAGS.beta2,
"epsilon": FLAGS.epsilon,
"match_mlperf": FLAGS.ml_perf,
"use_xla_for_gpu": FLAGS.use_xla_for_gpu,
"use_estimator": FLAGS.use_estimator,
}
if FLAGS.use_estimator:
train_estimator, eval_estimator = construct_estimator(
num_gpus=num_gpus, model_dir=FLAGS.model_dir,
iterations=num_train_steps, params=params,
batch_size=flags.FLAGS.batch_size, eval_batch_size=eval_batch_size)
else:
runner = model_runner.NcfModelRunner(ncf_dataset, params, num_train_steps,
num_eval_steps, FLAGS.use_while_loop)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size, # for ExamplesPerSecondHook
tensors_to_log={"cross_entropy": "cross_entropy"}
)
run_params = {
"batch_size": FLAGS.batch_size,
"eval_batch_size": eval_batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(
model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
eval_input_fn = None
target_reached = False
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_LOOP)
for cycle_index in range(total_training_cycle):
assert FLAGS.epochs_between_evals == 1 or not mlperf_helper.LOGGER.enabled
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_EPOCH,
value=cycle_index)
# Train the model
if FLAGS.use_estimator:
train_input_fn, train_record_dir, batch_count = \
data_preprocessing.make_input_fn(
ncf_dataset=ncf_dataset, is_training=True)
if batch_count != num_train_steps:
raise ValueError(
"Step counts do not match. ({} vs. {}) The async process is "
"producing incorrect shards.".format(batch_count, num_train_steps))
train_estimator.train(input_fn=train_input_fn, hooks=train_hooks,
steps=num_train_steps)
if train_record_dir:
tf.gfile.DeleteRecursively(train_record_dir)
tf.logging.info("Beginning evaluation.")
if eval_input_fn is None:
eval_input_fn, _, eval_batch_count = data_preprocessing.make_input_fn(
ncf_dataset=ncf_dataset, is_training=False)
if eval_batch_count != num_eval_steps:
raise ValueError(
"Step counts do not match. ({} vs. {}) The async process is "
"producing incorrect shards.".format(
eval_batch_count, num_eval_steps))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_START,
value=cycle_index)
eval_results = eval_estimator.evaluate(eval_input_fn,
steps=num_eval_steps)
tf.logging.info("Evaluation complete.")
else:
runner.train()
tf.logging.info("Beginning evaluation.")
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_START,
value=cycle_index)
eval_results = runner.eval()
tf.logging.info("Evaluation complete.")
hr = float(eval_results[rconst.HR_KEY])
ndcg = float(eval_results[rconst.NDCG_KEY])
mlperf_helper.ncf_print(
key=mlperf_helper.TAGS.EVAL_TARGET,
value={"epoch": cycle_index, "value": FLAGS.hr_threshold})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_ACCURACY,
value={"epoch": cycle_index, "value": hr})
mlperf_helper.ncf_print(
key=mlperf_helper.TAGS.EVAL_HP_NUM_NEG,
value={"epoch": cycle_index, "value": rconst.NUM_EVAL_NEGATIVES})
# Logged by the async process during record creation.
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_HP_NUM_USERS,
deferred=True)
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_STOP, value=cycle_index)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
tf.logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
target_reached = True
break
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_STOP,
value={"success": target_reached})
cleanup_fn() # Cleanup data construction artifacts and subprocess.
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_FINAL)
def run_loop(
estimator, schedule_manager, train_hooks=None, benchmark_logger=None,
bleu_source=None, bleu_ref=None, bleu_threshold=None, vocab_file=None):
"""Train and evaluate model, and optionally compute model's BLEU score.
**Step vs. Epoch vs. Iteration**
Steps and epochs are canonical terms used in TensorFlow and general machine
learning. They are used to describe running a single process (train/eval):
- Step refers to running the process through a single or batch of examples.
- Epoch refers to running the process through an entire dataset.
E.g. training a dataset with 100 examples. The dataset is
divided into 20 batches with 5 examples per batch. A single training step
trains the model on one batch. After 20 training steps, the model will have
trained on every batch in the dataset, or, in other words, one epoch.
Meanwhile, iteration is used in this implementation to describe running
multiple processes (training and eval).
- A single iteration:
1. trains the model for a specific number of steps or epochs.
2. evaluates the model.
3. (if source and ref files are provided) compute BLEU score.
This function runs through multiple train+eval+bleu iterations.
Args:
estimator: tf.Estimator containing model to train.
schedule_manager: A schedule.Manager object to guide the run loop.
train_hooks: List of hooks to pass to the estimator during training.
benchmark_logger: a BenchmarkLogger object that logs evaluation data
bleu_source: File containing text to be translated for BLEU calculation.
bleu_ref: File containing reference translations for BLEU calculation.
bleu_threshold: minimum BLEU score before training is stopped.
vocab_file: Path to vocab file that will be used to subtokenize bleu_source.
Raises:
ValueError: if both or none of single_iteration_train_steps and
single_iteration_train_epochs were defined.
NotFoundError: if the vocab file or bleu files don't exist.
"""
if bleu_source:
_validate_file(bleu_source)
if bleu_ref:
_validate_file(bleu_ref)
if vocab_file:
_validate_file(vocab_file)
evaluate_bleu = bleu_source is not None and bleu_ref is not None
if evaluate_bleu and schedule_manager.use_tpu:
raise ValueError("BLEU score can not be computed when training with a TPU, "
"as it requires estimator.predict which is not yet "
"supported.")
# Print details of training schedule.
tf.logging.info("Training schedule:")
tf.logging.info(
"\t1. Train for {}".format(schedule_manager.train_increment_str))
tf.logging.info("\t2. Evaluate model.")
if evaluate_bleu:
tf.logging.info("\t3. Compute BLEU score.")
if bleu_threshold is not None:
tf.logging.info("Repeat above steps until the BLEU score reaches %f" %
bleu_threshold)
if not evaluate_bleu or bleu_threshold is None:
tf.logging.info("Repeat above steps %d times." %
schedule_manager.train_eval_iterations)
if evaluate_bleu:
# Create summary writer to log bleu score (values can be displayed in
# Tensorboard).
bleu_writer = tf.summary.FileWriter(
os.path.join(estimator.model_dir, BLEU_DIR))
if bleu_threshold is not None:
# Change loop stopping condition if bleu_threshold is defined.
schedule_manager.train_eval_iterations = INF
# Loop training/evaluation/bleu cycles
for i in xrange(schedule_manager.train_eval_iterations):
tf.logging.info("Starting iteration %d" % (i + 1))
# Train the model for single_iteration_train_steps or until the input fn
# runs out of examples (if single_iteration_train_steps is None).
estimator.train(
dataset.train_input_fn,
steps=schedule_manager.single_iteration_train_steps,
hooks=train_hooks)
eval_results = estimator.evaluate(
input_fn=dataset.eval_input_fn,
steps=schedule_manager.single_iteration_eval_steps)
tf.logging.info("Evaluation results (iter %d/%d):" %
(i + 1, schedule_manager.train_eval_iterations))
tf.logging.info(eval_results)
benchmark_logger.log_evaluation_result(eval_results)
# The results from estimator.evaluate() are measured on an approximate
# translation, which utilize the target golden values provided. The actual
# bleu score must be computed using the estimator.predict() path, which
# outputs translations that are not based on golden values. The translations
# are compared to reference file to get the actual bleu score.
if evaluate_bleu:
uncased_score, cased_score = evaluate_and_log_bleu(
estimator, bleu_source, bleu_ref, vocab_file)
# Write actual bleu scores using summary writer and benchmark logger
global_step = get_global_step(estimator)
summary = tf.Summary(value=[
tf.Summary.Value(tag="bleu/uncased", simple_value=uncased_score),
tf.Summary.Value(tag="bleu/cased", simple_value=cased_score),
])
bleu_writer.add_summary(summary, global_step)
bleu_writer.flush()
benchmark_logger.log_metric(
"bleu_uncased", uncased_score, global_step=global_step)
benchmark_logger.log_metric(
"bleu_cased", cased_score, global_step=global_step)
# Stop training if bleu stopping threshold is met.
if model_helpers.past_stop_threshold(bleu_threshold, uncased_score):
bleu_writer.close()
break
def run_mnist(flags_obj):
"""Run MNIST training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
model_function = model_fn
if flags_obj.multi_gpu:
validate_batch_size_for_multi_gpu(flags_obj.batch_size)
# There are two steps required if using multi-GPU: (1) wrap the model_fn,
# and (2) wrap the optimizer. The first happens here, and (2) happens
# in the model_fn itself when the optimizer is defined.
model_function = tf.contrib.estimator.replicate_model_fn(
model_fn, loss_reduction=tf.losses.Reduction.MEAN)
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
mnist_classifier = tf.estimator.Estimator(model_fn=model_function,
model_dir=flags_obj.model_dir,
params={
'data_format': data_format,
'multi_gpu':
flags_obj.multi_gpu
})
# Set up training and evaluation input functions.
def train_input_fn():
"""Prepare data for training."""
# When choosing shuffle buffer sizes, larger sizes result in better
# randomness, while smaller sizes use less memory. MNIST is a small
# enough dataset that we can easily shuffle the full epoch.
ds = dataset.train(flags_obj.data_dir)
ds = ds.cache().shuffle(buffer_size=50000).batch(flags_obj.batch_size)
# Iterate through the dataset a set number (`epochs_between_evals`) of times
# during each training session.
ds = ds.repeat(flags_obj.epochs_between_evals)
return ds
def eval_input_fn():
return dataset.test(flags_obj.data_dir).batch(
flags_obj.batch_size).make_one_shot_iterator().get_next()
# Set up hook that outputs training logs every 100 steps.
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
batch_size=flags_obj.batch_size)
# Train and evaluate model.
for _ in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
mnist_classifier.train(input_fn=train_input_fn, hooks=train_hooks)
eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
# print('\nEvaluation results:\n\t%s\n' % eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
image = tf.placeholder(tf.float32, [None, 28, 28])
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'image':
image,
})
mnist_classifier.export_savedmodel(flags_obj.export_dir, input_fn)
def run_ncf(_):
"""Run NCF training and eval loop."""
if FLAGS.download_if_missing and not FLAGS.use_synthetic_data:
movielens.download(FLAGS.dataset, FLAGS.data_dir)
if FLAGS.seed is not None:
np.random.seed(FLAGS.seed)
num_gpus = flags_core.get_num_gpus(FLAGS)
batch_size = distribution_utils.per_device_batch_size(
int(FLAGS.batch_size), num_gpus)
eval_per_user = rconst.NUM_EVAL_NEGATIVES + 1
eval_batch_size = int(FLAGS.eval_batch_size
or max([FLAGS.batch_size, eval_per_user]))
if eval_batch_size % eval_per_user:
eval_batch_size = eval_batch_size // eval_per_user * eval_per_user
tf.logging.warning(
"eval examples per user does not evenly divide eval_batch_size. "
"Overriding to {}".format(eval_batch_size))
if FLAGS.use_synthetic_data:
ncf_dataset = None
cleanup_fn = lambda: None
num_users, num_items = data_preprocessing.DATASET_TO_NUM_USERS_AND_ITEMS[
FLAGS.dataset]
num_train_steps = data_preprocessing.SYNTHETIC_BATCHES_PER_EPOCH
num_eval_steps = data_preprocessing.SYNTHETIC_BATCHES_PER_EPOCH
else:
ncf_dataset, cleanup_fn = data_preprocessing.instantiate_pipeline(
dataset=FLAGS.dataset,
data_dir=FLAGS.data_dir,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_neg=FLAGS.num_neg,
epochs_per_cycle=FLAGS.epochs_between_evals,
match_mlperf=FLAGS.ml_perf,
deterministic=FLAGS.seed is not None,
use_subprocess=FLAGS.use_subprocess,
cache_id=FLAGS.cache_id)
num_users = ncf_dataset.num_users
num_items = ncf_dataset.num_items
num_train_steps = int(
np.ceil(FLAGS.epochs_between_evals *
ncf_dataset.num_train_positives * (1 + FLAGS.num_neg) /
FLAGS.batch_size))
num_eval_steps = int(
np.ceil((1 + rconst.NUM_EVAL_NEGATIVES) * ncf_dataset.num_users /
eval_batch_size))
model_helpers.apply_clean(flags.FLAGS)
train_estimator, eval_estimator = construct_estimator(
num_gpus=num_gpus,
model_dir=FLAGS.model_dir,
params={
"use_seed": FLAGS.seed is not None,
"hash_pipeline": FLAGS.hash_pipeline,
"batch_size": batch_size,
"eval_batch_size": eval_batch_size,
"learning_rate": FLAGS.learning_rate,
"num_users": num_users,
"num_items": num_items,
"mf_dim": FLAGS.num_factors,
"model_layers": [int(layer) for layer in FLAGS.layers],
"mf_regularization": FLAGS.mf_regularization,
"mlp_reg_layers": [float(reg) for reg in FLAGS.mlp_regularization],
"num_neg": FLAGS.num_neg,
"use_tpu": FLAGS.tpu is not None,
"tpu": FLAGS.tpu,
"tpu_zone": FLAGS.tpu_zone,
"tpu_gcp_project": FLAGS.tpu_gcp_project,
"beta1": FLAGS.beta1,
"beta2": FLAGS.beta2,
"epsilon": FLAGS.epsilon,
"match_mlperf": FLAGS.ml_perf,
"use_xla_for_gpu": FLAGS.use_xla_for_gpu,
},
batch_size=flags.FLAGS.batch_size,
eval_batch_size=eval_batch_size)
# Create hooks that log information about the training and metric values
train_hooks = hooks_helper.get_train_hooks(
FLAGS.hooks,
model_dir=FLAGS.model_dir,
batch_size=FLAGS.batch_size, # for ExamplesPerSecondHook
tensors_to_log={"cross_entropy": "cross_entropy"})
run_params = {
"batch_size": FLAGS.batch_size,
"eval_batch_size": eval_batch_size,
"number_factors": FLAGS.num_factors,
"hr_threshold": FLAGS.hr_threshold,
"train_epochs": FLAGS.train_epochs,
}
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info(model_name="recommendation",
dataset_name=FLAGS.dataset,
run_params=run_params,
test_id=FLAGS.benchmark_test_id)
pred_input_fn = None
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
target_reached = False
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_LOOP)
for cycle_index in range(total_training_cycle):
assert FLAGS.epochs_between_evals == 1 or not mlperf_helper.LOGGER.enabled
tf.logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_EPOCH,
value=cycle_index)
# Train the model
train_input_fn, train_record_dir, batch_count = \
data_preprocessing.make_input_fn(
ncf_dataset=ncf_dataset, is_training=True)
if batch_count != num_train_steps:
raise ValueError(
"Step counts do not match. ({} vs. {}) The async process is "
"producing incorrect shards.".format(batch_count,
num_train_steps))
train_estimator.train(input_fn=train_input_fn,
hooks=train_hooks,
steps=num_train_steps)
if train_record_dir:
tf.gfile.DeleteRecursively(train_record_dir)
tf.logging.info("Beginning evaluation.")
if pred_input_fn is None:
pred_input_fn, _, eval_batch_count = data_preprocessing.make_input_fn(
ncf_dataset=ncf_dataset, is_training=False)
if eval_batch_count != num_eval_steps:
raise ValueError(
"Step counts do not match. ({} vs. {}) The async process is "
"producing incorrect shards.".format(
eval_batch_count, num_eval_steps))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_START,
value=cycle_index)
eval_results = eval_estimator.evaluate(pred_input_fn,
steps=num_eval_steps)
hr = float(eval_results[rconst.HR_KEY])
ndcg = float(eval_results[rconst.NDCG_KEY])
tf.logging.info("Evaluation complete.")
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_TARGET,
value={
"epoch": cycle_index,
"value": FLAGS.hr_threshold
})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_ACCURACY,
value={
"epoch": cycle_index,
"value": hr
})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_HP_NUM_NEG,
value={
"epoch": cycle_index,
"value": rconst.NUM_EVAL_NEGATIVES
})
# Logged by the async process during record creation.
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_HP_NUM_USERS,
deferred=True)
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_STOP,
value=cycle_index)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
tf.logging.info("Iteration {}: HR = {:.4f}, NDCG = {:.4f}".format(
cycle_index + 1, hr, ndcg))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
target_reached = True
break
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_STOP,
value={"success": target_reached})
cleanup_fn() # Cleanup data construction artifacts and subprocess.
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_FINAL)
def run_mnist(flags_obj):
"""Run MNIST training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
model_helpers.apply_clean(flags_obj)
model_function = model_fn
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
run_config = tf.estimator.RunConfig(
train_distribute=distribution_strategy, session_config=session_config)
data_format = flags_obj.data_format
if data_format is None:
data_format = ('channels_first'
if tf.test.is_built_with_cuda() else 'channels_last')
mnist_classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'data_format': data_format,
})
# Set up training and evaluation input functions.
def train_input_fn():
"""Prepare data for training."""
# When choosing shuffle buffer sizes, larger sizes result in better
# randomness, while smaller sizes use less memory. MNIST is a small
# enough dataset that we can easily shuffle the full epoch.
ds = dataset.train(flags_obj.data_dir)
ds = ds.cache().shuffle(buffer_size=50000).batch(flags_obj.batch_size)
# Iterate through the dataset a set number (`epochs_between_evals`) of times
# during each training session.
ds = ds.repeat(flags_obj.epochs_between_evals)
return ds
def eval_input_fn():
return dataset.test(flags_obj.data_dir).batch(
flags_obj.batch_size).make_one_shot_iterator().get_next()
# Set up hook that outputs training logs every 100 steps.
train_hooks = hooks_helper.get_train_hooks(
flags_obj.hooks, model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
# Train and evaluate model.
for _ in range(flags_obj.train_epochs // flags_obj.epochs_between_evals):
mnist_classifier.train(input_fn=train_input_fn, hooks=train_hooks)
eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
print('\nEvaluation results:\n\t%s\n' % eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
# Export the model
if flags_obj.export_dir is not None:
image = tf.placeholder(tf.float32, [None, 28, 28])
input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn({
'image': image,
})
mnist_classifier.export_savedmodel(flags_obj.export_dir, input_fn)
def run_ncf(_):
"""Run NCF training and eval loop."""
params = ncf_common.parse_flags(FLAGS)
num_users, num_items, num_train_steps, num_eval_steps, producer = (
ncf_common.get_inputs(params))
params["num_users"], params["num_items"] = num_users, num_items
producer.start()
model_helpers.apply_clean(flags.FLAGS)
estimator = construct_estimator(model_dir=FLAGS.model_dir, params=params)
benchmark_logger, train_hooks = log_and_get_hooks(
params["eval_batch_size"])
total_training_cycle = FLAGS.train_epochs // FLAGS.epochs_between_evals
target_reached = False
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_LOOP)
for cycle_index in range(total_training_cycle):
assert FLAGS.epochs_between_evals == 1 or not mlperf_helper.LOGGER.enabled
logging.info("Starting a training cycle: {}/{}".format(
cycle_index + 1, total_training_cycle))
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.TRAIN_EPOCH,
value=cycle_index)
train_input_fn = producer.make_input_fn(is_training=True)
estimator.train(input_fn=train_input_fn,
hooks=train_hooks,
steps=num_train_steps)
logging.info("Beginning evaluation.")
eval_input_fn = producer.make_input_fn(is_training=False)
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_START,
value=cycle_index)
eval_results = estimator.evaluate(eval_input_fn, steps=num_eval_steps)
logging.info("Evaluation complete.")
hr = float(eval_results[rconst.HR_KEY])
ndcg = float(eval_results[rconst.NDCG_KEY])
loss = float(eval_results["loss"])
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_TARGET,
value={
"epoch": cycle_index,
"value": FLAGS.hr_threshold
})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_ACCURACY,
value={
"epoch": cycle_index,
"value": hr
})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_HP_NUM_NEG,
value={
"epoch": cycle_index,
"value": rconst.NUM_EVAL_NEGATIVES
})
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.EVAL_STOP,
value=cycle_index)
# Benchmark the evaluation results
benchmark_logger.log_evaluation_result(eval_results)
# Log the HR and NDCG results.
logging.info(
"Iteration {}: HR = {:.4f}, NDCG = {:.4f}, Loss = {:.4f}".format(
cycle_index + 1, hr, ndcg, loss))
# If some evaluation threshold is met
if model_helpers.past_stop_threshold(FLAGS.hr_threshold, hr):
target_reached = True
break
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_STOP,
value={"success": target_reached})
producer.stop_loop()
producer.join()
# Clear the session explicitly to avoid session delete error
tf.keras.backend.clear_session()
mlperf_helper.ncf_print(key=mlperf_helper.TAGS.RUN_FINAL)
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None):
"""Shared main loop for ResNet Models.
Args:
flags_obj: An object containing parsed flags. See define_resnet_flags()
for details.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
dataset_name: the name of the dataset for training and evaluation. This is
used for logging purpose.
shape: list of ints representing the shape of the images used for training.
This is only used if flags_obj.export_dir is passed.
Dict of results of the run. Contains the keys `eval_results` and
`train_hooks`. `eval_results` contains accuracy (top_1) and accuracy_top_5.
`train_hooks` is a list the instances of hooks used during training.
"""
model_helpers.apply_clean(flags.FLAGS)
# Ensures flag override logic is only executed if explicitly triggered.
if flags_obj.tf_gpu_thread_mode:
override_flags_and_set_envars_for_gpu_thread_pool(flags_obj)
# Configures cluster spec for distribution strategy.
num_workers = distribution_utils.configure_cluster(flags_obj.worker_hosts,
flags_obj.task_index)
# Creates session config. allow_soft_placement = True, is required for
# multi-GPU and is not harmful for other modes.
session_config = tf.compat.v1.ConfigProto(
inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
allow_soft_placement=True)
distribution_strategy = distribution_utils.get_distribution_strategy(
distribution_strategy=flags_obj.distribution_strategy,
num_gpus=flags_core.get_num_gpus(flags_obj),
num_workers=num_workers,
all_reduce_alg=flags_obj.all_reduce_alg,
num_packs=flags_obj.num_packs)
# Creates a `RunConfig` that checkpoints every 24 hours which essentially
# results in checkpoints determined only by `epochs_between_evals`.
run_config = tf.estimator.RunConfig(train_distribute=distribution_strategy,
session_config=session_config,
save_checkpoints_secs=60 * 60 * 24,
save_checkpoints_steps=None)
# Initializes model with all but the dense layer from pretrained ResNet.
if flags_obj.pretrained_model_checkpoint_path is not None:
warm_start_settings = tf.estimator.WarmStartSettings(
flags_obj.pretrained_model_checkpoint_path,
vars_to_warm_start='^(?!.*dense)')
else:
warm_start_settings = None
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
warm_start_from=warm_start_settings,
params={
'resnet_size': int(flags_obj.resnet_size),
'data_format': flags_obj.data_format,
'batch_size': flags_obj.batch_size,
'resnet_version': int(flags_obj.resnet_version),
'loss_scale': flags_core.get_loss_scale(flags_obj),
'dtype': flags_core.get_tf_dtype(flags_obj),
'fine_tune': flags_obj.fine_tune,
'num_workers': num_workers,
})
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
'num_workers': num_workers,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs, input_context=None):
return input_function(
is_training=True,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_replica_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=num_epochs,
dtype=flags_core.get_tf_dtype(flags_obj),
datasets_num_private_threads=flags_obj.
datasets_num_private_threads,
num_parallel_batches=flags_obj.datasets_num_parallel_batches,
input_context=input_context)
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags_obj.data_dir,
batch_size=distribution_utils.per_replica_batch_size(
flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
num_epochs=1,
dtype=flags_core.get_tf_dtype(flags_obj))
train_epochs = (0 if flags_obj.eval_only or not flags_obj.train_epochs else
flags_obj.train_epochs)
use_train_and_evaluate = flags_obj.use_train_and_evaluate or num_workers > 1
if use_train_and_evaluate:
train_spec = tf.estimator.TrainSpec(
input_fn=lambda input_context=None: input_fn_train(
train_epochs, input_context=input_context),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
eval_spec = tf.estimator.EvalSpec(input_fn=input_fn_eval)
tf.compat.v1.logging.info('Starting to train and evaluate.')
tf.estimator.train_and_evaluate(classifier, train_spec, eval_spec)
# tf.estimator.train_and_evalute doesn't return anything in multi-worker
# case.
return {}
else:
if train_epochs == 0:
# If --eval_only is set, perform a single loop with zero train epochs.
schedule, n_loops = [0], 1
else:
# Compute the number of times to loop while training. All but the last
# pass will train for `epochs_between_evals` epochs, while the last will
# train for the number needed to reach `training_epochs`. For instance if
# train_epochs = 25 and epochs_between_evals = 10
# schedule will be set to [10, 10, 5]. That is to say, the loop will:
# Train for 10 epochs and then evaluate.
# Train for another 10 epochs and then evaluate.
# Train for a final 5 epochs (to reach 25 epochs) and then evaluate.
n_loops = math.ceil(train_epochs / flags_obj.epochs_between_evals)
schedule = [
flags_obj.epochs_between_evals for _ in range(int(n_loops))
]
schedule[-1] = train_epochs - sum(schedule[:-1]) # over counting.
for cycle_index, num_train_epochs in enumerate(schedule):
tf.compat.v1.logging.info('Starting cycle: %d/%d', cycle_index,
int(n_loops))
if num_train_epochs:
# Since we are calling classifier.train immediately in each loop, the
# value of num_train_epochs in the lambda function will not be changed
# before it is used. So it is safe to ignore the pylint error here
# pylint: disable=cell-var-from-loop
classifier.train(
input_fn=lambda input_context=None: input_fn_train(
num_train_epochs, input_context=input_context),
hooks=train_hooks,
max_steps=flags_obj.max_train_steps)
# flags_obj.max_train_steps is generally associated with testing and
# profiling. As a result it is frequently called with synthetic data,
# which will iterate forever. Passing steps=flags_obj.max_train_steps
# allows the eval (which is generally unimportant in those circumstances)
# to terminate. Note that eval will run for max_train_steps each loop,
# regardless of the global_step count.
tf.compat.v1.logging.info('Starting to evaluate.')
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags_obj.max_train_steps)
benchmark_logger.log_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(flags_obj.stop_threshold,
eval_results['accuracy']):
break
if flags_obj.export_dir is not None:
# Exports a saved model for the given classifier.
export_dtype = flags_core.get_tf_dtype(flags_obj)
if flags_obj.image_bytes_as_serving_input:
input_receiver_fn = functools.partial(image_bytes_serving_input_fn,
shape,
dtype=export_dtype)
else:
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags_obj.batch_size, dtype=export_dtype)
classifier.export_savedmodel(flags_obj.export_dir,
input_receiver_fn,
strip_default_attrs=True)
stats = {}
stats['eval_results'] = eval_results
stats['train_hooks'] = train_hooks
return stats
def resnet_main(seed, flags, model_function, input_function, shape=None):
"""Shared main loop for ResNet Models.
Args:
flags: FLAGS object that contains the params for running. See
ResnetArgParser for created flags.
model_function: the function that instantiates the Model and builds the
ops for train/eval. This will be passed directly into the estimator.
input_function: the function that processes the dataset and returns a
dataset that the estimator can train on. This will be wrapped with
all the relevant flags for running and passed to estimator.
shape: list of ints representing the shape of the images used for training.
This is only used if flags.export_dir is passed.
"""
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
session_config = tf.ConfigProto(
inter_op_parallelism_threads=flags.inter_op_parallelism_threads,
intra_op_parallelism_threads=flags.intra_op_parallelism_threads,
allow_soft_placement=True)
if flags.num_gpus == 0:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:CPU:0')
elif flags.num_gpus == 1:
distribution = tf.contrib.distribute.OneDeviceStrategy('device:GPU:0')
else:
distribution = tf.contrib.distribute.MirroredStrategy(
num_gpus=flags.num_gpus
)
run_config = tf.estimator.RunConfig(train_distribute=distribution,
session_config=session_config, tf_random_seed=seed)
classifier = tf.estimator.Estimator(
model_fn=model_function, model_dir=flags.model_dir, config=run_config,
params={
'resnet_size': flags.resnet_size,
'data_format': flags.data_format,
'batch_size': flags.batch_size,
'version': flags.version,
'loss_scale': flags.loss_scale,
'dtype': flags.dtype
})
if flags.benchmark_log_dir is not None:
benchmark_logger = logger.BenchmarkLogger(flags.benchmark_log_dir)
benchmark_logger.log_run_info('resnet')
else:
benchmark_logger = None
for _ in range(flags.train_epochs // flags.epochs_between_evals):
train_hooks = hooks_helper.get_train_hooks(
flags.hooks,
batch_size=flags.batch_size,
benchmark_log_dir=flags.benchmark_log_dir)
print('Starting a training cycle.')
def input_fn_train():
return input_function(
is_training=True,
data_dir=flags.data_dir,
batch_size=per_device_batch_size(flags.batch_size, flags.num_gpus),
num_epochs=flags.epochs_between_evals,
)
classifier.train(input_fn=input_fn_train, hooks=train_hooks,
max_steps=flags.max_train_steps)
print('Starting to evaluate.')
# Evaluate the model and print results
def input_fn_eval():
return input_function(
is_training=False,
data_dir=flags.data_dir,
batch_size=per_device_batch_size(flags.batch_size, flags.num_gpus),
num_epochs=1,
)
# flags.max_train_steps is generally associated with testing and profiling.
# As a result it is frequently called with synthetic data, which will
# iterate forever. Passing steps=flags.max_train_steps allows the eval
# (which is generally unimportant in those circumstances) to terminate.
# Note that eval will run for max_train_steps each loop, regardless of the
# global_step count.
eval_results = classifier.evaluate(input_fn=input_fn_eval,
steps=flags.max_train_steps)
print(eval_results)
if benchmark_logger:
benchmark_logger.log_estimator_evaluation_result(eval_results)
if model_helpers.past_stop_threshold(
flags.stop_threshold, eval_results['accuracy']):
break
if flags.export_dir is not None:
# Exports a saved model for the given classifier.
input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
shape, batch_size=flags.batch_size)
classifier.export_savedmodel(flags.export_dir, input_receiver_fn)