def testTrainReplicated(self):
if ipu_utils.running_on_ipu_model():
self.skipTest(
"Replicated top level graphs are not supported on the "
"IPU_MODEL target")
def my_model_fn(features, labels, mode): # pylint: disable=unused-argument
self.assertEqual(model_fn_lib.ModeKeys.TRAIN, mode)
loss = ipu.ops.cross_replica_ops.cross_replica_sum(features,
name="loss")
train_op = array_ops.identity(loss)
return model_fn_lib.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
def my_input_fn():
dataset = tu.create_dual_increasing_dataset(10,
data_shape=[1],
label_shape=[1])
dataset = dataset.batch(batch_size=1, drop_remainder=True)
return dataset
ipu_options = ipu_utils.create_ipu_config()
ipu_options = ipu_utils.auto_select_ipus(ipu_options, 4)
config = ipu_run_config.RunConfig(
ipu_run_config=ipu_run_config.IPURunConfig(
iterations_per_loop=2, num_replicas=4,
ipu_options=ipu_options),
log_step_count_steps=1,
save_summary_steps=1)
estimator = ipu_estimator.IPUEstimator(model_fn=my_model_fn,
config=config)
session_run_counter = _SessionRunCounter()
num_steps = 6
estimator.train(input_fn=my_input_fn,
steps=num_steps,
hooks=[session_run_counter])
self.assertEqual(
session_run_counter.num_session_runs,
num_steps // config.ipu_run_config.iterations_per_loop)
model_dir = estimator.model_dir
events_file = glob.glob(model_dir + "/*tfevents*")
assert len(events_file) == 1
events_file = events_file[0]
loss_output = list()
for e in summary_iterator.summary_iterator(events_file):
for v in e.summary.value:
if "loss" in v.tag:
loss_output.append(v.simple_value)
# loss is averaged across iterations per loop
self.assertEqual(loss_output, [14.0, 16.0, 18.0])
def testTrainWithAutomaticSharding(self):
if ipu_utils.running_on_ipu_model():
self.skipTest(
"Replicated top level graphs are not supported on the "
"IPU_MODEL target")
def my_model_fn(features, labels, mode):
self.assertEqual(model_fn_lib.ModeKeys.TRAIN, mode)
with variable_scope.variable_scope("vs", use_resource=True):
predictions = layers.Dense(units=1)(features)
loss = losses.mean_squared_error(labels=labels,
predictions=predictions)
sharded_optimizer_obj = sharded_optimizer.ShardedOptimizer(
gradient_descent.GradientDescentOptimizer(0.1))
train_op = sharded_optimizer_obj.minimize(loss)
return model_fn_lib.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
def my_input_fn():
dataset = dataset_ops.Dataset.from_tensor_slices(
_create_regression_dataset(num_samples=1000, num_features=5))
dataset = dataset.batch(batch_size=2, drop_remainder=True).repeat()
return dataset
ipu_options = ipu_utils.create_ipu_config()
ipu_options = ipu_utils.auto_select_ipus(ipu_options, 4)
config = ipu_run_config.RunConfig(
ipu_run_config=ipu_run_config.IPURunConfig(
iterations_per_loop=2,
num_shards=4,
autosharding=True,
ipu_options=ipu_options),
log_step_count_steps=1,
save_summary_steps=1)
estimator = ipu_estimator.IPUEstimator(model_fn=my_model_fn,
config=config)
estimator.train(input_fn=my_input_fn, steps=10)
model_dir = estimator.model_dir
events_file = glob.glob(model_dir + "/*tfevents*")
assert len(events_file) == 1
events_file = events_file[0]
loss_output = list()
for e in summary_iterator.summary_iterator(events_file):
for v in e.summary.value:
if "loss" in v.tag:
loss_output.append(v.simple_value)
self.assertTrue(loss_output[0] > loss_output[-1])
def testReplicatedEvaluationOnHost(self):
if ipu_utils.running_on_ipu_model():
self.skipTest(
"Replicated top level graphs are not supported on the "
"IPU_MODEL target")
def my_input_fn():
features = [0, 0, 0, 1, 0, 0, 0, 1]
labels = [0, 1, 0, 1, 0, 1, 0, 1]
return dataset_ops.Dataset.from_tensor_slices(
(features, labels)).batch(2, drop_remainder=True)
def my_metrics_fn(features, labels):
labels64 = math_ops.cast(labels, np.int64)
return {
"accuracy": metrics_impl.accuracy(labels, features),
"precision": metrics_impl.precision(labels, features),
"recall": metrics_impl.recall(labels, features),
"recall_at_1": metrics_impl.recall_at_k(labels64,
features,
k=1),
"recall_at_2": metrics_impl.recall_at_k(labels64,
features,
k=2),
"mse": metrics_impl.mean_squared_error(labels, features),
"rmse": metrics_impl.root_mean_squared_error(labels, features),
}
def my_model_fn(features, labels, mode):
loss = math_ops.cast(replication_ops.replication_index(),
np.float32)
eval_metrics = (my_metrics_fn, [features, labels])
return ipu_estimator.IPUEstimatorSpec(mode,
loss=loss,
eval_metrics=eval_metrics)
ipu_options = ipu_utils.create_ipu_config()
ipu_options = ipu_utils.auto_select_ipus(ipu_options, num_ipus=4)
config = ipu_run_config.RunConfig(
ipu_run_config=ipu_run_config.IPURunConfig(
iterations_per_loop=1, num_replicas=4,
ipu_options=ipu_options))
estimator = ipu_estimator.IPUEstimator(model_fn=my_model_fn,
config=config)
scores = estimator.evaluate(my_input_fn, steps=1)
self.assertEqual(0.75, scores["accuracy"])
self.assertEqual(1.0, scores["precision"])
self.assertEqual(0.5, scores["recall"])
self.assertEqual(0.5, scores["recall_at_1"])
self.assertEqual(1.0, scores["recall_at_2"])
self.assertEqual(0.25, scores["mse"])
self.assertEqual(0.5, scores["rmse"])
self.assertEqual(1.5, scores[model_fn_lib.LOSS_METRIC_KEY])
def testReplicatedPrediction(self):
if ipu_utils.running_on_ipu_model():
self.skipTest(
"Replicated top level graphs are not supported on the "
"IPU_MODEL target")
def my_input_fn():
features = [
[1.0], # IPU0
[3.0], # IPU0
[5.0], # IPU1
[3.0], # IPU1
[7.0], # IPU2
[3.0], # IPU2
[9.0], # IPU3
[3.0], # IPU3
]
return dataset_ops.Dataset.from_tensor_slices(features).batch(
batch_size=2, drop_remainder=True)
hook = ipu_session_run_hooks.IPULoggingTensorHook(every_n_iter=1,
replication_factor=4)
def my_model_fn(features, mode):
logging_op = hook.log({"features": features})
with ops.control_dependencies([logging_op]):
predictions = math_ops.reduce_max(features)
return model_fn_lib.EstimatorSpec(
mode,
predictions=predictions,
)
ipu_options = ipu_utils.create_ipu_config()
ipu_options = ipu_utils.auto_select_ipus(ipu_options, num_ipus=4)
config = ipu_run_config.RunConfig(
ipu_run_config=ipu_run_config.IPURunConfig(
iterations_per_loop=1, num_replicas=4,
ipu_options=ipu_options))
estimator = ipu_estimator.IPUEstimator(model_fn=my_model_fn,
config=config)
outputs = estimator.predict(input_fn=my_input_fn,
yield_single_examples=True)
self.assertEqual(3.0, next(outputs))
self.assertEqual(5.0, next(outputs))
outputs = estimator.predict(input_fn=my_input_fn,
yield_single_examples=False,
hooks=[hook])
np.testing.assert_array_equal([3.0, 5.0, 7.0, 9.0], next(outputs))
def testReplicatedEvaluation(self):
if ipu_utils.running_on_ipu_model():
self.skipTest(
"Replicated top level graphs are not supported on the "
"IPU_MODEL target")
def my_input_fn():
# IPU0 mean: 2, max: 3
# IPU1 mean: 4, max: 5
features = [
[1.0], # IPU0
[3.0], # IPU0
[5.0], # IPU1
[3.0], # IPU1
[1.0], # IPU2
[3.0], # IPU2
[5.0], # IPU3
[3.0], # IPU3
]
return dataset_ops.Dataset.from_tensor_slices(features).batch(
batch_size=2, drop_remainder=True)
def my_model_fn(features, mode):
loss = math_ops.reduce_max(features)
eval_metric_ops = {
"feature_mean": metrics_impl.mean(features),
}
return model_fn_lib.EstimatorSpec(mode,
loss=loss,
eval_metric_ops=eval_metric_ops)
ipu_options = ipu_utils.create_ipu_config()
ipu_options = ipu_utils.auto_select_ipus(ipu_options, num_ipus=4)
config = ipu_run_config.RunConfig(
ipu_run_config=ipu_run_config.IPURunConfig(
iterations_per_loop=1, num_replicas=4,
ipu_options=ipu_options))
estimator = ipu_estimator.IPUEstimator(model_fn=my_model_fn,
config=config)
scores = estimator.evaluate(my_input_fn, steps=1)
self.assertEqual(3., scores["feature_mean"])
self.assertEqual(4., scores[model_fn_lib.LOSS_METRIC_KEY])
def testReplicatedTrainingWithoutCrossReplicaSumShouldThrow(self):
def my_input_fn():
return dataset_ops.Dataset.from_tensor_slices([])
def my_model_fn(features, mode):
loss = math_ops.reduce_sum(features)
train_op = array_ops.identity(loss)
return model_fn_lib.EstimatorSpec(mode,
loss=loss,
train_op=train_op)
ipu_options = ipu_utils.create_ipu_config()
ipu_options = ipu_utils.auto_select_ipus(ipu_options, num_ipus=4)
config = ipu_run_config.RunConfig(
ipu_run_config=ipu_run_config.IPURunConfig(
iterations_per_loop=1, num_replicas=4,
ipu_options=ipu_options))
estimator = ipu_estimator.IPUEstimator(model_fn=my_model_fn,
config=config)
with self.assertRaisesRegex(
ValueError, "This is not a valid replicated training graph"):
estimator.train(input_fn=my_input_fn, steps=1)