def test_direct_parquet_train(self):
with spark_session('test_direct_parquet_train') as spark:
df = create_noisy_xor_data(spark)
backend = CallbackBackend()
with local_store() as store:
store.get_train_data_path = lambda v=None: store._train_path
store.get_val_data_path = lambda v=None: store._val_path
with util.prepare_data(backend.num_processes(),
store,
df,
feature_columns=['features'],
label_columns=['y'],
validation=0.2):
model = create_xor_model()
for inmemory_cache_all in [False, True]:
est = hvd_spark.TorchEstimator(
backend=backend,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=1,
epochs=3,
verbose=2,
inmemory_cache_all=inmemory_cache_all)
transformer = est.fit_on_parquet()
predictions = transformer.transform(df)
assert predictions.count() == df.count()
def test_fit_model(self):
model = create_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = F.binary_cross_entropy
with spark_session('test_fit_model') as spark:
df = create_xor_data(spark)
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
input_shapes=[[2]],
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
random_seed=1,
verbose=2,
sample_weight_col='weight')
torch_model = torch_estimator.fit(df)
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_fit_model(self):
if sys.version_info < (3, 0, 0) and is_gloo_used():
self.skipTest(
'Horovod on Spark over Gloo only supported on Python3')
model = create_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = F.binary_cross_entropy
with spark_session('test_fit_model') as spark:
df = create_xor_data(spark)
with local_store() as store:
torch_estimator = hvd.TorchEstimator(
num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
input_shapes=[[2]],
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2,
sample_weight_col='weight')
torch_model = torch_estimator.fit(df)
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_torch_direct_parquet_train(self):
with spark_session('test_torch_direct_parquet_train') as spark:
df = create_xor_data(spark)
backend = CallbackBackend()
with local_store() as store:
store.get_train_data_path = lambda v=None: store._train_path
store.get_val_data_path = lambda v=None: store._val_path
with util.prepare_data(backend.num_processes(),
store,
df,
feature_columns=['features'],
label_columns=['y']):
model = create_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = nn.BCELoss()
est = hvd_spark.TorchEstimator(backend=backend,
store=store,
model=model,
optimizer=optimizer,
input_shapes=[[2]],
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2)
# To make sure that setLoss works with non-list loss.
est.setLoss(loss)
transformer = est.fit_on_parquet()
predictions = transformer.transform(df)
assert predictions.count() == df.count()
def test_fit_model(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
model = create_xor_model()
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
random_seed=1,
verbose=2)
torch_model = torch_estimator.fit(df)
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_fit_model(self):
model = create_xor_model()
optimizer = tf.keras.optimizers.SGD(lr=0.1)
loss = 'binary_crossentropy'
with spark_session('test_fit_model') as spark:
df = create_xor_data(spark)
with local_store() as store:
keras_estimator = hvd.KerasEstimator(
num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2)
keras_model = keras_estimator.fit(df)
trained_model = keras_model.getModel()
pred = trained_model.predict([np.ones([1, 2], dtype=np.float32)])
assert len(pred) == 1
assert pred.dtype == np.float32
def test_train_with_pytorch_infinite_async_data_loader(self):
from horovod.spark.data_loaders.pytorch_data_loaders import PytorchInfiniteAsyncDataLoader
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
model = create_xor_model()
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
data_loader_class=PytorchInfiniteAsyncDataLoader)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_fit_model(self):
if sys.version_info < (3, 0, 0) and is_gloo_used():
self.skipTest(
'Horovod on Spark over Gloo only supported on Python3')
model = create_xor_model()
optimizer = tf.keras.optimizers.SGD(lr=0.1)
loss = 'binary_crossentropy'
with spark_session('test_fit_model') as spark:
df = create_xor_data(spark)
with local_store() as store:
keras_estimator = hvd.KerasEstimator(num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2)
keras_model = keras_estimator.fit(df)
trained_model = keras_model.getModel()
pred = trained_model.predict(
[np.ones([1, 2], dtype=np.float32)])
assert len(pred) == 1
assert pred.dtype == np.float32
def test_model_override_trainer_args(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
model = create_xor_model()
with tempdir() as dir:
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
trainer_args={'stochastic_weight_avg': True})
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_train_with_inmemory_cache_all(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
model = create_xor_model()
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=
1, # Normally inmem dataloader is for single worker training with small data
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
inmemory_cache_all=True)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_keras_direct_parquet_train(self, mock_fit_fn, mock_pin_gpu_fn):
mock_fit_fn.return_value = get_mock_fit_fn()
mock_pin_gpu_fn.return_value = mock.Mock()
with spark_session('test_keras_direct_parquet_train') as spark:
df = create_xor_data(spark)
backend = CallbackBackend()
with local_store() as store:
store.get_train_data_path = lambda v=None: store._train_path
store.get_val_data_path = lambda v=None: store._val_path
with util.prepare_data(backend.num_processes(),
store,
df,
feature_columns=['features'],
label_columns=['y']):
model = create_xor_model()
optimizer = tf.keras.optimizers.SGD(lr=0.1)
loss = 'binary_crossentropy'
est = hvd.KerasEstimator(backend=backend,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2)
transformer = est.fit_on_parquet()
predictions = transformer.transform(df)
assert predictions.count() == df.count()
def test_fit_model_multiclass(self):
model = create_mnist_model()
optimizer = tf.keras.optimizers.Adadelta(1.0)
loss = tf.keras.losses.categorical_crossentropy
for num_cores in [2, constants.TOTAL_BUFFER_MEMORY_CAP_GIB + 1]:
with spark_session('test_fit_model_multiclass', cores=num_cores) as spark:
df = create_mnist_data(spark)
with local_store() as store:
keras_estimator = hvd.KerasEstimator(
num_proc=num_cores,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
metrics=['accuracy'],
feature_cols=['features'],
label_cols=['label_vec'],
batch_size=2,
epochs=2,
verbose=2)
keras_model = keras_estimator.fit(df).setOutputCols(['label_prob'])
pred_df = keras_model.transform(df)
argmax = udf(lambda v: float(np.argmax(v)), returnType=T.DoubleType())
pred_df = pred_df.withColumn('label_pred', argmax(pred_df.label_prob))
preds = pred_df.collect()
assert len(preds) == df.count()
row = preds[0]
label_prob = row.label_prob.toArray().tolist()
assert label_prob[int(row.label_pred)] == max(label_prob)
def test_early_stop_callback(self):
from pytorch_lightning.callbacks.early_stopping import EarlyStopping
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
model = create_xor_model()
early_stop_callback = EarlyStopping(monitor='val_loss',
min_delta=0.00,
patience=3,
verbose=True,
mode='max')
callbacks = [early_stop_callback]
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
callbacks=callbacks)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_model_checkpoint_callback(self):
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
model = create_xor_model()
with tempdir() as dir:
checkpoint_callback = ModelCheckpoint(dirpath=dir)
callbacks = [checkpoint_callback]
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
callbacks=callbacks)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_fit_model(self):
model = create_xor_model()
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2)
torch_model = torch_estimator.fit(df)
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_restore_from_checkpoint(self):
model = create_xor_model()
with spark_session('test_restore_from_checkpoint') as spark:
df = create_noisy_xor_data(spark)
ctx = CallbackBackend()
run_id = 'run01'
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
backend=ctx,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
run_id=run_id)
torch_estimator._read_checkpoint = Mock(
side_effect=torch_estimator._read_checkpoint)
ckpt_path = store.get_checkpoint_path(run_id)
assert not store.exists(ckpt_path)
torch_estimator._read_checkpoint.assert_not_called()
torch_estimator.fit(df)
assert store.exists(ckpt_path)
torch_estimator.fit(df)
torch_estimator._read_checkpoint.assert_called()
def test_legacy_fit_model(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
model = create_legacy_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = F.binary_cross_entropy
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
batch_size=4,
epochs=2,
verbose=2,
sample_weight_col='weight')
torch_model = torch_estimator.fit(df)
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_dummy_callback(self):
from pytorch_lightning.callbacks import Callback
model = create_xor_model()
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
for num_proc in [1, 2]:
for epochs in [2, 3]:
class MyDummyCallback(Callback):
def __init__(self):
self.epcoh_end_counter = 0
self.train_epcoh_end_counter = 0
def on_init_start(self, trainer):
print('Starting to init trainer!')
def on_init_end(self, trainer):
print('Trainer is initialized.')
def on_epoch_end(self, trainer, model):
print('A epoch ended.')
self.epcoh_end_counter += 1
def on_train_epoch_end(self, trainer, model, unused=None):
print('A train epoch ended.')
self.train_epcoh_end_counter += 1
def on_train_end(self, trainer, model):
print('Training ends')
assert self.train_epcoh_end_counter == epochs
dm_callback = MyDummyCallback()
callbacks = [dm_callback]
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=num_proc,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=epochs,
verbose=2,
callbacks=callbacks)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_lr_scheduler_callback(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
from pytorch_lightning.callbacks import LearningRateMonitor
class LRTestingModel(XOR):
def configure_optimizers(self):
optimizer = torch.optim.Adam(model.parameters(), lr=0.02)
def lambda_func(epoch):
return epoch // 30
lr_scheduler = {
'scheduler':
torch.optim.lr_scheduler.LambdaLR(optimizer,
lr_lambda=lambda_func),
'name':
'my_logging_name'
}
return [optimizer], [lr_scheduler]
model = LRTestingModel()
with spark_session('test_fit_model') as spark:
df = create_noisy_xor_data(spark)
lr_monitor = LearningRateMonitor(logging_interval='step')
callbacks = [lr_monitor]
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
callbacks=callbacks)
torch_model = torch_estimator.fit(df)
# TODO: Find a way to pass log metrics from remote, and assert base on the logger.
trained_model = torch_model.getModel()
pred = trained_model(torch.ones([1, 2], dtype=torch.int32))
assert len(pred) == 1
assert pred.dtype == torch.float32
def test_torch_direct_parquet_train(self):
with spark_session('test_torch_direct_parquet_train') as spark:
df = create_xor_data_with_val(spark)
backend = CallbackBackend()
with local_store() as store:
store.get_train_data_path = lambda v=None: store._train_path
store.get_val_data_path = lambda v=None: store._val_path
# Make sure we cover validation dataloader as well
for validation in [None, 'val']:
# Need validation ratio to split data
with util.prepare_data(backend.num_processes(),
store,
df,
feature_columns=['features'],
label_columns=['y'],
validation=validation):
model = create_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = nn.BCELoss()
for inmemory_cache_all in [False, True]:
for reader_pool_type in ['process', 'thread']:
est = hvd_spark.TorchEstimator(
backend=backend,
store=store,
model=model,
optimizer=optimizer,
input_shapes=[[2]],
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2,
reader_pool_type=reader_pool_type,
inmemory_cache_all=inmemory_cache_all,
validation=validation)
# To make sure that setLoss works with non-list loss.
est.setLoss(loss)
transformer = est.fit_on_parquet()
predictions = transformer.transform(df)
assert predictions.count() == df.count()
def test_keras_direct_parquet_train(self, mock_fit_fn, mock_pin_gpu_fn):
mock_fit_fn.return_value = get_mock_fit_fn()
mock_pin_gpu_fn.return_value = mock.Mock()
with spark_session('test_keras_direct_parquet_train') as spark:
df = create_xor_data_with_val(spark)
backend = CallbackBackend()
with local_store() as store:
store.get_train_data_path = lambda v=None: store._train_path
store.get_val_data_path = lambda v=None: store._val_path
# Make sure we cover val dataloader cases
for validation in [None, 'val']:
with util.prepare_data(backend.num_processes(),
store,
df,
feature_columns=['features'],
label_columns=['y'],
validation=validation):
model = create_xor_model()
optimizer = tf.keras.optimizers.SGD(lr=0.1)
loss = 'binary_crossentropy'
for inmemory_cache_all in [False, True]:
for reader_pool_type in ['process', 'thread']:
est = hvd.KerasEstimator(
backend=backend,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
reader_pool_type=reader_pool_type,
validation=validation,
inmemory_cache_all=inmemory_cache_all,
verbose=2)
transformer = est.fit_on_parquet()
predictions = transformer.transform(df)
assert predictions.count() == df.count()
def test_restore_from_checkpoint(self, mock_fit_fn, mock_pin_gpu_fn):
mock_fit_fn.return_value = get_mock_fit_fn()
mock_pin_gpu_fn.return_value = mock.Mock()
model = create_xor_model()
optimizer = tf.keras.optimizers.SGD(lr=0.1)
loss = 'binary_crossentropy'
with spark_session('test_restore_from_checkpoint') as spark:
df = create_xor_data(spark)
backend = CallbackBackend()
run_id = 'run01'
with local_store() as store:
keras_estimator = hvd.KerasEstimator(
backend=backend,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2,
run_id=run_id)
keras_estimator._load_model_from_checkpoint = mock.Mock(
side_effect=keras_estimator._load_model_from_checkpoint)
ckpt_path = store.get_checkpoint_path(run_id)
assert not store.exists(ckpt_path)
keras_estimator._load_model_from_checkpoint.assert_not_called()
keras_model = keras_estimator.fit(df)
trained_model = keras_model.getModel()
pred = trained_model.predict([np.ones([1, 2], dtype=np.float64)])
assert len(pred) == 1
assert store.exists(ckpt_path)
keras_estimator.fit(df)
keras_estimator._load_model_from_checkpoint.assert_called()
def test_direct_parquet_train(self):
if skip_lightning_tests:
self.skipTest(
'Spark PyTorch Lightning tests conflict with Tensorflow 2.5.x: '
'https://github.com/horovod/horovod/pull/3263')
with spark_session('test_direct_parquet_train') as spark:
df = create_noisy_xor_data_with_val(spark)
backend = CallbackBackend()
with local_store() as store:
store.get_train_data_path = lambda v=None: store._train_path
store.get_val_data_path = lambda v=None: store._val_path
# Make sure to cover val dataloader cases
for validation in [None, 'val']:
with util.prepare_data(backend.num_processes(),
store,
df,
feature_columns=['features'],
label_columns=['y'],
validation=validation):
model = create_xor_model()
for inmemory_cache_all in [False, True]:
for reader_pool_type in ['process', 'thread']:
est = hvd_spark.TorchEstimator(
backend=backend,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=validation,
batch_size=1,
epochs=3,
verbose=2,
inmemory_cache_all=inmemory_cache_all,
reader_pool_type=reader_pool_type)
transformer = est.fit_on_parquet()
predictions = transformer.transform(df)
assert predictions.count() == df.count()
def test_legacy_restore_from_checkpoint(self):
self.skipTest(
'There is a bug in current lightning version for checkpoint'
'call back. Will add this test back when it is solved.')
model = create_legacy_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = nn.BCELoss()
with spark_session('test_restore_from_checkpoint') as spark:
df = create_noisy_xor_data(spark)
ctx = CallbackBackend()
run_id = 'run01'
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
backend=ctx,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
run_id=run_id)
torch_estimator._read_checkpoint = mock.Mock(
side_effect=torch_estimator._read_checkpoint)
ckpt_path = store.get_checkpoint_path(run_id)
assert not store.exists(ckpt_path)
torch_estimator._read_checkpoint.assert_not_called()
torch_estimator.fit(df)
assert store.exists(ckpt_path)
torch_estimator.fit(df)
torch_estimator._read_checkpoint.assert_called()
def test_model_serialization(self, mock_remote_trainer):
model = create_xor_model()
optimizer = tf.keras.optimizers.SGD(lr=0.1)
loss = 'binary_crossentropy'
def train(serialized_model, train_rows, val_rows, avg_row_size):
return None, serialized_model, 2
mock_remote_trainer.return_value = train
with spark_session('test_model_serialization') as spark:
df = create_xor_data(spark)
keras_estimator = hvd.KerasEstimator(model=model,
optimizer=optimizer,
loss=loss,
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=3,
verbose=2)
backend = CallbackBackend()
with local_store() as store:
with temppath() as saved_path:
keras_estimator.save(saved_path)
keras_estimator_loaded = hvd.KerasEstimator.load(
saved_path)
keras_model = keras_estimator_loaded.fit(
df,
params={
keras_estimator_loaded.backend: backend,
keras_estimator_loaded.store: store
})
trained_model = keras_model.getModel()
pred = trained_model.predict(
[np.ones([1, 2], dtype=np.float32)])
assert len(pred) == 1
assert pred.dtype == np.float32
def test_terminate_on_nan_flag(self):
model = create_xor_model()
with spark_session('test_terminate_on_nan_flag') as spark:
df = create_noisy_xor_data(spark)
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
num_proc=2,
store=store,
model=model,
input_shapes=[[-1, 2]],
feature_cols=['features'],
label_cols=['y'],
validation=0.2,
batch_size=4,
epochs=2,
verbose=2,
terminate_on_nan=True,
profiler="pytorch")
assert torch_estimator.getTerminateOnNan() == True
def test_restore_from_checkpoint(self):
model = create_xor_model()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
loss = nn.BCELoss()
with spark_session('test_restore_from_checkpoint') as spark:
df = create_xor_data(spark)
ctx = CallbackBackend()
run_id = 'run01'
with local_store() as store:
torch_estimator = hvd_spark.TorchEstimator(
backend=ctx,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
input_shapes=[[2]],
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
epochs=1,
verbose=2,
run_id=run_id)
torch_estimator._load_checkpoint = mock.Mock(
side_effect=torch_estimator._load_checkpoint)
ckpt_path = store.get_checkpoint_path(run_id)
assert not store.exists(ckpt_path)
torch_estimator._load_checkpoint.assert_not_called()
torch_estimator.fit(df)
assert store.exists(ckpt_path)
torch_estimator.fit(df)
torch_estimator._load_checkpoint.assert_called()
def test_fit_model(self):
model = create_xor_model()
optimizer = tf.keras.optimizers.SGD(lr=0.1)
loss = 'binary_crossentropy'
with spark_session('test_fit_model') as spark:
df = create_xor_data(spark)
with local_store() as store:
keras_estimator = hvd.KerasEstimator(num_proc=2,
store=store,
model=model,
optimizer=optimizer,
loss=loss,
feature_cols=['features'],
label_cols=['y'],
batch_size=1,
random_seed=1,
epochs=3,
verbose=2,
use_gpu=False,
mp_start_method='spawn')
assert not keras_estimator.getUseGpu()
assert 'spawn' == keras_estimator.getMpStartMethod()
keras_estimator.setMpStartMethod('forkserver')
assert 'forkserver' == keras_estimator.getMpStartMethod()
keras_model = keras_estimator.fit(df)
trained_model = keras_model.getModel()
pred = trained_model.predict(
[np.ones([1, 2], dtype=np.float32)])
assert len(pred) == 1
assert pred.dtype == np.float32
def test_df_cache(self):
# Clean the cache before starting the test
util.clear_training_cache()
util._training_cache.get_dataset = mock.Mock(
side_effect=util._training_cache.get_dataset)
with spark_session('test_df_cache') as spark:
with local_store() as store:
df = create_xor_data(spark)
df2 = create_xor_data(spark)
df3 = create_xor_data(spark)
key = util._training_cache.create_key(df, store, None)
key2 = util._training_cache.create_key(df2, store, None)
key3 = util._training_cache.create_key(df3, store, None)
# All keys are distinct
assert key != key2
assert key != key3
assert key2 != key3
# The cache should be empty to start
assert not util._training_cache.is_cached(key, store)
assert not util._training_cache.is_cached(key2, store)
assert not util._training_cache.is_cached(key3, store)
# First insertion into the cache
with util.prepare_data(num_processes=2,
store=store,
df=df,
feature_columns=['features'],
label_columns=['y']) as dataset_idx:
train_rows, val_rows, metadata, avg_row_size = util.get_dataset_properties(
dataset_idx)
util._training_cache.get_dataset.assert_not_called()
assert len(util._training_cache._key_to_dataset) == 1
assert util._training_cache.is_cached(key, store)
assert dataset_idx == 0
# The first dataset is still in use, so we assign the next integer in sequence to this
# dataset
assert not util._training_cache.is_cached(key2, store)
with util.prepare_data(num_processes=2,
store=store,
df=df2,
feature_columns=['features'],
label_columns=['y'
]) as dataset_idx2:
util._training_cache.get_dataset.assert_not_called()
assert len(util._training_cache._key_to_dataset) == 2
assert util._training_cache.is_cached(key2, store)
assert dataset_idx2 == 1
# Even though the first dataset is no longer in use, it is still cached
with util.prepare_data(num_processes=2,
store=store,
df=df,
feature_columns=['features'],
label_columns=['y']) as dataset_idx1:
train_rows1, val_rows1, metadata1, avg_row_size1 = util.get_dataset_properties(
dataset_idx1)
util._training_cache.get_dataset.assert_called()
assert train_rows == train_rows1
assert val_rows == val_rows1
assert metadata == metadata1
assert avg_row_size == avg_row_size1
assert dataset_idx1 == 0
# The first dataset is no longer in use, so we can reclaim its dataset index
assert not util._training_cache.is_cached(key3, store)
with util.prepare_data(num_processes=2,
store=store,
df=df3,
feature_columns=['features'],
label_columns=['y']) as dataset_idx3:
train_rows3, val_rows3, metadata3, avg_row_size3 = util.get_dataset_properties(
dataset_idx3)
assert train_rows == train_rows3
assert val_rows == val_rows3
assert metadata == metadata3
assert avg_row_size == avg_row_size3
assert dataset_idx3 == 0
# Same dataframe, different validation
bad_key = util._training_cache.create_key(df, store, 0.1)
assert not util._training_cache.is_cached(bad_key, store)
def test_prepare_data_compress_sparse(self):
util.clear_training_cache()
expected_metadata = \
{
'float': {
'spark_data_type': FloatType,
'is_sparse_vector_only': False,
'intermediate_format': constants.NOCHANGE,
'max_size': 1,
'shape': 1
},
'dense': {
'spark_data_type': DenseVector,
'is_sparse_vector_only': False,
'intermediate_format': constants.ARRAY,
'max_size': 2,
'shape': 2
},
'sparse': {
'spark_data_type': SparseVector,
'is_sparse_vector_only': True,
'intermediate_format': constants.CUSTOM_SPARSE,
'max_size': 1,
'shape': 2
},
'mixed': {
'spark_data_type': DenseVector,
'is_sparse_vector_only': False,
'intermediate_format': constants.ARRAY,
'max_size': 2,
'shape': 2
},
}
with mock.patch('horovod.spark.common.util._get_metadata',
side_effect=util._get_metadata) as mock_get_metadata:
with spark_session('test_prepare_data') as spark:
data = [[
0.0,
DenseVector([1.0, 1.0]),
SparseVector(2, {1: 1.0}),
DenseVector([1.0, 1.0])
],
[
1.0,
DenseVector([1.0, 1.0]),
SparseVector(2, {1: 1.0}),
SparseVector(2, {1: 1.0})
]]
schema = StructType([
StructField('float', FloatType()),
StructField('dense', VectorUDT()),
StructField('sparse', VectorUDT()),
StructField('mixed', VectorUDT())
])
df = create_test_data_from_schema(spark, data, schema)
with local_store() as store:
with util.prepare_data(
num_processes=2,
store=store,
df=df,
feature_columns=['dense', 'sparse', 'mixed'],
label_columns=['float'],
compress_sparse=True) as dataset_idx:
mock_get_metadata.assert_called()
assert dataset_idx == 0
train_rows, val_rows, metadata, avg_row_size = util.get_dataset_properties(
dataset_idx)
self.assertDictEqual(metadata, expected_metadata)