def test_reshape(self):
metadata = \
{
'col1': {
'dtype': float,
'intermediate_format': constants.NOCHANGE,
'max_size': 1,
'shape': 1
},
'col2': {
'dtype': SparseVector,
'intermediate_format': constants.CUSTOM_SPARSE,
'max_size': 5,
'shape': 10
},
'label': {
'dtype': float,
'intermediate_format': constants.NOCHANGE,
'max_size': 1,
'shape': 1
},
}
feature_columns = ['col1', 'col2']
label_columns = ['label']
sample_weight_col = 'sample_weight'
Row = collections.namedtuple(
'row', ['col1', 'col2', 'sample_weight', 'label'])
col11 = tf.constant([3.])
col21 = tf.constant([3., 1., 3., 6., 10., 30., 60., 0, 0, 0, 0])
label1 = tf.constant([1.])
sw1 = tf.constant([.06])
row1 = Row(col1=col11, col2=col21, label=label1, sample_weight=sw1)
reshape_fn = TFKerasUtil._reshape_fn(sample_weight_col,
feature_columns, label_columns,
metadata)
reshaped_row = reshape_fn(row1)
reshaped_row_value = self.evaluate(reshaped_row)
assert np.allclose(reshaped_row_value['sample_weight'],
np.array([0.06]))
assert np.allclose(reshaped_row_value['col1'], np.array([3.]))
assert np.allclose(
reshaped_row_value['col2'],
np.array([[0., 10., 0., 30., 0., 0., 60., 0., 0., 0.]]))
assert np.allclose(reshaped_row_value['label'], np.array([1.]))
def read_serialized_keras_model(self, ckpt_path, model):
"""
Returns serialized keras model. On Databricks, only TFKeras is supported, not BareKeras.
The parameter `model` is for providing the model structure when the checkpoint file only
contains model weights.
"""
import tensorflow
from tensorflow import keras
from horovod.spark.keras.util import TFKerasUtil
if LooseVersion(tensorflow.__version__) < LooseVersion("2.0.0"):
model.load_weights(ckpt_path)
else:
model = keras.models.load_model(ckpt_path)
return TFKerasUtil.serialize_model(model)
def read_serialized_keras_model(self, ckpt_path, model, custom_objects):
"""
Returns serialized keras model.
The parameter `model` is for providing the model structure when the checkpoint file only
contains model weights.
"""
import tensorflow
from tensorflow import keras
from horovod.spark.keras.util import TFKerasUtil
if LooseVersion(tensorflow.__version__) < LooseVersion("2.0.0"):
model.load_weights(ckpt_path)
else:
with keras.utils.custom_object_scope(custom_objects):
model = keras.models.load_model(ckpt_path)
return TFKerasUtil.serialize_model(model)
def test_prep_data_tf_keras_fn_without_sparse_col(self):
has_sparse_col = False
feature_columns = ['col1', 'col2']
label_columns = ['label1', 'label2']
sample_weight_col = 'sample_weight'
col1 = tf.constant([3.])
col2 = tf.constant([float(i) for i in range(10)])
label1 = tf.constant([1., 2., 3., 4.])
label2 = tf.constant([1., 2., 3., 4.])
sw1 = tf.constant([.06])
input_shapes = [[-1, 1], [-1, 2, 5]]
output_shapes = [[-1, 4], [-1, 2, 2]]
output_names = ['label1', 'label2']
prep_data_tf_keras = \
TFKerasUtil._prep_data_fn(has_sparse_col, sample_weight_col,
feature_columns, label_columns, input_shapes,
output_shapes, output_names)
Row = collections.namedtuple(
'row', ['col1', 'col2', sample_weight_col, 'label1', 'label2'])
row = Row(col1=col1,
col2=col2,
label1=label1,
label2=label2,
sample_weight=sw1)
prepped_row = prep_data_tf_keras(row)
prepped_row_vals = self.evaluate(prepped_row)
assert np.array_equal(prepped_row_vals[0][0], np.array([[3.]]))
assert np.array_equal(
prepped_row_vals[0][1],
np.array([[[0., 1., 2., 3., 4.], [5., 6., 7., 8., 9.]]]))
assert np.array_equal(prepped_row_vals[1][0],
np.array([[1., 2., 3., 4.]]))
assert np.array_equal(prepped_row_vals[1][1],
np.array([[[1., 2.], [3., 4.]]]))
assert np.allclose(prepped_row_vals[2]['label1'], np.array([0.06]))
assert np.allclose(prepped_row_vals[2]['label2'], np.array([0.06]))
def test_prep_data_tf_keras_fn_with_sparse_col(self):
has_sparse_col = True
feature_columns = ['col1', 'col2']
label_columns = ['label1', 'label2']
sample_weight_col = 'sample_weight'
col1 = tf.constant([3.])
col2 = tf.constant([3., 1., 3., 6., 10., 30., 60., 0, 0, 0])
label1 = tf.constant([1., 2., 3., 4.])
label2 = tf.constant([1., 2., 3., 4.])
sw1 = tf.constant([.06])
input_shapes = [[-1, 1], [-1, 2, 5]]
output_shapes = [[-1, 4], [-1, 2, 2]]
output_names = ['label1', 'label2']
prep_data_tf_keras = \
TFKerasUtil._prep_data_fn(has_sparse_col, sample_weight_col,
feature_columns, label_columns, input_shapes,
output_shapes, output_names)
row = {
'col1': col1,
'col2': col2,
'label1': label1,
'label2': label2,
sample_weight_col: sw1
}
prepped_row = prep_data_tf_keras(row)
prepped_row_vals = self.evaluate(prepped_row)
assert np.array_equal(prepped_row_vals[0][0], np.array([[3.]]))
assert np.array_equal(
prepped_row_vals[0][1],
np.array([[[3., 1., 3., 6., 10.], [30., 60., 0., 0., 0.]]]))
assert np.array_equal(prepped_row_vals[1][0],
np.array([[1., 2., 3., 4.]]))
assert np.array_equal(prepped_row_vals[1][1],
np.array([[[1., 2.], [3., 4.]]]))
assert np.allclose(prepped_row_vals[2]['label1'], np.array([0.06]))
assert np.allclose(prepped_row_vals[2]['label2'], np.array([0.06]))
def read_serialized_keras_model(self, ckpt_path, model, custom_objects):
"""Reads the checkpoint file of the keras model into model bytes and returns the base 64
encoded model bytes.
:param ckpt_path: A string of path to the checkpoint file.
:param model: A keras model. This parameter will be used in DBFSLocalStore\
.read_serialized_keras_model() when the ckpt_path only contains model weights.
:param custom_objects: This parameter will be used in DBFSLocalStore\
.read_serialized_keras_model() when loading the keras model.
:return: the base 64 encoded model bytes of the checkpoint model.
"""
from horovod.runner.common.util import codec
import tensorflow
from tensorflow import keras
from horovod.spark.keras.util import TFKerasUtil
if LooseVersion(tensorflow.__version__) < LooseVersion("2.0.0"):
model_bytes = self.read(ckpt_path)
return codec.dumps_base64(model_bytes)
else:
with keras.utils.custom_object_scope(custom_objects):
model = keras.models.load_model(ckpt_path)
return TFKerasUtil.serialize_model(model)