def get_keras_sub_version():
try:
from keras import __version__
type = int(__version__.split('.')[1])
except:
type = 2
return type
def keras_version():
""" Retrieves the Keras major version.
"""
from keras import __version__ # pylint: disable=import-error
return int(__version__.split('.')[0])
def get_keras_sub_version():
from keras import __version__
type = int(__version__.split('.')[1])
return type
from keras import __version__ as keras_version
from keras.models import load_model, model_from_json, model_from_yaml
from keras.layers import Layer
from keras import backend
from dlhub_sdk.models.servables.python import BasePythonServableModel
from dlhub_sdk.utils.types import compose_argument_block
_keras_version_tuple = tuple(int(i) for i in keras_version.split("."))
_summary_limit = 10000
def _detect_backend(output):
"""Add the backend
Args:
output (KerasModel): Current description of Keras model, will be modified
"""
# Determine the name of the object
my_backend = backend.backend().lower()
# Add it as a requirement
output.add_requirement(my_backend, 'detect')
class KerasModel(BasePythonServableModel):
"""Servable based on a Keras Model object.
Assumes that the model has been saved to an hdf5 file"""
@classmethod
def keras_version():
""" Retrieves the Keras major version.
"""
from keras import __version__ # pylint: disable=import-error
return int(__version__.split('.')[0])
# weight mask constructor functions
#'construct_efn_weight_mask', 'construct_pfn_weight_mask',
# network consstructor functions
#'construct_distributed_dense', 'construct_latent', 'construct_dense',
# full model classes
'EFN', 'PFN'
]
###############################################################################
# Keras 2.2.5 fixes bug in 2.2.4 that affects our usage of the Dot layer
###############################################################################
keras_version_tuple = tuple(map(int, __keras_version__.split('.')))
DOT_AXIS = 0 if keras_version_tuple <= (2, 2, 4) else 1
###############################################################################
# INPUT FUNCTIONS
###############################################################################
def construct_efn_input(input_dim, zs_name=None, phats_name=None):
# construct input tensors
zs_input = Input(batch_shape=(None, None), name=zs_name)
phats_input = Input(batch_shape=(None, None, input_dim), name=phats_name)
return [zs_input, phats_input]
def construct_pfn_input(input_dim, name=None):