class Conv2DSchema(BaseLayerSchema):
filters = fields.Int()
kernel_size = ObjectOrListObject(fields.Int, min=2, max=2)
strides = ObjectOrListObject(fields.Int, min=2, max=2, default=(1, 1), missing=(1, 1))
padding = fields.Str(default='valid', missing='valid',
validate=validate.OneOf(['same', 'valid']))
data_format = fields.Str(default=None, missing=None,
validate=validate.OneOf('channels_first', 'channels_last'))
dilation_rate = ObjectOrListObject(fields.Int, min=2, max=2, default=(1, 1), missing=(1, 1))
activation = StrOrFct(allow_none=True, validate=validate.OneOf(ACTIVATION_VALUES))
use_bias = fields.Bool(default=True, missing=True)
kernel_initializer = fields.Nested(InitializerSchema, allow_none=True)
bias_initializer = fields.Nested(InitializerSchema, allow_none=True)
kernel_regularizer = fields.Nested(RegularizerSchema, allow_none=True)
bias_regularizer = fields.Nested(RegularizerSchema, allow_none=True)
activity_regularizer = fields.Nested(RegularizerSchema, allow_none=True)
kernel_constraint = fields.Nested(ConstraintSchema, allow_none=True)
bias_constraint = fields.Nested(ConstraintSchema, allow_none=True)
class Meta:
ordered = True
@post_load
def make_load(self, data):
return Conv2DConfig(**data)
class ConvRecurrent2DSchema(RecurrentSchema):
filters = fields.Int()
kernel_size = ObjectOrListObject(fields.Int, min=2, max=2)
strides = ObjectOrListObject(fields.Int,
min=2,
max=2,
default=(1, 1),
missing=(1, 1))
padding = fields.Str(default='valid',
missing='valid',
validate=validate.OneOf(['same', 'valid']))
data_format = fields.Str(allow_none=True,
validate=validate.OneOf('channels_first',
'channels_last'))
dilation_rate = ObjectOrListObject(fields.Int,
min=2,
max=2,
default=(1, 1),
missing=(1, 1))
return_sequences = fields.Bool(default=False, missing=False)
go_backwards = fields.Bool(default=False, missing=False)
stateful = fields.Bool(default=False, missing=False)
class Meta:
ordered = True
@post_load
def make_load(self, data):
return ConvRecurrent2DConfig(**data)
class AveragePooling3DSchema(BaseLayerSchema):
pool_size = ObjectOrListObject(fields.Int,
min=3,
max=3,
default=(2, 2, 2),
missing=(2, 2, 2))
strides = ObjectOrListObject(fields.Int,
min=3,
max=3,
default=None,
missing=None)
padding = fields.Str(default='valid',
missing='valid',
validate=validate.OneOf(['same', 'valid']))
data_format = fields.Str(default=None,
missing=None,
validate=validate.OneOf('channels_first',
'channels_last'))
class Meta:
ordered = True
@post_load
def make(self, data):
return AveragePooling3DConfig(**data)
@post_dump
def unmake(self, data):
return AveragePooling3DConfig.remove_reduced_attrs(data)
class MaxPooling2DSchema(BaseLayerSchema):
pool_size = ObjectOrListObject(fields.Int,
min=2,
max=2,
default=(2, 2),
missing=(2, 2))
strides = ObjectOrListObject(fields.Int,
min=2,
max=2,
default=None,
missing=None)
padding = fields.Str(default='valid',
missing='valid',
validate=validate.OneOf(['same', 'valid']))
data_format = fields.Str(default=None,
missing=None,
validate=validate.OneOf('channels_first',
'channels_last'))
class Meta:
ordered = True
@post_load
def make_load(self, data):
return MaxPooling2DConfig(**data)
class GraphSchema(BaseSchema):
input_layers = ObjectOrListObject(Tensor)
output_layers = ObjectOrListObject(Tensor)
layers = fields.Nested(LayerSchema, many=True)
name = fields.Str(allow_none=True)
class Meta:
unknown = EXCLUDE
@staticmethod
def schema_config():
return GraphConfig
class Cropping2DSchema(BaseLayerSchema):
cropping = ObjectOrListObject(ObjectOrListObject(fields.Int, min=2, max=2), min=2, max=2,
default=((0, 0), (0, 0)), missing=((0, 0), (0, 0)))
data_format = fields.Str(default=None, missing=None,
validate=validate.OneOf('channels_first', 'channels_last'))
class Meta:
ordered = True
@post_load
def make_load(self, data):
return Cropping2DConfig(**data)
class GraphSchema(Schema):
input_layers = ObjectOrListObject(Tensor)
output_layers = ObjectOrListObject(Tensor)
layers = fields.Nested(LayerSchema, many=True)
name = fields.Str(allow_none=True)
class Meta:
ordered = True
@post_load
def make(self, data):
return GraphConfig(**data)
class LocallyConnected1DSchema(BaseLayerSchema):
filters = fields.Int()
kernel_size = ObjectOrListObject(fields.Int, min=1, max=1)
strides = ObjectOrListObject(fields.Int,
min=1,
max=1,
default=1,
missing=1)
padding = fields.Str(default='valid',
missing='valid',
validate=validate.OneOf(['same', 'valid']))
data_format = fields.Str(default=None,
missing=None,
validate=validate.OneOf('channels_first',
'channels_last'))
activation = StrOrFct(allow_none=True,
validate=validate.OneOf(ACTIVATION_VALUES))
use_bias = fields.Bool(default=True, missing=True)
kernel_initializer = fields.Nested(InitializerSchema,
default=None,
missing=None)
bias_initializer = fields.Nested(InitializerSchema,
default=None,
missing=None)
kernel_regularizer = fields.Nested(RegularizerSchema,
default=None,
missing=None)
bias_regularizer = fields.Nested(RegularizerSchema,
default=None,
missing=None)
activity_regularizer = fields.Nested(RegularizerSchema,
default=None,
missing=None)
kernel_constraint = fields.Nested(RegularizerSchema,
default=None,
missing=None)
bias_constraint = fields.Nested(RegularizerSchema,
default=None,
missing=None)
class Meta:
ordered = True
@post_load
def make(self, data):
return LocallyConnected1DConfig(**data)
@post_dump
def unmake(self, data):
return LocallyConnected1DConfig.remove_reduced_attrs(data)
class DotSchema(BaseLayerSchema):
axes = ObjectOrListObject(fields.Int)
normalize = fields.Bool(allow_none=True)
@staticmethod
def schema_config():
return DotConfig
class EpisodeLoggingTensorHookSchema(BaseSchema):
tensors = ObjectOrListObject(fields.Str)
every_n_episodes = fields.Int()
@staticmethod
def schema_config():
return EpisodeLoggingTensorHookConfig
class BaseLayerSchema(Schema):
name = fields.Str(allow_none=True)
trainable = fields.Bool(default=True, missing=True)
dtype = DType(allow_none=True)
inbound_nodes = ObjectOrListObject(Tensor, allow_none=True)
def get_attribute(self, attr, obj, default):
return get_value(attr, obj, default)
class StepLoggingTensorHookSchema(BaseSchema):
tensors = ObjectOrListObject(fields.Str)
every_n_iter = fields.Int(allow_none=True)
every_n_secs = fields.Int(allow_none=True)
@staticmethod
def schema_config():
return StepLoggingTensorHookConfig
class UpSampling1DSchema(BaseLayerSchema):
size = ObjectOrListObject(fields.Int, min=2, max=2, default=2, missing=2)
class Meta:
ordered = True
@post_load
def make_load(self, data):
return UpSampling1DConfig(**data)
class ZeroPadding1DSchema(BaseLayerSchema):
padding = ObjectOrListObject(fields.Int, min=1, max=1, default=1, missing=1)
class Meta:
ordered = True
@post_load
def make_load(self, data):
return ZeroPadding1DConfig(**data)
class Cropping1DSchema(BaseLayerSchema):
cropping = ObjectOrListObject(fields.Int, min=2, max=2, default=(1, 1), missing=(1, 1))
class Meta:
ordered = True
@post_load
def make_load(self, data):
return Cropping1DConfig(**data)
class FinalOpsHookSchema(Schema):
final_ops = ObjectOrListObject(fields.Str)
class Meta:
ordered = True
@post_load
def make_load(self, data):
return FinalOpsHookConfig(**data)
class LocallyConnected2DSchema(BaseLayerSchema):
filters = fields.Int()
kernel_size = ObjectOrListObject(fields.Int, min=2, max=2)
strides = ObjectOrListObject(fields.Int,
min=2,
max=2,
default=(1, 1),
missing=(1, 1))
padding = fields.Str(default='valid',
missing='valid',
validate=validate.OneOf(['same', 'valid']))
data_format = fields.Str(default=None,
missing=None,
validate=validate.OneOf('channels_first',
'channels_last'))
activation = StrOrFct(allow_none=True,
validate=validate.OneOf(ACTIVATION_VALUES))
use_bias = fields.Bool(default=True, missing=True)
kernel_initializer = fields.Nested(InitializerSchema,
default=None,
missing=None)
bias_initializer = fields.Nested(InitializerSchema,
default=None,
missing=None)
kernel_regularizer = fields.Nested(RegularizerSchema,
default=None,
missing=None)
bias_regularizer = fields.Nested(RegularizerSchema,
default=None,
missing=None)
activity_regularizer = fields.Nested(RegularizerSchema,
default=None,
missing=None)
kernel_constraint = fields.Nested(RegularizerSchema,
default=None,
missing=None)
bias_constraint = fields.Nested(RegularizerSchema,
default=None,
missing=None)
@staticmethod
def schema_config():
return LocallyConnected2DConfig
class EpisodeLoggingTensorHookSchema(Schema):
tensors = ObjectOrListObject(fields.Str)
every_n_episodes = fields.Int()
class Meta:
ordered = True
@post_load
def make_load(self, data):
return EpisodeLoggingTensorHookConfig(**data)
class StepLoggingTensorHookSchema(Schema):
tensors = ObjectOrListObject(fields.Str)
every_n_iter = fields.Int(allow_none=True)
every_n_secs = fields.Int(allow_none=True)
class Meta:
ordered = True
@post_load
def make_load(self, data):
return StepLoggingTensorHookConfig(**data)
class ZeroPadding3DSchema(BaseLayerSchema):
padding = ObjectOrListObject(fields.Int, min=3, max=3, default=(1, 1, 1), missing=(1, 1, 1))
data_format = fields.Str(default=None, missing=None,
validate=validate.OneOf('channels_first', 'channels_last'))
class Meta:
ordered = True
@post_load
def make_load(self, data):
return ZeroPadding3DConfig(**data)
class MaxPooling2DSchema(BaseLayerSchema):
pool_size = ObjectOrListObject(fields.Int,
min=2,
max=2,
default=(2, 2),
missing=(2, 2))
strides = ObjectOrListObject(fields.Int,
min=2,
max=2,
default=None,
missing=None)
padding = fields.Str(default='valid',
missing='valid',
validate=validate.OneOf(['same', 'valid']))
data_format = fields.Str(default=None,
missing=None,
validate=validate.OneOf('channels_first',
'channels_last'))
@staticmethod
def schema_config():
return MaxPooling2DConfig
class BaseModelSchema(BaseSchema):
graph = fields.Nested(GraphSchema)
loss = fields.Nested(LossSchema, allow_none=True)
optimizer = fields.Nested(OptimizerSchema, allow_none=True)
metrics = fields.Nested(MetricSchema, many=True, allow_none=True)
summaries = ObjectOrListObject(fields.Str, allow_none=True)
clip_gradients = fields.Float(allow_none=True)
clip_embed_gradients = fields.Float(allow_none=True)
name = fields.Str(allow_none=True)
@staticmethod
def schema_config():
return BaseModelConfig
class FinalOpsHookSchema(Schema):
final_ops = ObjectOrListObject(fields.Str)
class Meta:
ordered = True
@post_load
def make(self, data):
return FinalOpsHookConfig(**data)
@post_dump
def unmake(self, data):
return FinalOpsHookConfig.remove_reduced_attrs(data)
class DotSchema(BaseLayerSchema):
axes = ObjectOrListObject(fields.Int)
normalize = fields.Bool(allow_none=True)
class Meta:
ordered = True
@post_load
def make(self, data):
return DotConfig(**data)
@post_dump
def unmake(self, data):
return DotConfig.remove_reduced_attrs(data)
class EpisodeLoggingTensorHookSchema(Schema):
tensors = ObjectOrListObject(fields.Str)
every_n_episodes = fields.Int()
class Meta:
ordered = True
@post_load
def make(self, data):
return EpisodeLoggingTensorHookConfig(**data)
@post_dump
def unmake(self, data):
return EpisodeLoggingTensorHookConfig.remove_reduced_attrs(data)
class PReLUSchema(BaseLayerSchema):
alpha_initializer = fields.Nested(InitializerSchema,
default=None,
missing=None)
alpha_regularizer = fields.Nested(RegularizerSchema,
default=None,
missing=None)
alpha_constraint = fields.Nested(ConstraintSchema,
default=None,
missing=None)
shared_axes = ObjectOrListObject(fields.Int, default=None, missing=None)
@staticmethod
def schema_config():
return PReLUConfig
class BaseModelSchema(Schema):
graph = fields.Nested(GraphSchema)
loss = fields.Nested(LossSchema, allow_none=True)
optimizer = fields.Nested(OptimizerSchema, allow_none=True)
metrics = fields.Nested(MetricSchema, many=True, allow_none=True)
summaries = ObjectOrListObject(fields.Str, allow_none=True)
clip_gradients = fields.Float(allow_none=True)
clip_embed_gradients = fields.Float(allow_none=True)
name = fields.Str(allow_none=True)
class Meta:
ordered = True
@post_load
def make_load(self, data):
return BaseModelConfig(**data)
class PReLUSchema(BaseLayerSchema):
alpha_initializer = fields.Nested(InitializerSchema,
default=None,
missing=None)
alpha_regularizer = fields.Nested(RegularizerSchema,
default=None,
missing=None)
alpha_constraint = fields.Nested(ConstraintSchema,
default=None,
missing=None)
shared_axes = ObjectOrListObject(fields.Int, default=None, missing=None)
class Meta:
ordered = True
@post_load
def make_load(self, data):
return PReLUConfig(**data)
class BaseBridgeSchema(Schema):
state_size = ObjectOrListObject(fields.Int, allow_none=True)
name = fields.Str(allow_none=True)
class FinalOpsHookSchema(BaseSchema):
final_ops = ObjectOrListObject(fields.Str)
@staticmethod
def schema_config():
return FinalOpsHookConfig