def dense(params):
imap = initializer_map(params)
return kl.Dense(get_units(params),
dtype=tf.float32,
activation=activation_function(params),
use_bias=True,
kernel_initializer=imap['weights'],
bias_initializer=imap['offset'])
def __init__(self, params, path_names, block_type):
super(BlockLayer, self).__init__()
self._block_name = block_type
self._path_names = path_names
self._paths = []
for name in path_names:
self._paths.append(make_sequence(params[name], SIMPLE_LAYERS))
self._activator = kl.Activation(activation_function(params))
def depthwise_conv_2d(params):
imap = initializer_map(params)
nfilters, kernel_dimensions = get_shape_params(params)
return kl.DepthwiseConv2D(kernel_size=kernel_dimensions,
strides=params['strides'],
dtype=tf.float32,
padding=params['padding'],
use_bias=params.get('bias', None) is not None,
depth_multiplier=params['depth_multiplier'],
activation=activation_function(params),
depthwise_initializer=imap['kernel'],
bias_initializer=imap['bias'])
def conv_2d(params):
imap = initializer_map(params)
nfilters, kernel_dimensions = get_shape_params(params)
return kl.Conv2D(filters=nfilters,
kernel_size=kernel_dimensions,
dtype=tf.float32,
strides=params['strides'],
padding=params['padding'],
use_bias=params.get('bias', None) is not None,
activation=activation_function(params),
kernel_initializer=imap['kernel'],
kernel_regularizer=tf.keras.regularizers.l2(0.0005),
bias_initializer=imap['bias'])
def separable_conv_2d(params):
imap = initializer_map(params)
nfilters, kernel_dimensions = get_shape_params(params)
return kl.SeparableConv2D(
filters=nfilters,
kernel_size=kernel_dimensions,
dtype=tf.float32,
strides=params["strides"],
padding=params["padding"],
use_bias=params.get("bias", None) is not None,
depth_multiplier=params["depth_multiplier"],
activation=activation_function(params),
depthwise_initializer=imap["depth_kernel"],
pointwise_initializer=imap["point_kernel"],
bias_initializer=imap["bias"],
)
def dense_with_weights(params):
imap = {}
for key in ["weights", "offset"]:
if isinstance(params[key], str):
imap[key] = params[key]
else:
imap[key] = tf.constant_initializer(np.array(params[key]))
return kl.Dense(
get_units(params),
dtype=tf.float32,
activation=activation_function(params),
use_bias=True,
kernel_initializer=imap["weights"],
bias_initializer=imap["offset"],
)
def with_popped_activation(params):
afn = activation_function(params)
params_copy = dict(params)
params_copy.pop('activation_function', None)
return params_copy, afn
def __init__(self, params, path_names, block_type):
self._paths = [params[name] for name in path_names]
self._activator = kl.Activation(activation_function(params))
def activation(params):
return kl.Activation(activation_function(params))