def _make_unique_name(self, name_uid_map=None, avoid_names=None,
namespace='', zero_based=False):
base_name = base_layer.to_snake_case(self.__class__.__name__)
name = base_layer.unique_layer_name(base_name,
name_uid_map=name_uid_map,
avoid_names=avoid_names,
namespace=namespace,
zero_based=zero_based)
return (name, base_name)
def _make_unique_name(self, name_uid_map=None, avoid_names=None,
namespace='', zero_based=False):
base_name = base_layer.to_snake_case(self.__class__.__name__)
name = base_layer.unique_layer_name(base_name,
name_uid_map=name_uid_map,
avoid_names=avoid_names,
namespace=namespace,
zero_based=zero_based)
return (name, base_name)
def conv2d(inputs,
filters,
kernel_size,
strides,
padding,
name="conv2d",
rate=1,
scope=None,
normalizer_fn=None,
return_preact=False,
**args):
activation = False
if "activation_fn" in args:
activation_fn = args["activation_fn"]
del args["activation_fn"]
activation = True
if "use_bias" in args and args["use_bias"]:
assert not args["use_bias"]
del args["use_bias"]
if "kernel_initializer" in args:
del args["kernel_initializer"]
if "data_format" in args:
args["data_format"] = "NCHW" if args[
"data_format"] == "channels_first" else "NHWC"
else:
args["data_format"] = "NHWC"
# kernel
if not isinstance(kernel_size, list):
kernel_size = [kernel_size] * 2
# data_format dependent kernel_size, strides, dilations
if args["data_format"] == "NHWC":
num_channels = getshape(inputs)[-1]
if not isinstance(strides, list):
strides = [1] + [strides] * 2 + [1]
dilations = [1, rate, rate, 1]
else:
num_channels = getshape(inputs)[0]
if not isinstance(strides, list):
strides = [1, 1] + [strides] * 2
dilations = [1, 1] + [rate, rate]
# get filter
with tf.variable_scope(unique_layer_name(name, zero_based=True)) as vc:
weight_s, weight_p = variableFromSettings(kernel_size +
[num_channels, filters])
# conv with sampled filter
conv_args = {
"input": inputs,
"strides": strides,
"padding": padding,
"dilations": dilations,
**args
}
conv = tf.nn.conv2d(filter=weight_s, name=name, **conv_args)
if normalizer_fn is not None:
conv = normalizer_fn(conv, name="bn")
# check if an activation needs some work beforehand
if activation and activation_fn is not None:
if "activation_prepare" in GLOBAL:
preact = GLOBAL["activation_prepare"](weight_s, weight_p, inputs,
conv, conv_args)
if return_preact:
return preact
conv = activation_fn(conv)
# just return that
return conv