def get_possible_initializer_keys(cls, use_bias=True):
return {"w_dw", "w_pw", "b"} if use_bias else {"w_dw", "w_pw"}
def _build(self, inputs):
Connects the module into the graph, with input Tensor `inputs`.
Args:
inputs: A 4D Tensor of shape:
[batch_size, input_height, input_width, input_channels]
and of type `tf.float16` or `tf.float32`.
Returns:
A 4D Tensor of shape:
[batch_size, output_height, output_width, output_channels]
with the same dtype as `inputs`.
Raises:
ValueError: If connecting the module into the graph any time after the
first time and the inferred input size does not match previous
invocations.
base.IncompatibleShapeError: If the input tensor has the wrong number
of dimensions.
base.UnderspecifiedError: If the channel dimension of `inputs` isn"t
defined.
ValueError: If `channel_multiplier` * `input_channels` >
`output_channels`, which means that the separable convolution is
overparameterized.
TypeError: If input Tensor dtype is not compatible with either
`tf.float16` or `tf.float32`.
_verify_inputs(inputs, self._channel_index, self._data_format)
self._input_shape = tuple(inputs.get_shape().as_list())
self._input_channels = self._input_shape[self._channel_index]
depthwise_weight_shape = (self._kernel_shape[0], self._kernel_shape[1],
self._input_channels, self._channel_multiplier)
pointwise_input_size = self._channel_multiplier * self._input_channels
pointwise_weight_shape = (1, 1, pointwise_input_size, self._output_channels)
if "w_dw" not in self._initializers:
fan_in_shape = depthwise_weight_shape[:2]
self._initializers["w_dw"] = create_weight_initializer(fan_in_shape,
dtype=inputs.dtype)
if "w_pw" not in self._initializers:
fan_in_shape = pointwise_weight_shape[:3]
self._initializers["w_pw"] = create_weight_initializer(fan_in_shape,
dtype=inputs.dtype)
self._w_dw = tf.get_variable(
"w_dw",
shape=depthwise_weight_shape,
dtype=inputs.dtype,
initializer=self._initializers["w_dw"],
partitioner=self._partitioners.get("w_dw", None),
regularizer=self._regularizers.get("w_dw", None))
self._w_pw = tf.get_variable(
"w_pw",
shape=pointwise_weight_shape,
dtype=inputs.dtype,
initializer=self._initializers["w_pw"],
partitioner=self._partitioners.get("w_pw", None),
regularizer=self._regularizers.get("w_pw", None))
outputs = tf.nn.separable_conv2d(inputs,
self._w_dw,
self._w_pw,
strides=self._stride,
padding=self._padding,
data_format=self._data_format)
if self._use_bias:
self._b, outputs = _apply_bias(
inputs, outputs, self._channel_index, self._data_format,
self._output_channels, self._initializers, self._partitioners,
self._regularizers)
return outputs
@property
