23cb8d9355951a3ca48b9b023678832e209d1fb2,sonnet/python/modules/conv_gpu_test.py,Conv3DTransposeTestDataFormats,testConv3DTransposeDataFormats,#Conv3DTransposeTestDataFormats#,568

Before Change

      ("WithBias_Stride2", True, 2), ("WithoutBias_Stride2", False, 2))
  def testConv3DTransposeDataFormats(self, use_bias, stride):
    Check the module produces the same result for supported data formats.
    input_shape = (self.INPUT_SHAPE.input_batch,
                   np.ceil(self.INPUT_SHAPE.input_depth / stride),
                   np.ceil(self.INPUT_SHAPE.input_height / stride),
                   np.ceil(self.INPUT_SHAPE.input_width / stride),
                   self.INPUT_SHAPE.input_channels)

    func = functools.partial(
        snt.Conv3DTranspose,
        output_channels=self.OUT_CHANNELS,
        kernel_shape=self.KERNEL_SHAPE,
        output_shape=(self.INPUT_SHAPE.input_depth,
                      self.INPUT_SHAPE.input_height,
                      self.INPUT_SHAPE.input_width),
        use_bias=use_bias,
        stride=stride,
        initializers=create_initializers(use_bias))

    conv_ndhwc = func(name="NDHWC", data_format="NDHWC")
    x = tf.constant(np.random.random(input_shape).astype(np.float32))
    result_ndhwc = conv_ndhwc(x)

    // We will force both modules to share the same weights by creating
    // a custom getter that returns the weights from the first conv module when
    // tf.get_variable is called.
    custom_getter = {"w": create_custom_field_getter(conv_ndhwc, "w"),
                     "b": create_custom_field_getter(conv_ndhwc, "b")}
    conv_ncdhw = func(name="NCDHW", data_format="NCDHW",
                      custom_getter=custom_getter)
    x_transpose = tf.transpose(x, perm=(0, 4, 1, 2, 3))
    result_ncdhw = tf.transpose(conv_ncdhw(x_transpose), perm=(0, 2, 3, 4, 1))

    self.checkEquality(result_ndhwc, result_ncdhw)

After Change

      ("WithBias_Stride2", True, 2), ("WithoutBias_Stride2", False, 2))
  def testConv3DTransposeDataFormats(self, use_bias, stride):
    Check the module produces the same result for supported data formats.
    input_shape = (self.INPUT_SHAPE.input_batch,
                   int(np.ceil(self.INPUT_SHAPE.input_depth / stride)),
                   int(np.ceil(self.INPUT_SHAPE.input_height / stride)),
                   int(np.ceil(self.INPUT_SHAPE.input_width / stride)),
                   self.INPUT_SHAPE.input_channels)

    func = functools.partial(
        snt.Conv3DTranspose,
        output_channels=self.OUT_CHANNELS,
        kernel_shape=self.KERNEL_SHAPE,
        output_shape=(self.INPUT_SHAPE.input_depth,
                      self.INPUT_SHAPE.input_height,
                      self.INPUT_SHAPE.input_width),
        use_bias=use_bias,
        stride=stride,
        initializers=create_initializers(use_bias))

    conv_ndhwc = func(name="NDHWC", data_format="NDHWC")
    x = tf.constant(np.random.random(input_shape).astype(np.float32))
    result_ndhwc = conv_ndhwc(x)

    // We will force both modules to share the same weights by creating
    // a custom getter that returns the weights from the first conv module when
    // tf.get_variable is called.
    custom_getter = {"w": create_custom_field_getter(conv_ndhwc, "w"),
                     "b": create_custom_field_getter(conv_ndhwc, "b")}
    conv_ncdhw = func(name="NCDHW", data_format="NCDHW",
                      custom_getter=custom_getter)
    x_transpose = tf.transpose(x, perm=(0, 4, 1, 2, 3))
    result_ncdhw = tf.transpose(conv_ncdhw(x_transpose), perm=(0, 2, 3, 4, 1))

    self.checkEquality(result_ndhwc, result_ncdhw)

Instances