def _core_depthwise_conv2d(self, **layer_kwargs):
filters_in = 3
input_shape = [2, 6, 6, filters_in] # channels last
if isinstance(layer_kwargs['kernel_size'], int):
kernel_size_in = (layer_kwargs['kernel_size'],) * 2
else:
kernel_size_in = layer_kwargs['kernel_size']
filters_out = layer_kwargs.get('depth_multiplier', 1)
kernel = np.random.normal(kernel_size_in +
(filters_in, filters_out))
initializer = tf.keras.initializers.Constant(kernel)
base_kwargs = {
"strides": 2,
"depthwise_initializer": initializer,
}
kwargs = {**base_kwargs, **layer_kwargs}
agreement_test(tfe.keras.layers.DepthwiseConv2D,
kwargs=kwargs,
input_shape=input_shape,
atol=1e-2)
layer_test(tfe.keras.layers.DepthwiseConv2D,
kwargs=kwargs,
batch_input_shape=input_shape)
def _core_conv2d(self, **layer_kwargs):
filters_in = 3
input_shape = [2, 6, 6, filters_in] # channels last
filters = 5
if isinstance(layer_kwargs['kernel_size'], int):
kernel_size_in = (layer_kwargs['kernel_size'],) * 2
else:
kernel_size_in = layer_kwargs['kernel_size']
kernel = np.random.normal(kernel_size_in +
(filters_in, filters))
initializer = tf.keras.initializers.Constant(kernel)
base_kwargs = {
"filters": filters,
"strides": 2,
"kernel_initializer": initializer,
}
kwargs = {**base_kwargs, **layer_kwargs}
agreement_test(tfe.keras.layers.Conv2D,
kwargs=kwargs,
input_shape=input_shape)
layer_test(tfe.keras.layers.Conv2D,
kwargs=kwargs,
batch_input_shape=input_shape)
def _core_flatten(self, **layer_kwargs):
input_shape = layer_kwargs['input_shape']
agreement_test(tfe.keras.layers.Flatten,
kwargs=layer_kwargs,
input_shape=input_shape,
atol=.1)
def _core_activation(self, **layer_kwargs):
agreement_test(tfe.keras.layers.Activation,
kwargs=layer_kwargs,
input_shape=[1, 5])
layer_test(tfe.keras.layers.Activation,
kwargs=layer_kwargs,
batch_input_shape=[1, 5])
def _core_reshape(self, input_shape, **layer_kwargs):
kwargs = {**layer_kwargs}
agreement_test(tfe.keras.layers.Reshape,
kwargs=kwargs,
input_shape=input_shape)
layer_test(tfe.keras.layers.Reshape,
kwargs=kwargs,
batch_input_shape=input_shape)
def _core_batchnorm(self, input_shape, **layer_kwargs):
base_kwargs = {'fused': False}
kwargs = {**base_kwargs, **layer_kwargs}
agreement_test(tfe.keras.layers.BatchNormalization,
kwargs=kwargs,
input_shape=input_shape)
layer_test(tfe.keras.layers.BatchNormalization,
kwargs=kwargs,
batch_input_shape=input_shape)
def _core_global_avgpooling2d(self, **layer_kwargs):
channel_in = 2
input_shape = [2, 8, 8, channel_in] # channels last
base_kwargs = {}
kwargs = {**base_kwargs, **layer_kwargs}
agreement_test(tfe.keras.layers.GlobalAveragePooling2D,
kwargs=kwargs,
input_shape=input_shape)
layer_test(tfe.keras.layers.GlobalAveragePooling2D,
kwargs=kwargs,
batch_input_shape=input_shape)
def _core_dense(self, **layer_kwargs):
input_shape = [4, 5]
kernel = np.random.normal(input_shape[::-1])
initializer = tf.keras.initializers.Constant(kernel)
base_kwargs = {
"units": 4,
"kernel_initializer": initializer,
}
kwargs = {**base_kwargs, **layer_kwargs}
agreement_test(tfe.keras.layers.Dense,
kwargs=kwargs,
input_shape=input_shape)
def _core_avgpooling2d(self, **layer_kwargs):
channel_in = 2
input_shape = [2, 8, 8, channel_in] # channels last
pool_size_in = 2
base_kwargs = {
"pool_size": pool_size_in,
}
kwargs = {**base_kwargs, **layer_kwargs}
agreement_test(tfe.keras.layers.AveragePooling2D,
kwargs=kwargs,
input_shape=input_shape)
layer_test(tfe.keras.layers.AveragePooling2D,
kwargs=kwargs,
batch_input_shape=input_shape)
def _core_relu(self, **layer_kwargs):
agreement_test(tfe.keras.layers.ReLU,
kwargs=layer_kwargs,
input_shape=[1, 5],
rtol=.1)
