def _define_model(
self, model_type, data_dim, num_hidden, kernel_initializer='ones',
use_weight_norm=True, data_init=True):
if use_weight_norm:
base_layer = tfkl.Dense(num_hidden, kernel_initializer=kernel_initializer)
layer = weight_norm.WeightNorm(
base_layer, data_init=data_init, name='maybe_norm_layer')
else:
layer = tfkl.Dense(num_hidden, kernel_initializer=kernel_initializer,
name='maybe_norm_layer')
if model_type == 'layer':
return layer
elif model_type == 'sequential':
return tfk.Sequential(
[tfkl.InputLayer((data_dim,)), layer,
tfkl.Dense(1, kernel_initializer=kernel_initializer)])
elif model_type == 'sequential_no_input':
return tfk.Sequential(
[layer, tfkl.Dense(1, kernel_initializer=kernel_initializer)])
elif model_type == 'functional':
inputs = tfkl.Input(shape=(data_dim,))
net = layer(inputs)
outputs = tfkl.Dense(1, kernel_initializer=kernel_initializer)(net)
return tfk.Model(inputs=inputs, outputs=outputs)
else:
raise ValueError('{} is not a valid model type'.format(model_type))
def testVariableCreationNoBias(self):
conv_layer = tfkl.Conv2D(filters=self.num_conv_filters,
kernel_size=(2, 2),
kernel_initializer='ones',
use_bias=False)
norm_layer = weight_norm.WeightNorm(conv_layer, name='norm_layer')
norm_layer.build(self.conv_random_input.shape)
self.assertLen(norm_layer.trainable_variables, 2)
def testConv2DInitializedCorrectly(self, transpose):
conv = tfkl.Conv2DTranspose if transpose else tfkl.Conv2D
conv_layer = conv(filters=self.num_conv_filters, kernel_size=(2, 2),
kernel_initializer='ones')
norm_layer = weight_norm.WeightNorm(conv_layer, name='norm_layer')
norm_layer.build(self.conv_random_input.shape)
self.evaluate([v.initializer for v in norm_layer.weights])
self.evaluate(norm_layer(self.conv_random_input))
true_init_g, true_init_bias = self._calculate_true_initial_variables_conv(
self.conv_random_input, self.num_conv_filters, transpose=transpose)
self.assertAllClose(true_init_g, self.evaluate(norm_layer.g))
self.assertAllClose(true_init_bias, self.evaluate(norm_layer.layer.bias))
def wrapped_layer(*args, **kwargs):
return weight_norm.WeightNorm(
layer(*args, **kwargs), data_init=use_data_init)
