def test_linear(self):
'''test linear'''
inputs = tf.random_uniform(shape=[4, 5], dtype=tf.float32,
maxval=1.0) # A 2D tensor
shape = [5, 4]
output = cl.linear(inputs, 'test_linear0', shape)
output_shape = [4, 4]
self.assertAllEqual(tf.shape(output), output_shape)
inputs = tf.random_uniform(shape=[2, 4, 5],
dtype=tf.float32,
maxval=1.0) # A 3D tensor
shape = [5, 4]
output = cl.linear(inputs, 'test_linear1', shape)
output_shape = [2, 4, 4]
self.assertAllEqual(tf.shape(output), output_shape)
# A 4D tensor [B, C, H, W]
inputs = tf.random_uniform(shape=[2, 3, 4, 5],
dtype=tf.float32,
maxval=1.0)
shape = [5, 4]
output = cl.linear(inputs, 'test_linear2', shape)
output_shape = [2, 3, 4, 4]
self.assertAllEqual(tf.shape(output), output_shape)
def test_tdnn(self):
'''test tdnn'''
#A 3D Tensor [batch, in_width, in_channels]
inputs = tf.random_uniform(shape=[2, 5, 3],
dtype=tf.float32,
maxval=1.0)
in_dim = inputs.get_shape().as_list()[2]
out_dim = 4
context = [-2, -1, 0, 1, 2]
output = cl.tdnn(inputs,
'test_tdnn0',
in_dim,
context,
out_dim,
method='splice_layer')
out_shape = [2, 5, 4]
self.assertAllEqual(tf.shape(output), out_shape)
context = 2
#output = cl.tdnn(inputs, 'test_tdnn1', in_dim, context, out_dim, method='splice_op')
#self.assertAllEqual(tf.shape(output), out_shape)
output = cl.tdnn(inputs,
'test_tdnn2',
in_dim,
context,
out_dim,
method='conv1d')
self.assertAllEqual(tf.shape(output), out_shape)
def embedding_look_up(text_inputs, vocab_size, embedding_size):
"""Embedding layer."""
with tf.variable_scope("embedding"):
W = tf.get_variable(
name='W',
initializer=tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0))
embedding_chars = tf.nn.embedding_lookup(W, text_inputs)
embedding_chars_expanded = tf.expand_dims(embedding_chars, -1)
return embedding_chars_expanded
def test_conv2d(self):
'''test conv2d'''
inputs = tf.random_uniform(shape=[2, 5, 5, 3],
dtype=tf.float32,
maxval=1.0) #A 4D Tensor
filter_size = [3, 3]
in_channels = inputs.get_shape().as_list()[3]
out_channels = 4
strides = [1, 1]
output = cl.conv2d(inputs, 'test_conv2d', filter_size, in_channels,
out_channels, strides)
output_shape = [2, 5, 5, 4]
self.assertAllEqual(tf.shape(output), output_shape)
def test_conv_pool(self):
'''test conv pool'''
# A 4D tensor [B, H, W, C]
embedded_chars_expanded = tf.random_uniform(shape=[2, 7, 7, 1],
dtype=tf.float32,
maxval=1.0)
filter_sizes = [3, 5]
embedding_size = 3
num_filters = 3
sequence_length = 5
output = cl.conv_pool(embedded_chars_expanded, filter_sizes,
embedding_size, num_filters, sequence_length)
output_shape = [30, 6]
self.assertAllEqual(tf.shape(output), output_shape)
def test_attention(self):
'''test attention'''
# A 3D tensor [B, T, D]
inputs = tf.random_uniform(shape=[2, 100, 512],
dtype=tf.float32,
maxval=1.0)
attention_size = 256
output, alpha = cl.attention(inputs,
attention_size,
return_alphas=True)
output_shape = [2, 512]
alpha_shape = [2, 100, 1]
self.assertAllEqual(tf.shape(output), output_shape)
self.assertAllEqual(tf.shape(alpha), alpha_shape)