def testEluGradGradWRTinputs(self):
inputs = constant_op.constant([[-2, -1, 1, 3], [5, 7, 8, 9]],
dtype=dtypes.float32)
dummy = constant_op.constant([[3, 1, -1, -2], [9, 8, 7, 6]],
dtype=dtypes.float32)
elu = gen_nn_ops.elu(inputs)
elu_grad = gradients_impl.gradients(elu, inputs, grad_ys=dummy)[0]
with self.cached_session():
error = gradient_checker.compute_gradient_error(
inputs, inputs.shape, elu_grad, elu_grad.shape)
self.assertLess(error, 1e-4)
def dilated_conv1d_layer(x, kernel, num_filters, name, dilation=1, stride=1, activation=True, padding='VALID'):
with tf.name_scope(name) as scope:
kernel = tf.get_variable(f'{name}/weights', shape=[kernel, x.get_shape()[-1].value, num_filters],
initializer=tf.truncated_normal_initializer(stddev=1e-2))
post_out = atrous_conv1d(x, kernel, dilation, padding=padding)
biases = tf.get_variable(f'{name}/biases', shape=[num_filters],
initializer=tf.constant_initializer(0.0))
post_out = tf.nn.bias_add(post_out, biases)
if activation:
post_out = elu(post_out, name=name)
# tf.summary.histogram(f'{name}/kernel', kernel)
# tf.summary.histogram(f'{name}/biases', biases)
return post_out
def testEluGradGradWRTinputs(self):
inputs = constant_op.constant(
[[-2, -1, 1, 3], [5, 7, 8, 9]], dtype=dtypes.float32)
dummy = constant_op.constant(
[[3, 1, -1, -2], [9, 8, 7, 6]], dtype=dtypes.float32)
elu = gen_nn_ops.elu(inputs)
elu_grad = gradients_impl.gradients(elu, inputs, grad_ys=dummy)[0]
with self.cached_session():
error = gradient_checker.compute_gradient_error(
inputs,
inputs.shape,
elu_grad,
elu_grad.shape)
self.assertLess(error, 1e-4)
def conv2d_layer(x, filter, units, name, strides=None, padding='SAME'):
if strides is None:
strides = [1, 1]
with tf.name_scope(name) as scope:
kernel = tf.get_variable(
f'{name}/weights',
shape=[filter[0], filter[1],
x.get_shape()[-1].value, units],
initializer=tf.random_normal_initializer(0, 0.1))
conv = tf.nn.conv2d(x,
filter=kernel,
strides=[1, strides[0], strides[1], 1],
padding=padding)
biases = tf.get_variable('%s/biases' % name,
shape=[units],
initializer=tf.constant_initializer(0.0))
out = tf.nn.bias_add(conv, biases)
post_out = elu(out, name=name)
# post_out = out
tf.summary.histogram(f'{name}/kernel', kernel)
# tf.summary.histogram('%s/biases' % name, biases)
return post_out, kernel, biases
def predict(self, inputTensor, _):
result = gen_nn_ops.elu(inputTensor)
return (result)