def capsules_init(inputs, shape, strides, padding, pose_shape, add_bias,
name): #nets, shape=[1, 1, 32, 16], strides=[1, 1, 1, 1],
# padding='VALID', pose_shape=16, add_bias=True, name='primary')
# nets = capsule_conv_layer(nets, shape=[3, 1, 16, 16], strides=[1, 1, 1, 1],
# iterations=3, name='conv2'
with tf.variable_scope(name):
poses = _conv2d_wrapper(
inputs,
shape=shape[0:-1] + [shape[-1] * pose_shape], # [1, 1, 32, 16*16]
strides=strides,
padding=padding,
add_bias=add_bias,
activation_fn=None,
name='pose_stacked')
poses_shape = poses.get_shape().as_list() #[batch , len-2, 1, 16*16]
poses = tf.reshape(
poses, [-1, poses_shape[1], poses_shape[2], shape[-1], pose_shape
]) # [batch , len-2 , 1 , 16, 16]
beta_a = _get_weights_wrapper(name='beta_a', shape=[1, shape[-1]])
poses = squash_v1(poses, axis=-1) # [batch , len-2 , 1 , 16, 16]
activations = K.sqrt(K.sum(
K.square(poses), axis=-1)) + beta_a # [batch , len-2 , 1 , 16]
tf.logging.info("prim poses dimension:{}".format(poses.get_shape()))
return poses, activations # [batch , len-2 , 1 , 16, 16] # [batch , len-2 , 1 , 16]
def capsule_model_B(X, num_classes):
poses_list = []
for _, ngram in enumerate([3, 4, 5]):
with tf.variable_scope('capsule_' + str(ngram)):
nets = _conv2d_wrapper(X,
shape=[ngram, 300, 1, 32],
strides=[1, 2, 1, 1],
padding='VALID',
add_bias=True,
activation_fn=tf.nn.relu,
name='conv1')
tf.logging.info('output shape: {}'.format(nets.get_shape()))
nets = capsules_init(nets,
shape=[1, 1, 32, 16],
strides=[1, 1, 1, 1],
padding='VALID',
pose_shape=16,
add_bias=True,
name='primary')
nets = capsule_conv_layer(nets,
shape=[3, 1, 16, 16],
strides=[1, 1, 1, 1],
iterations=3,
name='conv2')
nets = capsule_flatten(nets)
poses, activations = capsule_fc_layer(nets, num_classes, 3, 'fc2')
poses_list.append(poses)
poses = tf.reduce_mean(tf.convert_to_tensor(poses_list), axis=0)
activations = K.sqrt(K.sum(K.square(poses), 2))
return poses, activations
def capsule_model_A(X, num_classes):
#input dimension:(25, 200, 300, 1)
with tf.variable_scope('capsule_' + str(3)):
nets = _conv2d_wrapper(X,
shape=[3, 300, 1, 32],
strides=[1, 2, 1, 1],
padding='VALID',
add_bias=True,
activation_fn=tf.nn.relu,
name='conv1') #output shape: (25, 99, 1, 32)
tf.logging.info('output shape: {}'.format(nets.get_shape()))
nets = capsules_init(nets,
shape=[1, 1, 32, 16],
strides=[1, 1, 1, 1],
padding='VALID',
pose_shape=16,
add_bias=True,
name='primary') # (25, 99, 1, 16, 16)
nets = capsule_conv_layer(
nets,
shape=[3, 1, 16, 16],
strides=[1, 1, 1, 1],
iterations=3,
name='conv2') #25, 97, 1, 16, 16 #25, 97, 1, 16
nets = capsule_flatten(nets) #25, 97*16, 16 #25, 97*16
poses, activations = capsule_fc_layer(nets, num_classes, 3, 'fc2')
return poses, activations
def capsules_init(inputs, shape, strides, padding, pose_shape, add_bias, name):
with tf.variable_scope(name):
poses = _conv2d_wrapper(inputs,
shape=shape[0:-1] + [shape[-1] * pose_shape],
strides=strides,
padding=padding,
add_bias=add_bias,
activation_fn=None,
name='pose_stacked')
poses_shape = poses.get_shape().as_list()
poses = tf.reshape(
poses, [-1, poses_shape[1], poses_shape[2], shape[-1], pose_shape])
beta_a = _get_weights_wrapper(name='beta_a', shape=[1, shape[-1]])
if False:
u_hat_vecs = tf.reshape(poses, [-1, shape[-1], 1, pose_shape])
tf.logging.info("prim votes dimension:{}".format(
u_hat_vecs.get_shape()))
poses, activations = routing(u_hat_vecs, beta_a, 1, shape[3], None)
poses = tf.reshape(
poses,
[-1, poses_shape[1], poses_shape[2], shape[-1], pose_shape])
activations = tf.reshape(
activations, [-1, poses_shape[1], poses_shape[2], shape[-1]])
else:
poses = squash_v1(poses, axis=-1)
activations = K.sqrt(K.sum(K.square(poses), axis=-1)) + beta_a
tf.summary.histogram('activations', activations)
tf.logging.info("prim poses dimension:{}".format(poses.get_shape()))
tf.logging.info("prim activations dimension:{}".format(
activations.get_shape()))
return poses, activations
def capsule_model_A(X, num_classes):
with tf.variable_scope('capsule_' + str(3)):
print('capsule_' + str(3))
print('X ', X)
cnnout = _conv2d_wrapper(X, shape=[3, 300, 1, 32], strides=[1, 2, 1, 1], padding='VALID',
add_bias=True, activation_fn=tf.nn.relu, name='conv1')
print('cnnout', cnnout.shape)
tf.logging.info('output shape: {}'.format(cnnout.get_shape()))
poses_init, activations_init = capsules_init(cnnout, shape=[1, 1, 32, 16], strides=[1, 1, 1, 1],
padding='VALID', pose_shape=16, add_bias=True, name='primary')
print('poses_init', poses_init.shape)
print('activations_init', activations_init.get_shape())
poses_conv, activations_conv = capsule_conv_layer(poses_init, activations_init, shape=[3, 1, 16, 16], strides=[1, 1, 1, 1],
iterations=3, name='conv2')
print('poses_conv', poses_conv.shape)
print('activations_conv', activations_conv.shape)
poses_flat, activations_flat = capsule_flatten(poses_conv, activations_conv)
print('capsule_flatten', poses_flat.shape)
print('activations_flat', activations_flat.shape)
poses, activations = capsule_fc_layer(poses_flat, activations_flat, num_classes, 3, 'fc2')
print('poses ', poses.shape)
print('activations ', activations.shape)
return poses, activations
def baseline_model_cnn(X, num_classes):
nets = _conv2d_wrapper(
X, shape=[3, 300, 1, 32], strides=[1, 1, 1, 1], padding='VALID',
add_bias=False, activation_fn=tf.nn.relu, name='conv1'
)
nets = slim.flatten(nets)
tf.logging.info('flatten shape: {}'.format(nets.get_shape()))
nets = slim.fully_connected(nets, 128, scope='relu_fc3', activation_fn=tf.nn.relu)
tf.logging.info('fc shape: {}'.format(nets.get_shape()))
activations = tf.sigmoid(slim.fully_connected(nets, num_classes, scope='final_layer', activation_fn=None))
tf.logging.info('fc shape: {}'.format(activations.get_shape()))
return tf.zeros([0]), activations
def capsules_init(inputs, shape, strides, padding, pose_shape, add_bias, name):
with tf.variable_scope(name):
poses = _conv2d_wrapper(inputs,
shape=shape[0:-1] + [shape[-1] * pose_shape],
strides=strides,
padding=padding,
add_bias=add_bias,
activation_fn=None,
name='pose_stacked')
poses_shape = poses.get_shape().as_list()
poses = tf.reshape(
poses, [-1, poses_shape[1], poses_shape[2], shape[-1], pose_shape])
beta_a = _get_weights_wrapper(name='beta_a', shape=[1, shape[-1]])
poses = squash_v1(poses, axis=-1)
activations = K.sqrt(K.sum(K.square(poses), axis=-1)) + beta_a
tf.logging.info("prim poses dimension:{}".format(poses.get_shape()))
return poses, activations
def capsule_model_B(X, num_classes):
print('X.shape', X.shape) # (25, 200, 300, 1)
print('num_classes: ', num_classes.shape)
poses_list = []
for _, ngram in enumerate([3, 4, 5]):
with tf.variable_scope('capsule_' + str(ngram)):
print('capsule_' + str(ngram))
cnnout = _conv2d_wrapper(
X, shape=[ngram, 300, 1, 32], strides=[1, 2, 1, 1], padding='VALID',
add_bias=True, activation_fn=tf.nn.relu, name='conv1'
)
print('cnnout', cnnout.shape) # (25, 99, 1, 32)
tf.logging.info('output shape: {}'.format(cnnout.get_shape()))
poses_init, activations_init = capsules_init(cnnout, shape=[1, 1, 32, 16], strides=[1, 1, 1, 1],
padding='VALID', pose_shape=16, add_bias=True, name='primary')
print('poses_init', poses_init.shape) # (25, 99, 1, 16, 16)
print('activations_init', activations_init.get_shape()) # (25, 99, 1, 16)
poses_conv, activations_conv = capsule_conv_layer(poses_init, activations_init, shape=[3, 1, 16, 16], strides=[1, 1, 1, 1],
iterations=3, name='conv2')
print('poses_conv', poses_conv.shape) # (25, 97, 1, 16, 16)
print('activations_conv', activations_conv.shape) # (25, 97, 1, 16)
poses_flat, activations_flat = capsule_flatten(poses_conv, activations_conv)
print('capsule_flatten', poses_flat.shape) # (25, 1552, 16)
print('activations_flat', activations_flat.shape) # (25, 1552)
poses, activations = capsule_fc_layer(poses_flat, activations_flat, num_classes, 3, 'fc2')
print('poses ', poses.shape) # (25, 9, 16)
print('activations ', activations.shape) # (25, 9)
poses_list.append(poses)
print('-------------------------------')
poses = tf.reduce_mean(tf.convert_to_tensor(poses_list), axis=0)
print('poses ', poses.shape)
activations = K.sqrt(K.sum(K.square(poses), 2))
print('activations ', activations.shape)
return poses, activations
def capsules_init(inputs, shape, strides, padding, pose_shape, add_bias, name):
with tf.variable_scope(name):
poses = _conv2d_wrapper(
inputs,
shape=shape[0:-1] + [shape[-1] * pose_shape],
strides=strides,
padding=padding,
add_bias=add_bias,
activation_fn=None,
name='pose_stacked'
)
poses_shape = poses.get_shape().as_list()
poses = tf.reshape(
poses, [
-1, poses_shape[1], poses_shape[2], shape[-1], pose_shape
])
beta_a = _get_weights_wrapper(
name='beta_a', shape=[1, shape[-1]]
)
poses = squash_v1(poses, axis=-1)
activations = K.sqrt(K.sum(K.square(poses), axis=-1)) + beta_a
tf.logging.info("prim poses dimension:{}".format(poses.get_shape()))
return poses, activations
