def build_sum_network(roles,
fillers,
dual_roles,
max_depth,
number_sum_blocks=1):
filler_len = fillers[0].shape[0]
input_num_elements, flattened_tree_num_elements = unshift_matrix(
roles[0], filler_len, max_depth - 1).shape
shape = (flattened_tree_num_elements + filler_len, 1)
flattened_decrementing_input = tf.keras.layers.Input(shape=(*shape, ),
batch_size=1,
name='left_operand')
flattened_incrementing_input = tf.keras.layers.Input(shape=(*shape, ),
batch_size=1,
name='right_operand')
block_id = 0
shift_input, increment_input, filler_input = constant_inputs_for_increment_block(
roles, fillers, max_depth, block_id)
left_shift_input, right_shift_input = shift_input
tmp_reshaped_increment, const_increment = increment_input
tmp_reshaped_fake_filler, const_filler = filler_input
target_elements, _ = unshift_matrix(roles[0], filler_len,
max_depth - 1).shape
tmp_reshaped_fake, const_one = custom_constant_layer(
const_size=target_elements + filler_len, name='const_one')
all_sum_const_inputs = []
incremented = flattened_incrementing_input
decremented = flattened_decrementing_input
for i in range(number_sum_blocks):
sum_const_inputs, incremented, decremented = sum_block(
incrementing_input=incremented,
decrementing_input=decremented,
increment_value=tmp_reshaped_increment,
roles=roles,
dual_roles=dual_roles,
filler_len=filler_len,
max_depth=max_depth,
block_id=block_id + i * 5,
left_shift_input=left_shift_input,
right_shift_input=right_shift_input,
constant_input_filler=tmp_reshaped_fake_filler,
constant_for_decrementing_input=tmp_reshaped_fake)
all_sum_const_inputs.extend(sum_const_inputs)
return tf.keras.Model(inputs=[
left_shift_input,
right_shift_input,
const_increment,
const_filler,
const_one,
*all_sum_const_inputs,
flattened_decrementing_input,
flattened_incrementing_input,
],
outputs=[decremented, incremented])
def crop_tensor(layer, role, filler_len, stop_level):
_, flattened_num_elements = unshift_matrix(role, filler_len, stop_level).shape
return custom_cropping_layer(input_layer=layer,
crop_from_beginning=0,
crop_from_end=flattened_num_elements,
input_tensor_length=flattened_num_elements + filler_len,
final_tensor_length=filler_len)
def build_decode_model_2_tuple_network(filler_len, dual_roles, max_depth,
model_2_tuple_has_weights):
input_num_elements, flattened_tree_num_elements = unshift_matrix(
dual_roles[0], filler_len, max_depth - 1).shape
shape = (flattened_tree_num_elements + filler_len, 1)
flattened_input = tf.keras.layers.Input(shape=(*shape, ))
index_const_inputs, index_raw_output, _ = build_universal_extraction_branch(
model_input=flattened_input,
roles=dual_roles,
filler_len=filler_len,
max_depth=max_depth - 1,
stop_level=max_depth - 1,
role_extraction_order=[0],
prefix='extracting_index')
index_raw_output = tf.keras.layers.Lambda(
lambda x: tf.keras.backend.reshape(x, (filler_len, )))(
index_raw_output)
alpha_const_inputs, alpha_raw_output, _ = build_universal_extraction_branch(
model_input=flattened_input,
roles=dual_roles,
filler_len=filler_len,
max_depth=max_depth - 1,
stop_level=max_depth - 1,
role_extraction_order=[1],
prefix='extracting_alpha')
alpha_raw_output = tf.keras.layers.Lambda(
lambda x: tf.keras.backend.reshape(x, (filler_len, )))(
alpha_raw_output)
if model_2_tuple_has_weights:
weight_const_inputs, weight_raw_output, _ = build_universal_extraction_branch(
model_input=flattened_input,
roles=dual_roles,
filler_len=filler_len,
max_depth=max_depth - 1,
stop_level=max_depth - 1,
role_extraction_order=[2],
prefix='extracting_weight')
weight_raw_output = tf.keras.layers.Lambda(
lambda x: tf.keras.backend.reshape(x, (filler_len, )))(
weight_raw_output)
return tf.keras.Model(
inputs=[
*index_const_inputs, *alpha_const_inputs, *weight_const_inputs,
flattened_input
],
outputs=[index_raw_output, alpha_raw_output, weight_raw_output])
return tf.keras.Model(
inputs=[*index_const_inputs, *alpha_const_inputs, flattened_input],
outputs=[
index_raw_output,
alpha_raw_output,
])
def make_output_same_length_as_input(layer_to_crop, role, filler_len, max_depth):
target_num_elements, flattened_num_elements = unshift_matrix(role, filler_len, max_depth).shape
return custom_cropping_layer(
input_layer=layer_to_crop,
crop_from_beginning=0,
crop_from_end=flattened_num_elements + filler_len - target_num_elements,
input_tensor_length=flattened_num_elements + filler_len,
final_tensor_length=target_num_elements
)
def decode_most_nested_element_numpy(tree, depth):
fillers, roles, dual_roles = input_data()
current_input = tree
max_depth = depth - 1
stop_level = 0
filler_len = fillers[0].shape[0]
role_extraction_order = [1 for _ in range(max_depth)]
for level_index, role_index in zip(range(max_depth, stop_level - 1, -1),
role_extraction_order):
left_shift_input = unshift_matrix(roles[role_index], filler_len,
level_index)
current_input = left_shift_input.dot(current_input[filler_len:])
return current_input
def check_if_not_zero_branch(decrementing_input, role, filler_len, max_depth, block_id):
const_inputs, one_tensor_output = build_extract_branch(
input_layer=decrementing_input,
extract_role=role,
filler_len=filler_len,
max_depth=max_depth - 1,
block_id=block_id
)
target_elements, _ = unshift_matrix(role, filler_len, max_depth - 1).shape
reshape_for_pool = tf.keras.layers.Lambda(lambda x: tf.keras.backend.reshape(x, (1, target_elements, 1)))(
one_tensor_output)
global_max_pool = tf.keras.layers.GlobalMaxPooling1D()(reshape_for_pool)
return const_inputs, tf.keras.layers.Lambda(normalization)(global_max_pool), one_tensor_output
def build_decode_number_network(fillers, dual_roles, max_depth):
filler_len = fillers[0].shape[0]
input_num_elements, flattened_tree_num_elements = unshift_matrix(
dual_roles[0], filler_len, max_depth - 1).shape
shape = (flattened_tree_num_elements + filler_len, 1)
flattened_input = tf.keras.layers.Input(shape=(*shape, ), batch_size=1)
const_inputs, is_not_zero, unshifted = check_if_not_zero_branch(
decrementing_input=flattened_input,
role=dual_roles[0],
filler_len=filler_len,
max_depth=max_depth,
block_id=1)
return tf.keras.Model(inputs=[*const_inputs, flattened_input],
outputs=[unshifted, is_not_zero])
def build_increment_network(roles, dual_roles, fillers, max_depth):
filler_len = fillers[0].shape[0]
# Number to be incremented
input_num_elements, flattened_tree_num_elements = unshift_matrix(roles[0], filler_len, max_depth - 1).shape
shape = (flattened_tree_num_elements + filler_len, 1)
flattened_incrementing_input = tf.keras.layers.Input(shape=(*shape,), batch_size=1)
block_id = 0
shift_input, increment_input, filler_input = constant_inputs_for_increment_block(roles, fillers, max_depth,
block_id)
left_shift_input, right_shift_input = shift_input
tmp_reshaped_increment, const_increment = increment_input
tmp_reshaped_fake_filler, const_filler = filler_input
increment_const_inputs, output = increment_block(
incrementing_input=flattened_incrementing_input,
increment_value=tmp_reshaped_increment,
roles=roles,
dual_roles=dual_roles,
filler_len=filler_len,
max_depth=max_depth,
block_id=block_id,
left_shift_input=left_shift_input,
right_shift_input=right_shift_input,
constant_input_filler=tmp_reshaped_fake_filler
)
return tf.keras.Model(
inputs=[
left_shift_input,
right_shift_input,
const_increment,
const_filler,
*increment_const_inputs,
flattened_incrementing_input,
],
outputs=output)
def extract_semantic_tree_from_passive_voice_branch(input_layer, roles,
dual_roles, filler_len,
max_depth):
stop_level_for_verb = 0
verb_branch = build_one_level_extraction_branch(
model_input=input_layer,
roles=dual_roles,
filler_len=filler_len,
max_depth=max_depth,
stop_level=stop_level_for_verb,
role_extraction_order=[1, 0, 1],
prefix='passive_verb_extract')
verb_extraction_const_inputs, verb_extraction_output, _ = verb_branch
stop_level_for_agent = 0
agent_branch = build_one_level_extraction_branch(
model_input=input_layer,
roles=dual_roles,
filler_len=filler_len,
max_depth=max_depth,
stop_level=stop_level_for_agent,
role_extraction_order=[1, 1, 1],
prefix='passive_agent_extract')
agent_extraction_const_inputs, agent_extraction_output, _ = agent_branch
stop_level_for_p = max_depth - 1
p_branch = build_one_level_extraction_branch(model_input=input_layer,
roles=dual_roles,
filler_len=filler_len,
max_depth=max_depth,
stop_level=stop_level_for_p,
role_extraction_order=[0],
prefix='passive_p_extract')
p_extraction_const_inputs, p_raw_output, current_num_elements = p_branch
_, flattened_num_elements = unshift_matrix(roles[0], filler_len,
stop_level_for_p).shape
# TODO: insert cropping here
reshape_for_crop = tf.keras.layers.Lambda(
lambda x: tf.keras.backend.reshape(x, (1, flattened_num_elements +
filler_len, 1)))(p_raw_output)
clip_first_level = tf.keras.layers.Cropping1D(
cropping=(0, flattened_num_elements))(reshape_for_crop)
p_extraction_output = tf.keras.layers.Lambda(
lambda x: tf.keras.backend.reshape(x, (1, filler_len, 1)))(
clip_first_level)
# TODO: define how to tackle extractions not till the bottom of structure
# given that we have all fillers maximum joining depth is equal to 1
agentxr0_pxr1_const_inputs, agentxr0_pxr1_output = build_join_branch(
roles=roles,
filler_len=filler_len,
max_depth=1,
inputs=[agent_extraction_output, p_extraction_output],
prefix='passive_join(agent,p)')
# later we have to join two subtrees of different depth. for that we have to
# make filler of verb of the same depth - make fake constant layer
np_constant = np.zeros((filler_len, 1))
const_fake_extender = keras_constant_layer(
np_constant, name='passive_fake_extender_verb_agent')
concatenate_verb = tf.keras.layers.Concatenate(axis=0)(
[verb_extraction_output, const_fake_extender, const_fake_extender])
# TODO: why is there a constant 3?
reshaped_verb = tf.keras.layers.Lambda(lambda x: tf.keras.backend.reshape(
x, (1, filler_len * 3, 1)))(concatenate_verb)
# TODO: reshape by 2, why is there a constant 2?
tmp_reshaped_agentxr0_pxr1 = tf.keras.layers.Lambda(
lambda x: tf.keras.backend.reshape(x, (filler_len * 2, 1)))(
agentxr0_pxr1_output)
# TODO: reshaping constant input??
tmp_reshaped_fake = tf.keras.layers.Lambda(
lambda x: tf.keras.backend.reshape(x, (filler_len, 1)))(
const_fake_extender)
concatenate_agentxr0_pxr1 = tf.keras.layers.Concatenate(axis=0)(
[tmp_reshaped_fake, tmp_reshaped_agentxr0_pxr1])
# TODO: why is there a constant 3?
reshaped_agentxr0_pxr1 = tf.keras.layers.Lambda(
lambda x: tf.keras.backend.reshape(x, (1, filler_len * 3, 1)))(
concatenate_agentxr0_pxr1)
semantic_tree_const_inputs, semantic_tree_output = build_join_branch(
roles=roles,
filler_len=filler_len,
max_depth=2,
inputs=[reshaped_verb, reshaped_agentxr0_pxr1],
prefix='passive_join(verb, join(agent,p))')
return [
*verb_extraction_const_inputs, *agent_extraction_const_inputs,
*p_extraction_const_inputs, *agentxr0_pxr1_const_inputs,
*semantic_tree_const_inputs, const_fake_extender
], semantic_tree_output