def __init__(self, type_content_data):
super(SequenceDecodeModel, self).__init__(type_content_data)
self.token_lstm = YLSTMCell(0)
number_of_tokens = self.type_content_data[all_token_summary][
TokenHitNum]
self.linear_token_output_w = tf.Variable(
random_uniform_variable_initializer(256, 566,
[number_of_tokens, num_units]))
self.one_hot_token_embedding = tf.Variable(
random_uniform_variable_initializer(256, 56,
[number_of_tokens, num_units]))
self.token_embedder = TokenAtomEmbed(self.type_content_data,
self.one_hot_token_embedding)
self.token_attention = None
if decode_attention_way > decode_no_attention:
self.token_attention = YAttention(10)
self.dup_token_lstm, self.dup_one_hot_token_embedding, self.dup_token_embedder, self.dup_token_pointer = None, None, None, None
# if use_dup_model:
# self.dup_token_lstm = YLSTMCell(125)
# self.dup_one_hot_token_embedding = tf.Variable(random_uniform_variable_initializer(25, 56, [number_of_tokens, num_units]))
# self.dup_token_embedder = TokenAtomEmbed(self.type_content_data, self.dup_one_hot_token_embedding)
# self.dup_token_pointer = PointerNetwork(222)
self.token_decoder = TokenDecoder(self.type_content_data,
self.metrics_index,
self.linear_token_output_w,
self.token_attention,
self.dup_token_pointer)
def __init__(self, type_content_data):
super(TreeDecodeModel, self).__init__(type_content_data)
self.direct_descedent_lstm = YLSTMCell(0)
if tree_decode_way == tree_decode_2d:
self.two_dimen_lstm = Y2DLSTMCell(1)
elif tree_decode_way == tree_decode_embed:
self.one_dimen_lstm = YLSTMCell(2)
number_of_tokens = self.type_content_data[all_token_summary][
TokenHitNum]
self.linear_token_output_w = tf.Variable(
random_uniform_variable_initializer(256, 566,
[number_of_tokens, num_units]))
self.one_hot_token_embedding = tf.Variable(
random_uniform_variable_initializer(256, 56,
[number_of_tokens, num_units]))
self.one_hot_token_cell_embedding = tf.Variable(
random_uniform_variable_initializer(25, 56,
[number_of_tokens, num_units]))
self.token_embedder = TokenAtomEmbed(self.type_content_data,
self.one_hot_token_embedding)
self.token_cell_embedder = TokenAtomEmbed(
self.type_content_data, self.one_hot_token_cell_embedding)
self.encode_tree = EncodeOneAST(self.type_content_data,
self.token_embedder,
self.token_cell_embedder)
def __init__(self, num_desc):
self.merger = EmbedMerger(num_desc)
self.memory_update_lstm = YLSTMCell(30 + num_desc)
self.initial_cell = tf.Variable(
random_uniform_variable_initializer(175, 525 + num_desc,
[1, num_units]))
self.initial_h = tf.Variable(
random_uniform_variable_initializer(1755, 525 + num_desc,
[1, num_units]))
def __init__(self, type_content_data, vocab_embeddings):
self.type_content_data = type_content_data
self.vocab_embeddings = vocab_embeddings
self.sword_embed = SwordAtomEmbed(self.type_content_data,
self.vocab_embeddings)
self.merger = EmbedMerger()
self.forward_lstm = YLSTMCell(15)
self.backward_lstm = YLSTMCell(25)
self.ini_forward_cell = tf.Variable(
random_uniform_variable_initializer(881, 882, [1, num_units]))
self.ini_forward_h = tf.Variable(
random_uniform_variable_initializer(883, 884, [1, num_units]))
self.ini_backward_cell = tf.Variable(
random_uniform_variable_initializer(885, 886, [1, num_units]))
self.ini_backward_h = tf.Variable(
random_uniform_variable_initializer(887, 888, [1, num_units]))
def __init__(self, num_desc):
self.w = tf.Variable(random_uniform_variable_initializer(222, 333+num_desc, [2 * num_units, 5 * num_units], ini_range=lstm_initialize_range))
self.b = tf.Variable(zero_variable_initializer([1, 5 * num_units]))
if embed_merger_use_layer_norm:
self.norm_weights = []
self.norm_biases = []
for _ in range(6):
self.norm_weights.append(tf.Variable(one_variable_initializer([num_units])))
self.norm_biases.append(tf.Variable(zero_variable_initializer([num_units])))
def initialize_parameters(self):
self.skeleton_lstm_cell = YLSTMCell(1)
self.one_hot_skeleton_embedding = tf.Variable(
random_uniform_variable_initializer(258, 578,
[self.vocab_num, num_units]))
self.skeleton_embedder = SkeletonAtomEmbed(
self.type_content_data, self.one_hot_skeleton_embedding,
self.vocab_num)
self.linear_skeleton_output_w = tf.Variable(
random_uniform_variable_initializer(257, 576,
[self.vocab_num, num_units]))
if skeleton_multi_decode_mode_on:
self.skeleton_multi_decode_transfer = tf.Variable(
random_uniform_variable_initializer(
527, 567,
[skeleton_multi_decode_num, num_units, num_units]))
def __init__(self, type_content_data):
super(StatementDupModel, self).__init__(type_content_data)
number_of_tokens = self.type_content_data[all_token_summary][TokenHitNum]
self.dup_skeleton_forward_cell_h = tf.Variable(random_uniform_variable_initializer(155, 572, [1, 2, num_units]))
self.dup_skeleton_backward_cell_h = tf.Variable(random_uniform_variable_initializer(152, 572, [1, 2, num_units]))
if atom_decode_mode == token_decode:
self.one_dup_hot_token_embedding = tf.Variable(random_uniform_variable_initializer(252, 226, [number_of_tokens, num_units]))
self.dup_token_embedder = TokenAtomEmbed(self.type_content_data, self.one_dup_hot_token_embedding)
self.dup_token_lstm = YLSTMCell(9)
self.dup_token_pointer = PointerNetwork(655)
self.integrate_computer = None
if compute_memory_in_memory_mode:
self.integrate_computer = Y2DirectLSTMCell(105)
if compute_token_memory:
self.dup_mem_nn = NTMOneDirection(800)
self.dup_forward_token_lstm = YLSTMCell(10)
self.dup_backward_token_lstm = YLSTMCell(11)
self.dup_token_decoder = DupTokenDecoder(type_content_data, self.metrics_index, self.dup_token_pointer)
else:
assert False, "Wrong atom_decode_mode"
def __init__(self, type_content_data, token_embedder, token_cell_embedder):
self.type_content_data = type_content_data
self.token_embedder = token_embedder
self.token_cell_embedder = token_cell_embedder
self.y2direct_lstm = Y2DirectLSTMCell(60)
self.y2d_lstm = Y2DLSTMCell(61)
self.y_lstm = YLSTMCell(62)
if tree_leaf_one_more_lstm_step:
self.y_leaf_lstm = YLSTMCell(63)
self.y_forward = YLSTMCell(64)
self.y_backward = YLSTMCell(65)
self.tree_leaf_one_more_lstm_begin_cell_h = tf.Variable(
random_uniform_variable_initializer(25, 5, [2, num_units]))
def __init__(self, type_content_data):
self.type_content_data = type_content_data
self.statistical_metrics_meta = default_metrics_meta + self.create_extra_default_metrics_meta(
) + special_handle_metrics_meta
self.metrics_meta = self.statistical_metrics_meta + self.create_in_use_tensors_meta(
)
self.metrics_name = [metric_m[0] for metric_m in self.metrics_meta]
self.metrics_shape = [metric_m[1] for metric_m in self.metrics_meta]
self.index_metrics = dict(
(k, v)
for k, v in zip(range(len(self.metrics_name)), self.metrics_name))
self.metrics_index = {
value: key
for key, value in self.index_metrics.items()
}
self.metrics_contingent_index = create_metrics_contingent_index(
self.metrics_meta)
self.contingent_parameters = tf.Variable(
random_uniform_variable_initializer(
2, 5, [contingent_parameters_num, 1, num_units]))
self.contingent_parameters_for_idle = tf.Variable(
random_uniform_variable_initializer(20, 50, [2, 1, num_units]))
def __init__(self, num_desc):
if is_dup_mode == simple_is_dup:
self.is_dup_w = tf.Variable(random_uniform_variable_initializer(200, 1080 + num_desc, [num_units, num_units], ini_range=lstm_initialize_range))
self.is_not_dup_w = tf.Variable(random_uniform_variable_initializer(205, 1080 + num_desc, [num_units, num_units], ini_range=lstm_initialize_range))
elif is_dup_mode == mlp_is_dup:
self.is_dup_w = tf.Variable(random_uniform_variable_initializer(200, 1080 + num_desc, [2 * num_units, 2 * num_units], ini_range=lstm_initialize_range))
self.is_dup_h = tf.Variable(random_uniform_variable_initializer(200, 1050 + num_desc, [1, 2 * num_units]))
self.is_not_dup_w = tf.Variable(random_uniform_variable_initializer(205, 1080 + num_desc, [2 * num_units, 2 * num_units], ini_range=lstm_initialize_range))
self.is_not_dup_h = tf.Variable(random_uniform_variable_initializer(205, 1060 + num_desc, [1, 2 * num_units]))
elif is_dup_mode == sigmoid_is_dup:
self.is_dup_h = tf.Variable(random_uniform_variable_initializer(200, 1050 + num_desc, [1, 2 * num_units]))
if dup_use_two_poles:
self.two_poles_merge_w = tf.Variable(random_uniform_variable_initializer(200, 1060 + num_desc, [2 * num_units, 1 * num_units], ini_range=lstm_initialize_range))
self.two_poles_merge_b = tf.Variable(random_uniform_variable_initializer(200, 1070 + num_desc, [1, 1 * num_units]))
else:
assert False, "Unrecognized is_dup_mode!"
self.dup_point_atten = YAttention(100 + num_desc)
if repetition_mode == attention_repetition_mode:
if dup_share_parameters:
self.is_dup_point_atten = self.dup_point_atten
else:
self.is_dup_point_atten = YAttention(200 + num_desc)
def __init__(self, num_desc):
if attention_algorithm == v_attention:
self.v = tf.Variable(random_uniform_variable_initializer(203, 21 + num_desc, [1, num_units]))
self.w_v = tf.Variable(random_uniform_variable_initializer(2033, 231 + num_desc, [num_units, num_units], ini_range=lstm_initialize_range))
self.w_h = tf.Variable(random_uniform_variable_initializer(2033, 131 + num_desc, [num_units, num_units], ini_range=lstm_initialize_range))
self.w_ctx_c = tf.Variable(random_uniform_variable_initializer(203, 145 + num_desc, [2*num_units, num_units], ini_range=lstm_initialize_range))
elif attention_algorithm == stand_attention:
self.w = tf.Variable(random_uniform_variable_initializer(2033, 132 + num_desc, [num_units, num_units], ini_range=lstm_initialize_range))
else:
print("Strange Error! Unrecognized attention algorithm")
self.w_ctx_h = tf.Variable(random_uniform_variable_initializer(203, 135 + num_desc, [2*num_units, num_units], ini_range=lstm_initialize_range))
def __init__(self, num_desc, forget_bias=0.0, activation=tf.nn.tanh):
self.forget_bias = forget_bias
self.activation = activation
self.w = tf.Variable(
random_uniform_variable_initializer(
9,
88 + num_desc, [2 * num_units, 4 * num_units],
ini_range=lstm_initialize_range))
self.b = tf.Variable(zero_variable_initializer([1, 4 * num_units]))
if use_layer_norm:
self.norm_weights = []
self.norm_biases = []
for _ in range(5):
self.norm_weights.append(
tf.Variable(one_variable_initializer([num_units])))
self.norm_biases.append(
tf.Variable(zero_variable_initializer([num_units])))
def __init__(self, type_content_data):
super(LinearDupModel, self).__init__(type_content_data)
number_of_tokens = self.type_content_data[all_token_summary][
TokenHitNum]
self.dup_token_lstm = YLSTMCell(125)
self.dup_one_hot_token_embedding = tf.Variable(
random_uniform_variable_initializer(25, 56,
[number_of_tokens, num_units]))
self.dup_token_embedder = TokenAtomEmbed(
self.type_content_data, self.dup_one_hot_token_embedding)
self.dup_token_pointer = PointerNetwork(222)
self.integrate_computer = None
if compute_memory_in_memory_mode:
self.integrate_computer = Y2DirectLSTMCell(105)
self.dup_token_decoder = DupTokenDecoder(self.type_content_data,
self.metrics_index,
self.dup_token_pointer)
def __init__(self, type_content_data):
super(StatementDecodeModel, self).__init__(type_content_data)
self.ini_metrics = list(create_empty_tensorflow_tensors(self.metrics_meta, self.contingent_parameters, self.metrics_contingent_index))
self.skeleton_forward_cell_h = tf.Variable(random_uniform_variable_initializer(255, 572, [1, 2, num_units]))
self.skeleton_backward_cell_h = tf.Variable(random_uniform_variable_initializer(252, 572, [1, 2, num_units]))
if compute_token_memory:
self.mem_nn = NTMOneDirection(500)
self.forward_token_lstm = YLSTMCell(3)
self.backward_token_lstm = YLSTMCell(4)
if compose_tokens_of_a_statement:
if compose_mode == compose_one_way_lstm and (compose_one_way_lstm_mode == one_way_two_way_compose or compose_one_way_lstm_mode == one_way_three_way_compose):
self.tokens_merger = Y2DirectLSTMCell(150)
else:
self.tokens_merger = EmbedMerger(150)
if compose_mode == compose_one_way_lstm and compose_one_way_lstm_mode == one_way_three_way_compose:
self.compose_lstm_cell = Y3DirectLSTMCell(5)
else:
self.compose_lstm_cell = YLSTMCell(5)
if compose_mode == compose_bi_way_lstm:
self.bi_way_merger = Y2DirectLSTMCell(155)
self.bi_way_ini_cell = tf.Variable(random_uniform_variable_initializer(855, 55, [1, num_units]))
self.bi_way_ini_h = tf.Variable(random_uniform_variable_initializer(850, 50, [1, num_units]))
self.bi_way_lstm = YLSTMCell(52)
self.token_attention = None
if decode_attention_way > decode_no_attention:
self.token_attention = YAttention(10)
self.token_lstm = YLSTMCell(0)
r_token_embedder_mode = token_embedder_mode
number_of_tokens = self.type_content_data[all_token_summary][TokenHitNum]
number_of_subwords = self.type_content_data[all_token_summary][SwordHitNum]
if atom_decode_mode == token_decode:
self.linear_token_output_w = tf.Variable(random_uniform_variable_initializer(256, 566, [number_of_tokens, num_units]))
self.dup_token_embedder, self.dup_token_lstm, self.dup_token_pointer = None, None, None
# if use_dup_model:
# self.one_dup_hot_token_embedding = tf.Variable(random_uniform_variable_initializer(252, 226, [number_of_tokens, num_units]))
# self.dup_token_embedder = TokenAtomEmbed(self.type_content_data, self.one_dup_hot_token_embedding)
# self.dup_token_lstm = YLSTMCell(9)
# self.dup_token_pointer = PointerNetwork(655)
# self.dup_skeleton_forward_cell_h = tf.Variable(random_uniform_variable_initializer(155, 572, [number_of_skeletons, 2, num_units]))
# self.dup_skeleton_backward_cell_h = tf.Variable(random_uniform_variable_initializer(152, 572, [number_of_skeletons, 2, num_units]))
# if compute_token_memory:
# self.dup_mem_nn = NTMOneDirection(800)
# self.dup_forward_token_lstm = YLSTMCell(10)
# self.dup_backward_token_lstm = YLSTMCell(11)
self.token_decoder = TokenDecoder(type_content_data, self.metrics_index, self.linear_token_output_w, self.token_attention, self.dup_token_pointer)
elif atom_decode_mode == sword_decode:
self.linear_sword_output_w = tf.Variable(random_uniform_variable_initializer(256, 566, [number_of_subwords, num_units]))
self.sword_lstm = YLSTMCell(12)
r_token_embedder_mode = swords_compose_mode
else:
assert False, "Wrong atom_decode_mode"
if r_token_embedder_mode == token_only_mode:
self.one_hot_token_embedding = tf.Variable(random_uniform_variable_initializer(256, 56, [number_of_tokens, num_units]))
self.token_embedder = TokenAtomEmbed(self.type_content_data, self.one_hot_token_embedding)
elif r_token_embedder_mode == swords_compose_mode:
self.one_hot_sword_embedding = tf.Variable(random_uniform_variable_initializer(256, 56, [number_of_subwords, num_units]))
self.sword_embedder = SwordAtomEmbed(self.type_content_data, self.one_hot_sword_embedding)
self.token_embedder = BiLSTMEmbed(self.type_content_data, self.one_hot_sword_embedding)
else:
assert False, "Wrong token_embedder_mode"