def __init__(self):
# self.node_embedding = Embedding(config.node_num, config.node_embed_dim, name='node_embed')
self.query_embedding = Embedding(config.source_vocab_size, config.word_embed_dim, name='query_embed')
if config.encoder == 'bilstm':
self.query_encoder_lstm = BiLSTM(config.word_embed_dim, config.encoder_hidden_dim / 2, return_sequences=True,
name='query_encoder_lstm')
else:
self.query_encoder_lstm = LSTM(config.word_embed_dim, config.encoder_hidden_dim, return_sequences=True,
name='query_encoder_lstm')
self.decoder_lstm = CondAttLSTM(config.rule_embed_dim + config.node_embed_dim + config.rule_embed_dim,
config.decoder_hidden_dim, config.encoder_hidden_dim, config.attention_hidden_dim,
name='decoder_lstm')
self.src_ptr_net = PointerNet()
self.terminal_gen_softmax = Dense(config.decoder_hidden_dim, 2, activation='softmax', name='terminal_gen_softmax')
self.rule_embedding_W = initializations.get('normal')((config.rule_num, config.rule_embed_dim), name='rule_embedding_W', scale=0.1)
self.rule_embedding_b = shared_zeros(config.rule_num, name='rule_embedding_b')
self.node_embedding = initializations.get('normal')((config.node_num, config.node_embed_dim), name='node_embed', scale=0.1)
self.vocab_embedding_W = initializations.get('normal')((config.target_vocab_size, config.rule_embed_dim), name='vocab_embedding_W', scale=0.1)
self.vocab_embedding_b = shared_zeros(config.target_vocab_size, name='vocab_embedding_b')
# decoder_hidden_dim -> action embed
self.decoder_hidden_state_W_rule = Dense(config.decoder_hidden_dim, config.rule_embed_dim, name='decoder_hidden_state_W_rule')
# decoder_hidden_dim -> action embed
self.decoder_hidden_state_W_token= Dense(config.decoder_hidden_dim + config.encoder_hidden_dim, config.rule_embed_dim,
name='decoder_hidden_state_W_token')
# self.rule_encoder_lstm.params
self.params = self.query_embedding.params + self.query_encoder_lstm.params + \
self.decoder_lstm.params + self.src_ptr_net.params + self.terminal_gen_softmax.params + \
[self.rule_embedding_W, self.rule_embedding_b, self.node_embedding, self.vocab_embedding_W, self.vocab_embedding_b] + \
self.decoder_hidden_state_W_rule.params + self.decoder_hidden_state_W_token.params
self.srng = RandomStreams()
def __init__(self):
# self.node_embedding = Embedding(config.node_num, config.node_embed_dim, name='node_embed')
self.query_embedding = Embedding(config.source_vocab_size, config.word_embed_dim, name='query_embed')
if config.encoder_lstm == 'bilstm':
self.query_encoder_lstm = BiLSTM(config.word_embed_dim, config.encoder_hidden_dim / 2, return_sequences=True,
name='query_encoder_lstm')
else:
self.query_encoder_lstm = LSTM(config.word_embed_dim, config.encoder_hidden_dim, return_sequences=True,
name='query_encoder_lstm')
self.decoder_lstm = CondAttLSTM(config.rule_embed_dim + config.node_embed_dim + config.rule_embed_dim,
config.decoder_hidden_dim, config.encoder_hidden_dim, config.attention_hidden_dim,
name='decoder_lstm')
self.src_ptr_net = PointerNet()
self.terminal_gen_softmax = Dense(config.decoder_hidden_dim, 2, activation='softmax', name='terminal_gen_softmax')
self.rule_embedding_W = initializations.get('normal')((config.rule_num, config.rule_embed_dim), name='rule_embedding_W', scale=0.1)
self.rule_embedding_b = shared_zeros(config.rule_num, name='rule_embedding_b')
self.node_embedding = initializations.get('normal')((config.node_num, config.node_embed_dim), name='node_embed', scale=0.1)
self.vocab_embedding_W = initializations.get('normal')((config.target_vocab_size, config.rule_embed_dim), name='vocab_embedding_W', scale=0.1)
self.vocab_embedding_b = shared_zeros(config.target_vocab_size, name='vocab_embedding_b')
# decoder_hidden_dim -> action embed
self.decoder_hidden_state_W_rule = Dense(config.decoder_hidden_dim, config.rule_embed_dim, name='decoder_hidden_state_W_rule')
# decoder_hidden_dim -> action embed
self.decoder_hidden_state_W_token= Dense(config.decoder_hidden_dim + config.encoder_hidden_dim, config.rule_embed_dim,
name='decoder_hidden_state_W_token')
# self.rule_encoder_lstm.params
self.params = self.query_embedding.params + self.query_encoder_lstm.params + \
self.decoder_lstm.params + self.src_ptr_net.params + self.terminal_gen_softmax.params + \
[self.rule_embedding_W, self.rule_embedding_b, self.node_embedding, self.vocab_embedding_W, self.vocab_embedding_b] + \
self.decoder_hidden_state_W_rule.params + self.decoder_hidden_state_W_token.params
self.srng = RandomStreams()
def __init__(self, input_dim, output_dim, init='uniform', name=None):
super(Embedding, self).__init__()
self.init = initializations.get(init)
self.input_dim = input_dim
self.output_dim = output_dim
self.W = self.init((self.input_dim, self.output_dim), scale=0.1)
self.params = [self.W]
if name is not None:
self.set_name(name)
def __init__(self, input_dim, output_dim, init='uniform', name=None):
super(Embedding, self).__init__()
self.init = initializations.get(init)
self.input_dim = input_dim
self.output_dim = output_dim
self.W = self.init((self.input_dim, self.output_dim), scale=0.1)
self.params = [self.W]
if name is not None:
self.set_name(name)
def __init__(self, input_dim, output_dim, init='glorot_uniform', activation='tanh', name='Dense'):
super(Dense, self).__init__()
self.init = initializations.get(init)
self.activation = activations.get(activation)
self.input_dim = input_dim
self.output_dim = output_dim
self.input = T.matrix()
self.W = self.init((self.input_dim, self.output_dim))
self.b = shared_zeros((self.output_dim))
self.params = [self.W, self.b]
if name is not None:
self.set_name(name)
def __init__(self, embed_size, unfixed_embed_size, embed_dim, init='uniform', name='HybridEmbedding'):
super(HybridEmbedding, self).__init__()
self.init = initializations.get(init)
self.unfixed_embed_size = unfixed_embed_size
self.W_unfixed = self.init((embed_size, embed_dim))
self.W_fixed = self.init((embed_size, embed_dim))
self.W_fixed.name = 'HybridEmbedding_fiexed_embed_matrix'
# print W_fixed
# for id, row in enumerate(self.W_fixed.get_value()):
# if id >= 400: print '[word %d]' % id, row
self.params = [self.W_unfixed]
if name is not None:
self.set_name(name)
def __init__(self, max_sent_len, word_embed_dim, filter_num, filter_window_size,
border_mode='valid', activation='relu', name='Convolution2d'):
super(Convolution2d, self).__init__()
self.init = initializations.get('uniform')
self.activation = activations.get(activation)
self.border_mode = border_mode
self.W = self.init((filter_num, 1, filter_window_size, word_embed_dim), scale=0.01, name='W')
self.b = shared_zeros((filter_num), name='b')
self.params = [self.W, self.b]
if self.border_mode == 'valid':
self.ds = (max_sent_len - filter_window_size + 1, 1)
elif self.border_mode == 'full':
self.ds = (max_sent_len + filter_window_size - 1, 1)
if name is not None:
self.set_name(name)
def __init__(self,
embed_size,
unfixed_embed_size,
embed_dim,
init='uniform',
name='HybridEmbedding'):
super(HybridEmbedding, self).__init__()
self.init = initializations.get(init)
self.unfixed_embed_size = unfixed_embed_size
self.W_unfixed = self.init((embed_size, embed_dim))
self.W_fixed = self.init((embed_size, embed_dim))
self.W_fixed.name = 'HybridEmbedding_fiexed_embed_matrix'
# print W_fixed
# for id, row in enumerate(self.W_fixed.get_value()):
# if id >= 400: print '[word %d]' % id, row
self.params = [self.W_unfixed]
if name is not None:
self.set_name(name)
def __init__(self):
# self.node_embedding = Embedding(config.node_num, config.node_embed_dim, name='node_embed')
self.query_embedding = Embedding(config.source_vocab_size,
config.word_embed_dim,
name='query_embed')
encoder_dim = config.word_embed_dim
logging.info("Concatenation type: %s" % config.concat_type)
logging.info("Include canon_id matrix: %s" % config.include_cid)
if config.concat_type == 'basic':
encoder_dim += 2
if config.include_cid == True:
encoder_dim += 1
else:
# define layers
self.query_phrase_embedding = Embedding(14,
8,
name='query_phrase_embed')
self.query_pos_embedding = Embedding(44,
32,
name='query_pos_embed')
self.query_canon_embedding = Embedding(
102, 64, name='query_canon_embedding')
aug_dim = 8 + 32
if config.include_cid == True:
aug_dim += 64
self.projector = Dense(config.word_embed_dim + aug_dim,
config.word_embed_dim,
activation='linear',
name='concat_projector')
if config.encoder == 'bilstm':
self.query_encoder_lstm = BiLSTM(encoder_dim,
config.encoder_hidden_dim / 2,
return_sequences=True,
name='query_encoder_lstm')
else:
self.query_encoder_lstm = LSTM(encoder_dim,
config.encoder_hidden_dim,
return_sequences=True,
name='query_encoder_lstm')
self.decoder_lstm = CondAttLSTM(config.rule_embed_dim +
config.node_embed_dim +
config.rule_embed_dim,
config.decoder_hidden_dim,
config.encoder_hidden_dim,
config.attention_hidden_dim,
name='decoder_lstm')
self.src_ptr_net = PointerNet()
self.terminal_gen_softmax = Dense(config.decoder_hidden_dim,
2,
activation='softmax',
name='terminal_gen_softmax')
self.rule_embedding_W = initializations.get('normal')(
(config.rule_num, config.rule_embed_dim),
name='rule_embedding_W',
scale=0.1)
self.rule_embedding_b = shared_zeros(config.rule_num,
name='rule_embedding_b')
self.node_embedding = initializations.get('normal')(
(config.node_num, config.node_embed_dim),
name='node_embed',
scale=0.1)
self.vocab_embedding_W = initializations.get('normal')(
(config.target_vocab_size, config.rule_embed_dim),
name='vocab_embedding_W',
scale=0.1)
self.vocab_embedding_b = shared_zeros(config.target_vocab_size,
name='vocab_embedding_b')
# decoder_hidden_dim -> action embed
self.decoder_hidden_state_W_rule = Dense(
config.decoder_hidden_dim,
config.rule_embed_dim,
name='decoder_hidden_state_W_rule')
# decoder_hidden_dim -> action embed
self.decoder_hidden_state_W_token = Dense(
config.decoder_hidden_dim + config.encoder_hidden_dim,
config.rule_embed_dim,
name='decoder_hidden_state_W_token')
# self.rule_encoder_lstm.params
self.params = self.query_embedding.params + self.query_encoder_lstm.params + \
self.decoder_lstm.params + self.src_ptr_net.params + self.terminal_gen_softmax.params + \
[self.rule_embedding_W, self.rule_embedding_b, self.node_embedding, self.vocab_embedding_W, self.vocab_embedding_b] + \
self.decoder_hidden_state_W_rule.params + self.decoder_hidden_state_W_token.params
self.srng = RandomStreams()
def __init__(self,
input_dim,
output_dim,
context_dim,
att_hidden_dim,
init='glorot_uniform',
inner_init='orthogonal',
forget_bias_init='one',
activation='tanh',
inner_activation='sigmoid',
name='CondAttLSTM'):
super(CondAttLSTM, self).__init__()
self.output_dim = output_dim
self.init = initializations.get(init)
self.inner_init = initializations.get(inner_init)
self.forget_bias_init = initializations.get(forget_bias_init)
self.activation = activations.get(activation)
self.inner_activation = activations.get(inner_activation)
self.context_dim = context_dim
self.input_dim = input_dim
# regular LSTM layer
self.W_i = self.init((input_dim, self.output_dim))
self.U_i = self.inner_init((self.output_dim, self.output_dim))
self.C_i = self.inner_init((self.context_dim, self.output_dim))
self.H_i = self.inner_init((self.output_dim, self.output_dim))
self.P_i = self.inner_init((self.output_dim, self.output_dim))
self.b_i = shared_zeros((self.output_dim))
self.W_f = self.init((input_dim, self.output_dim))
self.U_f = self.inner_init((self.output_dim, self.output_dim))
self.C_f = self.inner_init((self.context_dim, self.output_dim))
self.H_f = self.inner_init((self.output_dim, self.output_dim))
self.P_f = self.inner_init((self.output_dim, self.output_dim))
self.b_f = self.forget_bias_init((self.output_dim))
self.W_c = self.init((input_dim, self.output_dim))
self.U_c = self.inner_init((self.output_dim, self.output_dim))
self.C_c = self.inner_init((self.context_dim, self.output_dim))
self.H_c = self.inner_init((self.output_dim, self.output_dim))
self.P_c = self.inner_init((self.output_dim, self.output_dim))
self.b_c = shared_zeros((self.output_dim))
self.W_o = self.init((input_dim, self.output_dim))
self.U_o = self.inner_init((self.output_dim, self.output_dim))
self.C_o = self.inner_init((self.context_dim, self.output_dim))
self.H_o = self.inner_init((self.output_dim, self.output_dim))
self.P_o = self.inner_init((self.output_dim, self.output_dim))
self.b_o = shared_zeros((self.output_dim))
self.params = [
self.W_i,
self.U_i,
self.b_i,
self.C_i,
self.H_i,
self.P_i,
self.W_c,
self.U_c,
self.b_c,
self.C_c,
self.H_c,
self.P_c,
self.W_f,
self.U_f,
self.b_f,
self.C_f,
self.H_f,
self.P_f,
self.W_o,
self.U_o,
self.b_o,
self.C_o,
self.H_o,
self.P_o,
]
# attention layer
self.att_ctx_W1 = self.init((context_dim, att_hidden_dim))
self.att_h_W1 = self.init((output_dim, att_hidden_dim))
self.att_b1 = shared_zeros((att_hidden_dim))
self.att_W2 = self.init((att_hidden_dim, 1))
self.att_b2 = shared_zeros((1))
self.params += [
self.att_ctx_W1, self.att_h_W1, self.att_b1, self.att_W2,
self.att_b2
]
# attention over history
self.hatt_h_W1 = self.init((output_dim, att_hidden_dim))
self.hatt_hist_W1 = self.init((output_dim, att_hidden_dim))
self.hatt_b1 = shared_zeros((att_hidden_dim))
self.hatt_W2 = self.init((att_hidden_dim, 1))
self.hatt_b2 = shared_zeros((1))
self.params += [
self.hatt_h_W1, self.hatt_hist_W1, self.hatt_b1, self.hatt_W2,
self.hatt_b2
]
self.set_name(name)
def __init__(self, input_dim, output_dim,
context_dim, att_hidden_dim,
init='glorot_uniform', inner_init='orthogonal', forget_bias_init='one',
activation='tanh', inner_activation='sigmoid', name='CondAttLSTM'):
super(CondAttLSTM, self).__init__()
self.output_dim = output_dim
self.init = initializations.get(init)
self.inner_init = initializations.get(inner_init)
self.forget_bias_init = initializations.get(forget_bias_init)
self.activation = activations.get(activation)
self.inner_activation = activations.get(inner_activation)
self.context_dim = context_dim
self.input_dim = input_dim
# regular LSTM layer
self.W_i = self.init((input_dim, self.output_dim))
self.U_i = self.inner_init((self.output_dim, self.output_dim))
self.C_i = self.inner_init((self.context_dim, self.output_dim))
self.H_i = self.inner_init((self.output_dim, self.output_dim))
self.P_i = self.inner_init((self.output_dim, self.output_dim))
self.b_i = shared_zeros((self.output_dim))
self.W_f = self.init((input_dim, self.output_dim))
self.U_f = self.inner_init((self.output_dim, self.output_dim))
self.C_f = self.inner_init((self.context_dim, self.output_dim))
self.H_f = self.inner_init((self.output_dim, self.output_dim))
self.P_f = self.inner_init((self.output_dim, self.output_dim))
self.b_f = self.forget_bias_init((self.output_dim))
self.W_c = self.init((input_dim, self.output_dim))
self.U_c = self.inner_init((self.output_dim, self.output_dim))
self.C_c = self.inner_init((self.context_dim, self.output_dim))
self.H_c = self.inner_init((self.output_dim, self.output_dim))
self.P_c = self.inner_init((self.output_dim, self.output_dim))
self.b_c = shared_zeros((self.output_dim))
self.W_o = self.init((input_dim, self.output_dim))
self.U_o = self.inner_init((self.output_dim, self.output_dim))
self.C_o = self.inner_init((self.context_dim, self.output_dim))
self.H_o = self.inner_init((self.output_dim, self.output_dim))
self.P_o = self.inner_init((self.output_dim, self.output_dim))
self.b_o = shared_zeros((self.output_dim))
self.params = [
self.W_i, self.U_i, self.b_i, self.C_i, self.H_i, self.P_i,
self.W_c, self.U_c, self.b_c, self.C_c, self.H_c, self.P_c,
self.W_f, self.U_f, self.b_f, self.C_f, self.H_f, self.P_f,
self.W_o, self.U_o, self.b_o, self.C_o, self.H_o, self.P_o,
]
# attention layer
self.att_ctx_W1 = self.init((context_dim, att_hidden_dim))
self.att_h_W1 = self.init((output_dim, att_hidden_dim))
self.att_b1 = shared_zeros((att_hidden_dim))
self.att_W2 = self.init((att_hidden_dim, 1))
self.att_b2 = shared_zeros((1))
self.params += [
self.att_ctx_W1, self.att_h_W1, self.att_b1,
self.att_W2, self.att_b2
]
# attention over history
self.hatt_h_W1 = self.init((output_dim, att_hidden_dim))
self.hatt_hist_W1 = self.init((output_dim, att_hidden_dim))
self.hatt_b1 = shared_zeros((att_hidden_dim))
self.hatt_W2 = self.init((att_hidden_dim, 1))
self.hatt_b2 = shared_zeros((1))
self.params += [
self.hatt_h_W1, self.hatt_hist_W1, self.hatt_b1,
self.hatt_W2, self.hatt_b2
]
self.set_name(name)
