def __init__(self, name, x, y, lr, init_emb, vocab_size, emb_dim,
hidden_dim, output_dim, window, opt):
assert window % 2 == 1, 'Window size must be odd'
""" input """
self.name = name
self.x = x
self.y = y
self.lr = lr
self.input = [self.x, self.y, self.lr]
n_words = x.shape[0]
""" params """
if init_emb is not None:
self.emb = theano.shared(init_emb)
else:
self.emb = theano.shared(sample_weights(vocab_size, emb_dim))
self.W_in = theano.shared(
sample_weights(hidden_dim, 1, window, emb_dim))
self.W_out = theano.shared(sample_weights(hidden_dim, output_dim))
self.b_in = theano.shared(sample_weights(hidden_dim, 1))
self.b_y = theano.shared(sample_weights(output_dim))
self.params = [self.W_in, self.W_out, self.b_in, self.b_y]
""" pad """
self.zero = theano.shared(
np.zeros(shape=(1, 1, window / 2, emb_dim),
dtype=theano.config.floatX))
""" look up embedding """
self.x_emb = self.emb[self.x] # x_emb: 1D: n_words, 2D: n_emb
""" convolution """
self.x_in = self.conv(self.x_emb)
""" feed-forward computation """
self.h = relu(
self.x_in.reshape((self.x_in.shape[1], self.x_in.shape[2])) +
T.repeat(self.b_in, T.cast(self.x_in.shape[2], 'int32'), 1)).T
self.o = T.dot(self.h, self.W_out) + self.b_y
self.p_y_given_x = T.nnet.softmax(self.o)
""" prediction """
self.y_pred = T.argmax(self.o, axis=1)
self.result = T.eq(self.y_pred, self.y)
""" cost function """
self.nll = -T.sum(T.log(self.p_y_given_x)[T.arange(n_words), self.y])
self.cost = self.nll
if opt == 'sgd':
self.updates = sgd(self.cost, self.params, self.emb, self.x_emb,
self.lr)
else:
self.updates = ada_grad(self.cost, self.params, self.emb,
self.x_emb, self.x, self.lr)
def __init__(self, name, x, y, lr, init_emb, vocab_size, emb_dim, hidden_dim, output_dim, window, opt):
assert window % 2 == 1, 'Window size must be odd'
""" input """
self.name = name
self.x = x
self.y = y
self.lr = lr
self.input = [self.x, self.y, self.lr]
n_words = x.shape[0]
""" params """
if init_emb is not None:
self.emb = theano.shared(init_emb)
else:
self.emb = theano.shared(sample_weights(vocab_size, emb_dim))
self.W_in = theano.shared(sample_weights(hidden_dim, 1, window, emb_dim))
self.W_out = theano.shared(sample_weights(hidden_dim, output_dim))
self.b_in = theano.shared(sample_weights(hidden_dim, 1))
self.b_y = theano.shared(sample_weights(output_dim))
self.params = [self.W_in, self.W_out, self.b_in, self.b_y]
""" pad """
self.zero = theano.shared(np.zeros(shape=(1, 1, window / 2, emb_dim), dtype=theano.config.floatX))
""" look up embedding """
self.x_emb = self.emb[self.x] # x_emb: 1D: n_words, 2D: n_emb
""" convolution """
self.x_in = self.conv(self.x_emb)
""" feed-forward computation """
self.h = relu(self.x_in.reshape((self.x_in.shape[1], self.x_in.shape[2])) + T.repeat(self.b_in, T.cast(self.x_in.shape[2], 'int32'), 1)).T
self.o = T.dot(self.h, self.W_out) + self.b_y
self.p_y_given_x = T.nnet.softmax(self.o)
""" prediction """
self.y_pred = T.argmax(self.o, axis=1)
self.result = T.eq(self.y_pred, self.y)
""" cost function """
self.nll = -T.sum(T.log(self.p_y_given_x)[T.arange(n_words), self.y])
self.cost = self.nll
if opt == 'sgd':
self.updates = sgd(self.cost, self.params, self.emb, self.x_emb, self.lr)
else:
self.updates = ada_grad(self.cost, self.params, self.emb, self.x_emb, self.x, self.lr)
def __init__(self, n_h, pooling, k=2):
self.k = k
# self.W_m = theano.shared(sample_weights(n_h, n_h))
self.W_m1 = theano.shared(sample_weights(n_h, n_h))
self.W_m2 = theano.shared(sample_weights(n_h, n_h))
self.W_m3 = theano.shared(sample_weights(n_h, n_h))
# self.W_c = theano.shared(sample_weights(n_h * (k+1), n_h))
self.W_c = theano.shared(sample_weights(n_h * 2, n_h))
# self.W_k = theano.shared(sample_weights(n_h, n_h))
self.pooling = pooling
# self.params = [self.W_m, self.W_c]
self.params = [self.W_m1, self.W_m2, self.W_m3, self.W_c]
def __init__(self, x, y, n_words, batch_size, lr, init_emb, vocab_size,
emb_dim, hidden_dim, output_dim, window, opt):
assert window % 2 == 1, 'Window size must be odd'
""" input """
self.x = x # 1D: n_words * batch_size, 2D: window; elem=word id
self.x_v = x.flatten(
) # 1D: n_words * batch_size * window; elem=word id
self.y = y
self.batch_size = batch_size
self.n_words = n_words
self.lr = lr
""" params """
if init_emb is not None:
self.emb = theano.shared(init_emb)
else:
self.emb = theano.shared(sample_weights(vocab_size, emb_dim))
self.W_in = theano.shared(sample_weights(emb_dim * window, hidden_dim))
self.W_out = theano.shared(sample_weights(hidden_dim, output_dim))
self.b_in = theano.shared(sample_weights(hidden_dim))
self.b_y = theano.shared(sample_weights(output_dim))
self.params = [self.W_in, self.W_out, self.b_in, self.b_y]
""" look up embedding """
self.x_emb = self.emb[
self.x_v] # x_emb: 1D: batch_size * n_words * window, 2D: emb_dim
""" forward """
self.h = relu(
T.dot(self.x_emb.reshape((batch_size * n_words, emb_dim *
window)), self.W_in) + self.b_in)
self.o = T.dot(self.h, self.W_out) + self.b_y
self.p_y_given_x = T.nnet.softmax(self.o)
""" predict """
self.y_pred = T.argmax(self.o, axis=1)
self.result = T.eq(self.y_pred, self.y)
""" loss """
self.log_p = T.log(self.p_y_given_x)[T.arange(batch_size * n_words),
self.y]
self.nll = -T.sum(self.log_p)
self.cost = self.nll
if opt == 'sgd':
self.updates = sgd(self.cost, self.params, self.emb, self.x_emb,
self.lr)
else:
self.updates = ada_grad(self.cost, self.params, self.emb,
self.x_emb, self.x, self.lr)
def __init__(self, x, y, n_words, batch_size, lr, init_emb, vocab_size, emb_dim, hidden_dim, output_dim, window, opt):
assert window % 2 == 1, 'Window size must be odd'
""" input """
self.x = x # 1D: n_words * batch_size, 2D: window; elem=word id
self.x_v = x.flatten() # 1D: n_words * batch_size * window; elem=word id
self.y = y
self.batch_size = batch_size
self.n_words = n_words
self.lr = lr
""" params """
if init_emb is not None:
self.emb = theano.shared(init_emb)
else:
self.emb = theano.shared(sample_weights(vocab_size, emb_dim))
self.W_in = theano.shared(sample_weights(emb_dim * window, hidden_dim))
self.W_out = theano.shared(sample_weights(hidden_dim, output_dim))
self.b_in = theano.shared(sample_weights(hidden_dim))
self.b_y = theano.shared(sample_weights(output_dim))
self.params = [self.W_in, self.W_out, self.b_in, self.b_y]
""" look up embedding """
self.x_emb = self.emb[self.x_v] # x_emb: 1D: batch_size * n_words * window, 2D: emb_dim
""" forward """
self.h = relu(T.dot(self.x_emb.reshape((batch_size * n_words, emb_dim * window)), self.W_in) + self.b_in)
self.o = T.dot(self.h, self.W_out) + self.b_y
self.p_y_given_x = T.nnet.softmax(self.o)
""" predict """
self.y_pred = T.argmax(self.o, axis=1)
self.result = T.eq(self.y_pred, self.y)
""" loss """
self.log_p = T.log(self.p_y_given_x)[T.arange(batch_size * n_words), self.y]
self.nll = -T.sum(self.log_p)
self.cost = self.nll
if opt == 'sgd':
self.updates = sgd(self.cost, self.params, self.emb, self.x_emb, self.lr)
else:
self.updates = ada_grad(self.cost, self.params, self.emb, self.x_emb, self.x, self.lr)
def layers(x, window, dim_emb, dim_hidden, n_layers, activation=tanh):
params = []
zero = T.zeros((1, dim_emb * window), dtype=theano.config.floatX)
def zero_pad_gate(matrix):
return T.neq(T.sum(T.eq(matrix, zero), 1, keepdims=True), dim_emb * window)
for i in xrange(n_layers):
if i == 0:
W = theano.shared(sample_weights(dim_emb * window, dim_hidden))
# h = zero_pad_gate(x) * relu(T.dot(x, W))
h = relu(T.dot(x, W))
else:
W = theano.shared(sample_weights(dim_hidden, dim_hidden))
h = activation(T.dot(h, W))
params.append(W)
return h, params
def __init__(self, n_i=32, n_h=32, activation=tanh):
self.activation = activation
self.W = theano.shared(sample_weights(n_i, n_h))
self.W_xr = theano.shared(sample_weights(n_h, n_h))
self.W_hr = theano.shared(sample_weights(n_h, n_h))
self.W_xz = theano.shared(sample_weights(n_h, n_h))
self.W_hz = theano.shared(sample_weights(n_h, n_h))
self.W_xh = theano.shared(sample_weights(n_h, n_h))
self.W_hh = theano.shared(sample_weights(n_h, n_h))
self.params = [self.W, self.W_xr, self.W_hr, self.W_xz, self.W_hz, self.W_xh, self.W_hh]
def __init__(self, n_i=32, n_h=32, activation=tanh):
self.activation = activation
self.W = theano.shared(sample_weights(n_i, n_h))
self.W_xr = theano.shared(sample_weights(n_h, n_h))
self.W_hr = theano.shared(sample_weights(n_h, n_h))
self.W_xz = theano.shared(sample_weights(n_h, n_h))
self.W_hz = theano.shared(sample_weights(n_h, n_h))
self.W_xh = theano.shared(sample_weights(n_h, n_h))
self.W_hh = theano.shared(sample_weights(n_h, n_h))
self.params = [self.W, self.W_xr, self.W_hr, self.W_xz, self.W_hz, self.W_xh, self.W_hh]
def __init__(self, n_i, n_labels):
self.W = theano.shared(sample_weights(n_i, n_labels))
self.W_trans = theano.shared(sample_weights(n_labels, n_labels))
self.params = [self.W, self.W_trans]
def __init__(self, name, w, c, b, y, lr,
init_w_emb, vocab_w_size, vocab_c_size,
w_emb_dim, c_emb_dim, w_hidden_dim, c_hidden_dim, output_dim,
window, opt):
assert window % 2 == 1, 'Window size must be odd'
""" input """
self.name = name
self.w = w
self.c = c
self.b = b
self.y = y
self.lr = lr
self.input = [self.w, self.c, self.b, self.y, self.lr]
n_phi = w_emb_dim + c_emb_dim * window
n_words = w.shape[0]
""" params """
if init_w_emb is not None:
self.emb = theano.shared(init_w_emb)
else:
self.emb = theano.shared(sample_weights(vocab_w_size, w_emb_dim))
self.emb_c = theano.shared(sample_norm_dist(vocab_c_size, c_emb_dim))
self.W_in = theano.shared(sample_weights(w_hidden_dim, 1, window, n_phi))
self.W_c = theano.shared(sample_weights(c_hidden_dim, 1, window, c_emb_dim))
self.W_out = theano.shared(sample_weights(w_hidden_dim, output_dim))
self.b_in = theano.shared(sample_weights(w_hidden_dim, 1))
self.b_c = theano.shared(sample_weights(c_hidden_dim))
self.b_y = theano.shared(sample_weights(output_dim))
""" pad """
self.zero = theano.shared(np.zeros(shape=(1, 1, window / 2, n_phi), dtype=theano.config.floatX))
self.zero_c = theano.shared(np.zeros(shape=(1, 1, window / 2, c_emb_dim), dtype=theano.config.floatX))
self.params = [self.emb_c, self.W_in, self.W_c, self.W_out, self.b_in, self.b_c, self.b_y]
""" look up embedding """
x_emb = self.emb[self.w] # x_emb: 1D: n_words, 2D: w_emb_dim
c_emb = self.emb_c[self.c] # c_emb: 1D: n_chars, 2D: c_emb_dim
""" create feature """
c_phi = self.create_char_feature(self.b, c_emb, self.zero_c) + self.b_c # 1D: n_words, 2D: c_hidden_dim(50)
x_phi = T.concatenate([x_emb, c_phi], axis=1) # 1D: n_words, 2D: w_emb_dim(100) + c_hidden_dim(50)
""" convolution """
x_padded = T.concatenate([self.zero, x_phi.reshape((1, 1, x_phi.shape[0], x_phi.shape[1])), self.zero], axis=2) # x_padded: 1D: n_words + n_pad, 2D: n_phi
x_in = conv2d(input=x_padded, filters=self.W_in) # 1D: 1, 2D: w_hidden_dim(300), 3D: n_words, 4D: 1
""" feed-forward computation """
h = relu(x_in.reshape((x_in.shape[1], x_in.shape[2])) + T.repeat(self.b_in, T.cast(x_in.shape[2], 'int32'), 1)).T
self.o = T.dot(h, self.W_out) + self.b_y
self.p_y_given_x = T.nnet.softmax(self.o)
""" prediction """
self.y_pred = T.argmax(self.o, axis=1)
self.result = T.eq(self.y_pred, self.y)
""" cost function """
self.nll = -T.sum(T.log(self.p_y_given_x)[T.arange(n_words), self.y])
self.cost = self.nll
if opt == 'sgd':
self.updates = sgd(self.cost, self.params, self.emb, x_emb, self.lr)
else:
self.updates = ada_grad(self.cost, self.params, self.emb, x_emb, self.w, self.lr)
def __init__(self, n_i=32, n_h=32, activation=tanh):
self.activation = activation
self.W = theano.shared(sample_weights(n_i, n_h))
self.params = [self.W]
def __init__(self, x, c, y, n_words, batch_size, lr, init_emb,
vocab_w_size, w_emb_dim, w_hidden_dim, c_emb_dim,
c_hidden_dim, output_dim, vocab_c_size, window, opt):
assert window % 2 == 1, 'Window size must be odd'
""" input """
self.x = x # 1D: n_words * batch_size, 2D: window; elem=word id
self.x_v = x.flatten(
) # 1D: n_words * batch_size * window; elem=word id
self.c = c # 1D: n_words * batch_size, 2D: window, 3D: max_len_char, 4D: window; elem=char id
self.y = y
self.batch_size = batch_size
self.n_words = n_words
self.lr = lr
n_phi = (w_emb_dim + c_hidden_dim) * window
max_len_char = T.cast(self.c.shape[2], 'int32')
""" params """
if init_emb is not None:
self.emb = theano.shared(init_emb)
else:
self.emb = theano.shared(sample_weights(vocab_w_size, w_emb_dim))
self.pad = build_shared_zeros((1, c_emb_dim))
self.e_c = theano.shared(sample_norm_dist(vocab_c_size - 1, c_emb_dim))
self.emb_c = T.concatenate([self.pad, self.e_c], 0)
self.W_in = theano.shared(sample_weights(n_phi, w_hidden_dim))
self.W_c = theano.shared(
sample_weights(c_emb_dim * window, c_hidden_dim))
self.W_out = theano.shared(sample_weights(w_hidden_dim, output_dim))
self.b_in = theano.shared(sample_weights(w_hidden_dim))
self.b_c = theano.shared(sample_weights(c_hidden_dim))
self.b_y = theano.shared(sample_weights(output_dim))
self.params = [
self.e_c, self.W_in, self.W_c, self.W_out, self.b_in, self.b_c,
self.b_y
]
""" look up embedding """
self.x_emb = self.emb[
self.x_v] # 1D: batch_size*n_words * window, 2D: emb_dim
self.c_emb = self.emb_c[
self.
c] # 1D: batch_size*n_words, 2D: window, 3D: max_len_char, 4D: window, 5D: n_c_emb
self.x_emb_r = self.x_emb.reshape((x.shape[0], x.shape[1], -1))
""" convolution """
self.c_phi = T.max(
T.dot(
self.c_emb.reshape(
(batch_size * n_words, window, max_len_char, -1)),
self.W_c) + self.b_c, 2) # 1D: n_words, 2D: window, 3D: n_h_c
self.x_phi = T.concatenate([self.x_emb_r, self.c_phi], axis=2)
""" forward """
self.h = relu(
T.dot(self.x_phi.reshape((batch_size * n_words,
n_phi)), self.W_in) + self.b_in)
self.o = T.dot(self.h, self.W_out) + self.b_y
self.p_y_given_x = T.nnet.softmax(self.o)
""" predict """
self.y_pred = T.argmax(self.o, axis=1)
self.result = T.eq(self.y_pred, self.y)
""" loss """
self.log_p = T.log(self.p_y_given_x)[T.arange(batch_size * n_words),
self.y]
self.nll = -T.sum(self.log_p)
self.cost = self.nll
if opt == 'sgd':
self.updates = sgd(self.cost, self.params, self.emb, self.x_emb,
self.lr)
else:
self.updates = ada_grad(self.cost, self.params, self.emb,
self.x_emb, self.x, self.lr)
def __init__(self, x_span, x_word, x_ctx, x_dist, x_slen, y, init_emb, n_vocab, dim_w_p, dim_d, dim_h, L2_reg):
"""
:param x_span: 1D: batch, 2D: limit * 2 (10); elem=word id
:param x_word: 1D: batch, 2D: 4 (m_first, m_last, a_first, a_last); elem=word id
:param x_ctx : 1D: batch, 2D: window * 2 * 2 (20); elem=word id
:param x_dist: 1D: batch; 2D: 2; elem=[sent dist, ment dist]
:param x_slen: 1D: batch; 2D: 3; elem=[m_span_len, a_span_len, head_match]
:param y : 1D: batch
"""
self.input = [x_span, x_word, x_ctx, x_dist, x_slen, y]
self.x_span = x_span
self.x_word = x_word
self.x_ctx = x_ctx
self.x_dist = x_dist
self.x_slen = x_slen
self.y = y
""" Dimensions """
dim_w_a = dim_w_p / 5
dim_x_a = dim_w_a * (5 + 2 + 2 + 1)
dim_x_p = dim_w_p * (10 + 4 + 4 + 2 + 3) + dim_x_a
batch = y.shape[0]
""" Hyper Parameters for Cost Function """
self.a1 = 0.5
self.a2 = 1.2
self.a3 = 1.0
""" Params """
if init_emb is None:
self.W_a_w = theano.shared(sample_weights(n_vocab, dim_w_a))
self.W_p_w = theano.shared(sample_weights(n_vocab, dim_w_p))
else:
self.W_a_w = theano.shared(init_emb)
self.W_p_w = theano.shared(init_emb)
self.W_a_l = theano.shared(sample_weights(5, dim_w_a))
self.W_a_o = theano.shared(sample_weights(dim_x_a, 1))
self.W_p_d = theano.shared(sample_weights(dim_d, dim_w_p))
self.W_p_l = theano.shared(sample_weights(7, dim_w_p))
self.W_p_h = theano.shared(sample_weights(dim_x_p, dim_h))
self.W_p_o = theano.shared(sample_weights(dim_h))
self.params = [self.W_p_d, self.W_p_l, self.W_a_l, self.W_p_h, self.W_p_o, self.W_a_o]
""" Anaphoric Layer """
x_vec_a = T.concatenate(
[x_span[0][: x_span.shape[1] / 2], x_word[0][: x_word.shape[1] / 2], x_ctx[0][: x_ctx.shape[1] / 2]]
)
x_a_w = self.W_a_w[x_vec_a] # 1D: batch, 2D: (limit * 1 + 2 + ctx), 3D: dim_w_a
x_a_l = self.W_a_l[x_slen[0][0]] # 1D: dim_w_a
h_a = T.concatenate([x_a_w.flatten(), x_a_l])
""" Pair Layer """
x_p_w_in = T.concatenate([x_span, x_word, x_ctx], 1).flatten() # 1D: batch * (limit * 2 + 4 + 20)
x_p_w = self.W_p_w[x_p_w_in] # 1D: batch, 2D: (limit * 2 + 4 + ctx * 2), 3D: dim_w
x_p_l = self.W_p_l[x_slen] # 1D: batch, 2D: 3, 3D: dim_w
x_p_d = self.W_p_d[x_dist] # 1D: batch, 2D: 2, 3D: dim_w
h_p = T.concatenate([x_p_w.reshape((batch, -1)), x_p_d.reshape((batch, -1)), x_p_l.reshape((batch, -1))], 1)
g_p = tanh(T.dot(T.concatenate([h_p, T.repeat(h_a.dimshuffle("x", 0), batch, 0)], 1), self.W_p_h))
""" Output Layer """
p_y_a = T.dot(h_a, self.W_a_o) # p_y_a: 1D: 1; elem=scalar
p_y_p = T.dot(g_p, self.W_p_o) # p_y_p: 1D: batch
p_y = T.concatenate([p_y_a, p_y_p])
""" Label Set """
y_0 = T.switch(T.sum(y), 0, 1) # y_0: 1 if the mention is non-anaph else 0
y_all = T.concatenate([y_0.dimshuffle("x"), y])
""" Predicts """
self.y_hat = T.argmax(p_y)
self.p_y_hat = p_y[T.argmax(p_y - T.min(p_y) * y_all)]
""" Cost Function """
self.nll = T.max(self.miss_cost(T.arange(y_all.shape[0]), y_all) * (1 + p_y - self.p_y_hat))
self.cost = self.nll + L2_reg * L2_sqr(params=self.params) / 2
""" Optimization """
self.updates = sgd_w(self.cost, self.params, self.W_p_w, x_p_w, self.W_a_w, x_a_w)
""" Check Results """
self.total_p = T.switch(self.y_hat, 1, 0)
self.total_r = 1 - y_0
self.correct = y_all[self.y_hat]
self.correct_t = T.switch(self.correct, T.switch(y_0, 0, 1), 0)
self.correct_f = T.switch(self.correct, T.switch(y_0, 1, 0), 0)
def __init__(self, x_span, x_word, x_ctx, x_dist, x_slen, y, init_emb, n_vocab, dim_w_p, dim_d, dim_h, L2_reg):
"""
:param x_span: 1D: batch, 2D: limit * 2 (10); elem=word id
:param x_word: 1D: batch, 2D: 4 (m_first, m_last, a_first, a_last); elem=word id
:param x_ctx : 1D: batch, 2D: window * 2 * 2 (20); elem=word id
:param x_dist: 1D: batch; 2D: 2; elem=[sent dist, ment dist]
:param x_slen: 1D: batch; 2D: 3; elem=[m_span_len, a_span_len, head_match]
:param y : 1D: batch
"""
self.input = [x_span, x_word, x_ctx, x_dist, x_slen, y]
self.x_span = x_span
self.x_word = x_word
self.x_ctx = x_ctx
self.x_dist = x_dist
self.x_slen = x_slen
self.y = y
""" Dimensions """
dim_w_a = dim_w_p / 5
dim_x_a = dim_w_a * (5 + 2 + 2 + 1)
dim_x_p = dim_w_p * (10 + 4 + 4 + 2 + 3) + dim_x_a
batch = y.shape[0]
""" Hyper Parameters for Cost Function """
self.a1 = 0.5
self.a2 = 1.2
self.a3 = 1.
""" Params """
if init_emb is None:
self.W_a_w = theano.shared(sample_weights(n_vocab, dim_w_a))
self.W_p_w = theano.shared(sample_weights(n_vocab, dim_w_p))
else:
self.W_a_w = theano.shared(init_emb)
self.W_p_w = theano.shared(init_emb)
self.W_a_l = theano.shared(sample_weights(5, dim_w_a))
self.W_a_o = theano.shared(sample_weights(dim_x_a, 1))
self.W_p_d = theano.shared(sample_weights(dim_d, dim_w_p))
self.W_p_l = theano.shared(sample_weights(7, dim_w_p))
self.W_p_h = theano.shared(sample_weights(dim_x_p, dim_h))
self.W_p_o = theano.shared(sample_weights(dim_h))
self.params = [self.W_p_d, self.W_p_l, self.W_a_l, self.W_p_h, self.W_p_o, self.W_a_o]
""" Anaphoric Layer """
x_vec_a = T.concatenate([x_span[0][:x_span.shape[1]/2],
x_word[0][:x_word.shape[1]/2],
x_ctx[0][:x_ctx.shape[1]/2]])
x_a_w = self.W_a_w[x_vec_a] # 1D: batch, 2D: (limit * 1 + 2 + ctx), 3D: dim_w_a
x_a_l = self.W_a_l[x_slen[0][0]] # 1D: dim_w_a
h_a = T.concatenate([x_a_w.flatten(), x_a_l])
""" Pair Layer """
x_p_w_in = T.concatenate([x_span, x_word, x_ctx], 1).flatten() # 1D: batch * (limit * 2 + 4 + 20)
x_p_w = self.W_p_w[x_p_w_in] # 1D: batch, 2D: (limit * 2 + 4 + ctx * 2), 3D: dim_w
x_p_l = self.W_p_l[x_slen] # 1D: batch, 2D: 3, 3D: dim_w
x_p_d = self.W_p_d[x_dist] # 1D: batch, 2D: 2, 3D: dim_w
h_p = T.concatenate([x_p_w.reshape((batch, -1)), x_p_d.reshape((batch, -1)), x_p_l.reshape((batch, -1))], 1)
g_p = tanh(T.dot(T.concatenate([h_p, T.repeat(h_a.dimshuffle('x', 0), batch, 0)], 1), self.W_p_h))
""" Output Layer """
p_y_a = T.dot(h_a, self.W_a_o) # p_y_a: 1D: 1; elem=scalar
p_y_p = T.dot(g_p, self.W_p_o) # p_y_p: 1D: batch
p_y = T.concatenate([p_y_a, p_y_p])
""" Label Set """
y_0 = T.switch(T.sum(y), 0, 1) # y_0: 1 if the mention is non-anaph else 0
y_all = T.concatenate([y_0.dimshuffle('x'), y])
""" Predicts """
self.y_hat = T.argmax(p_y)
self.p_y_hat = p_y[T.argmax(p_y - T.min(p_y) * y_all)]
""" Cost Function """
self.nll = T.max(self.miss_cost(T.arange(y_all.shape[0]), y_all) * (1 + p_y - self.p_y_hat))
self.cost = self.nll + L2_reg * L2_sqr(params=self.params) / 2
""" Optimization """
self.updates = sgd_w(self.cost, self.params, self.W_p_w, x_p_w, self.W_a_w, x_a_w)
""" Check Results """
self.total_p = T.switch(self.y_hat, 1, 0)
self.total_r = 1 - y_0
self.correct = y_all[self.y_hat]
self.correct_t = T.switch(self.correct, T.switch(y_0, 0, 1), 0)
self.correct_f = T.switch(self.correct, T.switch(y_0, 1, 0), 0)
def __init__(self, n_in, n_h, activation=tanh):
self.activation = activation
self.W_xi = theano.shared(sample_weights(n_in, n_h))
self.W_hi = theano.shared(sample_weights(n_h, n_h))
self.W_ci = theano.shared(sample_weights(n_h))
self.W_xf = theano.shared(sample_weights(n_in, n_h))
self.W_hf = theano.shared(sample_weights(n_h, n_h))
self.W_cf = theano.shared(sample_weights(n_h))
self.W_xc = theano.shared(sample_weights(n_in, n_h))
self.W_hc = theano.shared(sample_weights(n_h, n_h))
self.W_xo = theano.shared(sample_weights(n_in, n_h))
self.W_ho = theano.shared(sample_weights(n_h, n_h))
self.W_co = theano.shared(sample_weights(n_h))
self.params = [self.W_xi, self.W_hi, self.W_ci, self.W_xf, self.W_hf, self.W_cf,
self.W_xc, self.W_hc, self.W_xo, self.W_ho, self.W_co]
def __init__(self, x, c, y, n_words, batch_size, lr, init_emb, vocab_w_size, w_emb_dim, w_hidden_dim,
c_emb_dim, c_hidden_dim, output_dim, vocab_c_size, window, opt):
assert window % 2 == 1, 'Window size must be odd'
""" input """
self.x = x # 1D: n_words * batch_size, 2D: window; elem=word id
self.x_v = x.flatten() # 1D: n_words * batch_size * window; elem=word id
self.c = c # 1D: n_words * batch_size, 2D: window, 3D: max_len_char, 4D: window; elem=char id
self.y = y
self.batch_size = batch_size
self.n_words = n_words
self.lr = lr
n_phi = (w_emb_dim + c_hidden_dim) * window
max_len_char = T.cast(self.c.shape[2], 'int32')
""" params """
if init_emb is not None:
self.emb = theano.shared(init_emb)
else:
self.emb = theano.shared(sample_weights(vocab_w_size, w_emb_dim))
self.pad = build_shared_zeros((1, c_emb_dim))
self.e_c = theano.shared(sample_norm_dist(vocab_c_size - 1, c_emb_dim))
self.emb_c = T.concatenate([self.pad, self.e_c], 0)
self.W_in = theano.shared(sample_weights(n_phi, w_hidden_dim))
self.W_c = theano.shared(sample_weights(c_emb_dim * window, c_hidden_dim))
self.W_out = theano.shared(sample_weights(w_hidden_dim, output_dim))
self.b_in = theano.shared(sample_weights(w_hidden_dim))
self.b_c = theano.shared(sample_weights(c_hidden_dim))
self.b_y = theano.shared(sample_weights(output_dim))
self.params = [self.e_c, self.W_in, self.W_c, self.W_out, self.b_in, self.b_c, self.b_y]
""" look up embedding """
self.x_emb = self.emb[self.x_v] # 1D: batch_size*n_words * window, 2D: emb_dim
self.c_emb = self.emb_c[self.c] # 1D: batch_size*n_words, 2D: window, 3D: max_len_char, 4D: window, 5D: n_c_emb
self.x_emb_r = self.x_emb.reshape((x.shape[0], x.shape[1], -1))
""" convolution """
self.c_phi = T.max(T.dot(self.c_emb.reshape((batch_size * n_words, window, max_len_char, -1)), self.W_c) + self.b_c, 2) # 1D: n_words, 2D: window, 3D: n_h_c
self.x_phi = T.concatenate([self.x_emb_r, self.c_phi], axis=2)
""" forward """
self.h = relu(T.dot(self.x_phi.reshape((batch_size * n_words, n_phi)), self.W_in) + self.b_in)
self.o = T.dot(self.h, self.W_out) + self.b_y
self.p_y_given_x = T.nnet.softmax(self.o)
""" predict """
self.y_pred = T.argmax(self.o, axis=1)
self.result = T.eq(self.y_pred, self.y)
""" loss """
self.log_p = T.log(self.p_y_given_x)[T.arange(batch_size * n_words), self.y]
self.nll = -T.sum(self.log_p)
self.cost = self.nll
if opt == 'sgd':
self.updates = sgd(self.cost, self.params, self.emb, self.x_emb, self.lr)
else:
self.updates = ada_grad(self.cost, self.params, self.emb, self.x_emb, self.x, self.lr)
def __init__(self, n_i, n_h, activation=tanh):
self.activation = activation
self.W = theano.shared(sample_weights(n_i, n_h))
"""input gate parameters"""
self.W_xi = theano.shared(sample_weights(n_h, n_h))
self.W_hi = theano.shared(sample_weights(n_h, n_h))
self.W_ci = theano.shared(sample_weights(n_h))
"""forget gate parameters"""
self.W_xf = theano.shared(sample_weights(n_h, n_h))
self.W_hf = theano.shared(sample_weights(n_h, n_h))
self.W_cf = theano.shared(sample_weights(n_h))
"""cell parameters"""
self.W_xc = theano.shared(sample_weights(n_h, n_h))
self.W_hc = theano.shared(sample_weights(n_h, n_h))
"""output gate parameters"""
self.W_xo = theano.shared(sample_weights(n_h, n_h))
self.W_ho = theano.shared(sample_weights(n_h, n_h))
self.W_co = theano.shared(sample_weights(n_h))
self.params = [self.W, self.W_xi, self.W_hi, self.W_ci, self.W_xf, self.W_hf, self.W_cf,
self.W_xc, self.W_hc, self.W_xo, self.W_ho, self.W_co]
def __init__(self, n_in, n_h, activation=tanh):
self.activation = activation
self.W_xi = theano.shared(sample_weights(n_in, n_h))
self.W_hi = theano.shared(sample_weights(n_h, n_h))
self.W_ci = theano.shared(sample_weights(n_h))
self.W_xf = theano.shared(sample_weights(n_in, n_h))
self.W_hf = theano.shared(sample_weights(n_h, n_h))
self.W_cf = theano.shared(sample_weights(n_h))
self.W_xc = theano.shared(sample_weights(n_in, n_h))
self.W_hc = theano.shared(sample_weights(n_h, n_h))
self.W_xo = theano.shared(sample_weights(n_in, n_h))
self.W_ho = theano.shared(sample_weights(n_h, n_h))
self.W_co = theano.shared(sample_weights(n_h))
self.params = [
self.W_xi, self.W_hi, self.W_ci, self.W_xf, self.W_hf, self.W_cf,
self.W_xc, self.W_hc, self.W_xo, self.W_ho, self.W_co
]
def __init__(self, n_i, n_h):
self.W = theano.shared(sample_weights(n_i, n_h))
self.W_t = theano.shared(sample_weights(n_h, n_h))
self.BOS = theano.shared(sample_weights(n_h))
self.params = [self.W, self.W_t, self.BOS]
def __init__(self, x_span, x_word, x_ctx, x_dist, y, init_emb, n_vocab, dim_w, dim_d, dim_h, L2_reg):
"""
:param x_span: 1D: batch, 2D: limit * 2 (10); elem=word id
:param x_word: 1D: batch, 2D: 4 (m_first, m_last, a_first, a_last); elem=word id
:param x_ctx : 1D: batch, 2D: window * 2 * 2 (20); elem=word id
:param x_dist: 1D: batch; elem=distance between sentences of ant and ment
:param y : 1D: batch
"""
self.input = [x_span, x_word, x_ctx, x_dist, y]
self.x_span = x_span
self.x_word = x_word
self.x_ctx = x_ctx
self.x_dist = x_dist
self.y = y
dim_x = dim_w * (2 + 4 + 20) + 1
batch = y.shape[0]
""" Params """
if init_emb is None:
self.emb = theano.shared(sample_weights(n_vocab, dim_w))
else:
self.emb = theano.shared(init_emb)
self.W_d = theano.shared(sample_weights(dim_d))
self.W_i = theano.shared(sample_weights(dim_x, dim_h*3))
self.W_h = theano.shared(sample_weights(dim_h*3, dim_h))
self.W_o = theano.shared(sample_weights(dim_h))
self.params = [self.W_d, self.W_i, self.W_h, self.W_o]
""" Input Layer """
x_s = self.emb[x_span] # 1D: batch, 2D: limit * 2, 3D: dim_w
x_w = self.emb[x_word] # 1D: batch, 2D: 4, 3D: dim_w
x_c = self.emb[x_ctx] # 1D: batch, 2D: window * 2 * 2, 3D: dim_w
x_d = self.W_d[x_dist] # 1D: batch
x_s_avg = T.concatenate([T.mean(x_s[:, :x_s.shape[1]/2], 1), T.mean(x_s[:, x_s.shape[1]/2:], 1)], 1)
x = T.concatenate([x_s_avg, x_w.reshape((batch, -1)), x_c.reshape((batch, -1)), x_d.reshape((batch, 1))], 1)
""" Intermediate Layers """
h1 = relu(T.dot(x, self.W_i)) # h1: 1D: batch, 2D: dim_h
h2 = relu(T.dot(h1, self.W_h)) # h2: 1D: batch, 2D: dim_h
""" Output Layer """
p_y = sigmoid(T.dot(h2, self.W_o)) # p_y: 1D: batch
""" Predicts """
self.thresholds = theano.shared(np.asarray([0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9], dtype=theano.config.floatX))
self.y_hat = self.binary_predict(p_y) # 1D: batch, 2D: 9 (thresholds)
self.y_hat_index = T.argmax(p_y)
self.p_y_hat = p_y[self.y_hat_index]
""" Cost Function """
self.nll = - T.sum(y * T.log(p_y) + (1. - y) * T.log((1. - p_y))) # TODO: ranking criterion
self.cost = self.nll + L2_reg * L2_sqr(params=self.params) / 2
""" Update """
self.grad = T.grad(self.cost, self.params)
self.updates = adam(self.params, self.grad)
""" Check Results """
self.result = T.eq(self.y_hat, y.reshape((y.shape[0], 1))) # 1D: batch, 2D: 9 (thresholds)
self.total_p = T.sum(self.y_hat, 0)
self.total_r = T.sum(y, keepdims=True)
self.correct = T.sum(self.result, 0)
self.correct_t, self.correct_f = correct_tf(self.result, y.reshape((y.shape[0], 1)))
def __init__(self, n_i, n_h, activation=tanh):
self.activation = activation
self.c0 = build_shared_zeros(n_h)
self.h0 = self.activation(self.c0)
self.W = theano.shared(sample_weights(n_i, n_h))
"""input gate parameters"""
self.W_xi = theano.shared(sample_weights(n_h, n_h))
self.W_hi = theano.shared(sample_weights(n_h, n_h))
self.W_ci = theano.shared(sample_weights(n_h))
"""forget gate parameters"""
self.W_xf = theano.shared(sample_weights(n_h, n_h))
self.W_hf = theano.shared(sample_weights(n_h, n_h))
self.W_cf = theano.shared(sample_weights(n_h))
"""cell parameters"""
self.W_xc = theano.shared(sample_weights(n_h, n_h))
self.W_hc = theano.shared(sample_weights(n_h, n_h))
"""output gate parameters"""
self.W_xo = theano.shared(sample_weights(n_h, n_h))
self.W_ho = theano.shared(sample_weights(n_h, n_h))
self.W_co = theano.shared(sample_weights(n_h))
self.params = [self.W, self.W_xi, self.W_hi, self.W_ci, self.W_xf, self.W_hf, self.W_cf,
self.W_xc, self.W_hc, self.W_xo, self.W_ho, self.W_co]
def create_posit_emb(n_posit, dim_posit):
return theano.shared(sample_weights(n_posit, dim_posit))
def __init__(self, n_i):
self.W = theano.shared(sample_weights(n_i * 2, n_i))
self.W_a = theano.shared(sample_weights(n_i, n_i))
self.W_p = theano.shared(sample_weights(n_i, n_i))
self.params = [self.W_a, self.W_p, self.W]
def create_word_emb(n_vocab, init_emb, dim_emb):
if init_emb is None:
return theano.shared(sample_weights(n_vocab - 1, dim_emb))
return theano.shared(init_emb)
def __init__(self, n_i, n_h, activation=tanh):
self.activation = activation
self.c0 = build_shared_zeros(n_h)
self.h0 = self.activation(self.c0)
self.W = theano.shared(sample_weights(n_i, n_h))
# input gate parameters
self.W_xi = theano.shared(sample_weights(n_h, n_h))
self.W_hi = theano.shared(sample_weights(n_h, n_h))
self.W_ci = theano.shared(sample_weights(n_h))
# forget gate parameters
self.W_xf = theano.shared(sample_weights(n_h, n_h))
self.W_hf = theano.shared(sample_weights(n_h, n_h))
self.W_cf = theano.shared(sample_weights(n_h))
# cell parameters
self.W_xc = theano.shared(sample_weights(n_h, n_h))
self.W_hc = theano.shared(sample_weights(n_h, n_h))
# output gate parameters
self.W_xo = theano.shared(sample_weights(n_h, n_h))
self.W_ho = theano.shared(sample_weights(n_h, n_h))
self.W_co = theano.shared(sample_weights(n_h))
self.params = [
self.W, self.W_xi, self.W_hi, self.W_ci, self.W_xf, self.W_hf,
self.W_cf, self.W_xc, self.W_hc, self.W_xo, self.W_ho, self.W_co
]
def create_emb(dim_row, dim_column):
return theano.shared(sample_weights(dim_row, dim_column))
def __init__(self, x_span, x_word, x_ctx, x_dist, x_slen, y, init_emb, n_vocab, dim_w, dim_d, dim_h, L2_reg):
"""
:param x_span: 1D: batch, 2D: limit * 2 (10); elem=word id
:param x_word: 1D: batch, 2D: 4 (m_first, m_last, a_first, a_last); elem=word id
:param x_ctx : 1D: batch, 2D: window * 2 * 2 (20); elem=word id
:param x_dist: 1D: batch; 2D: 2; elem=[sent dist, ment dist]
:param x_slen: 1D: batch; 2D: 3; elem=[m_span_len, a_span_len, head_match]
:param y : 1D: batch
"""
self.input = [x_span, x_word, x_ctx, x_dist, y]
self.x_span = x_span
self.x_word = x_word
self.x_ctx = x_ctx
self.x_dist = x_dist
self.x_slen = x_slen
self.y = y
dim_x = dim_w * (10 + 4 + 4 + 2 + 3)
batch = y.shape[0]
""" Params """
if init_emb is None:
self.emb = theano.shared(sample_weights(n_vocab, dim_w))
else:
self.emb = theano.shared(init_emb)
self.W_d = theano.shared(sample_weights(dim_d, dim_w))
self.W_l = theano.shared(sample_weights(7, dim_w))
self.W_i = theano.shared(sample_weights(dim_x, dim_h))
self.W_h = theano.shared(sample_weights(dim_h, dim_h))
self.W_o = theano.shared(sample_weights(dim_h))
self.params = [self.W_d, self.W_l, self.W_i, self.W_h, self.W_o]
""" Input Layer """
x_vec = T.concatenate([x_span, x_word, x_ctx], 1).flatten() # 1D: batch * (limit * 2 + 4 + 20)
x_in = self.emb[x_vec] # 1D: batch, 2D: limit * 2, 3D: dim_w
x_d = self.W_d[x_dist] # 1D: batch, 2D: 2, 3D: dim_w
x_l = self.W_l[x_slen] # 1D: batch, 2D: 2, 3D: dim_w
x = T.concatenate([x_in.reshape((batch, -1)), x_d.reshape((batch, -1)), x_l.reshape((batch, -1))], 1)
""" Intermediate Layers """
h1 = relu(T.dot(x, self.W_i)) # h1: 1D: batch, 2D: dim_h
h2 = relu(T.dot(h1, self.W_h)) # h2: 1D: batch, 2D: dim_h
""" Output Layer """
p_y = sigmoid(T.dot(h2, self.W_o)) # p_y: 1D: batch
""" Cost Function """
self.nll = - T.sum(y * T.log(p_y) + (1. - y) * T.log((1. - p_y))) # TODO: ranking criterion
self.cost = self.nll + L2_reg * L2_sqr(params=self.params) / 2
""" Update """
self.updates = sgd(self.cost, self.params, self.emb, x_in)
""" Predicts """
self.thresholds = theano.shared(np.asarray([0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9], dtype=theano.config.floatX))
self.y_hat = self.binary_predict(p_y) # 1D: batch, 2D: 9 (thresholds)
self.y_hat_index = T.argmax(p_y)
self.p_y_hat = p_y[self.y_hat_index]
""" Check Results """
self.result = T.eq(self.y_hat, y.reshape((y.shape[0], 1))) # 1D: batch, 2D: 9 (thresholds)
self.total_p = T.sum(self.y_hat, 0)
self.total_r = T.sum(y, keepdims=True)
self.correct = T.sum(self.result, 0)
self.correct_t, self.correct_f = correct_tf(self.result, y.reshape((y.shape[0], 1)))
def __init__(self, x_span, x_word, x_ctx, x_dist, y, init_emb, n_vocab,
dim_w, dim_d, dim_h, L2_reg):
"""
:param x_span: 1D: batch, 2D: limit * 2 (10); elem=word id
:param x_word: 1D: batch, 2D: 4 (m_first, m_last, a_first, a_last); elem=word id
:param x_ctx : 1D: batch, 2D: window * 2 * 2 (20); elem=word id
:param x_dist: 1D: batch; elem=distance between sentences of ant and ment
:param y : 1D: batch
"""
self.input = [x_span, x_word, x_ctx, x_dist, y]
self.x_span = x_span
self.x_word = x_word
self.x_ctx = x_ctx
self.x_dist = x_dist
self.y = y
dim_x = dim_w * (2 + 4 + 20) + 1
batch = y.shape[0]
""" Params """
if init_emb is None:
self.emb = theano.shared(sample_weights(n_vocab, dim_w))
else:
self.emb = theano.shared(init_emb)
self.W_d = theano.shared(sample_weights(dim_d))
self.W_i = theano.shared(sample_weights(dim_x, dim_h * 3))
self.W_h = theano.shared(sample_weights(dim_h * 3, dim_h))
self.W_o = theano.shared(sample_weights(dim_h))
self.params = [self.W_d, self.W_i, self.W_h, self.W_o]
""" Input Layer """
x_s = self.emb[x_span] # 1D: batch, 2D: limit * 2, 3D: dim_w
x_w = self.emb[x_word] # 1D: batch, 2D: 4, 3D: dim_w
x_c = self.emb[x_ctx] # 1D: batch, 2D: window * 2 * 2, 3D: dim_w
x_d = self.W_d[x_dist] # 1D: batch
x_s_avg = T.concatenate([
T.mean(x_s[:, :x_s.shape[1] / 2], 1),
T.mean(x_s[:, x_s.shape[1] / 2:], 1)
], 1)
x = T.concatenate([
x_s_avg,
x_w.reshape((batch, -1)),
x_c.reshape((batch, -1)),
x_d.reshape((batch, 1))
], 1)
""" Intermediate Layers """
h1 = relu(T.dot(x, self.W_i)) # h1: 1D: batch, 2D: dim_h
h2 = relu(T.dot(h1, self.W_h)) # h2: 1D: batch, 2D: dim_h
""" Output Layer """
p_y = sigmoid(T.dot(h2, self.W_o)) # p_y: 1D: batch
""" Predicts """
self.thresholds = theano.shared(
np.asarray([0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9],
dtype=theano.config.floatX))
self.y_hat = self.binary_predict(p_y) # 1D: batch, 2D: 9 (thresholds)
self.y_hat_index = T.argmax(p_y)
self.p_y_hat = p_y[self.y_hat_index]
""" Cost Function """
self.nll = -T.sum(y * T.log(p_y) + (1. - y) * T.log(
(1. - p_y))) # TODO: ranking criterion
self.cost = self.nll + L2_reg * L2_sqr(params=self.params) / 2
""" Update """
self.grad = T.grad(self.cost, self.params)
self.updates = adam(self.params, self.grad)
""" Check Results """
self.result = T.eq(self.y_hat, y.reshape(
(y.shape[0], 1))) # 1D: batch, 2D: 9 (thresholds)
self.total_p = T.sum(self.y_hat, 0)
self.total_r = T.sum(y, keepdims=True)
self.correct = T.sum(self.result, 0)
self.correct_t, self.correct_f = correct_tf(self.result,
y.reshape((y.shape[0], 1)))
