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, 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, 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)