def check_backward(self, x_data, W_data, b_data, y_grad):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
if b_data is None:
y = functions.maxout(x, W)
else:
b = chainer.Variable(b_data)
y = functions.maxout(x, W, b)
y.grad = y_grad
y.backward()
func = y.creator
if b_data is None:
f = lambda: func.forward((x.data, W.data))
gx, gW = gradient_check.numerical_grad(f, (x.data, W.data),
(y.grad, ),
eps=1e-2)
else:
f = lambda: func.forward((x.data, W.data, b.data))
gx, gW, gb = gradient_check.numerical_grad(
f, (x.data, W.data, b.data), (y.grad, ), eps=1e-2)
gradient_check.assert_allclose(gx, x.grad, atol=1e-2)
gradient_check.assert_allclose(gW, W.grad, atol=1e-2)
if b_data is not None:
gradient_check.assert_allclose(gb, b.grad, atol=1e-2)
def check_forward(self, x_data, W_data, b_data, y_expect):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
if b_data is None:
y = functions.maxout(x, W)
else:
b = chainer.Variable(b_data)
y = functions.maxout(x, W, b)
gradient_check.assert_allclose(y_expect, y.data)
def check_forward(self, x_data, W_data, b_data, y_expect):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
if b_data is None:
y = functions.maxout(x, W)
else:
b = chainer.Variable(b_data)
y = functions.maxout(x, W, b)
gradient_check.assert_allclose(y_expect, y.data)
def __call__(self, s, q, s_mask, q_mask):
"""
s_bar, _, _ = self.pred_bilstm(None, None, s)
s_bar_new = F.concat(s_bar, axis=1)
q_bar, _, _ = self.pred_bilstm(None, None, q)
q_bar_new = F.concat(q_bar, axis=1)
"""
_, _, s_bar = self.pred_bilstm(None, None, s) # get list of [seq, dim]
s_bar_new = F.stack(s_bar, axis=0) # turn list to 3d tensor
_, _, q_bar = self.pred_bilstm(None, None, q) # get list of [seq, dim]
q_bar_new = F.stack(q_bar, axis=0) # turn list to 3d tensor
# mean-max pooling
s_sum = F.sum(s_mask, axis=-1)
q_sum = F.sum(q_mask, axis=-1)
s_batch, s_seq = s_mask.shape
s_mask_broad = F.broadcast_to(F.reshape(s_mask, (s_batch, s_seq, 1)),
(s_batch, s_seq, s_bar_new.shape[-1]))
s_broad = s_bar_new * s_mask_broad
"""
s_infinit_matrix = self.xp.ones((s_batch, s_seq, s_bar_new.shape[-1]), dtype=self.xp.float32) * -1 * self.xp.inf
s_cond = s_mask_broad.data.astype(self.xp.bool)
s_broad_max = F.where(s_cond, s_bar_new, s_infinit_matrix)
"""
s_mean = F.average(s_broad, axis=1) # [batch_size, dim]
s_max = F.maxout(
F.reshape(
s_bar_new,
(s_bar_new.shape[0], s_bar_new.shape[1] * s_bar_new.shape[2])),
s_bar_new.shape[-1]) # [batch_size, dim]
q_batch, q_seq = q_mask.shape
q_broad = q_bar_new * F.broadcast_to(
F.reshape(q_mask, (q_batch, q_seq, 1)),
(q_batch, q_seq, q_bar_new.shape[-1]))
q_mean = F.average(q_broad, axis=1) # [batch_size, dim]
q_max = F.maxout(
F.reshape(
q_bar_new,
(q_bar_new.shape[0], q_bar_new.shape[1] * q_bar_new.shape[2])),
q_bar_new.shape[-1]) # [batch_size, dim]
summarized_vector = F.concat([s_mean, s_max, q_mean, q_max], axis=1)
s_linear_output = self.gelu(self.L(summarized_vector))
y = F.softmax(s_linear_output)
return y
def check_backward(self, x_data, W_data, b_data, y_grad):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
if b_data is None:
y = functions.maxout(x, W)
else:
b = chainer.Variable(b_data)
y = functions.maxout(x, W, b)
y.grad = y_grad
y.backward()
func = y.creator
if b_data is None:
f = lambda: func.forward((x.data, W.data))
gx, gW = gradient_check.numerical_grad(f, (x.data, W.data), (y.grad,), eps=1e-2)
else:
f = lambda: func.forward((x.data, W.data, b.data))
gx, gW, gb = gradient_check.numerical_grad(f, (x.data, W.data, b.data), (y.grad,), eps=1e-2)
gradient_check.assert_allclose(gx, x.grad, atol=1e-2)
gradient_check.assert_allclose(gW, W.grad, atol=1e-2)
if b_data is not None:
gradient_check.assert_allclose(gb, b.grad, atol=1e-2)
def maxpooling(self,xs,neighbor): sources = defaultdict(list) for ee in neighbor: for i in neighbor[ee]: sources[i].append(xs[ee]) result = [] for i,xxs in sorted(sources.items(),key=lambda x:x[0]): if len(xxs)==1: result.append(xxs[0]) else: x = F.concat(xxs,axis=0) # -> (b,d) x = F.swapaxes(x,0,1) # -> (d,b) x = F.maxout(x,len(xxs)) # -> (d,1) x = F.swapaxes(x,0,1) # -> (1,d) result.append(x) return result
def __call__(self, y, cv, c, h):
""" @param:
y: y_{i-1}, last generated word
cv: context vector c_{i}
c: LSTM memory cell
h: LSTM hidden state
@return:
y: the weight of y_{i}
c: Updated LSTM memory cell
h: Updated LSTM hidden state
"""
e = self.ye(y)
c, h = F.lstm(c, self.eh(e) + self.hh(h) + self.ch(cv))
t = F.maxout(self.sm(h) + self.em(e) + self.cm(cv), self.POOL_SIZE)
y = self.my(t)
return y, c, h
def generateHyp(self, enc_states):
bosID = 1
bos = self.xp.array([bosID], dtype=self.xp.int32)
predicts = [bos]
while len(predicts) - 1 < self.gen_limit:
embedding = self.word2embedding(predicts[-1])
previous_hidden = self.gru.h
context = self.attention(previous_hidden, enc_states)
hidden = self.gru(embedding, context)
t = self.U_o(previous_hidden) + self.V_o(embedding) + self.C_o(
context)
t = chainFunc.maxout(t, 2)
score = self.W_o(t)
predict = chainFunc.argmax(score, axis=1)
predicts.append(predict)
del predicts[0]
return predicts
def __call__(self, x):
V = self.embed(x)
V_norm = F.normalize(V.transpose(0, 2, 1), axis=1)
C = self.embed_class.W
C_norm = F.normalize(C, axis=1)
G = F.matmul(
F.broadcast_to(
C_norm, (V_norm.shape[0], C_norm.shape[0], C_norm.shape[1])),
V_norm)
u = F.relu(self.conv1(G))
m = F.maxout(u, pool_size=self.n_class, axis=1)
beta = F.softmax(m, axis=2)
z = F.sum((V * F.broadcast_to(beta.transpose(0, 2, 1), V.shape)),
axis=1)
z = self.fc2(F.dropout(z))
z = F.sigmoid(z)
z_class = self.fc2(F.dropout(C))
out = F.concat([z, z_class], axis=0)
return out
def __call__(self, enc_states, batch_tgt):
loss = chainer.Variable(self.xp.zeros((), dtype=self.xp.float32))
predicts = list()
for previous_wordID, word in enumerate(batch_tgt[1:]):
previous_hidden = self.gru.h
embedding = self.word2embedding(batch_tgt[previous_wordID])
#util.trace('embedding size decside: {}'.format(embedding.shape))
context = self.attention(previous_hidden, enc_states)
hidden = chainFunc.dropout(self.gru(embedding, context),
self.dropoutr)
t = self.U_o(previous_hidden) + self.V_o(embedding) + self.C_o(
context)
t = chainFunc.maxout(t, 2)
score = self.W_o(t)
predict = chainFunc.argmax(score, axis=1)
loss += chainFunc.softmax_cross_entropy(score,
word,
ignore_label=-1)
predicts.append(predict.data)
return loss, predicts
def maxpooling(self, xs, neighbor):
sources = defaultdict(list)
for ee in neighbor:
for i in neighbor[ee]:
sources[i].append(xs[ee])
# sources:键值为实体编号i,值为与实体有关系的所有实体经转换函数后的embedding
# 将sources根据实体集的编号排序
# 最后得到的len(result)=len(entities)
result = []
for i, xxs in sorted(sources.items(), key=lambda x: x[0]):
if len(xxs) == 1:
x = xxs[0]
x = self.forwardAA(x, xxs) # attention
result.append(x)
else:
x = F.concat(xxs, axis=0) # -> (b,d)
x = F.swapaxes(x, 0, 1) # -> (d,b)
x = F.maxout(x, len(xxs)) # -> (d,1)
x = F.swapaxes(x, 0, 1) # -> (1,d)
x = self.forwardAA(x, xxs) # attention
result.append(x)
return result
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(
lambda x: functions.maxout(x, self.pool_size, self.axis),
x_data,
y_grad,
dtype=numpy.float64)
def __call__(self, x):
return F.maxout(x, self.pool_size, self.axis)
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.maxout(x, self.pool_size, self.axis)
gradient_check.assert_allclose(self.y, y.data)
def test_invalid_shape_gpu(self):
self.x.to_gpu()
with self.assertRaises(self.error):
functions.maxout(self.x, self.pool_size)
def forward(self, inputs, device):
x, = inputs
return functions.maxout(x, self.pool_size, self.axis),
def __call__(self, x): return F.maxout(x, self.pool_size, self.axis)
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.maxout(x, self.pool_size, self.axis)
self.assertEqual(y.data.dtype, self.dtype)
testing.assert_allclose(self.y, y.data)
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.maxout(x, self.pool_size, self.axis)
gradient_check.assert_allclose(self.y, y.data)
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(
lambda x: functions.maxout(x, self.pool_size, self.axis),
x_data, y_grad, dtype=numpy.float64)
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(
lambda x: functions.maxout(x, self.pool_size, self.axis),
x_data, y_grad, atol=1e-2)
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(
lambda x: functions.maxout(x, self.pool_size, self.axis),
x_data,
y_grad,
eps=0.125)
def test_invalid_shape_gpu(self):
self.x.to_gpu()
with self.assertRaises(self.error):
functions.maxout(self.x, self.pool_size)
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.maxout(x, self.pool_size, self.axis)
self.assertEqual(y.data.dtype, self.dtype)
testing.assert_allclose(self.y, y.data)
def __call__(self, x): return functions.maxout(x, self.pool_size)
def __call__(self, x): return functions.maxout(x, self.pool_size)
