def forward(x_data, y_data, model,train=True):
# Neural net architecture
#x, t = chainer.Variable(x_data), chainer.Variable(y_data)
t = chainer.Variable(y_data)
x = {}
for n in range(500):
x[n] = chainer.Variable(x_data[n])
h = {}
initial_V = {}
initial_V_relu = {}
for nameint in range(len(l_name)-2):
initial_V[nameint] = model[l_name[nameint]](x[nameint])
#initial_V_relu[nameint] = F.relu(initial_V[nameint])
#initial_V_relu[nameint] = F.sigmoid(initial_V[nameint])
initial_V_relu[nameint] = F.tanh(initial_V[nameint])
#h[nameint] = F.dropout(F.relu(initial_V[nameint]), train=train)
#h[nameint] = F.dropout(F.sigmoid(initial_V[nameint]), train=train)
h[nameint] = F.dropout(F.tanh(initial_V[nameint]), train=train)
#h[nameint] = F.relu(model[l_name[nameint]](x[nameint]))
#h6 = F.dropout(F.relu(model.l501(Returnharray(h))), train=train)
#h6 = F.dropout(F.sigmoid(model.l501(Returnharray(h))), train=train)
h6 = F.dropout(F.tanh(model.l501(Returnharray(h))), train=train)
y = model.l502(h6)
y_pre = (y.data.argmax(axis = 1))
return F.softmax_cross_entropy(y, t), F.accuracy(y, t),y_pre,initial_V,initial_V_relu
def translate(model, id2wd, jline):
result_words = []
gh = []
for w in jline:
wid = model.jvocab[w]
x_k = model.embedx(Variable(np.array([wid], dtype=np.int32), volatile='on'))
h = model.H(x_k)
gh.append(h.data[0])
x_k = model.embedx(Variable(np.array([model.jvocab[EOS]], dtype=np.int32), volatile='on'))
h = model.H(x_k)
ct = Variable(attention.mk_ct(gh, h.data[0]), volatile='on')
h2 = F.tanh(model.Wc1(ct) + model.Wc2(h))
wid = np.argmax(F.softmax(model.W(h2)).data[0])
result_words.append(id2wd.get(wid, wid))
loop = 0
while (wid != model.evocab[EOS]) and (loop <= 30):
x_k = model.embedy(Variable(np.array([wid], dtype=np.int32), volatile='on'))
h = model.H(x_k)
ct = Variable(attention.mk_ct(gh, h.data[0]), volatile='on')
h2 = F.tanh(model.Wc1(ct) + model.Wc2(h))
wid = np.argmax(F.softmax(model.W(h2)).data[0])
result_words.append(id2wd.get(wid, wid))
loop += 1
return ' '.join(result_words)
def forward(self, x_data, y_data, train=True):
#print y_data
batchsize = len(x_data)
csize = self.channel
x, t = chainer.Variable(x_data,volatile=not train), chainer.Variable(y_data.reshape(len(y_data),),volatile=not train)
x = F.reshape(x,(batchsize,csize,-1))
h = F.reshape(x,(batchsize,csize,-1,1))
h = self.conv1(h)
h = F.reshape(h,(batchsize,10,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,10,-1,1))
h = self.conv2(h)
h = F.reshape(h,(batchsize,10,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,10,-1,1))
h = self.conv3(h)
h = F.reshape(h,(batchsize,100,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,100,-1,1))
h = self.conv4(h)
h = F.reshape(h,(batchsize,100,-1))
h = F.tanh(h)
h = F.dropout(F.tanh(self.fc5(h)), train=train)
y = self.fc6(h)
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def __call__(self, a_list, b_list, p, sentence_length, window_size):
batch_size = p.data.shape[0]
SENTENCE_LENGTH = XP.fnonzeros((batch_size, 1),sentence_length)
e_list = []
sum_e = XP.fzeros((batch_size, 1))
s = functions.tanh(self.ts(p))
pos = SENTENCE_LENGTH * functions.sigmoid(self.sp(s))
# Develop batch logic to set to zero the components of a and b which are out of the window
# Big question: Do I have to iterate over each element in the batch? That would suck.
# One logic: Get global alignment matrix of (batch x) hidden size x sentence length and then another matrix of (batch x) sentence length which
# will essentially be a matrix containing the gaussian distrubution weight and there will be zeros where the sentence position falls out of the window
# Another logic: Create a matrix of (batch x) sentence length where there will be 1 for each position in the window
# Separate the attention weights for a and b cause forward is different from backward.
for a, b in zip(a_list, b_list):
w = functions.tanh(self.aw(a) + self.bw(b) + self.pw(p))
e = functions.exp(self.we(w))
e_list.append(e)
sum_e += e
ZEROS = XP.fzeros((batch_size, self.hidden_size))
aa = ZEROS
bb = ZEROS
for a, b, e in zip(a_list, b_list, e_list):
e /= sum_e
aa += a * e
bb += b * e
return aa, bb
def forward(self, x_data, y_data, train=True):
#print y_data
batchsize = len(x_data)
csize = self.channel
x, t = chainer.Variable(x_data,volatile=not train), chainer.Variable(y_data,volatile=not train)
x = F.reshape(x,(batchsize,csize,-1))
h = F.reshape(x,(batchsize,csize,-1,1))
h = self.conv1(h)
h = F.reshape(h,(batchsize,10,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,10,-1,1))
h = self.conv2(h)
h = F.reshape(h,(batchsize,10,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,10,-1,1))
h = self.conv3(h)
h = F.reshape(h,(batchsize,100,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,100,-1,1))
h = self.conv4(h)
h = F.reshape(h,(batchsize,100,-1))
h = F.tanh(h)
h = F.dropout(F.tanh(self.fc5(h)), train=train)
y = self.fc6(h)
return F.mean_squared_error(y, t)
def predict(self, x_data, train=False):
#print y_data
batchsize = len(x_data)
csize = self.channel
x = chainer.Variable(x_data,volatile=True)
x = F.reshape(x,(batchsize,csize,-1))
h = F.reshape(x,(batchsize,1,-1,1))
h = self.conv1(h)
h = F.reshape(h,(batchsize,10,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,10,-1,1))
h = self.conv2(h)
h = F.reshape(h,(batchsize,100,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,100,-1,1))
h = self.conv3(h)
h = F.reshape(h,(batchsize,256,-1))
h = F.tanh(h)
h = F.dropout(F.tanh(self.fc4(h)), train=train)
y = self.fc5(h)
return y
def __call__(self, jline, eline):
gh = []
self.H.reset_state()
for w in jline:
wid = self.jvocab[w]
x_k = self.embedx(Variable(np.array([wid], dtype=np.int32)))
h = self.H(x_k)
gh.append(np.copy(h.data[0]))
x_k = self.embedx(Variable(np.array([self.jvocab[EOS]], dtype=np.int32)))
tx = Variable(np.array([self.evocab[eline[0]]], dtype=np.int32))
h = self.H(x_k)
ct = Variable(mk_ct(gh, h.data[0]))
h2 = F.tanh(self.Wc1(ct) + self.Wc2(h))
accum_loss = F.softmax_cross_entropy(self.W(h2), tx)
for i in range(len(eline)):
wid = self.evocab[eline[i]]
x_k = self.embedy(Variable(np.array([wid], dtype=np.int32)))
next_w = eline[i + 1] if i < len(eline) - 1 else EOS
next_wid = self.evocab[next_w]
tx = Variable(np.array([next_wid], dtype=np.int32))
h = self.H(x_k)
ct = Variable(mk_ct(gh, h.data[0]))
h2 = F.tanh(self.Wc1(ct) + self.Wc2(h))
loss = F.softmax_cross_entropy(self.W(h2), tx)
accum_loss += loss
return accum_loss
def check_forward(self, x_data, use_cudnn='always'):
x = chainer.Variable(x_data)
with chainer.using_config('use_cudnn', use_cudnn):
y = functions.tanh(x)
self.assertEqual(y.data.dtype, self.dtype)
y_expect = functions.tanh(chainer.Variable(self.x))
testing.assert_allclose(y_expect.data, y.data)
def forward_one_step(c, h, cur_word, next_word):
i = Variable(np.array([cur_word], dtype=np.int32))
t = Variable(np.array([next_word], dtype=np.int32))
x = F.tanh(model.embed(i))
c, h = F.lstm(c, model.x_to_h(x) + model.h_to_h(h))
y = F.tanh(model.h_to_y(h))
return c, h, F.softmax_cross_entropy(y, t)
def __call__(self, xs):
xs = self.embed(xs)
batch_size, length, _ = xs.shape
h = F.sum(xs, axis=1) / length
h = F.tanh(self.linear1(h))
h = F.tanh(self.linear2(h))
return h
def __call__(self,x,y,state,train=True,target=True):
if train:
h = Variable(x.reshape(self.batchsize,12), volatile=not train)
else:
h = Variable(x, volatile=not train)
t = Variable(y.flatten(), volatile=not train)
h0 = F.relu(self.l0(h))
if target == False:
data = h0.data
self.data_first.append(data)
h1_in = self.l1_x(h0) + self.l1_h(state['h1'])
h1_in = F.dropout(F.tanh(h1_in),train=train)
c1, h1 = F.lstm(state['c1'], h1_in)
h2_in = F.dropout(F.tanh(self.l2_x(h1)), train=train) + self.l2_h(state['h2'])
c2, h2 = F.lstm(state['c2'], h2_in)
if target == False:
data = h1.data
self.data_hidden.append(data)
y = self.l3(h2)
if target ==False:
data = y.data
self.data_output.append(data)
state = {'c1': c1, 'h1': h1,'c2':c2,'h2':h2}
self.loss = F.softmax_cross_entropy(y,t)
return state,self.loss
def __call__(self, s):
accum_loss = None
_, k = self.embed.W.data.shape
h = Variable(np.zeros((1, k), dtype=np.float32))
c = Variable(np.zeros((1, k), dtype=np.float32))
s_length = len(s)
for i in range(s_length):
w1 = s[i]
w2 = s[i + 1] if i < s_length - 1 else self.eos_id
x_k = self.embed(Variable(np.array([w1], dtype=np.int32)))
tx = Variable(np.array([w2], dtype=np.int32))
z0 = self.Wz(x_k) + self.Rz(F.dropout(h))
z1 = F.tanh(z0)
i0 = self.Wi(x_k) + self.Ri(F.dropout(h))
i1 = F.sigmoid(i0)
f0 = self.Wf(x_k) + self.Rf(F.dropout(h))
f1 = F.sigmoid(f0)
c = i1 * z1 + f1 * c
o0 = self.Wo(x_k) + self.Ro(F.dropout(h))
o1 = F.sigmoid(o0)
y = o1 * F.tanh(c)
h = y
loss = F.softmax_cross_entropy(self.W(y), tx)
accum_loss = loss if accum_loss is None else accum_loss + loss
return accum_loss
def predict(node, neural_model_size, root=True):
if isinstance(node['node'], np.ndarray):
# leaf node
word = np.reshape(node['node'], (1, neural_model_size))
v = chainer.Variable(word)
else:
# internal node
left_node, right_node = node['node']
left = predict(left_node, neural_model_size, root=False)
right = predict(right_node, neural_model_size, root=False)
intermediate = F.tanh(model.h(F.concat((left, right))))
v = F.tanh(model.l(F.concat((left, right))))
y = model.w(v)
# evaluate root label
if root:
predicted = cuda.to_cpu(y.data).argmax(1)
try:
label = node['label']
return predicted[0], label
except:
pass
return predicted[0]
return v
def forward_one_step_rnn(model, h, cur_word, next_word, train=True):
word = Variable(cur_word, volatile=not train)
t = Variable(next_word, volatile=not train)
x = F.tanh(model.embed(word))
h = F.tanh(model.x_to_h(x) + model.h_to_h(h))
y = model.h_to_y(h)
loss = F.softmax_cross_entropy(y, t)
return h, loss
def predict(self, x_data, train=False):
# print y_data
x = chainer.Variable(x_data)
h1 = F.dropout(F.tanh(self.fc1(x)), train=train)
h2 = F.dropout(F.tanh(self.fc2(h1)), train=train)
h3 = F.dropout(F.tanh(self.fc3(h2)), train=train)
y = self.fc4(h3)
return y
def forward_one_step(h, cur_word, next_word, volatile=False):
word = Variable(cur_word, volatile=volatile)
t = Variable(next_word, volatile=volatile)
x = F.tanh(model.embed(word))
h = F.tanh(model.x_to_h(x) + model.h_to_h(h))
y = model.h_to_y(h)
loss = F.softmax_cross_entropy(y, t)
return h, loss
def forward(x_data, y_data, model, train=True):
# Neural net architecture
x, t = chainer.Variable(x_data), chainer.Variable(y_data)
initialV = model.l1(x)
h1 = F.dropout(F.tanh(initialV), train=train)
h2 = F.dropout(F.tanh(model.l2(h1)), train=train)
y = model.l3(h2)
y_pre = (y.data.argmax(axis = 1))
return F.softmax_cross_entropy(y, t), F.accuracy(y, t),y_pre,initialV,h1,h2
def forward_one_step(h, cur_word, next_word, volatile=False):
word = V(cur_word, volatile=volatile)
t = V(next_word, volatile=volatile)
x = F.tanh(model.embed(word))
h = F.tanh(model.Wx(x) + model.Wh(h))
y = model.Wy(h)
loss = F.softmax_cross_entropy(y, t)
pred = F.softmax(y)
return h, loss, np.argmax(pred.data)
def __call__(self, c1, c2, a, b, s1, s2, s3):
u1 = XP.dropout(functions.tanh(self.w_au1(a) + self.w_bu1(b) + self.w_s2u1(s2)))
u2 = XP.dropout(functions.tanh(self.w_s1u2(s1) + self.w_s3u2(s3)))
c, h = slstm(
c1, c2,
self.w_u1s1(u1) + self.w_s1s1(s1),
self.w_u2s2(u2) + self.w_s2s2(s2),
)
return c, XP.dropout(h)
def forward_one_step(model, h, cur_word, label, volatile=False):
word = Variable(cur_word)
t = Variable(label, volatile=volatile)
x = F.tanh(model.embed(word))
h = F.tanh(model.x_to_h(x) + model.h_to_h(h))
y = model.h_to_y(h)
loss = F.softmax_cross_entropy(y, t)
accuracy = F.accuracy(y,t)
return h, loss, accuracy
def predict(self, x_data, train=False):
#print y_data
x = chainer.Variable(x_data,volatile=True)
h = F.dropout(F.tanh(self.fc1(x)), train=train)
h = F.dropout(F.tanh(self.fc2(h)), train=train)
h = F.dropout(F.tanh(self.fc3(h)), train=train)
h = F.dropout(F.tanh(self.fc4(h)), train=train)
h = F.dropout(F.tanh(self.fc5(h)), train=train)
y = self.fc6(h)
return y
def test_forward(x_data):
# Variable(chainer独自の型)に変換
x = Variable(x_data)
# 第三引数がFalse場合は第一引数の値をそのまま返す(trainの場合のみDropoutを行い、testの場合は行わないようにする)
# hは前の層からの出力
h1 = F.tanh(model.l1(x))
h2 = F.tanh(model.l2(h1))
y = model.l3(h2)
# 出力データを返す
return y
def train_forward(x_data, y_data):
# Variable(chainer独自の型)に変換
x, t = Variable(x_data), Variable(y_data)
# 第三引数がFalse場合は第一引数の値をそのまま返す(trainの場合のみDropoutを行い、testの場合は行わないようにする)
# hは前の層からの出力
h1 = F.dropout(F.tanh(model.l1(x)))
h2 = F.dropout(F.tanh(model.l2(h1)))
y = model.l3(h2)
# 2乗平均誤差(MSE)を返す
return F.mean_squared_error(y, t)
def __call__(self, x,train = True):
x_batch1,x_batch2 = x
initial_V_concat_1 = self.l_polarity(x_batch1)
initial_V_concat_2 = self.l_polarity(x_batch2)
h_concat_1 = F.dropout(F.leaky_relu(initial_V_concat_1), train=False)
h_concat_2 = F.dropout(F.leaky_relu(initial_V_concat_2), train=False)
h_hidden_1 = F.dropout(F.tanh(self.l_hidden1(h_concat_1)), train=train)
h_hidden_2 = F.dropout(F.tanh(self.l_hidden2(h_concat_2)), train=train)
y = self.l_output(h_hidden_1 + h_hidden_2)
#y = self.l_output(h_concat_1 + h_concat_2)
return y, (initial_V_concat_1 + initial_V_concat_2)
def __call__(self, x):
h = {}
initial_V = {}
initial_V_relu = {}
for nameint in range(len(l_name)-2):
initial_V[nameint] = model[l_name[nameint]](x[nameint])
initial_V_relu[nameint] = F.tanh(initial_V[nameint])
h[nameint] = F.dropout(F.tanh(initial_V[nameint]), train=self.train)
h6 = F.dropout(F.tanh(model.l501(Returnharray(h))), train=self.train)
y = model.l502(h6)
y_pre = (y.data.argmax(axis = 1))
return self.l3(h2)
def __call__(self, x_batch,train = True):
initial_V_list = []
for index in range(DimentionN):
link_name = ("l_"+ str(index))
initial_V_list.append(self[link_name](x_batch[index]))
#initial_V_list.append(self[link_name](chainer.Variable(xp.array(x_batch[index]))))
initial_V_concat = F.concat(initial_V_list)
h_concat = F.dropout(F.tanh(initial_V_concat), train=train)
h_concat_tanh = F.dropout(F.tanh(self.l_hidden(h_concat)), train=train)
y = self.l_output(h_concat_tanh)
#y_pre = (y.data.argmax(axis = 1))
return y, initial_V_concat
def main():
args = parse_args()
trace('making vocabulary ...')
vocab, num_lines, num_words = make_vocab(args.corpus, args.vocab)
trace('initializing CUDA ...')
cuda.init()
trace('start training ...')
model = make_rnnlm_model(args.vocab, args.embed, args.hidden)
for epoch in range(args.epoch):
trace('epoch %d/%d: ' % (epoch + 1, args.epoch))
log_ppl = 0.0
trained = 0
opt = optimizers.SGD()
opt.setup(model)
for batch in generate_batch(args.corpus, args.minibatch):
batch = [[vocab[x] for x in words] for words in batch]
K = len(batch)
L = len(batch[0]) - 1
opt.zero_grads()
s_h = zeros((K, args.hidden))
for l in range(L):
s_x = make_var([batch[k][l] for k in range(K)], dtype=np.int32)
s_t = make_var([batch[k][l + 1] for k in range(K)], dtype=np.int32)
s_e = functions.tanh(model.w_xe(s_x))
s_h = functions.tanh(model.w_eh(s_e) + model.w_hh(s_h))
s_y = model.w_hy(s_h)
loss = functions.softmax_cross_entropy(s_y, s_t)
loss.backward()
log_ppl += get_data(loss).reshape(()) * K
opt.update()
trained += K
trace(' %d/%d' % (trained, num_lines))
log_ppl /= float(num_words)
trace(' log(PPL) = %.10f' % log_ppl)
trace(' PPL = %.10f' % math.exp(log_ppl))
trace(' writing model ...')
save_rnnlm_model(args.model + '.%d' % (epoch + 1), args.vocab, args.embed, args.hidden, vocab, model)
trace('training finished.')
def forward(self, x_data, y_data, train=True):
#print y_data
x, t = chainer.Variable(x_data,volatile=not train), chainer.Variable(y_data,volatile=not train)
h = F.dropout(F.tanh(self.fc1(x)), train=train)
h = F.dropout(F.tanh(self.fc2(h)), train=train)
h = F.dropout(F.tanh(self.fc3(h)), train=train)
h = F.dropout(F.tanh(self.fc4(h)), train=train)
h = F.dropout(F.tanh(self.fc5(h)), train=train)
y = self.fc6(h)
return F.mean_squared_error(y, t)
def forward_one_step_lstm(model, state, cur_word, next_word, train=True):
x = Variable(cur_word, volatile=not train)
t = Variable(next_word, volatile=not train)
h0 = model.embed(x)
h1_in = model.l1_x(F.tanh(h0)) + model.l1_h(state["h1"])
c1, h1 = F.lstm(state["c1"], h1_in)
h2_in = model.l2_x(F.tanh(h1)) + model.l2_h(state["h2"])
c2, h2 = F.lstm(state["c2"], h2_in)
y = model.l3(F.tanh(h2))
state = {"c1": c1, "h1": h1, "c2": c2, "h2":h2}
loss = F.softmax_cross_entropy(y, t)
return state, loss
def forward_dumb(x_data,train=True,level=1):
x = Variable(x_data)
y = Variable(x_data)
for d in range(level):
x = F.average_pooling_2d(x,2)
for d in range(level):
x = zoom_x2(x)
ret = (global_normalization**(-2))*F.mean_squared_error(F.tanh(y),F.tanh(x))
if(not train):
plot_img(x.data, 'd{}'.format(level),
'Lv {} dumb encoder, msqe={}'.format(level, ret.data))
return ret
def __call__(self, x):
h1 = F.leaky_relu(self.e0(x))
h2 = self.e1(h1)
h3 = self.e2(h2)
h4 = self.e3(h3)
h5 = self.e4(h4)
h6 = self.e5(h5)
h7 = self.e6(h6)
h8 = self.e7(h7)
h = self.d0(h8)
attn0 = self.a0(x=h7, g=h8)
h9 = F.concat([h, attn0], axis=1)
h = self.d1(h9)
attn1 = self.a1(x=h6, g=h9)
h10 = F.concat([h, attn1], axis=1)
h = self.d2(h10)
attn2 = self.a2(x=h5, g=h10)
h11 = F.concat([h, attn2], axis=1)
h = self.d3(h11)
attn3 = self.a3(x=h4, g=h11)
h12 = F.concat([h, attn3], axis=1)
h = self.d4(h12)
attn4 = self.a4(x=h3, g=h12)
h13 = F.concat([h, attn4], axis=1)
h = self.d5(h13)
attn5 = self.a5(x=h2, g=h13)
h14 = F.concat([h, attn5], axis=1)
h = self.d6(h14)
attn6 = self.a6(x=h1, g=h14)
h15 = F.concat([h, attn6], axis=1)
out = self.d7(h15)
out = F.tanh(out)
return out
def decoder_predict(self, start_word, enc_states, max_predict_len=MAX_PREDICT_LEN, sample=False):
xp = cuda.cupy if self.gpuid >= 0 else np
# __QUESTION -- Following code is to assist with ATTENTION
# alpha_arr should store the alphas for every predicted word
alpha_arr = xp.empty((0,enc_states.shape[0]), dtype=xp.float32)
# return list of predicted words
predicted_sent = []
# load start symbol
with chainer.no_backprop_mode():
pred_word = Variable(xp.asarray([start_word], dtype=np.int32))
pred_count = 0
# start prediction loop
while pred_count < max_predict_len and (int(pred_word.data) != (EOS_ID)):
self.decode(pred_word, train=False)
if self.attn == NO_ATTN:
prob = F.softmax(self.out(self[self.lstm_dec[-1]].h))
else:
# __QUESTION Add attention
hidden_decoder=self[self.lstm_dec[-1]].h
hidden_encoder=enc_states
a_t = F.matmul(hidden_decoder, hidden_encoder, transa=False, transb=True)
score = F.softmax(a_t)
context=F.matmul(score,hidden_encoder)
final_vector=F.concat((context,hidden_decoder))
h_t_p=F.tanh(final_vector)
predict = self.attention_layer(h_t_p)
predicted_out = self.out(predict)
prob = F.softmax(predicted_out)
alpha_arr = np.concatenate((alpha_arr,score.data),axis=0)
pred_word = self.select_word(prob, train=False, sample=sample)
# add integer id of predicted word to output list
predicted_sent.append(int(pred_word.data))
pred_count += 1
# __QUESTION Add attention
# When implementing attention, make sure to use alpha_array to store
# your attention vectors.
# The visualisation function in nmt_translate.py assumes such an array as input.
return predicted_sent, alpha_arr
def __call__(self, x, train):
self.train = train
if self.train:
test = False
else:
test = True
h = F.relu(self.bn0(self.fc0(x), test=test))
h = F.relu(self.bn1(self.fc1(h), test=test))
h = F.reshape(h, (h.data.shape[]))
h = F.relu(self.bn2(self.dc2(h), test=test))
l = F.tanh(self.dc3(h))
return l
def forward(self, z):
xp = self.xp
c = xp.ones((z.shape[0], 1)).astype("float32")
c = self.linear(c)
z = F.split_axis(z, 7, axis=1)
h = self.gen.G_linear(z[0]).reshape(-1, 4, 4,
16 * self.gen.ch).transpose(
0, 3, 1, 2)
h = self.BSA_linear(h)
h = self.gen.GBlock(h, z[1], c)
h = self.gen.GBlock_1(h, z[2], c)
h = self.gen.GBlock_2(h, z[3], c)
h = self.gen.GBlock_3(h, z[4], c)
h = self.gen.GBlock_4(h, z[5], c)
h = self.gen.attention(h)
h = self.gen.GBlock_5(h, z[6], c)
h = F.relu(self.gen.ScaledCrossReplicaBN(h))
h = F.tanh(self.gen.conv_2d(h))
return h
def run():
parser = argparse.ArgumentParser()
parser.add_argument('-io',
'--filename_obj',
type=str,
default='./examples/data/teapot.obj')
parser.add_argument('-ir',
'--filename_ref',
type=str,
default='./examples/data/example3_ref.png')
parser.add_argument('-or',
'--filename_output',
type=str,
default='./examples/data/example3_result.gif')
parser.add_argument('-g', '--gpu', type=int, default=0)
args = parser.parse_args()
working_directory = os.path.dirname(args.filename_output)
model = Model(args.filename_obj, args.filename_ref)
model.to_gpu()
optimizer = chainer.optimizers.Adam(alpha=0.1, beta1=0.5)
optimizer.setup(model)
loop = tqdm.tqdm(range(300))
for _ in loop:
loop.set_description('Optimizing')
optimizer.target.cleargrads()
loss = model()
loss.backward()
optimizer.update()
# draw object
loop = tqdm.tqdm(range(0, 360, 4))
for num, azimuth in enumerate(loop):
loop.set_description('Drawing')
model.renderer.eye = neural_renderer.get_points_from_angles(
2.732, 0, azimuth)
images = model.renderer.render(model.vertices, model.faces,
cf.tanh(model.textures))
image = images.data.get()[0].transpose((1, 2, 0))
scipy.misc.toimage(image, cmin=0, cmax=1).save(
'%s/_tmp_%04d.png' % (working_directory, num))
make_gif(working_directory, args.filename_output)
def autograd(X, W, b, initial_ct=None, use_tanh=False, mask_x=None):
batchsize, feature_dimension, seq_length = X.shape
if initial_ct is None:
initial_ct = chainer.Variable(
np.zeros((batchsize, feature_dimension), dtype=X.dtype))
if isinstance(X, chainer.Variable) is False:
X = chainer.Variable(X)
if mask_x is not None:
X *= mask_x[..., None]
U = functions.connection.convolution_2d.convolution_2d(
X[:, :, None, :], W[..., None, None])[:, :, 0]
Z, F, R = functions.split_axis(U, 3, axis=1)
H = None
C = None
bf = functions.broadcast_to(b[:feature_dimension],
(batchsize, feature_dimension))
br = functions.broadcast_to(b[feature_dimension:],
(batchsize, feature_dimension))
ct = initial_ct
for t in range(seq_length):
xt = X[..., t]
zt = Z[..., t]
ft = functions.sigmoid(F[..., t] + bf)
rt = functions.sigmoid(R[..., t] + br)
ct = ft * ct + (1 - ft) * zt
C = functions.expand_dims(ct, 2) if C is None else functions.concat(
(C, functions.expand_dims(ct, 2)), axis=2)
g_ct = ct
if use_tanh:
g_ct = functions.tanh(ct)
ht = rt * g_ct + (1 - rt) * xt
H = functions.expand_dims(ht, 2) if H is None else functions.concat(
(H, functions.expand_dims(ht, 2)), axis=2)
return H, C, C[..., -1]
def seq_encode(self, xs, cr):
if cr == "c":
embed_xs = self.embed_c(xs)
else:
embed_xs = self.embed_r(xs)
if self.wordvec_unchain:
embed_xs.unchain_backward()
batchsize, seq_length, dim = embed_xs.shape
sum_embed_xs = F.sum(embed_xs, axis=1)
embed_xs = F.reshape(embed_xs, (batchsize, 1, seq_length, dim))
# embed_avg = F.average_pooling_2d(embed_xs, ksize=(embed_xs.shape[2], 1))
# 1. wide_convolution
# 著者はnarrow?
xs_conv1 = F.tanh(self.conv1(embed_xs))
# xs_conv1_swap = F.reshape(F.swapaxes(xs_conv1, 1, 3),(batchsize, seq_length+3, 50))
xs_conv1_swap = F.swapaxes(
xs_conv1, 1,
3) # (batchsize, 50, seqlen, 1) --> (batchsize, 1, seqlen, 50)
return sum_embed_xs, xs_conv1, xs_conv1_swap
def __call__(self, x, h, c):
hy = []
cy = []
for i, name in enumerate(self.x_amps.layer_names):
hx_i = h[i]
cx_i = c[i]
gates = self.x_amps[name](x) + self.h_amps[name](hx_i)
i_gate, f_gate, c_gate, o_gate = F.split_axis(
gates, indices_or_sections=4, axis=1)
i_gate = F.sigmoid(i_gate)
f_gate = F.sigmoid(f_gate)
c_gate = F.tanh(c_gate)
o_gate = F.sigmoid(o_gate)
cy_i = (f_gate * cx_i) + (i_gate * c_gate)
hy_i = o_gate * F.sigmoid(cy_i)
cy.append(cy_i)
hy.append(hy_i)
x = self.dropout(hy_i)
return hy, cy
def __call__(self, x, adj):
"""
Args:
x: (batchsize, num_nodes, in_channels)
adj: (batchsize, num_edge_type, num_nodes, num_nodes)
Returns: (batchsize, hidden_channels)
"""
if x.dtype == self.xp.int32:
assert self.input_type == 'int'
else:
assert self.input_type == 'float'
h = self.embed(x) # (minibatch, max_num_atoms)
if self.scale_adj:
adj = rescale_adj(adj)
for rgcn_conv in self.rgcn_convs:
h = functions.tanh(rgcn_conv(h, adj))
h = self.rgcn_readout(h)
return h
def forward(self, x_list):
height = len(x_list[0])
batch_size = len(x_list)
xs = np.array(x_list).flatten()
itmp = self.x_i(_mkivar(xs))
i = F.reshape(itmp, (batch_size, 1, height, self.embed_size))
c1 = self.i_c1(i)
c2 = self.i_c2(i)
c3 = self.i_c3(i)
pc1 = F.max_pooling_2d(c1, (height, 1))
pc2 = F.max_pooling_2d(c2, (height, 1))
pc3 = F.max_pooling_2d(c3, (height, 1))
h1 = F.reshape(pc1, (batch_size, 128))
h2 = F.reshape(pc2, (batch_size, 128))
h3 = F.reshape(pc3, (batch_size, 128))
h = F.dropout(F.concat((h1, h2, h3), axis=1), 0.2)
h = F.dropout(F.tanh(self.h_h(h)), 0.2)
z = self.h_z(h)
return z
def __call__(self, x, c, test=False):
### text encoding
hc_mu = F.leaky_relu(self.lc_mu(c))
hc_var = F.leaky_relu(self.lc_var(c))
h_c = F.gaussian(hc_mu, hc_var)
h_c = F.expand_dims(h_c, axis=2)
h_c = F.expand_dims(h_c, axis=2)
h_c = F.tile(h_c, (1, 1, self.s16, self.s16))
### image encoder
h = self.c1(x, test=test)
h = self.c2(h, test=test)
h = self.c3(h, test=test)
### concate text and image
h = F.concat((h, h_c))
h = self.c_joint(h, test=test)
### residual block
h0 = self.cr1_0(h, test=test)
h0 = self.cr1_1(h0, test=test)
h = F.relu(h + h0)
h0 = self.cr2_0(h, test=test)
h0 = self.cr2_1(h0, test=test)
h = F.relu(h + h0)
h0 = self.cr3_0(h, test=test)
h0 = self.cr3_1(h0, test=test)
h = F.relu(h + h0)
h0 = self.cr4_0(h, test=test)
h0 = self.cr4_1(h0, test=test)
h = F.relu(h + h0)
### upsampling
h = self.dc1(h, test=test)
h = self.dc2(h, test=test)
h = self.dc3(h, test=test)
h = self.dc4(h, test=test)
h = F.tanh(self.c5(h))
if test:
return h
else:
return h, hc_mu, hc_var
def __call__(self, batchsize=64, z=None, y=None, **kwargs):
if z is None:
z = sample_continuous(self.dim_z, batchsize, distribution=self.distribution, xp=self.xp)
if y is None:
y = sample_categorical(self.n_classes, batchsize, distribution="uniform",
xp=self.xp) if self.n_classes > 0 else None
if (y is not None) and z.shape[0] != y.shape[0]:
raise Exception('z.shape[0] != y.shape[0], z.shape[0]={}, y.shape[0]={}'.format(z.shape[0], y.shape[0]))
h = z
h = self.l1(h)
h = F.reshape(h, (h.shape[0], -1, self.bottom_width, self.bottom_width))
h = self.block2(h, y, **kwargs)
h = self.block3(h, y, **kwargs)
h = self.block4(h, y, **kwargs)
h = self.block5(h, y, **kwargs)
h = self.b6(h)
h = self.activation(h)
h = F.tanh(self.l6(h))
return h
def forward(self, xs, ys):
#print(xs,ys)
#exit()
#xs = [x[::-1] for x in xs]
eos_dst = self.xp.array([self.v_eos_dst], np.int32)
ys_in = [F.concat([eos_dst, y], axis=0) for y in ys]
ys_out = [F.concat([y, eos_dst], axis=0) for y in ys]
# Both xs and ys_in are lists of arrays.
exs = sequence_embed(self.embed_x, xs)
eys = sequence_embed(self.embed_y, ys_in)
#print(list(map(lambda x: len(x),exs)))
#print(list(map(lambda x: len(x),eys)))
#exit()
batch = len(xs)
# None represents a zero vector in an encoder.
hx, cx, xs_states = self.encoder(None, None, exs)
hx = F.transpose(
F.reshape(F.transpose(hx, (1, 0, 2)),
(batch, self.n_layers, self.n_units * 2)), (1, 0, 2))
cx = F.transpose(
F.reshape(F.transpose(cx, (1, 0, 2)),
(batch, self.n_layers, self.n_units * 2)), (1, 0, 2))
_, _, os = self.decoder(hx, cx, eys)
ctxs = [self.att(xh, yh) for (xh, yh) in zip(xs_states, os)]
att_os = [
F.tanh(self.Wc(F.concat([ch, yh], axis=1)))
for (ch, yh) in zip(ctxs, os)
]
concat_os = F.concat(att_os, axis=0)
concat_ys_out = F.concat(ys_out, axis=0)
#print(concat_ys_out.dtype)
loss = F.sum(
F.softmax_cross_entropy(
self.Ws(concat_os), concat_ys_out, reduce='no')) / batch
chainer.report({'loss': loss}, self)
return loss
def __call__(self, bs, dim=100, y=None, test=False):
self.hiddens = []
z = self.generate_norm(bs, dim)
# Linear/BatchNorm/Branch/Nonlinear
h = self.linear0(z)
h = self.bn0(h, test)
z = self.generate_unif(bs, dim)
b = self.branch0(z, y, test)
h = h + b
h = self.act(h)
self.hiddens.append(h)
h = self.linear1(h)
h = self.bn1(h, test)
z = self.generate_unif(bs, dim)
b = self.branch1(z, y, test)
h = h + b
h = self.act(h)
self.hiddens.append(h)
h = self.linear2(h)
h = self.bn2(h, test)
z = self.generate_unif(bs, dim)
b = self.branch2(z, y, test)
h = h + b
h = self.act(h)
self.hiddens.append(h)
h = self.linear3(h)
h = self.bn3(h, test)
z = self.generate_unif(bs, dim)
b = self.branch3(z, y, test)
h = h + b
h = self.act(h)
self.hiddens.append(h)
h = self.linear4(h)
return h
#TODO: tanh?
return F.tanh(h)
def __call__(self, batchsize=64, z=None, y=None):
if z is None:
z = sample_continuous(self.dim_z, batchsize, distribution=self.distribution, xp=self.xp)
if y is None:
y = sample_categorical(self.n_classes, batchsize, distribution="uniform",
xp=self.xp) if self.n_classes > 0 else None
if (y is not None) and z.shape[0] != y.shape[0]:
raise ValueError('z.shape[0] != y.shape[0]')
print("B0", np.sum(z.data))
print("C2B0", np.sum(self.block2.c2.b.data))
h = z
h = self.l1(h)
h = F.reshape(h, (h.shape[0], -1, self.bottom_width, self.bottom_width))
print("B1", np.sum(h.data))
print("C2B1", np.sum(self.block2.c2.b.data))
h = self.block2(h, y)
print("B2", np.sum(h.data))
print("C2B2", np.sum(self.block2.c2.b.data))
h = self.block3(h, y)
print("B3", np.sum(h.data))
print("C2B3", np.sum(self.block2.c2.b.data))
h = self.block4(h, y)
print("B4", np.sum(h.data))
print("C2B4", np.sum(self.block2.c2.b.data))
h = self.b5(h)
print("B5", np.sum(h.data))
print("C2B5", np.sum(self.block2.c2.b.data))
h = self.activation(h)
print("B6", np.sum(h.data))
print("C2B6", np.sum(self.block2.c2.b.data))
h = F.tanh(self.c5(h))
print("B7", np.sum(h.data))
print("C2B7", np.sum(self.block2.c2.b.data))
return h
def compute_ctxt(previous_state, prev_word_embedding=None):
current_mb_size = previous_state.data.shape[0]
if current_mb_size < mb_size:
al_factor, _ = F.split_axis(precomputed_al_factor,
(current_mb_size, ), 0)
used_fb_concat, _ = F.split_axis(fb_concat,
(current_mb_size, ), 0)
if mask_length > 0:
used_concatenated_penalties = concatenated_penalties[:
current_mb_size]
else:
al_factor = precomputed_al_factor
used_fb_concat = fb_concat
if mask_length > 0:
used_concatenated_penalties = concatenated_penalties
state_al_factor = self.al_lin_s(previous_state)
#As suggested by Isao Goto
if prev_word_embedding is not None:
state_al_factor = state_al_factor + self.al_lin_y(
prev_word_embedding)
state_al_factor_bc = F.broadcast_to(
F.reshape(state_al_factor, (current_mb_size, 1, self.Ha)),
(current_mb_size, nb_elems, self.Ha))
a_coeffs = F.reshape(
self.al_lin_o(
F.reshape(F.tanh(state_al_factor_bc + al_factor),
(current_mb_size * nb_elems, self.Ha))),
(current_mb_size, nb_elems))
if mask_length > 0:
with cuda.get_device_from_array(used_concatenated_penalties):
a_coeffs = a_coeffs + used_concatenated_penalties # - 10000 * (1-used_concatenated_mask.data)
attn = F.softmax(a_coeffs)
ci = F.reshape(batch_matmul(attn, used_fb_concat, transa=True),
(current_mb_size, self.Hi))
return ci, attn
def run():
parser = argparse.ArgumentParser()
parser.add_argument('-io',
'--filename_obj',
type=str,
default='./examples/data/teapot.obj')
parser.add_argument('-ir',
'--filename_ref',
type=str,
default='./examples/data/example4_ref.png')
parser.add_argument('-or',
'--filename_output',
type=str,
default='./examples/data/example4_result.gif')
parser.add_argument('-mr', '--make_reference_image', type=int, default=0)
parser.add_argument('-g', '--gpu', type=int, default=0)
args = parser.parse_args()
working_directory = os.path.dirname(args.filename_output)
if args.make_reference_image:
make_reference_image(args.filename_ref, args.filename_obj)
model = Model(args.filename_obj, args.filename_ref)
model.to_gpu()
optimizer = chainer.optimizers.Adam(alpha=0.1)
optimizer.setup(model)
loop = tqdm.tqdm(range(1000))
for i in loop:
optimizer.target.cleargrads()
loss = model()
loss.backward()
optimizer.update()
images = model.renderer.render(model.vertices, model.faces,
cf.tanh(model.textures))
image = images.data.get()[0]
scipy.misc.toimage(image, cmin=0, cmax=1).save('%s/_tmp_%04d.png' %
(working_directory, i))
loop.set_description('Optimizing (loss %.4f)' % loss.data)
if loss.data < 70:
break
make_gif(working_directory, args.filename_output)
def __call__(self, fs, bs, h):
"""
Attentionの計算
:param fs: 順向きのEncoderの中間ベクトルが記録されたリスト
:param bs: 逆向きのEncoderの中間ベクトルが記録されたリスト
:param h: Decoderで出力された中間ベクトル
:return att_f: 順向きのEncoderの中間ベクトルの加重平均
:return att_b: 逆向きのEncoderの中間ベクトルの加重平均
:return att: 各中間ベクトルの重み
"""
# ミニバッチのサイズを記憶
batch_size = h.data.shape[0]
# ウェイトを記録するためのリストの初期化
ws = []
att = []
# ウェイトの合計値を計算するための値を初期化
sum_w = Variable(self.ARR.zeros((batch_size, 1), dtype='float32'))
# Encoderの中間ベクトルとDecoderの中間ベクトルを使ってウェイトの計算
for f, b in zip(fs, bs):
# 順向きEncoderの中間ベクトル、逆向きEncoderの中間ベクトル、Decoderの中間ベクトルを使ってウェイトの計算
w = self.hw(functions.tanh(self.fh(f) + self.bh(b) + self.hh(h)))
att.append(w)
# softmax関数を使って正規化する
w = functions.exp(w)
# 計算したウェイトを記録
ws.append(w)
sum_w += w
# 出力する加重平均ベクトルの初期化
att_f = Variable(
self.ARR.zeros((batch_size, self.hidden_size), dtype='float32'))
att_b = Variable(
self.ARR.zeros((batch_size, self.hidden_size), dtype='float32'))
for i, (f, b, w) in enumerate(zip(fs, bs, ws)):
# ウェイトの和が1になるように正規化
w /= sum_w
# ウェイト * Encoderの中間ベクトルを出力するベクトルに足していく
att_f += functions.reshape(functions.batch_matmul(f, w),
(batch_size, self.hidden_size))
att_b += functions.reshape(functions.batch_matmul(f, w),
(batch_size, self.hidden_size))
att = functions.concat(att, axis=1)
return att_f, att_b, att
def forward_one(x, target, label):
# make input window vector
distance = window // 2
char_vecs = list()
x = list(x)
for i in range(distance):
x.append('</s>')
x.insert(0,'<s>')
for i in range(-distance+1 , distance + 2):
char = x[target + i]
char_id = char2id[char]
char_vec = model.embed(get_onehot(char_id))
char_vecs.append(char_vec)
concat = F.concat(tuple(char_vecs))
hidden = model.hidden1(F.sigmoid(concat))
output = F.tanh(model.output(hidden))
dist = F.softmax(output)
#print(dist.data, label, np.argmax(dist.data))
correct = get_onehot(label)
return np.argmax(dist.data), F.softmax_cross_entropy(output, correct)
def _attention(self, h, s, batch_size, sequence_length):
decoder_hidden_size = self.decoder_hidden_size
encoder_hidden_size = self.encoder_hidden_size
input_shape = (batch_size, sequence_length, decoder_hidden_size)
weighted_h = F.reshape(self.W1(h), input_shape)
weighted_s = F.broadcast_to(F.expand_dims(self.W2(s), axis=1),
input_shape)
score = self.v(
F.reshape(F.tanh(weighted_s + weighted_h),
(batch_size * sequence_length, decoder_hidden_size)))
a = F.softmax(F.reshape(score, (batch_size, sequence_length)))
self.a = a
# c = F.matmul(F.reshape(h, (batch_size, encoder_hidden_size, sequence_length)),
# a[..., None])
c = F.batch_matmul(
F.reshape(h, (batch_size, encoder_hidden_size, sequence_length)),
a)
return F.reshape(c, (batch_size, encoder_hidden_size))
def __call__(self, z):
"""
Function that computs foward
Parametors
----------------
z: Variable
random vector drown from a uniform distribution,
this shape is (N, 100)
"""
h = F.relu(self.bn0(self.l0(z)))
h = F.reshape(h, (len(z), self.ch, self.bottom_width,
self.bottom_width)) # dataformat is NCHW
h = F.relu(self.bn1(self.dc1(h)))
h = F.relu(self.bn2(self.dc2(h)))
h = F.relu(self.bn3(self.dc3(h)))
h = F.relu(self.bn4(self.dc4(h)))
x = F.tanh(self.dc5(h))
return x
def forward(self, x):
#x = F.relu(self.linear(x))
#mb, _ = x.data.shape
#x = F.reshape(x, [mb, -1, int(img_height / 16), int(img_width / 16)])
#x = self.bn(x)
x = self.l1(x)
x = self.bn1(x)
x = F.relu(x)
x = self.l2(x)
x = self.bn2(x)
x = F.relu(x)
x = self.l3(x)
x = self.bn3(x)
x = F.relu(x)
x = self.l4(x)
x = self.bn4(x)
x = F.relu(x)
x = self.l5(x)
x = F.tanh(x)
return x
def get_scores(self,candidates,links,relations,edges,xp,mode,RC,EC): entities = set() for h,r,t,l in candidates: entities.add(h) entities.add(t) entities = list(entities) xe = self.get_context(entities,links,relations,edges,0,xp,RC,EC) xe = F.split_axis(xe,len(entities),axis=0) edict = dict() for e,x in zip(entities,xe): edict[e]=x diffs,rels = [],[] for h,r,t,l in candidates: rels.append(r) diffs.append(edict[h]-edict[t]) diffs = F.concat(diffs,axis=0) xr = self.embedR(xp.array(rels,'i')) if self.is_bound_wr: xr = F.tanh(xr) scores = F.batch_l2_norm_squared(diffs+xr) return scores
def decode(self, p, hiddens, t=None):
"""
@param p
@param t ground truth
"""
c = Seq2Seq.calculate_ct(hiddens, p.data)
vc = Variable(c)
q = F.tanh(self.w1(vc) + self.w2(p))
y = self.l3(q)
if self.phase is Seq2Seq.Train:
loss = F.mean_squared_error(y, t)
p = self.encode(y)
return p, loss
elif self.phase is Seq2Seq.Valid:
loss = F.mean_squared_error(y, t)
p = self.encode(y)
return p, loss
else: # Test
p = self.encode(y)
return p, y
def onestep(self, ys, hx, cx, oxs, hts):
bs = len(ys)
emb_ys = self.emb(ys)
if self.feeding:
hts = F.stack(hts)
emb_ys = F.expand_dims(emb_ys, axis=1)
emb_ys = F.concat((emb_ys, hts), axis=2)
hy, cy, oys = self.rnn(hx, cx, F.separate(emb_ys))
else:
emb_ys = F.split_axis(emb_ys, bs, 0)
hy, cy, oys = self.rnn(hx, cx, emb_ys)
oys = F.stack(oys)
oxs = F.stack(oxs)
cts = self.attn(oys, oxs)
cs = F.concat((oys, cts), axis=2)
hts = F.tanh(F.stack(sequence_linear(self.wc, cs)))
oys = self.wo(F.concat(hts, axis=0))
return hy, cy, oys, hts
def forward(self, x, index_array):
x = x.reshape(-1, self.dim)
if self.dropout_rate != 0:
x = F.dropout(x, ratio=self.dropout_rate)
z = self.x2z(x)
if self.activate == 'tanh':
z = F.tanh(z)
if self.activate == 'relu':
z = F.relu(z)
if self.is_residual:
z = z + x
split_array = F.split_axis(z, index_array, axis=0)[:-1]
a = []
for i in split_array:
if len(i) > 0:
a.append(F.average(i, axis=0))
else:
a.append(Variable(np.zeros(self.dim, dtype=np.float32)))
p = F.stack(a)
return p
def __call__(self, batchsize=64, z=None, **kwargs):
if z is None:
z = sample_continuous(self.dim_z,
batchsize,
distribution=self.distribution,
xp=self.xp)
h = z
h = self.l1(h)
h = F.reshape(h, (h.shape[0], -1, self.bottom_width,
self.bottom_width)) # (Batchsize, auto, 4, 4)
h = self.block2(h)
h = self.block3(h)
h = self.block4(h)
h = self.block5(h)
h = self.block6(h)
h = self.block7(h)
h = self.b7(h)
h = self.activation(h)
h = F.tanh(self.l7(h))
return h
def __call__(self, hy, ys):
context_vector = []
# reshape: (1, batch_size, hidden_size) -> (batch_size, hidden_size)
hy = F.reshape(hy, hy.shape[1:])
for h, y in zip(hy, ys):
'''
各文に対して、最後の隠れ状態h:(1, hidden_size)とその単語y:(単語数, hidden_size)とで要素積をとる
broadcast_to: (1, hidden_size) -> (単語数, hidden_size)
sum: (単語数, hidden_size) -> (hidden_size, )
'''
h = F.broadcast_to(h, y.shape)
m = F.sum(h * y, axis=0)
context_vector.append(m)
context_vector = F.stack([v for v in context_vector], axis=0)
sentence_vector = F.tanh(
self.linear(
F.dropout(F.concat((hy, context_vector), axis=1),
self.dropout)))
return sentence_vector
def decode_once(self, x, state, train=True):
c = state['c']
h = state['h']
h_tilde = state.get('h_tilde', None)
emb = self.trg_emb(x)
lstm_in = self.eh(emb) + self.hh(h)
if h_tilde is not None:
lstm_in += self.ch(h_tilde)
c, h = F.lstm(c, lstm_in)
a = self.attender(h, train=train)
h_tilde = F.concat([a, h])
h_tilde = F.tanh(self.w_c(h_tilde))
o = self.ho(h_tilde)
state['c'] = c
state['h'] = h
state['h_tilde'] = h_tilde
return o, state
def pi_and_v(self, x):
h = F.relu(self.conv1(x[:, 0:1, :, :]))
h = self.diconv2(h)
h = self.diconv3(h)
h = self.diconv4(h)
h_pi = self.diconv5_pi(h)
x_t = self.diconv6_pi(h_pi)
h_t1 = x[:, -64:, :, :]
z_t = F.sigmoid(self.conv7_Wz(x_t) + self.conv7_Uz(h_t1))
r_t = F.sigmoid(self.conv7_Wr(x_t) + self.conv7_Ur(h_t1))
h_tilde_t = F.tanh(self.conv7_W(x_t) + self.conv7_U(r_t * h_t1))
h_t = (1 - z_t) * h_t1 + z_t * h_tilde_t
pout = self.conv8_pi(h_t)
h_V = self.diconv5_V(h)
h_V = self.diconv6_V(h_V)
vout = self.conv7_V(h_V)
return pout, vout, h_t