def forward_one(x, target, label, train_flag):
# make input window vector
distance = window // 2
char_vecs = list()
char_type_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)
char_concat = F.concat(tuple(char_vecs))
for i in range(-distance+1 , distance + 2):
char = x[target + i]
char_type = make_char_type(char)
char_type_id = char_type2id[char_type]
char_type_vec = model.char_type_embed(get_onehot(char_type_id))
char_type_vecs.append(char_type_vec)
char_type_concat = F.concat(tuple(char_type_vecs))
#dropout_concat = F.dropout(concat, ratio=dropout_rate, train=train_flag)
concat = F.concat((char_concat, char_type_concat))
hidden = F.sigmoid(model.hidden1(concat))
output = model.output(hidden)
dist = F.softmax(output)
#print(dist.data, label, np.argmax(dist.data))
correct = get_onehot(label)
#print(output.data, correct.data)
return np.argmax(dist.data), F.softmax_cross_entropy(output, correct)
def forward(self, xs, ys):
xs = [x[::-1] for x in xs]
eos = self.xp.array([EOS], numpy.int32)
ys_in = [F.concat([eos, y], axis=0) for y in ys]
ys_out = [F.concat([y, eos], 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)
batch = len(xs)
# None represents a zero vector in an encoder.
hx, cx, _ = self.encoder(None, None, exs)
_, _, os = self.decoder(hx, cx, eys)
# It is faster to concatenate data before calculating loss
# because only one matrix multiplication is called.
concat_os = F.concat(os, axis=0)
concat_ys_out = F.concat(ys_out, axis=0)
loss = F.sum(F.softmax_cross_entropy(
self.W(concat_os), concat_ys_out, reduce='no')) / batch
chainer.report({'loss': loss}, self)
n_words = concat_ys_out.shape[0]
perp = self.xp.exp(loss.array * batch / n_words)
chainer.report({'perp': perp}, self)
return loss
def render(self, vertices, faces, textures):
# fill back
if self.fill_back:
faces = cf.concat((faces, faces[:, :, ::-1]), axis=1).data
textures = cf.concat((textures, textures.transpose((0, 1, 4, 3, 2, 5))), axis=1)
# lighting
faces_lighting = neural_renderer.vertices_to_faces(vertices, faces)
textures = neural_renderer.lighting(
faces_lighting,
textures,
self.light_intensity_ambient,
self.light_intensity_directional,
self.light_color_ambient,
self.light_color_directional,
self.light_direction)
# viewpoint transformation
if self.camera_mode == 'look_at':
vertices = neural_renderer.look_at(vertices, self.eye)
elif self.camera_mode == 'look':
vertices = neural_renderer.look(vertices, self.eye, self.camera_direction)
# perspective transformation
if self.perspective:
vertices = neural_renderer.perspective(vertices, angle=self.viewing_angle)
# rasterization
faces = neural_renderer.vertices_to_faces(vertices, faces)
images = neural_renderer.rasterize(
faces, textures, self.image_size, self.anti_aliasing, self.near, self.far, self.rasterizer_eps,
self.background_color)
return images
def __call__(self, src, is_train=False, xp=np):
# Some namings
B = len(src) # Batch Size
N = len(src[0]) # length of source
H = self.H
src_col = lambda x: Variable(self.xp.array([src[i][x] for i in range(B)], dtype=np.int32))
embed = lambda e, x: e(self.IE(x), is_train=is_train)
bi_rnn = lambda x, y: self.AE(F.concat((x[0], y[1]), axis=1))
concat_source = lambda S, s: s if S is None else F.concat((S, s), axis=2)
# State Reset
self.EF.reset_state()
self.EB.reset_state()
# Forward + backward encoding
s = []
for j in range(N):
s.append((
embed(self.EF, src_col(j)),
embed(self.EB, src_col(-j-1))
))
# Joining the encoding data together
S = None
for j in range(N):
s_j = bi_rnn(s[j], s[-j-1])
S = concat_source(S, F.reshape(s_j, (B, H, 1)))
S = F.swapaxes(S, 1, 2)
return S, s_j
def __call__(self, x):
for nth in range(self.layers):
if getattr(self, 'P' + str(nth)) is None:
setattr(self, 'P' + str(nth), variable.Variable(
self.xp.zeros(self.sizes[nth], dtype=x.data.dtype),
volatile='auto'))
E = [None] * self.layers
for nth in range(self.layers):
if nth == 0:
E[nth] = F.concat((F.relu(x - getattr(self, 'P' + str(nth))),
F.relu(getattr(self, 'P' + str(nth)) - x)))
else:
A = F.max_pooling_2d(F.relu(getattr(self, 'ConvA' + str(nth))(E[nth - 1])), 2, stride = 2)
E[nth] = F.concat((F.relu(A - getattr(self, 'P' + str(nth))),
F.relu(getattr(self, 'P' + str(nth)) - A)))
R = [None] * self.layers
for nth in reversed(range(self.layers)):
if nth == self.layers - 1:
R[nth] = getattr(self, self.rnn_module + str(nth))((E[nth],))
else:
upR = F.unpooling_2d(R[nth + 1], 2, stride = 2, cover_all=False)
R[nth] = getattr(self, self.rnn_module + str(nth))((E[nth], upR))
if nth == 0:
setattr(self, 'P' + str(nth), F.clipped_relu(getattr(self, 'ConvP' + str(nth))(R[nth]), 1.0))
else:
setattr(self, 'P' + str(nth), F.relu(getattr(self, 'ConvP' + str(nth))(R[nth])))
return self.P0
def forward(self, data):
self.reset_state()
x_list = [XP.iarray([d[0]]) for d in data]
ep_list = [self.p_embed(x) for x in x_list]
ec_list = [self.c_embed(x) for x in x_list]
er_list = [self.r_embed(x) for x in x_list]
p_list = self.p_encode(ep_list)
c_list = self.c_encode(ec_list)
r_list = self.r_encode(er_list)
P = functions.reshape(
functions.concat(p_list, 0),
(1, len(data), self.hidden_size))
C = functions.reshape(
functions.concat(c_list, 0),
(1, len(data), self.hidden_size))
R = functions.concat(r_list, 0)
parent_scores = functions.reshape(
functions.batch_matmul(C, P, transb=True),
(len(data), len(data)))
root_scores = functions.reshape(
self.r_scorer(R),
(1, len(data)))
return parent_scores, root_scores
def __call__(self, x, y):
"""
:param x: ミニバッチの入力データ
:param y: 入力データに対応するミニバッチの出力
:return: 誤差
"""
batch_size = len(x)
eos = self.xp.array([EOS], dtype='int32')
# EOS信号の埋め込み
y_in = [F.concat((eos, tmp), axis=0) for tmp in y]
y_out = [F.concat((tmp, eos), axis=0) for tmp in y]
# Embedding Layer
emb_x = [self.x_embed(tmp) for tmp in x]
emb_y = [self.y_embed(tmp) for tmp in y_in]
# Encoder, Decoderへの入力
h, c, a = self.encoder(None, None, emb_x) # h => hidden, c => cell, a => output(Attention)
_, _, dec_hs = self.decoder(h, c, emb_y) # dec_hs=> output
# Output Layerの計算
loss = 0
accuracy = 0
for dec_h, t, attention in zip(dec_hs, y_out, a):
# o = self.y(dec_h)
o = self.global_attention_layer(dec_h, attention) # Attention Layerの計算
loss += F.softmax_cross_entropy(o, t) # 誤差計算
accuracy += F.accuracy(o, t) # 精度計算
loss /= batch_size
accuracy /= batch_size
return loss, accuracy
def __call__(self, x, train=True):
# First Convolution
c0 = self.bnF1(x)
c0 = self.convF1(c0)
c0 = F.relu(c0)
c0 = self.bnF2(c0)
c0 = self.convF2(c0)
c0 = F.relu(c0)
c0 = self.bnF3(c0)
c0 = self.convF3(c0)
c0 = F.relu(c0)
# Atrous Convolution size 3
cn3 = self.bn1x1_D3(c0)
cn3 = self.conv1x1_D3(cn3)
cn3 = self.bnD3(cn3)
cn3 = self.dilate_conv6(cn3)
cn3 = F.relu(cn3)
# Atrous Convolution size 6
cn3_con = F.concat((c0, cn3), axis=1)
cn6 = self.bn1x1_D6(cn3_con)
cn6 = self.conv1x1_D6(cn6)
cn6 = self.bnD6(cn6)
cn6 = self.dilate_conv6(cn6)
cn6 = F.relu(cn6)
# Atrous Convolution size 12
cn6_con = F.concat((cn3_con, cn6), axis=1)
cn12 = self.bn1x1_D12(cn6_con)
cn12 = self.conv1x1_D12(cn12)
cn12 = self.bnD12(cn12)
cn12 = self.dilate_conv12(cn12)
cn12 = F.relu(cn12)
# Atrous Convolution size 18
cn12_con = F.concat((cn6_con, cn12), axis=1)
cn18 = self.bn1x1_D18(cn12_con)
cn18 = self.conv1x1_D18(cn18)
cn18 = self.bnD18(cn18)
cn18 = self.dilate_conv18(cn18)
cn18 = F.relu(cn18)
# Atrous Convolution size 24
cn18_con = F.concat((cn12_con, cn18), axis=1)
cn24 = self.bn1x1_D24(cn18_con)
cn24 = self.conv1x1_D24(cn24)
cn24 = self.bnD24(cn24)
cn24 = self.dilate_conv24(cn24)
cn24 = F.relu(cn24)
# Last convolution
cn24_con = F.concat((cn18_con, cn24), axis=1)
cL = self.bnL(cn24_con)
cL = self.convL(cL)
out = F.softmax(cL, axis=1)
return out
def forward_one(x, target, hidden, prev_c, train_flag):
# 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))
dropout_concat = F.dropout(concat, ratio=dropout_rate, train=train_flag)
concat = F.concat((concat, hidden))
i_gate = F.sigmoid(model.i_gate(concat))
f_gate = F.sigmoid(model.f_gate(concat))
o_gate = F.sigmoid(model.o_gate(concat))
concat = F.concat((hidden, i_gate, f_gate, o_gate))
prev_c, hidden = F.lstm(prev_c, concat)
output = model.output(hidden)
dist = F.softmax(output)
#print(dist.data, label, np.argmax(dist.data))
#correct = get_onehot(label)
#print(output.data, correct.data)
return dist
def forward(self, x):
n_batch, n_atom, n_feature = x.shape
atom_repeat = functions.reshape(x, (n_batch, 1, n_atom, n_feature))
atom_repeat = functions.broadcast_to(
atom_repeat, (n_batch, n_atom, n_atom, n_feature))
atom_repeat = functions.reshape(atom_repeat,
(n_batch, n_atom * n_atom, n_feature))
atom_tile = functions.reshape(x, (n_batch, n_atom, 1, n_feature))
atom_tile = functions.broadcast_to(
atom_tile, (n_batch, n_atom, n_atom, n_feature))
atom_tile = functions.reshape(atom_tile,
(n_batch, n_atom * n_atom, n_feature))
pair_x0 = functions.concat((atom_tile, atom_repeat), axis=2)
pair_x0 = functions.reshape(pair_x0,
(n_batch * n_atom * n_atom, n_feature * 2))
for l in self.linear_layers:
pair_x0 = l(pair_x0)
pair_x0 = functions.relu(pair_x0)
pair_x0 = functions.reshape(pair_x0,
(n_batch, n_atom * n_atom, self.n_channel))
pair_x1 = functions.concat((atom_repeat, atom_tile), axis=2)
pair_x1 = functions.reshape(pair_x1,
(n_batch * n_atom * n_atom, n_feature * 2))
for l in self.linear_layers:
pair_x1 = l(pair_x1)
pair_x1 = functions.relu(pair_x1)
pair_x1 = functions.reshape(pair_x1,
(n_batch, n_atom * n_atom, self.n_channel))
return pair_x0 + pair_x1
def channel_normalize(x, test=False):
s0, s1, s2, s3 = x.data.shape
cavg = F.reshape(F.sum(x, axis=1) / s1, (s0, 1, s2, s3))
xavg = F.concat(s1 * [cavg])
cvar = F.reshape(F.sum((x - xavg) ** 2, axis=1) / s1, (s0, 1, s2, s3))
xvar = F.concat(s1 * [cvar])
return (x - xavg) / (xvar + 1e-5) ** 0.5
def compute_vecs(self, word_ids, word_boundaries, phrase_num,
char_vecs=None):
word_ids = my_variable(word_ids, volatile=not self.train)
word_embs = self.emb(word_ids) # total_len x dim
word_embs_reshape = F.reshape(word_embs, (1, 1, -1, self.emb_dim))
if self.word_level_flag and char_vecs is not None:
# print char_vecs.data.shape
# print word_embs.data.shape
word_embs = F.concat([word_embs, char_vecs], axis=1)
# print word_embs.data.shape
dim = self.emb_dim + self.add_dim
word_embs_reshape = F.reshape(word_embs, (1, 1, -1, dim))
# 1 x 1 x total_len x dim
# convolution
word_emb_conv = self.conv(word_embs_reshape)
# 1 x dim x total_len x 1
word_emb_conv_reshape = F.reshape(word_emb_conv,
(self.hidden_dim, -1))
# max
word_emb_conv_reshape = F.split_axis(word_emb_conv_reshape,
word_boundaries, axis=1)
embs = [F.max(word_emb_conv_word, axis=1) for i, word_emb_conv_word in
enumerate(word_emb_conv_reshape) if i % 2 == 1]
embs = F.concat(embs, axis=0)
phrase_emb_conv = F.reshape(embs,
(phrase_num, self.hidden_dim))
return phrase_emb_conv
def forward_one(x, target, label, hidden_vec, prev_c):
# 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, distance + 1):
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))
concat = F.concat((concat, hidden_vec))
i_gate = F.sigmoid(model.i_gate(concat))
f_gate = F.sigmoid(model.f_gate(concat))
o_gate = F.sigmoid(model.o_gate(concat))
concat = F.concat((hidden_vec, i_gate, f_gate, o_gate))
prev_c, hidden_vec = F.lstm(prev_c, concat)
pred = F.softmax(model.output(hidden_vec))
#pred = add_delta(pred)
correct = get_onehot(label)
return np.argmax(pred), F.softmax_cross_entropy(pred, correct)
def __call__(self, x):
batch_size = len(x)
eos = np.array([EOS], dtype='int32')
# EOS信号の埋め込み
in_x = [F.concat((eos, tmp), axis=0) for tmp in x]
in_y = [F.concat((tmp, eos), axis=0) for tmp in x]
# Embedding Layer
emb_x = [self.embed(tmp) for tmp in in_x]
# LSTMへの入力
_, _, outputs = self.h(None, None, emb_x) # h => hidden, c => cell, a => output(Attention)
# Output Layerの計算
loss = 0
for output, t in zip(outputs, in_y):
o = self.y(output)
# print(o.shape[0])
# print(t[1:].shape[0])
loss += F.softmax_cross_entropy(o, t) # 誤差計算
loss /= batch_size
return 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 _encode(self, x_list):
batch_size = len(x_list[0])
source_length = len(x_list)
# Encoding
fc = bc = f = b = _zeros((batch_size, self.hidden_size))
i_list = [self.x_i(_mkivar(x)) for x in x_list]
f_list = []
b_list = []
for i in i_list:
fc, f = F.lstm(fc, self.i_f(i) + self.f_f(f))
f_list.append(f)
for i in reversed(i_list):
bc, b = F.lstm(bc, self.i_b(i) + self.b_b(b))
b_list.append(b)
b_list.reverse()
# Making concatenated matrix
# {f,b}_mat: shape = [batch, srclen, hidden]
f_mat = F.concat([F.expand_dims(f, 1) for f in f_list], 1)
b_mat = F.concat([F.expand_dims(b, 1) for b in b_list], 1)
# fb_mat: shape = [batch, srclen, 2 * hidden]
fb_mat = F.concat([f_mat, b_mat], 2)
# fbe_mat: shape = [batch * srclen, atten]
fbe_mat = self.fb_e(
F.reshape(fb_mat, [batch_size * source_length, 2 * self.hidden_size]))
return fb_mat, fbe_mat, fc, bc, f_list[-1], b_list[0]
def __call__(self, *inputs):
xs = inputs[:len(inputs) // 2]
ys = inputs[len(inputs) // 2:]
xs = [x[::-1] for x in xs]
batch = len(xs)
eos = self.xp.zeros(1, self.xp.int32)
ys_in = [F.concat([eos, y], axis=0) for y in ys]
ys_out = [F.concat([y, eos], axis=0) for y in ys]
eys = sequence_embed(self.embed_y, ys_in)
# Receive hidden states from encoder process and decode.
_, _, os, _ = self.mn_decoder(eys)
# It is faster to concatenate data before calculating loss
# because only one matrix multiplication is called.
concat_os = F.concat(os, axis=0)
concat_ys_out = F.concat(ys_out, axis=0)
loss = F.sum(F.softmax_cross_entropy(
self.W(concat_os), concat_ys_out, reduce='no')) / batch
reporter.report({'loss': loss.data}, self)
n_words = concat_ys_out.shape[0]
perp = self.xp.exp(loss.data * batch / n_words)
reporter.report({'perp': perp}, self)
return loss
def output_and_loss_from_seq_batch(self, y_seq_batch, t_seq_batch, normalize=None):
y = F.concat(y_seq_batch, axis=0)
y = F.dropout(y, ratio=self.dropout)
t = F.concat(t_seq_batch, axis=0)
loss = self.output.output_and_loss(y, t)
if normalize is not None:
loss *= 1. * t.shape[0] / normalize
else:
loss *= t.shape[0]
return loss
def forward_one(x,target, hidden, prev_c, model):
# make input window vector
distance = window // 2
char_vecs = list()
char_type_vecs = list()
x = list(x)
for i in range(distance):
x.append('</s>')
x.append('</s>')
x.insert(0,'<s>')
x.insert(0,'<s>')
for i in range(-distance , distance+1):
char = x[target+2 + i]
try:
char_id = char2id[char]
except(KeyError):
char_id = char2id['UNK']
char_vec = model.embed(get_onehot(char_id))
char_vecs.append(char_vec)
bi_gram = x[target+2+i] + x[target+2+i+1]
try:
bi_gram_id = char2id[bi_gram]
except(KeyError):
bi_gram_id = char2id['UNK']
bi_gram_char_vec = model.embed(get_onehot(bi_gram_id))
char_vecs.append(bi_gram_char_vec)
char_concat = F.concat(tuple(char_vecs))
for i in range(-distance, distance+1):
char = x[target+2+ i]
pre_char = x[target+2+ i + 1]
char_type = make_char_type(char)
pre_char_type = make_char_type(pre_char)
bi_gram_type = pre_char_type + char_type
char_type_id = char_type2id[char_type]
bigram_type_id = char_type2id[bi_gram_type]
char_type_vec = model.char_type_embed(get_onehot(char_type_id))
bigram_type_vec = model.char_type_embed(get_onehot(bigram_type_id))
char_type_vecs.append(char_type_vec)
char_type_vecs.append(bigram_type_vec)
char_type_concat = F.concat(tuple(char_type_vecs))
#dropout_concat = F.dropout(concat, ratio=dropout_rate, train=train_flag)
concat = F.concat((char_concat, char_type_concat))
concat = F.concat((concat, hidden))
i_gate = F.sigmoid(model.i_gate(concat))
f_gate = F.sigmoid(model.f_gate(concat))
o_gate = F.sigmoid(model.o_gate(concat))
concat = F.concat((hidden, i_gate, f_gate, o_gate))
prev_c, hidden = F.lstm(prev_c, concat)
output = model.output(hidden)
dist = F.softmax(output)
return np.argmax(dist.data)
def forward(self, data):
self.reset_state()
x_list = [XP.iarray([d[0]]) for d in data]
pe_list = [self.p_embed(x) for x in x_list]
ce_list = [self.c_embed(x) for x in x_list]
re_list = [self.r_embed(x) for x in x_list]
pf_list = []
for pe in pe_list:
pf_list.append(self.p_forward(pe))
cf_list = []
for ce in ce_list:
cf_list.append(self.c_forward(ce))
rf_list = []
for re in re_list:
rf_list.append(self.r_forward(re))
pb_list = []
for pe in reversed(pe_list):
pb_list.append(self.p_backward(pe))
cb_list = []
for ce in reversed(ce_list):
cb_list.append(self.c_backward(ce))
rb_list = []
for re in reversed(re_list):
rb_list.append(self.r_backward(re))
pc_list = [self.p_combine(pf, pb) for pf, pb in zip(pf_list, pb_list)]
cc_list = [self.c_combine(cf, cb) for cf, cb in zip(cf_list, cb_list)]
rc_list = [self.r_combine(rf, rb) for rf, rb in zip(rf_list, rb_list)]
P = functions.reshape(
functions.concat(pc_list, 0),
(1, len(data), self.hidden_size))
C = functions.reshape(
functions.concat(cc_list, 0),
(1, len(data), self.hidden_size))
R = functions.concat(rc_list, 0)
parent_scores = functions.reshape(
functions.batch_matmul(C, P, transb=True),
(len(data), len(data)))
root_scores = functions.reshape(
self.r_scorer(R),
(1, len(data)))
return parent_scores, root_scores
def single_step_forward(self, single_timestep_inputs, rel_rec, rel_send, single_timestep_rel_type):
# single_timestep_inputs: [batch_size, num_sequences, num_nodes, feature_dims]
# single_timestep_rel_type: [batch_size, num_sequences, num_edges, edge_types]
batch_size, num_sequences, num_edges, _ = single_timestep_rel_type.shape
_, num_nodes = rel_rec.shape
# Node2edge
# rel_rec: [num_edges, num_nodes]
# rel_send: [num_edges, num_nodes]
receivers = F.matmul(rel_rec, single_timestep_inputs)
senders = F.matmul(rel_send, single_timestep_inputs)
pre_msg = F.concat([receivers, senders], axis=-1)
# pre_msg: [batch_size, num_sequences, num_edges, 2 * feature_dims]
pre_msg = F.reshape(pre_msg, [batch_size * num_sequences * num_edges, -1])
all_msgs = chainer.Variable(
pre_msg.xp.zeros((batch_size, num_sequences, num_edges, self.msg_out_shape),
dtype=single_timestep_rel_type.dtype))
if self.skip_first_edge_type:
start_idx = 1
else:
start_idx = 0
# Run separate MLP for every edge type
# NOTE: To exlude one edge type, simply offset range by 1
for i in range(start_idx, len(self.msg_fc2)):
msg = F.relu(self.msg_fc1[i](pre_msg))
msg = F.dropout(msg, self.dropout_prob)
msg = F.relu(self.msg_fc2[i](msg))
# msg: [batch_size * num_sequences * num_edges, msg_hid]
msg = F.reshape(msg, [batch_size, num_sequences, num_edges, -1])
msg = msg * single_timestep_rel_type[:, :, :, i:i + 1]
all_msgs += msg
# Aggregate all msgs to receiver
# all_msgs: [batch_size, num_sequences, num_edges, msg_out_shape]
# rel_rec: [num_edges, num_nodes]
agg_msgs = F.matmul(rel_rec.T, all_msgs)
# Skip connection
aug_inputs = F.concat([single_timestep_inputs, agg_msgs], axis=-1)
# aug_inputs: [batch_size, num_sequences, num_nodes, msg_out_shape + feature_dims]
aug_inputs = F.reshape(aug_inputs, [batch_size * num_sequences * num_nodes, -1])
# Output MLP
pred = F.dropout(F.relu(self.out_fc1(aug_inputs)), self.dropout_prob)
pred = F.dropout(F.relu(self.out_fc2(pred)), self.dropout_prob)
pred = self.out_fc3(pred)
pred = F.reshape(pred, [batch_size, num_sequences, num_nodes, -1])
# Predict position/velocity difference
return single_timestep_inputs + pred
def generate(self, x, condition):
self.embed_queue = F.concat((self.embed_queue[:, :, 1:], x), axis=2)
x = self.embed(self.embed_queue)
if self.use_embed_tanh:
x = F.tanh(x)
x = F.relu(self.resnet.generate(x, condition))
self.proj1_queue = F.concat((self.proj1_queue[:, :, 1:], x), axis=2)
x = F.relu(self.proj1(self.proj1_queue))
self.proj2_queue3 = F.concat((self.proj2_queue3[:, :, 1:], x), axis=2)
x = self.proj2(self.proj2_queue3)
return x
def __call__(self, imgs, questions):
feat = self.feat_extractor(imgs)
# Append relative coordinates to each location in the feature maps.
n, c, h, w = feat.shape
spatial_area = h * w
xp = self.xp
coords_h = xp.linspace(-1, 1, h, dtype=feat.dtype)
coords_w = xp.linspace(-1, 1, w, dtype=feat.dtype)
coords_hh, coords_ww = xp.meshgrid(coords_h, coords_w)
coords_hh = coords_hh[None]
coords_ww = coords_ww[None]
coords = xp.concatenate((coords_hh, coords_ww), axis=0)
coords = coords.reshape(2, -1)
coords = coords[None] # (1, 2, spatial_area * spatial_area)
coords = xp.repeat(coords, n, axis=0)
# Coordinates may be cached here but the performance gain is not
# significant so it is skipped in favor of readability.
feat = feat.reshape(n, c, spatial_area)
h = F.concat((feat, coords), axis=1) # (n, c + 2, spatial_area)
# Create coordinate pairs (differentiable meshgrid).
h_hh = F.expand_dims(h, 2)
h_ww = F.expand_dims(h, 3)
h_hh = F.repeat(h_hh, spatial_area, axis=2)
h_ww = F.repeat(h_ww, spatial_area, axis=3)
h = F.concat((h_hh, h_ww), axis=1)
# Append questions to each coordinate pair.
questions = questions.astype(imgs.dtype)
questions = questions[:, :, None, None]
questions = F.tile(questions, (1, 1, spatial_area, spatial_area))
h = F.concat((h, questions), axis=1)
# (n, (c + 2) * 2 + questions_length, spatial_area, spatial_area)
# g.
h = F.transpose(h, (0, 2, 3, 1))
h = F.reshape(h, (n * spatial_area * spatial_area, -1))
h = self.g(h)
h = F.reshape(h, (n, spatial_area * spatial_area, -1))
h = F.sum(h, axis=1)
h = self.f(h)
# Logits.
h = self.fc(h)
return h
def forward(self, x):
y1 = self.model['conv1/7x7_s2'](x)
h = F.relu(y1)
h = F.local_response_normalization(self.pool_func(h, 3, stride=2), n=5)
h = F.relu(self.model['conv2/3x3_reduce'](h))
y2 = self.model['conv2/3x3'](h)
h = F.relu(y2)
h = self.pool_func(F.local_response_normalization(h, n=5), 3, stride=2)
out1 = self.model['inception_3a/1x1'](h)
out3 = self.model[
'inception_3a/3x3'](F.relu(self.model['inception_3a/3x3_reduce'](h)))
out5 = self.model[
'inception_3a/5x5'](F.relu(self.model['inception_3a/5x5_reduce'](h)))
pool = self.model[
'inception_3a/pool_proj'](self.pool_func(h, 3, stride=1, pad=1))
y3 = F.concat((out1, out3, out5, pool), axis=1)
h = F.relu(y3)
out1 = self.model['inception_3b/1x1'](h)
out3 = self.model[
'inception_3b/3x3'](F.relu(self.model['inception_3b/3x3_reduce'](h)))
out5 = self.model[
'inception_3b/5x5'](F.relu(self.model['inception_3b/5x5_reduce'](h)))
pool = self.model[
'inception_3b/pool_proj'](self.pool_func(h, 3, stride=1, pad=1))
y4 = F.concat((out1, out3, out5, pool), axis=1)
h = F.relu(y4)
h = self.pool_func(h, 3, stride=2)
out1 = self.model['inception_4a/1x1'](h)
out3 = self.model[
'inception_4a/3x3'](F.relu(self.model['inception_4a/3x3_reduce'](h)))
out5 = self.model[
'inception_4a/5x5'](F.relu(self.model['inception_4a/5x5_reduce'](h)))
pool = self.model[
'inception_4a/pool_proj'](self.pool_func(h, 3, stride=1, pad=1))
y5 = F.concat((out1, out3, out5, pool), axis=1)
h = F.relu(y5)
out1 = self.model['inception_4b/1x1'](h)
out3 = self.model[
'inception_4b/3x3'](F.relu(self.model['inception_4b/3x3_reduce'](h)))
out5 = self.model[
'inception_4b/5x5'](F.relu(self.model['inception_4b/5x5_reduce'](h)))
pool = self.model[
'inception_4b/pool_proj'](self.pool_func(h, 3, stride=1, pad=1))
y6 = F.concat((out1, out3, out5, pool), axis=1)
h = F.relu(y6)
return [y1, y2, y3, y4, y5, y6]
def zoom_x2(batch):
shape = batch.data.shape
channel_shape = shape[0:-2]
height, width = shape[-2:]
volume = reduce(operator.mul,shape,1)
b1 = F.reshape(batch,(volume,1))
b2 = F.concat([b1,b1],1)
b3 = F.reshape(b2,(volume/width,2*width))
b4 = F.concat([b3,b3],1)
return F.reshape(b4, channel_shape + (2*height ,) + (2*width ,))
def fit_partial(self, rsty_ids, raut_ids, rwrd_ids, window=5):
doc_idx, usr_idx, wrd_idx = move(self.xp, rsty_ids, raut_ids, rwrd_ids)
pivot = self.embed(next(move(self.xp, rwrd_ids[window: -window])))
sty_at_pivot = rsty_ids[window: -window]
aut_at_pivot = raut_ids[window: -window]
sty = self.mixture_stories(next(move(self.xp, sty_at_pivot)))
aut = self.mixture_authors(next(move(self.xp, aut_at_pivot)))
start, end = window, rwrd_ids.shape[0] - window
context = (F.dropout(sty, self.dropout_ratio) +
F.dropout(aut, self.dropout_ratio) +
F.dropout(pivot, self.dropout_ratio))
n_frame = 2 * window
# Precompute all neg samples since they're indep of frame
size = context.data.shape[0]
samples = self.sampler.sampler.sample((self.n_samples * n_frame, size))
samples = chainer.cuda.cupy.split(samples.ravel(), n_frame)
sources = []
targets = []
weights = []
for frame in range(-window, window + 1):
# Predict word given context and pivot word
# The target starts before the pivot
# Skip predicting the current pivot
if frame == 0:
continue
# Here we're creating a weight mask. We don't want to
# predict tokens that are outside this document or user
# scope.
wrd_at_target = rwrd_ids[start + frame: end + frame]
sty_at_target = rsty_ids[start + frame: end + frame]
aut_at_target = raut_ids[start + frame: end + frame]
sty_is_same = sty_at_target == sty_at_pivot
usr_is_same = aut_at_target == aut_at_pivot
is_same = sty_is_same & usr_is_same
weight, = move(self.xp, is_same.astype('float32'))
target, = move(self.xp, wrd_at_target)
sources.append(context)
targets.append(target)
weights.append(weight)
sample, = move(self.xp, samples.pop())
targets.append(sample)
for _ in range(self.n_samples):
# Note that the context is now negative
sources.append(-context)
weights.append(weight)
sources = F.concat(sources, axis=0)
targets = F.concat(targets, axis=0)
weights = F.concat(weights, axis=0)
loss = self.loss(sources, targets, weights)
return loss
def __call__(self, X):
h1 = F.leaky_relu(self.norm1(self.conv1(X)))
h2 = F.leaky_relu(self.norm2(self.conv2(h1)))
h3 = F.leaky_relu(self.norm3(self.conv3(h2)))
h4 = F.leaky_relu(self.norm4(self.conv4(h3)))
h5 = F.leaky_relu(self.norm5(self.conv5(h4)))
h6 = F.leaky_relu(self.norm6(self.conv6(h5)))
dh = F.relu(F.dropout(self.denorm1(self.deconv1(h6))))
dh = F.relu(F.dropout(self.denorm2(self.deconv2(F.concat((dh, h5))))))
dh = F.relu(F.dropout(self.denorm3(self.deconv3(F.concat((dh, h4))))))
dh = F.relu(self.denorm4(self.deconv4(F.concat((dh, h3)))))
dh = F.relu(self.denorm5(self.deconv5(F.concat((dh, h2)))))
dh = F.sigmoid(self.deconv6(F.concat((dh, h1))))
return dh
def shake_camera(img):
s0,s1,s2,s3 = img.data.shape
zerobar = Variable(xp.zeros((s0,s1,4,s3),dtype=np.float32))
img = F.concat([zerobar, img, zerobar],axis=2)
randshift=np.random.randint(1,8)
img = F.split_axis(img, [randshift,randshift+img_w],axis=2)[1]
zerobar = Variable(xp.zeros((s0,s1,s2,4,1),dtype=np.float32))
img = F.reshape(img,(s0,s1,s2,s3,1))
img = F.concat([zerobar, img, zerobar],axis=3)
randshift=np.random.randint(1,8)
img = F.split_axis(img, [randshift,randshift+img_w],axis=3)[1]
img = F.reshape(img,(s0,s1,s2,s3))
return img
def forward(self, atom_x, pair_x, atom_only=False):
a0 = self.atom_to_atom.forward(atom_x)
a1 = self.pair_to_atom.forward(pair_x)
a = functions.concat([a0, a1], axis=2)
next_atom = self.atom_layer.forward(a)
next_atom = functions.relu(next_atom)
if atom_only:
return next_atom
p0 = self.atom_to_pair.forward(atom_x)
p1 = self.pair_to_pair.forward(pair_x)
p = functions.concat([p0, p1], axis=2)
next_pair = self.pair_layer.forward(p)
next_pair = functions.relu(next_pair)
return next_atom, next_pair
def __call__(self, atom_array, adj, super_node, is_real_node=None):
"""
Describe a layer
Args:
atom_array (numpy.ndarray): mol-minibatch by node numpy.ndarray,
minibatch of molecular which is represented with atom IDs (representing C, O, S, ...)
atom_array[m, i] = a represents
m-th molecule's i-th node is value a (atomic number)
adj (numpy.ndarray): mol-minibatch by relation-types by node by node numpy.ndarray,
minibatch of multiple relational adjancency matrix with edge-type information
adj[i, j] = b represents
m-th molecule's edge from node i to node j has value b
super_node (numpy.ndarray): 1D array, the supernode hidden state
is_real_node:
Returns:
numpy.ndarray: final molecule representation
"""
if atom_array.dtype == self.xp.int32:
h = self.embed(atom_array) # (minibatch, max_num_atoms)
else:
h = atom_array
# end if-else
h0 = functions.copy(h, cuda.get_device_from_array(h.data).id)
self.gwm.GRU_local.reset_state()
self.gwm.GRU_super.reset_state()
# ebmbed super node
h_s = self.embed_super(super_node)
g_list = []
for step in range(self.n_message_layers):
message_layer_index = 0 if self.weight_tying else step
h2 = self.update_layers[message_layer_index](h, adj)
h, h_s = self.gwm(h, h2, h_s, message_layer_index)
if self.concat_hidden:
g = self.readout_layers[step](h, h0, is_real_node)
g_list.append(g)
if self.concat_hidden:
return functions.concat(g_list, axis=1)
else:
g = self.readout_layers[0](h, h0, is_real_node)
g2 = functions.concat( (g, h_s), axis=1 )
out_g = functions.relu(self.linear_for_concat_super(g2))
return out_g
channel_input = get_normalized_image_variable(t, w)
if channel_input is None:
no_image = True
continue
channel_inputs.append(channel_input)
channel_observed = get_normalized_image_variable(t + dt, w)
if channel_observed is None:
no_image = True
continue
channel_observeds.append(channel_observed)
if no_image:
continue
img_input = F.concat(channel_inputs)
img_observed = F.concat(channel_observeds)
img_predicted = predictor(img_input)
loss = F.sum(abs(img_predicted - img_observed))
predictor.cleargrads()
loss.backward()
optimizer_p.update()
"""
Train the generator and discriminator
"""
t2 = t
no_missing_image = True
img_forecast = img_input
if epoch >= start_dcgan_at_epoch:
def __call__(self, x, encoded):
h = F.unpooling_2d(x, ksize=2, outsize=encoded.shape[2:])
h = F.concat([h, encoded], axis=1)
h = super(UpBlock, self).__call__(h)
return h
def norm_embed(bn, xs):
x_len = [len(x.data) for x in xs]
x_section = np.cumsum(x_len[:-1])
ex = bn(F.concat(xs, axis=0))
exs = F.split_axis(ex, x_section, 0, force_tuple=True)
return exs
def __call__(self, x1, x2):
return F.concat((x1, x2))
y1 = second_net(vec1)
vec2 = [np.array([ord(char) - 96], dtype=np.int32) for char in word2]
y2 = second_net2(vec2)
loss = third_net(y1,y2,glove_vec)
loss_record.append(float(loss.data))
loss.backward(retain_grad=True)
optimizer4.update()
optimizer3.update()
optimizer2.update()
f1_loss = F.concat([second_net.x.grad,second_net2.x.grad])
f1_loss = F.sum(F.absolute(f1_loss))
f1_record.append(float(f1_loss.data))
f1.grad = (f1 * f1_loss.data).data
#f1.grad = (f1 * loss.data).data
f1.unchain_backward()
f1.backward(retain_grad=True)
optimizer1.update()
plt.plot(loss_record)
plt.show()
plt.plot(f1_record)
plt.show()
print(original_word,'is :')
print(word1,'+',word2)
def __call__(self, x, x_mask):
h_dict = {}
mask_dict = {}
#print("Encode stage")
#print("[new step]: input -> PConv_00")
#print("input shape:",x.shape)
#print("mask shape:",x_mask.shape)
h_dict['PConv_00'], mask_dict['PConv_00'] = self.enc_layers[
'PConv_00'](x, x_mask)
key_prev = 'PConv_00'
#print("PConv_00 sum: ",self.xp.sum(h_dict['PConv_00'].data))
for i in range(1, self.layer_size):
key = 'PConv_0' + str(i)
#print("[new step]: ",key_prev," -> ",key)
#print("input shape:",h_dict[key_prev].shape)
#print("mask shape:",mask_dict[key_prev].shape)
h_dict[key], mask_dict[key] = self.enc_layers[key](
h_dict[key_prev], mask_dict[key_prev])
key_prev = key
#print(key," sum: ",self.xp.sum(h_dict[key].data))
#print("Decode stage")
#key_prev should be PConv06
for i in reversed(range(self.layer_size - 1)):
enc_in_key = 'PConv_0' + str(i)
dec_out_key = "PConv_1" + str(i + 1)
#print("[new step]:")
#print("h_dict['",enc_in_key,"'] ---l")
#print("h_dict['",key_prev,"'] --- h_dict['",dec_out_key,"']")
#print("input enc shape:",h_dict[enc_in_key].shape)
#unpooling (original paper used unsampling)
h = F.unpooling_2d(h_dict[key_prev], 2, 2, 0, cover_all=False)
mask = F.unpooling_2d(mask_dict[key_prev],
2,
2,
0,
cover_all=False)
#print("unpooled input dec shape:",h.shape)
#print("unpooled input mask shape:",mask.shape)
h = F.concat([h_dict[enc_in_key], h], axis=1)
mask = F.concat([mask_dict[enc_in_key], mask], axis=1)
h_dict[dec_out_key], mask_dict[dec_out_key] = self.dec_layers[
dec_out_key](h, mask)
key_prev = dec_out_key
#print(dec_out_key," sum: ",self.xp.sum(h_dict[dec_out_key].data))
#last step
dec_out_key = "PConv_10"
#print("[new step]:")
#print(" input ---l")
#print("h_dict['",key_prev,"'] --- h_dict['PConv_10']")
#print("input shape:",x.shape)
#unpooling (original paper used unsampling)
h = F.unpooling_2d(h_dict[key_prev], 2, 2, 0, cover_all=False)
mask = F.unpooling_2d(mask_dict[key_prev], 2, 2, 0, cover_all=False)
#print("unpooled input dec shape:",h.shape)
#print("unpooled input mask shape:",mask.shape)
h = F.concat([x, h], axis=1)
mask = F.concat([x_mask, mask], axis=1)
h_dict[dec_out_key], mask_dict[dec_out_key] = self.dec_layers[
dec_out_key](h, mask)
#print(dec_out_key," sum: ",self.xp.sum(h_dict[dec_out_key].data))
return h_dict[dec_out_key]
def train():
# model
gen = Generator()
dis = Discriminator()
gan = chainer.Sequential(gen, dis)
if GPU >= 0:
chainer.cuda.get_device(GPU).use()
gen.to_gpu()
dis.to_gpu()
gan.to_gpu()
opt_d = chainer.optimizers.Adam(0.0002, beta1=0.5)
opt_d.setup(dis)
opt_g = chainer.optimizers.Adam(0.0002, beta1=0.5)
opt_g.setup(gen)
xs, paths = data_load('../Dataset/train/images/', hf=True, vf=True, rot=1)
# training
mb = 64
mbi = 0
train_ind = np.arange(len(xs))
np.random.seed(0)
np.random.shuffle(train_ind)
for ite in range(5000):
if mbi + mb > len(xs):
mb_ind = train_ind[mbi:]
np.random.shuffle(train_ind)
mb_ind = np.hstack((mb_ind, train_ind[:(mb - (len(xs) - mbi))]))
mbi = mb - (len(xs) - mbi)
else:
mb_ind = train_ind[mbi:mbi + mb]
mbi += mb
gen.cleargrads()
dis.cleargrads()
gan.cleargrads()
x = xs[mb_ind]
input_noise = np.random.uniform(-1, 1, size=(mb, 100, 1,
1)).astype(np.float32)
dt = np.array([1] * mb + [0] * mb, dtype=np.int32).reshape([mb * 2, 1])
gt = np.array([1] * mb, dtype=np.int32).reshape([mb, 1])
if GPU >= 0:
x = chainer.cuda.to_gpu(x)
input_noise = chainer.cuda.to_gpu(input_noise)
dt = chainer.cuda.to_gpu(dt)
gt = chainer.cuda.to_gpu(gt)
g_output = gen(input_noise)
#if GPU >= 0:
# g_output = chainer.cuda.to_cpu(g_output)
X = F.concat((x, g_output), axis=0)
y = dis(X)
loss_d = F.sigmoid_cross_entropy(y, dt)
loss_d.backward()
opt_d.update()
y = gan(input_noise)
loss_g = F.sigmoid_cross_entropy(y, gt)
loss_g.backward()
opt_g.update()
loss_d = loss_d.data
loss_g = loss_g.data
if GPU >= 0:
loss_d = chainer.cuda.to_cpu(loss_d)
loss_g = chainer.cuda.to_cpu(loss_g)
if ite % 500 == 0:
print("iter >>", ite + 1, ',G:loss >>', loss_g.item(),
', D:loss >>', loss_d.item())
chainer.serializers.save_npz('cnn.npz', gen)
rnd = np.random.randint(list_len)
dir_path = image_path + image_list[rnd]
for index in range(4, 12):
inp = dir_path + "/" + str(0) + ".png"
inp = prepare_dataset(inp)
input_box.append(inp)
img = dir_path + "/" + str(index) + ".png"
img = prepare_dataset(img)
frame_box.append(img)
x = chainer.as_variable(xp.array(input_box).astype(xp.float32))
t = chainer.as_variable(xp.array(frame_box).astype(xp.float32))
embed = feature_extractor(t) - feature_extractor(x)
c = feature_embed(embed)
z = F.concat([x, c], axis=1)
y = predictor(z)
y_dis = discriminator_content(y)
t_dis = discriminator_content(t)
dis_loss = F.mean(F.softplus(-t_dis)) + F.mean(F.softplus(y_dis))
c_g = feature_extractor(y) - feature_extractor(make_diff(y))
c_dis = discriminator_sequence(embed)
c_g_dis = discriminator_sequence(c_g)
dis_loss += F.mean(F.softplus(-c_dis)) + F.mean(F.softplus(c_g_dis))
c_g.unchain_backward()
discriminator_content.cleargrads()
discriminator_sequence.cleargrads()
dis_loss.backward()
def compute_ctxt(previous_state):
ci1, attn1 = compute_ctxt1(previous_state)
intermediate_state = F.concat((previous_state, ci1), axis=1)
ci2, attn2 = compute_ctxt2(intermediate_state)
return ci2, attn2
def __call__(self, h, o):
left, right = F.split_axis(self.fch(h) + o, 2, axis=1)
g = F.concat((F.relu(left), right), axis=1)
#g = F.leaky_relu(self.fch(h) + o, slope=0.5)
q = self.fcg(g)
return q
def __call__(self, x_data, lengths=None, d=None, first_step=False):
batchsize = len(x_data)
h_shape = (self.n_layers, batchsize, self.hidden_dim)
hx = None
cx = None
x_data = self.xp.concatenate(x_data, axis=0)
xs = self.word_embed(x_data)
# dropout
xs = F.dropout(xs, ratio=self.use_dropout)
adv_flag = self.train and (self.use_adv or self.args.use_semi_data)
if adv_flag:
def norm_vec_sentence_level(d,
nn_flag=False,
include_norm_term=False):
dim = d.shape[1]
d_list = F.split_axis(d, np.cumsum(lengths)[:-1], axis=0)
max_length = np.max(lengths)
d_pad = F.pad_sequence(d_list, length=max_length, padding=0.0)
d_flat = F.reshape(get_normalized_vector(d_pad, None),
(-1, dim))
split_size = np.cumsum(np.full(batchsize, max_length))[:-1]
d_list = F.split_axis(d_flat, split_size, axis=0)
d_list = [_d[:_length] for _d, _length in zip(d_list, lengths)]
d = F.concat(d_list, axis=0)
return d
if first_step:
if self.args.use_semi_data:
# Vat
d = self.xp.random.normal(size=xs.shape, dtype='f')
else:
# Adv
d = self.xp.zeros(xs.shape, dtype='f')
# Normalize at word-level
d = get_normalized_vector(d, self.xp)
d_var = Variable(d.astype(self.xp.float32))
self.d_var = d_var
xs = xs + self.args.xi_var_first * d_var
elif d is not None:
d_original = d.data if isinstance(d, Variable) else d
if self.args.norm_sentence_level:
# Normalize at sentence-level
d_variable = norm_vec_sentence_level(
d, include_norm_term=True)
d = d_variable.data
else:
# Normalize at word-level
d = get_normalized_vector(d_original, self.xp)
xs_noise_final = self.xi_var * d
xs = xs + xs_noise_final
split_size = np.cumsum(lengths)[:-1]
xs_f = F.split_axis(xs, split_size, axis=0)
hy_f, cy_f, ys_list = self.uni_lstm(hx=hx, cx=cx, xs=xs_f)
hy = [_h[-1] for _h in ys_list]
hy = F.concat(hy, axis=0)
hy = F.reshape(hy, (batchsize, -1))
self.hy = hy
output = self.output_mlp(hy)
return output
def zero_pads(self, x, pad, where):
sizes = list(x.data.shape)
sizes[where] = pad
pad_mat = chainer.Variable(chainer.cuda.to_gpu(self.xp.zeros(sizes, dtype=np.float32), device=chainer.cuda.get_device_from_array(x.data)))
return F.concat((pad_mat, x), axis=where)
def forward_one(x, target,model, label, hidden, prev_c, train_flag):
# make input window vector
distance = window // 2
s_num = 3-1 + window // 2
char_vecs = list()
char_type_vecs = list()
words = list()
for wp in x:
w = wp.split('/')[0]
words.append(w)
c_index_r = 0
for i in range(target+1):
c_index_r += len(words[i])
c_index_l = c_index_r - len(words[target])
x = words
target_word = words[i]
c = list(''.join(words))
x = c[:c_index_l] + c[c_index_r:]
for i in range(s_num):
x.append('</s>')
x.insert(0,'<s>')
for i in range(-distance, distance+1):
# make char vector
# import char
uni_gram = x[target+s_num+i]
bi_gram = x[target+s_num-1+i] + x[target+s_num+i]
tri_gram = x[target+s_num-2+i] + x[target+s_num-1+i] + x[target+s_num+i]
# char2id
uni_gram_id = char2id[uni_gram]
bi_gram_id = char2id[bi_gram]
tri_gram_id = char2id[tri_gram]
# id 2 embedding
uni_gram_vec = model.embed(get_onehot(uni_gram_id))
bi_gram_vec = model.embed(get_onehot(bi_gram_id))
tri_gram_vec = model.embed(get_onehot(tri_gram_id))
# add all char_vec
char_vecs.append(uni_gram_vec)
char_vecs.append(bi_gram_vec)
char_vecs.append(tri_gram_vec)
# make char type vector
# import char type
uni_gram_type = make_char_type(uni_gram)
bi_gram_type = make_char_type(x[target+s_num-1+i]) + make_char_type(x[target+s_num+i])
tri_gram_type = make_char_type(x[target+s_num-2+i]) + make_char_type(x[target+s_num+i] + make_char_type(x[target+s_num-2+i]))
# chartype 2 id
uni_gram_type_id = char_type2id[uni_gram_type]
bi_gram_type_id = char_type2id[bi_gram_type]
tri_gram_type_id = char_type2id[tri_gram_type]
# id 2 embedding
uni_gram_type_vec = model.char_type_embed(get_onehot(uni_gram_type_id))
bi_gram_type_vec = model.char_type_embed(get_onehot(bi_gram_type_id))
tri_gram_type_vec = model.char_type_embed(get_onehot(tri_gram_type_id))
# add all char_type_vec
char_type_vecs.append(uni_gram_type_vec)
char_type_vecs.append(bi_gram_type_vec)
char_type_vecs.append(tri_gram_type_vec)
# word feature
target_word_id = char2id['word:'+target_word]
word_vec = model.embed(get_onehot(target_word_id))
ct_word = list()
for c in target_word:
ct_word.append(make_char_type(c))
try:
ct_index = char_type2id[''.join(list(set(ct_word)))]
except:
ct_index = char_type2id['other']
one_hot = list()
one_hot = [0]*len(char_type2id)
one_hot[ct_index] = 1
#for i in range(len(char_type2id)):
# if i == ct_index:
# one_hot.append(1)
# else:
# one_hot.append(0)
pos_d = list()
for i in range(23):
if target_word+'/'+id2pos[i] in word_pos:
pos_d.append(1)
else:
pos_d.append(0)
#pos_d = [0]*23
ct_word_vec = chainer.Variable(np.array([one_hot], dtype=np.float32))
pos_d_vec = chainer.Variable(np.array([pos_d], dtype=np.float32))
char_concat = F.concat(tuple(char_vecs))
char_type_concat = F.concat(tuple(char_type_vecs))
#dropout_concat = F.dropout(concat, ratio=dropout_rate, train=train_flag)
concat = F.concat((char_concat, char_type_concat, word_vec))
concat = F.concat((concat, hidden))
i_gate = F.sigmoid(model.i_gate(concat))
f_gate = F.sigmoid(model.f_gate(concat))
o_gate = F.sigmoid(model.o_gate(concat))
concat = F.concat((hidden, i_gate, f_gate, o_gate))
prev_c, hidden = F.lstm(prev_c, concat)
output = model.output(F.concat((hidden, ct_word_vec, pos_d_vec)))
dist = F.softmax(output)
#print(dist.data, label, np.argmax(dist.data))
#print(output.dataect.data)
return np.argmax(dist.data)
def __call__(self, hs, ys):
"""Core function of Decoder layer.
Args:
hs (list of chainer.Variable | N-dimension array): Input variable from encoder.
ys (list of chainer.Variable | N-dimension array): Input variable of decoder.
Returns:
chainer.Variable: A variable holding a scalar array of the training loss.
chainer.Variable: A variable holding a scalar array of the accuracy.
"""
self.loss = None
# prepare input and output word sequences with sos/eos IDs
eos = self.xp.array([self.eos], 'i')
sos = self.xp.array([self.sos], 'i')
ys_in = [F.concat([sos, y], axis=0) for y in ys]
ys_out = [F.concat([y, eos], axis=0) for y in ys]
# padding for ys with -1
# pys: utt x olen
pad_ys_in = F.pad_sequence(ys_in, padding=self.eos)
pad_ys_out = F.pad_sequence(ys_out, padding=-1)
# get dim, length info
batch = pad_ys_out.shape[0]
olength = pad_ys_out.shape[1]
logging.info(self.__class__.__name__ + ' input lengths: ' +
str(self.xp.array([h.shape[0] for h in hs])))
logging.info(self.__class__.__name__ + ' output lengths: ' +
str(self.xp.array([y.shape[0] for y in ys_out])))
# initialization
c_list = [None] # list of cell state of each layer
z_list = [None] # list of hidden state of each layer
for _ in six.moves.range(1, self.dlayers):
c_list.append(None)
z_list.append(None)
att_w = None
z_all = []
self.att.reset() # reset pre-computation of h
# pre-computation of embedding
eys = self.embed(pad_ys_in) # utt x olen x zdim
eys = F.separate(eys, axis=1)
# loop for an output sequence
for i in six.moves.range(olength):
att_c, att_w = self.att(hs, z_list[0], att_w)
if i > 0 and random.random() < self.sampling_probability:
logging.info(' scheduled sampling ')
z_out = self.output(z_all[-1])
z_out = F.argmax(F.log_softmax(z_out), axis=1)
z_out = self.embed(z_out)
ey = F.hstack((z_out, att_c)) # utt x (zdim + hdim)
else:
ey = F.hstack((eys[i], att_c)) # utt x (zdim + hdim)
z_list, c_list = self.rnn_forward(ey, z_list, c_list, z_list,
c_list)
z_all.append(z_list[-1])
z_all = F.reshape(F.stack(z_all, axis=1),
(batch * olength, self.dunits))
# compute loss
y_all = self.output(z_all)
self.loss = F.softmax_cross_entropy(y_all, F.flatten(pad_ys_out))
# -1: eos, which is removed in the loss computation
self.loss *= (np.mean([len(x) for x in ys_in]) - 1)
acc = F.accuracy(y_all, F.flatten(pad_ys_out), ignore_label=-1)
logging.info('att loss:' + str(self.loss.data))
# show predicted character sequence for debug
if self.verbose > 0 and self.char_list is not None:
y_hat = F.reshape(y_all, (batch, olength, -1))
y_true = pad_ys_out
for (i, y_hat_), y_true_ in zip(enumerate(y_hat.data),
y_true.data):
if i == MAX_DECODER_OUTPUT:
break
idx_hat = self.xp.argmax(y_hat_[y_true_ != -1], axis=1)
idx_true = y_true_[y_true_ != -1]
seq_hat = [self.char_list[int(idx)] for idx in idx_hat]
seq_true = [self.char_list[int(idx)] for idx in idx_true]
seq_hat = "".join(seq_hat).replace('<space>', ' ')
seq_true = "".join(seq_true).replace('<space>', ' ')
logging.info("groundtruth[%d]: " % i + seq_true)
logging.info("prediction [%d]: " % i + seq_hat)
if self.labeldist is not None:
if self.vlabeldist is None:
self.vlabeldist = chainer.Variable(
self.xp.asarray(self.labeldist))
loss_reg = -F.sum(
F.scale(F.log_softmax(y_all), self.vlabeldist,
axis=1)) / len(ys_in)
self.loss = (
1. - self.lsm_weight) * self.loss + self.lsm_weight * loss_reg
return self.loss, acc
def forward_one(x, target, label, word_dict, train_flag):
# make dict feature vector
dict_vec = list()
L1 = L2 = L3 = L4 = R1 = R2 = R3 = R4 = I1 = I2 = I3 = I4 = 0
for i in range(len(x[:target])):
word_candidate = x[target-(i+1):target]
if word_candidate in word_dict:
if len(word_candidate) == 1:
L1 = 1
elif len(word_candidate) == 2:
L2 = 1
elif len(word_candidate) == 3:
L3 = 1
else:
L4 = 1
if i == 10:
break
for i in range(len(x[target:])):
word_candidate = x[target:target+i+1]
if word_candidate in word_dict:
if len(word_candidate) == 1:
R1 = 1
elif len(word_candidate) == 2:
R2 = 1
elif len(word_candidate) == 3:
R3 = 1
else:
R4 = 1
if i == 10:
break
for i in range(1, 6, 1):
for j in range(1, 6, 1):
word_candidate = x[target-i:target+j]
if word_candidate in word_dict:
if len(word_candidate) == 1:
I1 = 1
elif len(word_candidate) == 2:
I2 = 1
elif len(word_candidate) == 3:
I3 = 1
else:
I4 = 1
dict_vec = chainer.Variable(np.array([[L1,L2,L3,L4,R1,R2,R3,R4,I1,I2,I3,I4]], dtype=np.float32))
# dict_embed_vec = model.dict_embed(dict_vec)
# make input window vector
distance = window // 2
s_num = 3-1 + window // 2
char_vecs = list()
char_type_vecs = list()
x = list(x)
for i in range(s_num):
x.append('</s>')
x.insert(0,'<s>')
for i in range(-distance, distance+1):
# make char vector
# import char
uni_gram = x[target+s_num+i]
bi_gram = x[target+s_num-1+i] + x[target+s_num+i]
tri_gram = x[target+s_num-2+i] + x[target+s_num-1+i] + x[target+s_num+i]
# char2id
uni_gram_id = char2id[uni_gram]
bi_gram_id = char2id[bi_gram]
tri_gram_id = char2id[tri_gram]
# id 2 embedding
uni_gram_vec = model.embed(get_onehot(uni_gram_id))
bi_gram_vec = model.embed(get_onehot(bi_gram_id))
tri_gram_vec = model.embed(get_onehot(tri_gram_id))
# add all char_vec
char_vecs.append(uni_gram_vec)
char_vecs.append(bi_gram_vec)
char_vecs.append(tri_gram_vec)
# make char type vector
# import char type
uni_gram_type = make_char_type(uni_gram)
bi_gram_type = make_char_type(x[target+s_num-1+i]) + make_char_type(x[target+s_num+i])
tri_gram_type = make_char_type(x[target+s_num-2+i]) + make_char_type(x[target+s_num+i] + make_char_type(x[target+s_num-2+i]))
# chartype 2 id
uni_gram_type_id = char_type2id[uni_gram_type]
bi_gram_type_id = char_type2id[bi_gram_type]
tri_gram_type_id = char_type2id[tri_gram_type]
# id 2 embedding
uni_gram_type_vec = model.char_type_embed(get_onehot(uni_gram_type_id))
bi_gram_type_vec = model.char_type_embed(get_onehot(bi_gram_type_id))
tri_gram_type_vec = model.char_type_embed(get_onehot(tri_gram_type_id))
# add all char_type_vec
char_type_vecs.append(uni_gram_type_vec)
char_type_vecs.append(bi_gram_type_vec)
char_type_vecs.append(tri_gram_type_vec)
char_concat = F.concat(tuple(char_vecs))
char_type_concat = F.concat(tuple(char_type_vecs))
#dropout_concat = F.dropout(concat, ratio=dropout_rate, train=train_flag)
concat = F.concat((char_concat, char_type_concat))
hidden = model.hidden1(concat)
hidden2 = F.concat((hidden, dict_vec))
output = model.output(hidden2)
dist = F.softmax(output)
#print(dist.data, label, np.argmax(dist.data))
correct = get_onehot(label)
#print(output.data, correct.data)
return np.argmax(dist.data), F.softmax_cross_entropy(output, correct)
def posneg(x):
n, c, h, w = x.shape
y = F.concat((x, -x), axis=2)
return y.reshape((n, c * 2, h, w))
def __call__(self, x, **kwargs):
out = []
for name in self.layer_names:
out.append(self[name](x, **kwargs))
out = F.concat(tuple(out), axis=self.axis)
return out
def __call__(self, x, conf, loc, conf_mask, loc_mask):
h = F.relu(self.conv1_1(x))
h = F.max_pooling_2d(F.relu(self.conv1_2(h)), 2, 2)
h = F.relu(self.conv2_1(h))
h = F.max_pooling_2d(F.relu(self.conv2_2(h)), 2, 2)
h = F.relu(self.conv3_1(h))
h = F.relu(self.conv3_2(h))
h = F.max_pooling_2d(F.relu(self.conv3_3(h)), 2, 2)
h = F.relu(self.conv4_1(h))
h = F.relu(self.conv4_2(h))
h = F.relu(self.conv4_3(h))
self.h_conv4_3 = h
h = F.max_pooling_2d(h, 2, 2)
h = F.relu(self.conv5_1(h))
h = F.relu(self.conv5_2(h))
h = F.max_pooling_2d(F.relu(self.conv5_3(h)), 3, stride=1, pad=1)
h = F.relu(self.fc6(h))
h = F.relu(self.fc7(h))
self.h_fc7 = h
h = F.relu(self.conv6_1(h))
h = F.relu(self.conv6_2(h))
self.h_conv6_2 = h
h = F.relu(self.conv7_1(h))
h = F.relu(self.conv7_2(h))
self.h_conv7_2 = h
h = F.relu(self.conv8_1(h))
h = F.relu(self.conv8_2(h))
self.h_conv8_2 = h
h = F.average_pooling_2d(h, 3)
self.h_pool6 = h
batchsize, ch, hh, ww = self.h_conv4_3.shape
kari = F.reshape(self.h_conv4_3, (batchsize * ch, hh * ww))
kari = F.transpose(kari, (1, 0))
kari = F.normalize(kari)
kari = F.transpose(kari, (1, 0))
kari = F.reshape(kari, (batchsize, ch, hh, ww))
self.h_conv4_3_norm = self.normalize(kari)
self.h_conv4_3_norm_mbox_loc = self.conv4_3_norm_mbox_loc(
self.h_conv4_3_norm)
self.h_conv4_3_norm_mbox_conf = self.conv4_3_norm_mbox_conf(
self.h_conv4_3_norm)
self.h_conv4_3_norm_mbox_loc_perm = F.transpose(
self.h_conv4_3_norm_mbox_loc, (0, 2, 3, 1))
self.h_conv4_3_norm_mbox_conf_perm = F.transpose(
self.h_conv4_3_norm_mbox_conf, (0, 2, 3, 1))
self.h_conv4_3_norm_mbox_loc_flat = F.reshape(
self.h_conv4_3_norm_mbox_loc_perm,
(batchsize, self.c4_h, self.c4_w, self.c4_d, 4))
self.h_conv4_3_norm_mbox_conf_flat = F.reshape(
self.h_conv4_3_norm_mbox_conf_perm,
(batchsize, self.c4_h, self.c4_w, self.c4_d, 21))
self.h_fc7_mbox_loc = self.fc7_mbox_loc(self.h_fc7)
self.h_fc7_mbox_conf = self.fc7_mbox_conf(self.h_fc7)
self.h_fc7_mbox_loc_perm = F.transpose(self.h_fc7_mbox_loc,
(0, 2, 3, 1))
self.h_fc7_mbox_conf_perm = F.transpose(self.h_fc7_mbox_conf,
(0, 2, 3, 1))
self.h_fc7_mbox_loc_flat = F.reshape(
self.h_fc7_mbox_loc_perm,
(batchsize, self.f7_h, self.f7_w, self.f7_d, 4))
self.h_fc7_mbox_conf_flat = F.reshape(
self.h_fc7_mbox_conf_perm,
(batchsize, self.f7_h, self.f7_w, self.f7_d, 21))
self.h_conv6_2_mbox_loc = self.conv6_2_mbox_loc(self.h_conv6_2)
self.h_conv6_2_mbox_conf = self.conv6_2_mbox_conf(self.h_conv6_2)
self.h_conv6_2_mbox_loc_perm = F.transpose(self.h_conv6_2_mbox_loc,
(0, 2, 3, 1))
self.h_conv6_2_mbox_conf_perm = F.transpose(self.h_conv6_2_mbox_conf,
(0, 2, 3, 1))
self.h_conv6_2_mbox_loc_flat = F.reshape(
self.h_conv6_2_mbox_loc_perm,
(batchsize, self.c6_h, self.c6_w, self.c6_d, 4))
self.h_conv6_2_mbox_conf_flat = F.reshape(
self.h_conv6_2_mbox_conf_perm,
(batchsize, self.c6_h, self.c6_w, self.c6_d, 21))
self.h_conv7_2_mbox_loc = self.conv7_2_mbox_loc(self.h_conv7_2)
self.h_conv7_2_mbox_conf = self.conv7_2_mbox_conf(self.h_conv7_2)
self.h_conv7_2_mbox_loc_perm = F.transpose(self.h_conv7_2_mbox_loc,
(0, 2, 3, 1))
self.h_conv7_2_mbox_conf_perm = F.transpose(self.h_conv7_2_mbox_conf,
(0, 2, 3, 1))
self.h_conv7_2_mbox_loc_flat = F.reshape(
self.h_conv7_2_mbox_loc_perm,
(batchsize, self.c7_h, self.c7_w, self.c7_d, 4))
self.h_conv7_2_mbox_conf_flat = F.reshape(
self.h_conv7_2_mbox_conf_perm,
(batchsize, self.c7_h, self.c7_w, self.c7_d, 21))
self.h_conv8_2_mbox_loc = self.conv8_2_mbox_loc(self.h_conv8_2)
self.h_conv8_2_mbox_conf = self.conv8_2_mbox_conf(self.h_conv8_2)
self.h_conv8_2_mbox_loc_perm = F.transpose(self.h_conv8_2_mbox_loc,
(0, 2, 3, 1))
self.h_conv8_2_mbox_conf_perm = F.transpose(self.h_conv8_2_mbox_conf,
(0, 2, 3, 1))
self.h_conv8_2_mbox_loc_flat = F.reshape(
self.h_conv8_2_mbox_loc_perm,
(batchsize, self.c8_h, self.c8_w, self.c8_d, 4))
self.h_conv8_2_mbox_conf_flat = F.reshape(
self.h_conv8_2_mbox_conf_perm,
(batchsize, self.c8_h, self.c8_w, self.c8_d, 21))
self.h_pool6_mbox_loc = self.pool6_mbox_loc(self.h_pool6)
self.h_pool6_mbox_conf = self.pool6_mbox_conf(self.h_pool6)
self.h_pool6_mbox_loc_perm = F.transpose(self.h_pool6_mbox_loc,
(0, 2, 3, 1))
self.h_pool6_mbox_conf_perm = F.transpose(self.h_pool6_mbox_conf,
(0, 2, 3, 1))
self.h_pool6_mbox_loc_flat = F.reshape(
self.h_pool6_mbox_loc_perm,
(batchsize, self.p6_h, self.p6_w, self.p6_d, 4))
self.h_pool6_mbox_conf_flat = F.reshape(
self.h_pool6_mbox_conf_perm,
(batchsize, self.p6_h, self.p6_w, self.p6_d, 21))
self.mbox_loc = F.concat([
F.reshape(self.h_conv4_3_norm_mbox_loc_flat, [batchsize, -1, 4]),
F.reshape(self.h_fc7_mbox_loc_flat, [batchsize, -1, 4]),
F.reshape(self.h_conv6_2_mbox_loc_flat, [batchsize, -1, 4]),
F.reshape(self.h_conv7_2_mbox_loc_flat, [batchsize, -1, 4]),
F.reshape(self.h_conv8_2_mbox_loc_flat, [batchsize, -1, 4]),
F.reshape(self.h_pool6_mbox_loc_flat, [batchsize, -1, 4]),
],
axis=1)
self.mbox_conf = F.concat([
F.reshape(self.h_conv4_3_norm_mbox_conf_flat, [batchsize, -1, 21]),
F.reshape(self.h_fc7_mbox_conf_flat, [batchsize, -1, 21]),
F.reshape(self.h_conv6_2_mbox_conf_flat, [batchsize, -1, 21]),
F.reshape(self.h_conv7_2_mbox_conf_flat, [batchsize, -1, 21]),
F.reshape(self.h_conv8_2_mbox_conf_flat, [batchsize, -1, 21]),
F.reshape(self.h_pool6_mbox_conf_flat, [batchsize, -1, 21]),
],
axis=1)
self.mbox_conf_reahpe = F.reshape(self.mbox_conf,
(7308 * batchsize, 21))
self.mbox_conf_softmax = F.softmax(self.mbox_conf_reahpe)
self.mbox_conf_softmax_reahpe = F.reshape(self.mbox_conf_softmax,
(batchsize, 7308, 21))
if self.train:
mbox_conf = cuda.to_cpu(self.mbox_conf_softmax_reahpe.data)
dammy_label = np.zeros([batchsize, 7308, 21])
for i in range(batchsize):
self.conf_num = int(conf_mask[i].sum())
self.mask = conf_mask[i].copy()
negative_sample_num = int(
conf_mask[i].sum() *
5) if int(conf_mask[i].sum() * 5) < 4000 else 4000
self.num = negative_sample_num
negative_index = mbox_conf[i, :,
0].argsort()[:negative_sample_num]
self.ind = negative_index
self.conf_mask = conf_mask[i]
conf_mask[i, negative_index] = 1
dammy_label[i][np.where(conf_mask[i, :, 0] == 0)][0] = 100
t_conf_mask = chainer.Variable(cuda.cupy.array(conf_mask),
volatile=x.volatile)
t_loc_mask = chainer.Variable(cuda.cupy.array(loc_mask),
volatile=x.volatile)
dammy_label = cuda.cupy.array(dammy_label)
self.t_conf_mask = t_conf_mask
train_conf = self.mbox_conf * t_conf_mask
train_conf.data += dammy_label
self.train_conf = F.reshape(train_conf, (-1, 21))
#train_conf = F.reshape(self.mbox_conf * t_conf_mask, (-1, 21))
#print(type(dammy_label), type(self.train_conf.data))
self.val_conf = F.flatten(conf)
self.loss = F.softmax_cross_entropy(self.train_conf, self.val_conf)
self.loss += F.mean_squared_error(self.mbox_loc * t_loc_mask, loc)
self.accuracy = F.accuracy(self.train_conf, self.val_conf)
return self.loss
else:
return self.mbox_loc, self.mbox_conf_softmax_reahpe
def __call__(self, x):
store_activations = {}
# down convolution
for i in range(1, self.n_layers + 1):
if i == 1:
h = F.identity(x)
else:
h = F.max_pooling_nd(h, 2, stride=2)
h = self['down_unet_block_%d' % (i)](h)
h = self.down_conv_dropout(h)
store_activations['down_unet_block_%d' % (i)] = h
del h # clear hidden layer
# up convolution
for i in range(self.n_layers - 1, 0, -1):
if i == self.n_layers - 1:
h = store_activations['down_unet_block_%d' % (i + 1)]
del store_activations['down_unet_block_%d' % (i + 1)] # clear
else:
h = h
h = self['deconv_%d' % i](h)
if self.batch_norm:
h = self['bn_deconv_%d' % i](h)
h = self.up_conv_activate_function(h)
down_conv = store_activations['down_unet_block_%d' % (i)]
del store_activations['down_unet_block_%d' % (i)] # clear
if self.n_dims == 2:
h = F.concat([
h[:, :, 0:down_conv.shape[2], 0:down_conv.shape[3]],
down_conv
]) # fuse layer
elif self.n_dims == 3:
h = F.concat([
h[:, :, 0:down_conv.shape[2], 0:down_conv.shape[3],
0:down_conv.shape[4]], down_conv
]) # fuse layer
del down_conv
h = self['up_unet_block_%d' % i](h)
h = self.up_conv_dropout(h)
if i == 1:
o = self['up_conv%d_3' % i](h)
if self.n_dims == 2:
score = o[:, :, 0:x.shape[2], 0:x.shape[3]]
elif self.n_dims == 3:
score = o[:, :, 0:x.shape[2], 0:x.shape[3], 0:x.shape[4]]
self.score = score
del h, o # clear hidden layer
return self.score
def test_invlaid_axis_type(self):
inputs = [numpy.random.rand(3, 4), numpy.random.rand(3, 1)]
with self.assertRaises(TypeError):
functions.concat(inputs, 'a')
def output_from_seq_batch(self, y_seq_batch):
y = F.concat(y_seq_batch, axis=0)
y = F.dropout(y, ratio=self.dropout)
return self.output(y)
def _forward(self, *inputs):
return functions.concat(inputs, self.axis)
def sequence_embed(embed, xs):
x_len = [len(x) for x in xs]
x_section = np.cumsum(x_len[:-1])
ex = embed(F.concat(xs, axis=0))
exs = F.split_axis(ex, x_section, 0, force_tuple=True)
return exs
def __call__(self, x):
heatmaps = []
h = F.relu(self.conv1_1(x))
h = F.relu(self.conv1_2(h))
h = F.max_pooling_2d(h, ksize=2, stride=2)
h = F.relu(self.conv2_1(h))
h = F.relu(self.conv2_2(h))
h = F.max_pooling_2d(h, ksize=2, stride=2)
h = F.relu(self.conv3_1(h))
h = F.relu(self.conv3_2(h))
h = F.relu(self.conv3_3(h))
h = F.relu(self.conv3_4(h))
h = F.max_pooling_2d(h, ksize=2, stride=2)
h = F.relu(self.conv4_1(h))
h = F.relu(self.conv4_2(h))
h = F.relu(self.conv4_3(h))
h = F.relu(self.conv4_4(h))
h = F.relu(self.conv5_1(h))
h = F.relu(self.conv5_2(h))
h = F.relu(self.conv5_3_CPM(h))
feature_map = h
# stage1
h = F.relu(self.conv6_1_CPM(h))
h = self.conv6_2_CPM(h)
heatmaps.append(h)
# stage2
h = F.concat((h, feature_map), axis=1) # channel concat
h = F.relu(self.Mconv1_stage2(h))
h = F.relu(self.Mconv2_stage2(h))
h = F.relu(self.Mconv3_stage2(h))
h = F.relu(self.Mconv4_stage2(h))
h = F.relu(self.Mconv5_stage2(h))
h = F.relu(self.Mconv6_stage2(h))
h = self.Mconv7_stage2(h)
heatmaps.append(h)
# stage3
h = F.concat((h, feature_map), axis=1) # channel concat
h = F.relu(self.Mconv1_stage3(h))
h = F.relu(self.Mconv2_stage3(h))
h = F.relu(self.Mconv3_stage3(h))
h = F.relu(self.Mconv4_stage3(h))
h = F.relu(self.Mconv5_stage3(h))
h = F.relu(self.Mconv6_stage3(h))
h = self.Mconv7_stage3(h)
heatmaps.append(h)
# stage4
h = F.concat((h, feature_map), axis=1) # channel concat
h = F.relu(self.Mconv1_stage4(h))
h = F.relu(self.Mconv2_stage4(h))
h = F.relu(self.Mconv3_stage4(h))
h = F.relu(self.Mconv4_stage4(h))
h = F.relu(self.Mconv5_stage4(h))
h = F.relu(self.Mconv6_stage4(h))
h = self.Mconv7_stage4(h)
heatmaps.append(h)
# stage5
h = F.concat((h, feature_map), axis=1) # channel concat
h = F.relu(self.Mconv1_stage5(h))
h = F.relu(self.Mconv2_stage5(h))
h = F.relu(self.Mconv3_stage5(h))
h = F.relu(self.Mconv4_stage5(h))
h = F.relu(self.Mconv5_stage5(h))
h = F.relu(self.Mconv6_stage5(h))
h = self.Mconv7_stage5(h)
heatmaps.append(h)
# stage6
h = F.concat((h, feature_map), axis=1) # channel concat
h = F.relu(self.Mconv1_stage6(h))
h = F.relu(self.Mconv2_stage6(h))
h = F.relu(self.Mconv3_stage6(h))
h = F.relu(self.Mconv4_stage6(h))
h = F.relu(self.Mconv5_stage6(h))
h = F.relu(self.Mconv6_stage6(h))
h = self.Mconv7_stage6(h)
heatmaps.append(h)
return heatmaps
def __call__(self, x):
h = F.concat((x[:, 0, :, :, :], x[:, 1, :, :, :]), axis=1)
h = F.relu(self.conv1(h))
h = F.relu(self.conv2(h))
h = F.relu(self.conv3(h))
return h
def run_disco_update():
# read data
DataN = get_data_N_rand(DataA,
N_pic=Gv.BatchSize,
imgH=64,
imgW=64,
keys=['x'])
x_a = Variable(xp.asarray(DataN['x']))
DataN = get_data_N_rand(DataB,
N_pic=Gv.BatchSize,
imgH=64,
imgW=64,
keys=['x'])
x_b = Variable(xp.asarray(DataN['x']))
# conversion
x_ax_1 = Md['g_ab_c1'](x_a)
x_ax_2 = Md['g_ab_c2'](x_a)
x_ab_1 = Md['g_ab_d'](x_ax_1)
x_ab_2 = Md['g_ab_d'](x_ax_2)
x_ab = F.concat((x_ab_1, x_ab_2), axis=1) #3ch,3ch⇒6ch
x_b_1, x_b_2 = F.split_axis(x_b, 2, axis=1) #6ch⇒3ch,3ch
x_ba_1 = Md['g_ba_d1'](Md['g_ba_c'](x_b_1))
x_ba_2 = Md['g_ba_d2'](Md['g_ba_c'](x_b_2))
x_ba = x_ba_1 + x_ba_2
# reconversion
x_aba_1 = Md['g_ba_d1'](Md['g_ba_c'](x_ab_1))
x_aba_2 = Md['g_ba_d2'](Md['g_ba_c'](x_ab_2))
x_aba = x_aba_1 + x_aba_2
x_bab_1 = Md['g_ab_d'](Md['g_ab_c1'](x_ba))
x_bab_2 = Md['g_ab_d'](Md['g_ab_c2'](x_ba))
x_bab = F.concat((x_bab_1, x_bab_2), axis=1) #3ch,3ch⇒6ch
# reconstruction loss
recon_loss_a = F.mean_squared_error(x_a, x_aba)
recon_loss_b = F.mean_squared_error(x_b, x_bab)
# discriminate
y_a_real, feats_a_real = Md['d_a'](x_a)
y_a_fake, feats_a_fake = Md['d_a'](x_ba)
y_b_real, feats_b_real = Md['d_b'](x_b)
y_b_fake, feats_b_fake = Md['d_b'](x_ab)
# GAN loss
gan_loss_dis_a, gan_loss_gen_a = compute_loss_gan(y_a_real, y_a_fake)
feat_loss_a = compute_loss_feat(feats_a_real, feats_a_fake)
gan_loss_dis_b, gan_loss_gen_b = compute_loss_gan(y_b_real, y_b_fake)
feat_loss_b = compute_loss_feat(feats_b_real, feats_b_fake)
# compute loss
total_loss_gen_a = (1. - Gv.Recon_rate) * (
0.1 * gan_loss_gen_b +
0.9 * feat_loss_b) + Gv.Recon_rate * recon_loss_a
total_loss_gen_b = (1. - Gv.Recon_rate) * (
0.1 * gan_loss_gen_a +
0.9 * feat_loss_a) + Gv.Recon_rate * recon_loss_b
gen_loss = total_loss_gen_a + total_loss_gen_b
dis_loss = gan_loss_dis_a + gan_loss_dis_b
if Iteration % 3 == 0:
Md['d_a'].cleargrads()
Md['d_b'].cleargrads()
dis_loss.backward()
Op['d_a'].update()
Op['d_b'].update()
else:
Md['g_ab_c1'].cleargrads()
Md['g_ab_c2'].cleargrads()
Md['g_ab_d'].cleargrads()
Md['g_ba_c'].cleargrads()
Md['g_ba_d1'].cleargrads()
Md['g_ba_d2'].cleargrads()
gen_loss.backward()
Op['g_ab_c1'].update()
Op['g_ab_c2'].update()
Op['g_ab_d'].update()
Op['g_ba_c'].update()
Op['g_ba_d1'].update()
Op['g_ba_d2'].update()
def encode(self, x, y):
h1 = F.tanh(self.le1(x))
h2 = F.tanh(self.embed_e(y))
mu = self.le2_mu(F.concat([h1, h2]))
ln_var = self.le2_ln_var(F.concat([h1, h2])) # log(sigma**2)
return mu, ln_var
def find_closest_latent_state(real_o, generator, transition, classifier, args):
trials = 400
target = OptimizableLatentState(s_shape=(trials, 7), z_shape=(trials, 4))
if not args.gpu < 0:
target.to_gpu()
_, channels, height, width = real_o.shape
real_o = real_o.reshape((channels, height, width))
real_o = F.broadcast_to(real_o, (trials, ) + real_o.shape)
print('real_o shape: ', real_o.shape)
optimizer = optimizers.Adam(alpha=1e-2)
optimizer.setup(target)
iterations = 1000
def compute_loss(real_o, o_current):
concat_image = F.concat((real_o, o_current), axis=1)
classification_loss = classifier(concat_image)
classification_loss = F.squeeze(classification_loss)
l2_loss = F.batch_l2_norm_squared(real_o - o_current)
assert classification_loss.shape == l2_loss.shape
loss = l2_loss - classification_loss
return loss
s_current, z = target()
for i in range(iterations):
optimizer.target.cleargrads()
s_next, _ = transition(s_current)
# print('s_current shape: ', s_current.shape, 's_next shape: ', s_next.shape)
x = F.concat((z, s_current, s_next), axis=1)
x = F.reshape(x, shape=x.shape + (1, 1))
o = generator(x)
o_current, _ = F.split_axis(o, 2, axis=1, force_tuple=True)
# print('o shape: ', o_current.shape)
# print('real_o shape: ', real_o.shape)
loss = compute_loss(real_o, o_current)
mean_loss = F.mean(loss)
mean_loss.backward()
optimizer.update()
mean_loss.unchain_backward()
if i % 100 == 0:
index = F.argmin(loss).data
print('loss at: ', i, ' min index: ', index, ' min loss: ',
loss[index])
# Select s and z with min loss
s_current, z = target()
s_next, _ = transition(s_current)
x = F.concat((z, s_current, s_next), axis=1)
x = F.reshape(x, shape=x.shape + (1, 1))
o = generator(x)
o_current, _ = F.split_axis(o, 2, axis=1, force_tuple=True)
loss = compute_loss(real_o, o_current)
index = F.argmin(loss).data
print('min index: ', index, ' min loss: ', loss[index])
s_min = s_current.data[index]
print('s min: ', s_min)
z_min = z.data[index]
print('z min: ', z_min)
return chainer.Variable(s_min), chainer.Variable(z_min)
def __call__(self, x, rel_y, neighbor_entities, neighbor_dict, assign,
entities, relations, RC, EC):
if self.layer == 0:
return self.easy_case(x, neighbor_entities, neighbor_dict, assign,
entities, relations)
if len(neighbor_dict) == 1:
x = [x]
else:
x = F.split_axis(x, len(neighbor_dict), axis=0)
assignR = dict()
bundle = defaultdict(list)
for v, k in enumerate(neighbor_entities):
for i in assign[v]:
e = entities[i]
if (e, k) in relations: r = relations[(e, k)] * 2
else: r = relations[(k, e)] * 2 + 1
assignR[(r, len(bundle[r]))] = v
bundle[r].append(x[v])
result = [0 for i in range(len(neighbor_dict))]
rel_y = F.split_axis(rel_y, len(RC), axis=0)
for r in bundle:
rx = bundle[r]
r_rep = rel_y[r // 2]
if len(rx) == 1:
rx = rx[0]
tmp = F.pad(F.concat((rx, r_rep), axis=0), ((0, 0), (0, 1)),
'constant')
tmp = F.reshape(tmp, (1, 1, 2, -1))
if r % 2 == 0: rx = getattr(self, self.forwardH[0][0])(tmp)
else: rx = getattr(self, self.forwardT[0][0])(tmp)
rx = F.reshape(rx, (1, -1))
result[assignR[(r, 0)]] = rx
else:
size = len(rx)
#tempRx = list()
for i in range(size):
#print ('shape is:', rx[i].shape)
rx[i] = F.pad(F.concat((rx[i], r_rep), axis=0),
((0, 0), (0, 1)), 'constant')
#print ('shape is:', rx[i].shape)
rx = F.concat(rx, axis=0)
#print ('shape is:', rx.shape)
tmp = F.reshape(rx, (size, 1, 2, -1))
#print ('shape is:', tmp.shape)
if r % 2 == 0: rx = getattr(self, self.forwardH[0][0])(tmp)
else: rx = getattr(self, self.forwardT[0][0])(tmp)
#print ('shape is:', rx.shape)
#print rx.data
rx = F.reshape(rx, (size, -1))
#print ('shape is:', rx.shape, size)
rx = F.split_axis(rx, size, axis=0)
for i, x in enumerate(rx):
result[assignR[(r, i)]] = x
#if self.pooling_method=='val':
# result = self.valpooling(result,assign,RC,EC)
if self.pooling_method == 'weight':
sources = defaultdict(list)
for ee in assign:
for i in assign[ee]:
sources[i].append((ee, result[ee]))
resultT = []
for i, xxs in sorted(sources.items(), key=lambda x: x[0]):
ii = entities[i]
if len(xxs) == 1:
resultT.append(xxs[0][1])
else:
tmplist = []
for txxs in xxs:
k = txxs[0]
ke = txxs[1]
'''
if (ii,k) in relations:
r = relations[(ii,k)]
else if (k,ii) in relations:
r = relations[(k,ii)]'''
tmplist.append(EC[k])
tmplist2 = []
if (sum(tmplist) != 0.0):
for t in range(len(xxs)):
tmplist2.append(xxs[t][1] * tmplist[t])
resultT.append(sum(tmplist2) / sum(tmplist))
else:
for t in range(len(xxs)):
tmplist2.append(xxs[t][1])
resultT.append(sum(tmplist2) / len(xxs))
result = resultT
if self.pooling_method == 'max':
result = self.maxpooling(result, assign)
if self.pooling_method == 'avg':
result = self.averagepooling(result, assign)
if self.pooling_method == 'sum':
result = self.sumpooling(result, assign)
result = F.concat(result, axis=0)
#result = F.concat(resultT,axis=0)
return result
def __call__(self, x, z=None, mask=None):
# pass in x and z shape = (batch, D, T), T is sequence length
xp = self.xp
h = self.h
if self.is_self_attention:
# expand dimension to 3 times first, then split, thus information of Q and V comes from x in self attenion.
# self.W_QKV output (batchsize, n_units * 3, sentence_length), then split by axis=1
Q, K, V = F.split_axis(
self.W_QKV(x)
) # node_feature shape = (F,T,D), z shape = (F',T,D) indicate neighbor features, F' may not equals with F
# F' may be edge_number in each frame, because it may come from EdgeRNN outputV(x), 3, axis=1)
else:
Q = self.W_Q(
x
) # Query is come from x, the Value we want to extract is from z
K, V = F.split_axis(self.W_KV(z), 2, axis=1)
batch, n_units, n_querys = Q.shape # n_querys is count of query which can also named as "T_q"
_, _, n_keys = K.shape # n_keys is sentence_length which can also named as "T_k"
# Calculate Attention Scores with Mask for Zero-padded Areas
# Perform Multi-head Attention using pseudo batching: h * batch_size
# all together at once for efficiency
# batch, n_units, n_querys : split by axis=1 to cut n_heads slice,
# each slice shape is (batch, n_units//8, n_querys)
# then concat in axis=0
batch_Q = F.concat(F.split_axis(Q, h, axis=1), axis=0)
batch_K = F.concat(F.split_axis(K, h, axis=1), axis=0)
batch_V = F.concat(F.split_axis(V, h, axis=1), axis=0)
assert (batch_Q.shape == (batch * h, n_units // h, n_querys))
assert (batch_K.shape == (batch * h, n_units // h, n_keys))
assert (batch_V.shape == (batch * h, n_units // h, n_keys))
# Notice that this formula is different from paper Eqn 1., this time is transpose of Q matrix,
# This F.matmul actually perform batch_matmul
batch_A = F.matmul(batch_Q, batch_K, transa=True) \
* self.scale_score # shape = batch * h, T_q , T_k，matrix mat along the dimension of n_units//8
# if mask==False，use -np.inf, so above code:[mask] * h
if mask is not None:
mask = xp.concatenate([mask] * h, axis=0)
batch_A = F.where(mask, batch_A,
xp.full(batch_A.shape, -np.inf, 'f'))
batch_A = F.softmax(
batch_A, axis=2
) # axis=2 means attend along n_keys axis. Thus you can softly choose keys
batch_A = F.where(xp.isnan(batch_A.data), xp.zeros(batch_A.shape, 'f'),
batch_A) # push nan value to zero
assert (batch_A.shape == (batch * h, n_querys, n_keys))
# Calculate Weighted Sum before broad_cast (batch_A stores weights)
# batch_A shape = (batch * h, 1, n_querys, n_keys); batch_V shape = (batch * h, n_units//8, 1, n_keys)
batch_A, batch_V = F.broadcast(batch_A[:, None],
batch_V[:, :, None]) # n_units//8 就是d_v
# batch_C shape = batch * h, n_units//8, n_querys, n_keys, axis=3 means weighted sum along sequence
batch_C = F.sum(batch_A * batch_V,
axis=3) # shape = batch * h, n_units//8, n_querys
assert (batch_C.shape == (batch * h, n_units // h, n_querys))
# slice in h piece，then concat along axis=1, shape = (batch, n_units//8 * 8, n_querys), head = 8
C = F.concat(F.split_axis(batch_C, h, axis=0), axis=1)
# Notice that there is no n_keys in shape any more, because weighed sum already eliminated this dimension.
assert (C.shape == (batch, n_units, n_querys))
C = self.finishing_linear_layer(C)
return C