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 __call__(self, x):
if self.h is None:
self.h = variable.Variable(
self.xp.zeros(self.state_size, dtype=x[0].data.dtype),
volatile='auto')
zz = 0.
for nth in range(0, len(self.in_channels)):
zz += getattr(self, 'x_z' + str(nth))(x[nth])
zz += self.h_z(self.h)
zz = F.sigmoid(zz)
rr = 0.
for nth in range(0, len(self.in_channels)):
rr += getattr(self, 'x_r' + str(nth))(x[nth])
rr += self.h_r(self.h)
rr = F.sigmoid(rr)
hh = 0.
for nth in range(0, len(self.in_channels)):
hh += getattr(self, 'x_chr' + str(nth))(x[nth])
hh += self.h_chr(rr*self.h)
hh = F.tanh(hh)
y = (1-zz)*self.h + zz*hh
self.h = y
return y
def forward_eye_states(self, x_batch_curr, y_batch_curr, volatile):
current_sample = Variable(x_batch_curr, volatile=volatile)
y_batch_curr = np.asarray(y_batch_curr).reshape(32, -1)
current_output = Variable(y_batch_curr, volatile=volatile)
h1_current = F.sigmoid(self.model_to_use.x_h1(current_sample))
h2_current = F.sigmoid(self.model_to_use.h1_h2(h1_current))
h3_current = F.sigmoid(self.model_to_use.h2_h3(h2_current))
h4_current = F.sigmoid(self.model_to_use.h3_h4(h3_current))
h4 = h4_current
y = self.model_to_use.h4_y(h4)
y.data = y.data.reshape(32, -1)
loss = F.sigmoid_cross_entropy(y, current_output)
current_output.data = np.squeeze(current_output.data)
accuracy = F.accuracy(y, current_output)
return accuracy, loss, y
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 forward_one_step(self, h, x, computeOutput=True):
h=F.sigmoid(self.model.x_to_h(x) + self.model.h_to_h(h))
if computeOutput:
y=F.sigmoid(self.model.h_to_y(h))
return h, y
else:
return h
def _propagate(self, Y, dropout=0.):
blstm = self.blstm_layer(Y, dropout=dropout)
relu_1 = F.clipped_relu(self.relu_1(blstm, dropout=dropout))
relu_2 = F.clipped_relu(self.relu_2(relu_1, dropout=dropout))
N_mask = F.sigmoid(self.noise_mask_estimate(relu_2))
X_mask = F.sigmoid(self.speech_mask_estimate(relu_2))
return N_mask, X_mask
def check_forward(self, x_data, use_cudnn=True):
x = chainer.Variable(x_data)
y = functions.sigmoid(x, use_cudnn=use_cudnn)
self.assertEqual(y.data.dtype, numpy.float32)
y_expect = functions.sigmoid(chainer.Variable(self.x))
gradient_check.assert_allclose(y_expect.data, y.data)
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 tag(self, x, z, test=True):
a, b, h = self.forward(x, z, test=test)
tag_a = F.sigmoid(self.out_a_tag(a))
tag_b = F.sigmoid(self.out_b_tag(b))
tag = F.sigmoid(self.out_tag(h))
tag = tag * 0.8 + tag_a * 0.1 + tag_b * 0.1
return tag
def forward(self, x_data):
x = Variable(_as_mat(x_data))
t = Variable(_as_mat(x_data))
x = F.dropout(x)
h = F.sigmoid(self.encoder(x))
y = F.sigmoid(self.decoder(h))
loss = F.mean_squared_error(y, t)
return loss
def forward(self, x_data, train=True):
x = Variable(x_data)
t = Variable(x_data)
if train:
x = F.dropout(x)
h = F.sigmoid(self.encoder(x))
y = F.sigmoid(self.decoder(h))
return F.mean_squared_error(y, t)
def forward(x_data, y_data): x = chainer.Variable(x_data) y = chainer.Variable(y_data) X1 = model.l1(x) out1 = F.sigmoid(X1) X2 = model.l2(out1) out2 = F.sigmoid(X2) return F.mean_squared_error(out2, y), out2
def decode(self, x, layer=None, train=False):
if not train or layer == 2:
x = F.sigmoid(self.model.dec2(x))
if not train or layer == 1:
x = F.sigmoid(self.model.dec1(x))
return x
def __call__(self, x):
h1 = F.sigmoid(F.average_pooling_2d(self.conv1(x), 2))
h2 = F.sigmoid(F.average_pooling_2d(self.conv2(h1),2))
h3 = self.conv3(h2)
h4 = F.tanh(self.l1(h3))
p = self.l2(h4)
return p
def forward(self,x_data):
x = Variable(x_data)
x = F.dropout(x)
y = F.sigmoid(self.encoder(x))
y_hat = F.sigmoid(self.decoder(y))
Loss = F.mean_squared_error(y_hat,x)
return Loss
def check_forward(self, x_data, use_cudnn=True):
x = chainer.Variable(x_data)
y = functions.sigmoid(x, use_cudnn=use_cudnn)
self.assertEqual(y.data.dtype, self.dtype)
y_expect = functions.sigmoid(chainer.Variable(self.x))
testing.assert_allclose(
y_expect.data, y.data, **self.check_forward_options)
def predict(self, x_data, train=False):
x = chainer.Variable(x_data)
h = F.sigmoid(self.encode1(x))
h = F.sigmoid(self.encode2(h))
h = F.sigmoid(self.decode1(h))
y = F.sigmoid(self.decode2(h))
return y.data
def __call__(self, x):
f1 = F.sigmoid(self.beta1)
f2 = F.sigmoid(self.beta2)
#self.m = f1 * self.m + (1 - f1) * x
#self.v = f2 * self.v + (1 - f2) * x**2
self.m = self.beta1 * self.m + (1 - self.beta1) * x
self.v = self.beta2 * self.v + (1 - self.beta2) * x**2
g = 1e-3 * self.m / F.sqrt(self.v + 1e-8)
return g
def check_forward(self, x_data, use_cudnn='always'):
x = chainer.Variable(x_data)
with chainer.using_config('use_cudnn', use_cudnn):
y = functions.sigmoid(x)
self.assertEqual(y.data.dtype, self.dtype)
y_expect = functions.sigmoid(chainer.Variable(self.x))
testing.assert_allclose(
y_expect.data, y.data, **self.check_forward_options)
def forward(x_data):
x = Variable(x_data)
h1 = F.sigmoid(model.l1(x))
h2 = F.sigmoid(model.l2(h1))
y = model.l3(h2)
y2 = softmax(model.l3(h2))
return y2
def forward(self, x_data, y_data, dropout, train=True):
x, t = chainer.Variable(x_data), chainer.Variable(y_data)
h1 = F.dropout(F.sigmoid(self.model.l1(x)), ratio=dropout, train=train)
h2 = F.dropout(F.sigmoid(self.model.l2(h1)), ratio=dropout, train=train)
## softmax and accuracy for discrimination, mse for regression
y = F.dropout(self.model.l3(h2), ratio=dropout, train=train)
return F.mean_squared_error(y,t), t.data, y.data, y_data
def forward(self, x_data, t_data, train=True):
x = chainer.Variable(x_data)
t = chainer.Variable(t_data)
h = F.sigmoid(self.encode1(x))
h = F.sigmoid(self.encode2(h))
h = F.sigmoid(self.decode1(h))
y = F.sigmoid(self.decode2(h))
return F.mean_squared_error(y, t)
def predict(self, x_data, y_data, train=False):
# print y_data
x = chainer.Variable(x_data)
# x, t = Variable(x_data), Variable(y_data)#mnist
h1 = F.sigmoid(self.fc1(x))
h2 = F.sigmoid(self.fc2(h1))
y = self.fc3(h2)
# 最後はソフトマックスを通すのか？
# print y.data, t.data
# print y.data.shape, t.data.shape
return y
def __call__(self, d_x_gen, d_x=None):
#TODO: reverse trick
bs_d_x_gen = d_x_gen.shape[0]
if d_x is not None:
bs_d_x = d_x.shape[0]
loss = F.sum(F.log(F.sigmoid(d_x))) / bs_d_x \
+ F.sum(F.log(1 - F.sigmoid(d_x_gen))) / bs_d_x_gen
return - loss # to minimize
else:
loss = F.sum(F.log(1 - F.sigmoid(d_x_gen))) / bs_d_x_gen
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 _extract(self, inputs, layername):
if layername == 'prob':
h = self._forward(inputs, layername='conv6_4')
h = average_pooling_2d(h, ksize=7)
y = sigmoid(h)
return y.data
elif layername == 'encode1neuron':
h = self._forward(inputs, layername='encode1')
y = sigmoid(h)
return y.data
else:
y = self._forward(inputs, layername)
return y.data
def forward_one_step(model, x_data, u_io, u_fh, u_sh, tau_io, tau_fh, tau_sh, train=True):
#original MTRNN have only sigmoid activation functions
x = chainer.Variable(x_data)
fh = F.sigmoid(u_fh)
#fh = F.tanh(u_fh2)
sh = F.sigmoid(u_sh)
#sh = F.tanh(u_sh2)
y = F.sigmoid(u_io)
u_io2 = (1-1/tau_io)*u_io+(model.fh_to_y(fh))/tau_io
u_fh2 = (1-1/tau_fh)*u_fh+(model.x_to_fh(x)+model.fh_to_fh(fh)+model.sh_to_fh(sh))/tau_fh
u_sh2 = (1-1/tau_sh)*u_sh+(model.fh_to_sh(fh)+model.sh_to_sh(sh))/tau_sh
return u_io2, u_fh2, u_sh2, y
def _get_region_boxes(self, output, img_W, img_H):
conf_thresh = 0.1
B, C, H, W = output.shape
assert C == 19 + self.n_class
det_confs = F.sigmoid(output[:, 18]).data
cls_conf = F.softmax(output[:, 19:19 + self.n_class]).data
rpoints = output[:, :18].reshape(B, 9, 2, H, W)
rpoints0 = F.sigmoid(rpoints[:, 0])
rpoints = F.concat(
(rpoints0[:, None], rpoints[:, 1:]), axis=1)
points_img = rpoints_to_points(rpoints.data)
cls_max_ids = self.xp.argmax(cls_conf, axis=1)
# cls_max_confs = self.xp.max(cls_conf, axis=1)
points = []
labels = []
scores = []
for b in range(B):
point = []
label = []
score = []
for cy in range(H):
for cx in range(W):
# only_objectness == True
det_conf = det_confs[b, cy, cx]
conf = det_conf
if conf > conf_thresh:
# cls_max_conf = cls_max_confs[b, cy, cx]
cls_max_id = cls_max_ids[b, cy, cx]
pnt = self.xp.zeros((9, 2), dtype=np.float32)
pnt[:, 0] = points_img[b, :, 0, cy, cx] * img_W
pnt[:, 1] = points_img[b, :, 1, cy, cx] * img_H
# TODO: logic when only_objectness == False
point.append(pnt)
label.append(cls_max_id)
score.append(det_conf)
if len(point) == 0:
point = np.zeros((0, 9, 2), dtype=np.float32)
points.append(self.xp.array(point, dtype=np.float32))
labels.append(self.xp.array(label, dtype=np.int32))
scores.append(self.xp.array(score, dtype=np.float32))
return points, labels, scores
def forward(x_data):
x = Variable(x_data)
'''
h = F.max_pooling_2d(F.relu(model.l1(x)), ksize=5,stribe=2,pad=2)
h = F.max_pooling_2d(F.relu(model.l2(h)))
'''
h1 = F.sigmoid(model.l1(x))
h2 = F.sigmoid(model.l2(h1))
y = model.l3(h2)
y2 = softmax(model.l3(h2))
return y2
def forward(x_data, y_data, train=True):
# print("hoge", y_data)
# x, t = chainer.Variable(x_data), chainer.Variable(y_data)
x = chainer.Variable(x_data.reshape(batchsize, 400).astype(numpy.float32), volatile=False)
t = chainer.Variable(y_data.astype(numpy.int32), volatile=False)
# h1 = F.dropout(F.relu(model.l1(x)), train=train)
# y = F.dropout(F.relu(model.l2(h1)), train=train)
# y = model.l2(h1)
h1 = F.dropout(F.sigmoid(model.l1(x)), train=train)
h2 = F.dropout(F.sigmoid(model.l2(h1)), train=train)
y = F.dropout(F.sigmoid(model.l3(h2)), train=train)
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def decode(self, x):
h = F.relu(self.l5(x))
h = F.reshape(h, (-1, 16, 17, 17))
h = F.relu(self.deconv6(h))
h = F.relu(self.deconv7(h))
return F.sigmoid(self.deconv8(h))
def predict(self, x):
h1 = F.sigmoid(self.l1(x))
h2 = F.sigmoid(self.l2(h1))
h3 = self.l3(h2)
return h3
def _predict_heads_attn(self,
sent_states,
subword_embeds,
mask,
sub_lengths,
batch_stats,
extract_attn=False,
sorted_heads=None):
"""For each token in the sentence predict which token in the sentence
is its head."""
# sub_lengths is the mask for morph embeddings -- refer to the subword lengths
# start and end word are masked
batch_size, max_sent_len, col_lengths = batch_stats
calc_loss = sorted_heads is not None
# In order to predict which head is most probable for a given word
# For each token in the sentence we get a vector represention
# for that token as a head, and another for that token as a dependent.
# The idea here is that we only need to calculate these once
# and then reuse them to get all combinations
# ------------------------------------------------------
# In g(a_j, a_i) we note that we can precompute the matrix multiplications
# for each word, we consider all possible heads
# we can pre-calculate U * a_j , for all a_j
# head activations for each token lstm activation
# bs * max_sent x mlp_arc_units
h_arc = self.H_arc(sent_states)
# we transform results to be indexable by sentence index for upcoming for loop
# h_arc is now max_sent x bs x mlp_arc_units
h_arc = F.reshape(h_arc, (-1, batch_size, self.mlp_arc_units))
# bs * max_sent x mlp_arc_units
d_arc = self.D_arc(sent_states)
# max_sent x bs x mlp_arc_units
d_arc = F.reshape(d_arc, (-1, batch_size, self.mlp_arc_units))
# the values to use to mask softmax for head prediction
# e ^ -100 is ~ zero (can be changed from self.MIN_PAD)
mask_vals = Variable(
self.xp.full((batch_size, max_sent_len),
self.MIN_PAD,
dtype=self.xp.float32))
# reshape sub_lengths for attention computation
# max_sent_len x bs x self.unit_mult*self.encoder.num_units
sub_lengths = F.swapaxes(sub_lengths, axis1=0, axis2=1)
# subword_embeds shape is bs x max_sen x max_sub_len x units_dim
# reshape subword embeds to compute attention
# max_sent x bs x units_dim*max_sub_len
subword_embeds = F.reshape(
subword_embeds,
(-1, batch_size, self.units_dim * self.max_sub_len))
sent_arcs = []
sent_attn_vectors = []
sent_h_heads = []
# we start from 1 because we don't consider root
for i in range(1, max_sent_len):
num_active = col_lengths[i]
# if we are calculating loss create truth variables
if calc_loss:
# i-1 because sentence has root appended to beginning
gold_heads = sorted_heads[i - 1]
# ================== HEAD PREDICTION ======================
# NOTE Because some sentences may be shorter - only num_active of
# the batch have valid activations for this token. If in softmax
# we didn't limit arcs to [:num_active] we would need to replace
# embeddings that are out of sentence range with zeros - because
# otherwise when broadcasting and summing we will modify valid
# batch activations for earlier tokens of the sentence.
# ====================== Code for padding ==========================
# invalid_pad = ((0, int(batch_size - num_active)), (0, 0))
# d_arc_pad = F.pad(d_arc[i][:num_active],
# invalid_pad, 'constant', constant_values=0.)
# ==================================================================
# h_i is the current word and h_j is the candidate head
h_i = d_arc[i]
# we compute the attention for every possible head
h_heads = []
attn_vectors = []
# now, we start from 0 because we consider ROOT as head
for j in range(max_sent_len):
# f_j is the morph features of h_j
f_j = subword_embeds[j]
h_j = h_arc[j]
# compute the attention vector
k = self.V_attn(f_j, h_i)
# or.. (another option)
# k = self.V_attn(f_j, F.concat((h_i, h_j), axis=1))
# attention mask, shape is the same as k
attn_mask = Variable(
self.xp.full(k.shape, self.MIN_PAD, dtype=self.xp.float32))
# NOTE that we also put mask to the start and end symbol of the subword unit sequence
cond = sub_lengths[j]
k = F.where(cond, k, attn_mask)
k = F.softmax(k, axis=1)
attn_vectors.append(k)
# reshape k and f_j to compute m_j
k = F.reshape(k, (-1, 1))
f_j = F.reshape(f_j, (-1, self.units_dim))
# compute m_j
m_j = f_j * k.data
m_j = F.reshape(m_j,
(batch_size, self.max_sub_len, self.units_dim))
m_j = F.sum(m_j, axis=1)
# compute gating function
g = F.sigmoid(self.W_glob(h_j) + self.W_loc(h_i))
z_j = g * h_j + (1 - g) * m_j
h_heads.append(z_j)
# sent_attn_vectors store the attention vectors for each dependent word in the batch
# max_sent_len - 1 (because we don't consider root) x max_sent_len x bs x num of morph features
sent_attn_vectors.append(attn_vectors)
h_i = self.W_dependent(h_i)
h_heads = self.W_head(
F.reshape(F.stack(h_heads, axis=0), (-1, self.mlp_arc_units)))
h_heads = F.reshape(h_heads, (-1, batch_size, self.mlp_arc_units))
sent_h_heads.append(h_heads)
a_u, a_w = F.broadcast(h_heads, h_i)
arc_logit = F.reshape(F.tanh(a_u + a_w), (-1, self.mlp_arc_units))
if self.arc_dropout > 0.:
arc_logit = F.dropout(arc_logit, ratio=self.arc_dropout)
arc_logit = self.vT(arc_logit)
arcs = F.swapaxes(F.reshape(arc_logit, (-1, batch_size)), 0, 1)
arcs = F.where(mask, arcs, mask_vals)
# Calculate losses
if calc_loss:
# we don't want to average out over seen words yet
# NOTE: do not use ignore_label - in gpu mode gold_heads gets mutated
# and furthermore we would need to have padded invalid state of
# d_arc[i] with zeros before broadcasting.
# see NOTE above
head_loss = F.sum(
F.softmax_cross_entropy(arcs[:num_active],
gold_heads[:num_active],
reduce='no'))
self.loss += head_loss
sent_arcs.append(F.reshape(arcs, (batch_size, -1, 1)))
arcs = F.concat(sent_arcs, axis=2)
# sent_h_heads store the morphological representations
# for each possible heads, for each dependent word
# max_sent_len - 1 (num of dep) x max_sent_len (num of possible heads) x bs x mlp_arc_units
# in other words, each word has different views/representation of its possible heads
sent_h_heads = F.stack(sent_h_heads, axis=0)
return arcs, sent_h_heads, sent_attn_vectors
def __call__(self, x):
return F.sigmoid(x)
def __call__(self, x):
x = add_zero_pad(x, self.kernel // 2, 3)
A = F.tanh(self.W(x))
B = F.sigmoid(self.V(x))
return A * B
def fwd(self,x):
h1 = F.sigmoid(self.l1(x))
h2 = self.l2(h1)
#h3 = F.softmax(h2) :順伝播の最後にsoftmax関数必要なし
return h2
def mlp_forward(self, x):
out1 = self.l1(x)
out2 = F.sigmoid(out1)
return F.sigmoid(self.l2(out2))
def __call__(self, x_target):
# x_target: chainer.Variable of shape = [N, 3, H, W]
# There might be dimension mismatch due to uneven down/up-sampling
H, W = x_target.shape[2:]
normalizer = lambda z: z
h = x_target
h = self.activation(normalizer(self.c1(h)))
h = self.activation(normalizer(self.c1b(h)))
h_c1b = h
h = self.activation(normalizer(self.c2(h)))
h = self.activation(normalizer(self.c2b(h)))
h_c2b = h
h = self.activation(normalizer(self.c3(h)))
h = self.activation(normalizer(self.c3b(h)))
h_c3b = h
h = self.activation(normalizer(self.c4(h)))
h = self.activation(normalizer(self.c4b(h)))
h_c4b = h
h = self.activation(normalizer(self.c5(h)))
h = self.activation(normalizer(self.c5b(h)))
h_c5b = h
h = self.activation(normalizer(self.c6(h)))
h = self.activation(normalizer(self.c6b(h)))
h_c6b = h
h = self.activation(normalizer(self.c7(h)))
h = self.activation(normalizer(self.c7b(h)))
h = self.activation(normalizer(self.dc7(h)))
# There might be dimension mismatch due to uneven down/up-sampling
# Resize by bilinear interpolation.
# (by nearest neighbor sampling in the original implemntation.)
h = resize_like(h, h_c6b)
h = F.concat([h, h_c6b], axis=1)
h = self.activation(normalizer(self.idc7(h)))
h = self.activation(normalizer(self.dc6(h)))
h = resize_like(h, h_c5b)
h = F.concat([h, h_c5b], axis=1)
h = self.activation(normalizer(self.idc6(h)))
h = self.activation(normalizer(self.dc5(h)))
h = resize_like(h, h_c4b)
h = F.concat([h, h_c4b], axis=1)
h = self.activation(normalizer(self.idc5(h)))
h = self.activation(normalizer(self.dc4(h)))
h = F.concat([h, h_c3b], axis=1)
h = self.activation(normalizer(self.idc4(h)))
disp4 = DISP_SCALING * F.sigmoid(self.dispout4(h)) + MIN_DISP
disp4_up = F.resize_images(disp4, (H // 4, W // 4))
h = self.activation(normalizer(self.dc3(h)))
h = F.concat([h, h_c2b, disp4_up], axis=1)
h = self.activation(normalizer(self.idc3(h)))
disp3 = DISP_SCALING * F.sigmoid(self.dispout3(h)) + MIN_DISP
disp3_up = F.resize_images(disp3, (H // 2, W // 2))
h = self.activation(normalizer(self.dc2(h)))
h = F.concat([h, h_c1b, disp3_up], axis=1)
h = self.activation(normalizer(self.idc2(h)))
disp2 = DISP_SCALING * F.sigmoid(self.dispout2(h)) + MIN_DISP
disp2_up = F.resize_images(disp2, (H, W))
h = self.activation(normalizer(self.dc1(h)))
h = F.concat([h, disp2_up], axis=1)
h = self.activation(normalizer(self.idc1(h)))
disp1 = DISP_SCALING * F.sigmoid(self.dispout1(h)) + MIN_DISP
return [disp1, disp2, disp3, disp4]
def forward(self, x):
h = F.sigmoid(self.l1(x))
# import ipdb; ipdb.set_trace()
h = self.l2(h)
return h
def fwd(self,x):
# 活性化関数としてシグモイド関数を利用し出力ベクトルを得る
h1 = F.sigmoid(self.l1(x))
# TODO 活性化関数はいらない？
h2 = self.l2(h1)
return h2
def forward(self):
x = chainer.Variable(self.x)
return functions.sigmoid(x)
def __call__(self, x, y):
xy = x * y
a = self.a3(self.a0(x) + self.a1(y) + self.a2(xy))
b = self.b3(self.b0(x) + self.b1(y) + self.b2(xy))
return b + self.a4(F.sigmoid(a))
def __call__(self, x):
a, b = F.split_axis(x, 2, axis=1)
h = a * F.sigmoid(b)
return h
def __call__(self, fbs, ns):
"""
Attentionの計算
:param fbs: 順向き逆向きのEncoderの中間ベクトルが記録されたリスト
(4, 1000)
(3, 1000)
(2, 1000)
:param h: Decoderで出力された中間ベクトル
:return: 順向きのEncoderの中間ベクトルの加重平均と逆向きのEncoderの中間ベクトルの加重平均
"""
# ミニバッチのサイズを記憶
#batch_size = ns.data.shape[0]
# ウェイトを記録するためのリストの初期化
ws = []
# 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
#start_time_x = time.time()
for fb, n in zip(fbs, ns):
# 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
# start_time = time.time()
# print(fb.data.shape[0])
# print(n.shape)
# n_s = self.ARR.array(
# [n for _ in range(fb.data.shape[0])], dtype="float32")
# print(n_s)
n_s = self.ARR.tile(n, (fb.data.shape[0], 1))
# print(n_s)
# print(n_s.shape)
# n_s = self.ARR.empty((0, self.fnn_size), dtype=np.float32)
# for _ in range(fb.data.shape[0]):
# n_s = self.ARR.concatenate(
# [n_s, F.reshape(copy.deepcopy(n), (-1, self.fnn_size))], axis=0)
# print(n_s.shape)
# exit()
# interval = float(time.time() - start_time)
# print("n_s実行時間: {}sec".format(interval))
# start_time = time.time()
w = F.tanh(
F.dropout(self.fbh(fb), ratio=self.USE_DROPOUT) +
F.dropout(self.nh(n_s), ratio=self.USE_DROPOUT))
# w = F.tanh(F.dropout(self.fbh(fb), ratio=self.USE_DROPOUT) +
# F.dropout(self.nh(self.ARR.full(fb.data.shape[0], n)),
# ratio=self.USE_DROPOUT))
# interval = float(time.time() - start_time)
# print("tanh実行時間: {}sec".format(interval))
# start_time = time.time()
# softmax関数を使って正規化する
w = F.exp(F.dropout(self.hw(w), ratio=self.USE_DROPOUT))
# interval = float(time.time() - start_time)
# print("exp実行時間: {}sec".format(interval))
# 計算したウェイトを記録
# print(self.ARR.sum(w))
# sum_w += w
ws.append(w / F.sum(w).data)
#interval = float(time.time() - start_time_x)
#print("tanh_exp実行時間: {}sec".format(interval))
# 出力する加重平均ベクトルの初期化
#att_fb = []
#start_time_x = time.time()
att_fb = self.ARR.empty((0, self.hidden_size), dtype=np.float32)
if self.flag_local == 0:
for fb, w in zip(fbs, 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
# ここにローカルアテンション用の何かを入れる
att_fb += F.reshape(F.matmul(fb, w), (-1, self.hidden_size))
else:
D = self.local_window
for fb, w, n in zip(fbs, ws, ns):
# ウェイトの和が1になるように正規化
# w /= sum_w
#start_time = time.time()
w_local_input = fb.data.shape[0] * F.sigmoid(
F.dropout(self.tw(
F.tanh(
F.dropout(self.nt(F.reshape(n, (1, -1))),
ratio=self.USE_DROPOUT))),
ratio=self.USE_DROPOUT))
# ここにローカルアテンション用の何かを入れる
#interval = float(time.time() - start_time)
#print("local_input実行時間: {}sec".format(interval))
#start_time = time.time()
w_local_output = self.ARR.array([
self.ARR.exp(
-(float(s + 1) - float(w_local_input.data))**2 /
(((D / 2)**2) * 2)) for s in range(fb.data.shape[0])
],
dtype='float32')
# 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
#interval = float(time.time() - start_time)
#print("local_out_put実行時間: {}sec".format(interval))
#start_time = time.time()
w = w * F.reshape(w_local_output, (-1, 1))
#interval = float(time.time() - start_time)
#print("reshape実行時間: {}sec".format(interval))
#start_time = time.time()
# att_fb.append(self.ARR.sum(self.ARR.array(
# [fb_x.data * w_x.data for fb_x, w_x in zip(fb, w)], dtype='float32'), axis=0))
# fb_w = F.reshape(self.ARR.sum(self.ARR.array(
# [fb_x.data * w_x.data for fb_x, w_x in zip(fb, w)], dtype='float32'), axis=0), (1, -1))
# print(fb)
# print(w)
fb_w = F.reshape(F.matmul(F.transpose(w), fb), (1, -1))
# print(fb_w)
#interval = float(time.time() - start_time)
#print("kakezan実行時間: {}sec".format(interval))
#start_time = time.time()
# if self.FLAG_GPU:
# print(type(fb_w))
# fb_w = cuda.to_gpu(fb_w, device=0)
att_fb = F.concat((att_fb, fb_w), axis=0)
#print("concatenate実行時間: {}sec".format(interval))
#start_time = time.time()
#interval = float(time.time() - start_time_x)
#print("local実行時間: {}sec".format(interval))
return att_fb
def fwd(self, x):
h1 = F.sigmoid(self.l1(x))
h2 = self.l2(h1)
return h2
def fwd(self, x):
#接続
mid = F.sigmoid(self.l1(x))
result = self.l2(mid)
return result
def activation(self, x):
#return F.leaky_relu(x)
return x * F.sigmoid(x) # Swish activation function
def __call__(self, x):
image, steps = x
h = self.image2hidden(image) * F.sigmoid(self.embed(steps))
return self.hidden2out(h)
def __call__(self, encs, hiddens, batch_size, prev_image, num_masks,
color_channels):
"""
Learn through StatelessCDNA.
Args:
encs: An array of computed transformation
hiddens: An array of hidden layers
batch_size: Size of mini batches
prev_image: The image to transform
num_masks: Number of masks to apply
color_channels: Output color channels
Returns:
transformed: A list of masks to apply on the previous image
"""
logger = logging.getLogger(__name__)
enc0, enc1, enc2, enc3, enc4, enc5, enc6 = encs
hidden1, hidden2, hidden3, hidden4, hidden5, hidden6, hidden7 = hiddens
img_height = prev_image.shape[2]
img_width = prev_image.shape[3]
# CDNA specific
enc7 = self.enc7(enc6)
enc7 = F.relu(enc7)
transformed_list = list([F.sigmoid(enc7)])
# CDNA specific
# Predict kernels using linear function of last layer
cdna_input = F.reshape(hidden5, (int(batch_size), -1))
cdna_kerns = self.cdna_kerns(cdna_input)
# Reshape and normalize
# B x C x H x W => B x NUM_MASKS x 1 x H x W
cdna_kerns = F.reshape(
cdna_kerns,
(int(batch_size), self.num_masks, 1, DNA_KERN_SIZE, DNA_KERN_SIZE))
cdna_kerns = F.relu(cdna_kerns - RELU_SHIFT) + RELU_SHIFT
norm_factor = F.sum(cdna_kerns, (2, 3, 4), keepdims=True)
cdna_kerns = broadcasted_division(cdna_kerns, norm_factor)
# Treat the color channel dimension as the batch dimension since the same
# transformation is applied to each color channel.
# Treat the batch dimension as the channel dimension so that
# F.depthwise_convolution_2d can apply a different transformation to each sample.
cdna_kerns = F.reshape(
cdna_kerns,
(int(batch_size), self.num_masks, DNA_KERN_SIZE, DNA_KERN_SIZE))
cdna_kerns = F.transpose(cdna_kerns, (1, 0, 2, 3))
# Swap the batch and channel dimension.
prev_image = F.transpose(prev_image, (1, 0, 2, 3))
# Transform the image.
transformed = F.depthwise_convolution_2d(prev_image,
cdna_kerns,
stride=(1, 1),
pad=DNA_KERN_SIZE / 2)
# Transpose the dimensions where they belong.
transformed = F.reshape(transformed,
(color_channels, int(batch_size),
self.num_masks, img_height, img_width))
transformed = F.transpose(transformed, (2, 1, 0, 3, 4))
transformed = F.split_axis(transformed,
indices_or_sections=self.num_masks,
axis=0)
transformed = [F.squeeze(t, axis=0) for t in transformed]
transformed_list += transformed
return transformed_list, enc7
def __decode(x, layer=None, train=False):
if not train or layer == 2:
x = F.sigmoid(model.dec2(x))
if not train or layer == 1:
x = F.sigmoid(model.dec1(x))
return x
def __call__(self, x):
h1 = F.sigmoid(self.l1(x))
h2 = F.sigmoid(self.l2(h1))
return h2
def dec_forward(x_data, layer):
x = chainer.Variable(x_data.astype(np.float32))
if layer >= 2:
x = F.sigmoid(model.dec2(x))
y = model.dec1(x)
return y
def __call__(self, x):
out_plain = self.activate(self.plain(x))
out_transform = F.sigmoid(self.transform(x))
y = out_plain * out_transform + x * (1 - out_transform)
return y
def f(x):
y = functions.sigmoid(x)
return y * y
def zoneout(self, U):
if self._using_zoneout and chainer.config.train:
return 1- zoneout(functions.sigmoid(-U), self._zoneout)
return functions.sigmoid(U)
#初期化、設定
model = MyAE()
optimizer = optimizers.SGD()
optimizer.setup(model)
x = Variable(np.array(x, dtype=np.float32))
#学習
for a in range(10000):
model.cleargrads()
loss = model(x)
loss.backward()
optimizer.update()
#圧縮結果、6次元
y = F.sigmoid(model.l1(x))
ans = y.data
#6次元を2次元に圧縮するAE
class MyAE2(Chain):
def __init__(self):
super(MyAE2, self).__init__(
#ネットワーク
l1=L.Linear(6, 2),
l2=L.Linear(2, 6),
)
def __call__(self, x):
#誤差
out = self.fwd(x)
def __call__(self, input_x, t):
output = self.predictor(input_x)
batch_size, _, grid_h, grid_w = output.shape
self.seen += batch_size
x, y, w, h, conf, prob = F.split_axis(F.reshape(
output, (batch_size, self.predictor.n_boxes,
self.predictor.n_classes + 5, grid_h, grid_w)),
(1, 2, 3, 4, 5),
axis=2)
x = F.sigmoid(x) # xのactivation
y = F.sigmoid(y) # yのactivation
conf = F.sigmoid(conf) # confのactivation
prob = F.transpose(prob, (0, 2, 1, 3, 4))
prob = F.softmax(prob) # probablitiyのacitivation
# 教師データの用意
tw = np.zeros(
w.shape,
dtype=np.float32) # wとhが0になるように学習(e^wとe^hは1に近づく -> 担当するbboxの倍率1)
th = np.zeros(h.shape, dtype=np.float32)
tx = np.tile(0.5, x.shape).astype(np.float32) # 活性化後のxとyが0.5になるように学習()
ty = np.tile(0.5, y.shape).astype(np.float32)
if self.seen < self.unstable_seen: # centerの存在しないbbox誤差学習スケールは基本0.1
box_learning_scale = np.tile(0.1, x.shape).astype(np.float32)
else:
box_learning_scale = np.tile(0, x.shape).astype(np.float32)
tconf = np.zeros(
conf.shape, dtype=np.float32
) # confidenceのtruthは基本0、iouがthresh以上のものは学習しない、ただしobjectの存在するgridのbest_boxのみ真のIOUに近づかせる
conf_learning_scale = np.tile(0.1, conf.shape).astype(np.float32)
tprob = prob.data.copy() # best_anchor以外は学習させない(自身との二乗和誤差 = 0)
# 全bboxとtruthのiouを計算(batch単位で計算する)
x_shift = Variable(
np.broadcast_to(np.arange(grid_w, dtype=np.float32), x.shape[1:]))
y_shift = Variable(
np.broadcast_to(
np.arange(grid_h, dtype=np.float32).reshape(grid_h, 1),
y.shape[1:]))
w_anchor = Variable(
np.broadcast_to(
np.reshape(
np.array(self.anchors, dtype=np.float32)[:, 0],
(self.predictor.n_boxes, 1, 1, 1)), w.shape[1:]))
h_anchor = Variable(
np.broadcast_to(
np.reshape(
np.array(self.anchors, dtype=np.float32)[:, 1],
(self.predictor.n_boxes, 1, 1, 1)), h.shape[1:]))
x_shift.to_gpu(), y_shift.to_gpu(), w_anchor.to_gpu(), h_anchor.to_gpu(
)
best_ious = []
for batch in range(batch_size):
n_truth_boxes = len(t[batch])
box_x = (x[batch] + x_shift) / grid_w
box_y = (y[batch] + y_shift) / grid_h
box_w = F.exp(w[batch]) * w_anchor / grid_w
box_h = F.exp(h[batch]) * h_anchor / grid_h
ious = []
for truth_index in range(n_truth_boxes):
truth_box_x = Variable(
np.broadcast_to(
np.array(t[batch][truth_index]["x"], dtype=np.float32),
box_x.shape))
truth_box_y = Variable(
np.broadcast_to(
np.array(t[batch][truth_index]["y"], dtype=np.float32),
box_y.shape))
truth_box_w = Variable(
np.broadcast_to(
np.array(t[batch][truth_index]["w"], dtype=np.float32),
box_w.shape))
truth_box_h = Variable(
np.broadcast_to(
np.array(t[batch][truth_index]["h"], dtype=np.float32),
box_h.shape))
truth_box_x.to_gpu(), truth_box_y.to_gpu(), truth_box_w.to_gpu(
), truth_box_h.to_gpu()
ious.append(
multi_box_iou(
Box(box_x, box_y, box_w, box_h),
Box(truth_box_x, truth_box_y, truth_box_w,
truth_box_h)).data.get())
ious = np.array(ious)
best_ious.append(np.max(ious, axis=0))
best_ious = np.array(best_ious)
# 一定以上のiouを持つanchorに対しては、confを0に下げないようにする(truthの周りのgridはconfをそのまま維持)。
tconf[best_ious > self.thresh] = conf.data.get()[
best_ious > self.thresh]
conf_learning_scale[best_ious > self.thresh] = 0
# objectの存在するanchor boxのみ、x、y、w、h、conf、probを個別修正
abs_anchors = self.anchors / np.array([grid_w, grid_h])
for batch in range(batch_size):
for truth_box in t[batch]:
truth_w = int(float(truth_box["x"]) * grid_w)
truth_h = int(float(truth_box["y"]) * grid_h)
truth_n = 0
best_iou = 0.0
for anchor_index, abs_anchor in enumerate(abs_anchors):
iou = box_iou(
Box(0, 0, float(truth_box["w"]),
float(truth_box["h"])),
Box(0, 0, abs_anchor[0], abs_anchor[1]))
if best_iou < iou:
best_iou = iou
truth_n = anchor_index
# objectの存在するanchorについて、centerを0.5ではなく、真の座標に近づかせる。anchorのスケールを1ではなく真のスケールに近づかせる。学習スケールを1にする。
box_learning_scale[batch, truth_n, :, truth_h, truth_w] = 1.0
tx[batch, truth_n, :, truth_h,
truth_w] = float(truth_box["x"]) * grid_w - truth_w
ty[batch, truth_n, :, truth_h,
truth_w] = float(truth_box["y"]) * grid_h - truth_h
tw[batch, truth_n, :, truth_h, truth_w] = np.log(
float(truth_box["w"]) / abs_anchors[truth_n][0])
th[batch, truth_n, :, truth_h, truth_w] = np.log(
float(truth_box["h"]) / abs_anchors[truth_n][1])
tprob[batch, :, truth_n, truth_h, truth_w] = 0
tprob[batch,
int(truth_box["label"]), truth_n, truth_h, truth_w] = 1
# IOUの観測
full_truth_box = Box(float(truth_box["x"]),
float(truth_box["y"]),
float(truth_box["w"]),
float(truth_box["h"]))
predicted_box = Box(
(x[batch][truth_n][0][truth_h][truth_w].data.get() +
truth_w) / grid_w,
(y[batch][truth_n][0][truth_h][truth_w].data.get() +
truth_h) / grid_h,
np.exp(w[batch][truth_n][0][truth_h][truth_w].data.get()) *
abs_anchors[truth_n][0],
np.exp(h[batch][truth_n][0][truth_h][truth_w].data.get()) *
abs_anchors[truth_n][1])
predicted_iou = box_iou(full_truth_box, predicted_box)
tconf[batch, truth_n, :, truth_h, truth_w] = predicted_iou
conf_learning_scale[batch, truth_n, :, truth_h, truth_w] = 10.0
# debug prints
maps = F.transpose(prob[batch], (2, 3, 1, 0)).data
print(
"best confidences and best conditional probability and predicted class of each grid:"
)
for i in range(grid_h):
for j in range(grid_w):
print("%2d" %
(int(conf[batch, :, :, i, j].data.max() * 100)),
end=" ")
print(" ", end="")
for j in range(grid_w):
print("%2d" % (maps[i][j][int(
maps[i][j].max(axis=1).argmax())].argmax()),
end=" ")
print(" ", end="")
for j in range(grid_w):
print("%2d" % (maps[i][j][int(
maps[i][j].max(axis=1).argmax())].max() * 100),
end=" ")
print()
print(
"best default iou: %.2f predicted iou: %.2f confidence: %.2f class: %s"
% (best_iou, predicted_iou,
conf[batch][truth_n][0][truth_h][truth_w].data,
t[batch][0]["label"]))
print("-------------------------------")
print("seen = %d" % self.seen)
# loss計算
tx, ty, tw, th, tconf, tprob = Variable(tx), Variable(ty), Variable(
tw), Variable(th), Variable(tconf), Variable(tprob)
box_learning_scale, conf_learning_scale = Variable(
box_learning_scale), Variable(conf_learning_scale)
tx.to_gpu(), ty.to_gpu(), tw.to_gpu(), th.to_gpu(), tconf.to_gpu(
), tprob.to_gpu()
box_learning_scale.to_gpu()
conf_learning_scale.to_gpu()
x_loss = F.sum((tx - x)**2 * box_learning_scale) / 2
y_loss = F.sum((ty - y)**2 * box_learning_scale) / 2
w_loss = F.sum((tw - w)**2 * box_learning_scale) / 2
h_loss = F.sum((th - h)**2 * box_learning_scale) / 2
c_loss = F.sum((tconf - conf)**2 * conf_learning_scale) / 2
p_loss = F.sum((tprob - prob)**2) / 2
print(
"x_loss: %f y_loss: %f w_loss: %f h_loss: %f c_loss: %f p_loss: %f"
% (F.sum(x_loss).data, F.sum(y_loss).data, F.sum(w_loss).data,
F.sum(h_loss).data, F.sum(c_loss).data, F.sum(p_loss).data))
loss = x_loss + y_loss + w_loss + h_loss + c_loss + p_loss
return loss
def __call__(self, x, extractor):
h = F.relu(self.c0(F.concat([x, extractor])))
h = F.sigmoid(self.c1(h))
return h
def __call__(self, x):
self.chi = F.concat((x, self.r))
(self.nu, self.xi) = F.split_axis(self.l_dl(self.chi), [self.Y], 1)
(self.kr, self.betar, self.kw, self.betaw, self.e, self.v, self.f, self.ga, self.gw, self.pi) = \
F.split_axis(self.xi, self.xi_split_indices, 1)
self.kr = F.reshape(self.kr, (self.R, self.W)) # R * W
self.betar = 1 + F.softplus(self.betar) # 1 * R
# self.kw: 1 * W
self.betaw = 1 + F.softplus(self.betaw) # 1 * 1
self.e = F.sigmoid(self.e) # 1 * W
# self.v : 1 * W
self.f = F.sigmoid(self.f) # 1 * R
self.ga = F.sigmoid(self.ga) # 1 * 1
self.gw = F.sigmoid(self.gw) # 1 * 1
self.pi = F.softmax(F.reshape(self.pi,
(self.R, 3))) # R * 3 (softmax for 3)
# self.wr : N * R
self.psi_mat = 1 - F.broadcast_to(self.f,
(self.N, self.R)) * self.wr # N x R
self.psi = F.prod(self.psi_mat, 1).reshape(self.N, 1) # N x 1
# self.ww, self.u : N * 1
self.u = (self.u + self.ww - (self.u * self.ww)) * self.psi
self.a = u2a(self.u.data) # N * 1
self.cw = C(self.M.data, self.kw.data, self.betaw.data) # N * 1
self.ww = F.matmul(
F.matmul(self.a, self.ga) + F.matmul(self.cw, 1.0 - self.ga),
self.gw) # N * 1
self.M = self.M * (xp.ones(
(self.N, self.W)).astype(xp.float32) - F.matmul(
self.ww, self.e)) + F.matmul(self.ww, self.v) # N * W
if self.K > 0:
self.p = (1.0 - F.matmul(Variable(xp.ones((self.N, 1)).astype(xp.float32)), F.reshape(F.sum(self.ww), (1, 1)))) \
* self.p + self.ww # N * 1
self.p.data = xp.sort(self.p.data, 0)
self.p.data[0:-self.K] = 0.
self.p.data[-self.K:] = self.p.data[-self.K:] / xp.sum(
self.p.data[-self.K:])
self.ww.data = xp.sort(self.ww.data, 0)
self.ww.data[0:-self.K] = 0.
self.ww.data[-self.K:] = self.ww[-self.K:].data / xp.sum(
self.ww.data[-self.K:])
self.wwrep = F.matmul(
self.ww, Variable(xp.ones(
(1, self.N)).astype(xp.float32))) # N * N
self.ww_p_product = xp.zeros((self.N, self.N)).astype(xp.float32)
self.ww_p_product[-self.K:, -self.K:] = F.matmul(
self.ww[-self.K:, -self.K:],
F.transpose(self.p[-self.K:, -self.K:])).data
self.L = (1.0 - self.wwrep - F.transpose(
self.wwrep)) * self.L + self.ww_p_product # N * N
self.L = self.L * (xp.ones(
(self.N, self.N)) - xp.eye(self.N)) # force L[i,i] == 0
self.L.data[self.L.data < 1 / self.K] = 0.
else:
self.p = (1.0 - F.matmul(Variable(xp.ones((self.N, 1)).astype(xp.float32)),
F.reshape(F.sum(self.ww), (1, 1)))) \
* self.p + self.ww # N * 1
self.wwrep = F.matmul(
self.ww, Variable(xp.ones(
(1, self.N)).astype(xp.float32))) # N * N
self.L = (1.0 - self.wwrep -
F.transpose(self.wwrep)) * self.L + F.matmul(
self.ww, F.transpose(self.p)) # N * N
self.L = self.L * (xp.ones(
(self.N, self.N)) - xp.eye(self.N)) # force L[i,i] == 0
self.fo = F.matmul(self.L, self.wr) # N * R
self.ba = F.matmul(F.transpose(self.L), self.wr) # N * R
self.cr = C(self.M.data, self.kr.data, self.betar.data)
self.bacrfo = F.concat((
F.reshape(F.transpose(self.ba), (self.R, self.N, 1)),
F.reshape(F.transpose(self.cr), (self.R, self.N, 1)),
F.reshape(F.transpose(self.fo), (self.R, self.N, 1)),
), 2) # R * N * 3
self.pi = F.reshape(self.pi, (self.R, 3, 1)) # R * 3 * 1
self.wr = F.transpose(
F.reshape(F.batch_matmul(self.bacrfo, self.pi),
(self.R, self.N))) # N * R
self.r = F.reshape(F.matmul(F.transpose(self.M), self.wr),
(1, self.R * self.W)) # W * R (-> 1 * RW)
self.y = self.l_Wr(self.r) + self.nu # 1 * Y
return self.y
def predict(self, atoms, adjs):
with chainer.no_backprop_mode(), chainer.using_config('train', False):
x = self.__call__(atoms, adjs)
return F.sigmoid(x)