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 clipped_loss(x, t):
diff = x - t
abs_loss = abs(diff)
squared_loss = diff ** 2
abs_loss = F.expand_dims(abs_loss, 1)
squared_loss = F.expand_dims(squared_loss, 1)
return F.sum(F.min(F.concat((abs_loss, squared_loss), axis=1), axis=1))
def __call__(self, x):
input_shape = x.shape
x = F.average_pooling_2d(x, ksize=x.shape[2:])
x = F.reshape(x, shape=(x.shape[0], -1))
x = self.init_fc(x)
x = self.main_fcs(x)
x = self.final_fc(x)
x = F.broadcast_to(F.expand_dims(F.expand_dims(x, axis=2), axis=3),
input_shape)
return x
def attention_layer(self, features, features_proj, Xp):
h = F.expand_dims(self.w_att(Xp), 1)
features_proj = F.normalize(features_proj, axis=-1)
h = F.normalize(h, axis=-1)
h_att = F.relu(features_proj + F.broadcast_to(h, features_proj.shape)) # (N, self.D, self.C) + (N, 1, self.C)
out_att = self.w(F.reshape(h_att, (-1, self.C))) # (Nxself.D, self.C) -> (Nxself.D, 1)
out_att = F.reshape(out_att, (-1, self.D)) # (N, self.D)
alpha = F.softmax(out_att) # (N, self.D)
context = F.sum(features * F.broadcast_to(F.expand_dims(alpha, 1), features.shape), axis=2) # (N, self.C, self.D) * (N, 1, self.D)
return context, alpha
def __call__(self, x, t, index):
h = self.predict(x)
self.history = np.append(self.history, np.array([np.mean(h.data, axis=0)]), axis=0)
h = F.select_item(h, index) # choose the action[index] in each column
error_abs = abs(h - t)
error = F.concat((F.expand_dims(error_abs ** 2, 1), F.expand_dims(error_abs, 1)), axis=1)
# 1 < error_abs <=> error ** 2 > error, error < 1 <=> error ** 2 < error
self.loss = F.sum(F.min(error, axis=1)) / np.float32(len(error_abs))
return self.loss
def __call__(self, x):
w = F.average_pooling_2d(x, ksize=x.shape[2:])
w = self.fc1(w)
if self.use_conv2:
w = self.activ(w)
w = self.fc2(w)
w = self.sigmoid(w)
w = F.broadcast_to(F.expand_dims(F.expand_dims(w, axis=2), axis=3), x.shape)
x = x * w
return x
def pool(self, WX, skip_mask=None):
Z, F, O, I = None, None, None, None
# f-pooling
if len(self._pooling) == 1:
assert len(WX) == 2
Z, F = WX
Z = functions.tanh(Z)
F = self.zoneout(F)
# fo-pooling
if len(self._pooling) == 2:
assert len(WX) == 3
Z, F, O = WX
Z = functions.tanh(Z)
F = self.zoneout(F)
O = functions.sigmoid(O)
# ifo-pooling
if len(self._pooling) == 3:
assert len(WX) == 4
Z, F, O, I = WX
Z = functions.tanh(Z)
F = self.zoneout(F)
O = functions.sigmoid(O)
I = functions.sigmoid(I)
assert Z is not None
assert F is not None
T = Z.shape[2]
for t in xrange(T):
zt = Z[:, :, t]
ft = F[:, :, t]
ot = 1 if O is None else O[:, :, t]
it = 1 - ft if I is None else I[:, :, t]
xt = 1 if skip_mask is None else skip_mask[:, t,
None] # will be used for seq2seq to skip PAD
if self.ct is None:
self.ct = (1 - ft) * zt * xt
else:
self.ct = ft * self.ct + it * zt * xt
self.ht = self.ct if O is None else ot * self.ct
if self.H is None:
self.H = functions.expand_dims(self.ht, 2)
else:
self.H = functions.concat(
(self.H, functions.expand_dims(self.ht, 2)), axis=2)
if self._test:
self.H.unchain_backward()
return self.H
def __call__(self, x, enc_out=None, mask=None):
"""
args
x: paralleled main features in the model
Variable in (batch, hidden_dim, length)
u: hidden features from Encoder
Variable in (batch, hidden_dim, length)
mask: padding-mask or future-mask
xp-array in (batch, length, length)
an element takes 'False' when pad/future, otherwise 'True'
returns
"""
# ksize-1-convolution results in parallel linear projections
if self.self_attention:
qkv = F.squeeze(self.W(F.expand_dims(x, axis=3)), axis=3)
query, key, value = F.split_axis(qkv, 3, axis=1)
else:
query = F.squeeze(self.W_Q(F.expand_dims(x, axis=3)), axis=3)
kv = F.squeeze(self.W_KV(F.expand_dims(enc_out, axis=3)), axis=3)
key, value = F.split_axis(kv, 2, axis=1)
# make q,k,v into (batch*parallel, dim/parallel, length)shape
query = F.concat(F.split_axis(query, self.parallel_num, axis=1),
axis=0)
key = F.concat(F.split_axis(key, self.parallel_num, axis=1), axis=0)
value = F.concat(F.split_axis(value, self.parallel_num, axis=1),
axis=0)
mask = self.xp.concatenate([mask] * self.parallel_num, axis=0)
attention_weight = F.batch_matmul(query, key, transa=True) * self.scale
attention_weight = F.where(
mask, attention_weight,
self.xp.full(attention_weight.shape, -np.inf, dtype=np.float32))
attention_weight = F.softmax(attention_weight, axis=2)
attention_weight = F.dropout(attention_weight, self.dropout_rate)
attention_weight = F.where(
self.xp.isnan(attention_weight.data),
self.xp.full(attention_weight.shape, 0, dtype=np.float32),
attention_weight)
self.attention_weight = copy.deepcopy(attention_weight.data)
# attention: (batch, q-length, k-length) -> (batch, 1, q-length, k-length)
# value: (batch, dim/parallel, k-length) -> (batch, dim/parallel, 1, k-length)
attention_weight, value = F.broadcast(attention_weight[:, None],
value[:, :, None])
weighted_sum = F.sum(attention_weight * value, axis=3)
weighted_sum = F.concat(F.split_axis(weighted_sum,
self.parallel_num,
axis=0),
axis=1)
weighted_sum = F.squeeze(self.linear(
F.expand_dims(weighted_sum, axis=3)),
axis=3)
return weighted_sum
def kl_div(mu1, lv1, lv2):
# KL Divergence between given normal and prior at N(0, sigma_2)
# Prior assumes mean at zero
# lns2 - lns1 + (s2^2 + (u1 - u2)**2)/ 2s2**2 - 0.5
if len(lv1.shape) == 2:
lv1 = F.expand_dims(lv1, 0)
mu1 = F.expand_dims(mu1, 0)
lv2 = F.broadcast_to(lv2, lv1.shape)
v12 = F.exp(lv1)**2.0
v22 = F.exp(lv2)**2.0
return lv2 - lv1 + .5 * v12 / v22 + .5 * mu1**2. / v22 - .5
def get_gcam(self, end_output, activations, shape, label): self.cleargrads() class_id = self.set_init_grad(end_output, label) end_output.backward(retain_grad=True) grad = activations.grad_var grad = F.average_pooling_2d(grad, (grad.shape[-2], grad.shape[-1]), 1) grad = F.expand_dims(F.reshape(grad, (grad.shape[0]*grad.shape[1], grad.shape[2], grad.shape[3])), 0) weights = activations weights = F.expand_dims(F.reshape(weights, (weights.shape[0]*weights.shape[1], weights.shape[2], weights.shape[3])), 0) gcam = F.resize_images(F.relu(F.convolution_2d(weights, grad, None, 1, 0)), shape) return gcam, class_id
def calcLoss(G, X, S):
GXr = G*xp.real(X).astype(np.float32)
GXi = G*xp.imag(X).astype(np.float32)
Sr = xp.real(S).astype(np.float32)
Si = xp.imag(S).astype(np.float32)
gxL = [F.expand_dims(iDGTcf.iDGT(GXr[ii],GXi[ii],windowDG,shiftLenG,fftLenG), axis=0) for ii in range(len(G))]
sL = [F.expand_dims(iDGTcf.iDGT(Sr[ii],Si[ii],windowDG,shiftLenG,fftLenG), axis=0) for ii in range(len(G))]
gx = F.vstack(gxL)
s = F.vstack(sL)
loss = F.mean_absolute_error(gx,s)
return loss
def __call__(self, x):
att1 = F.average_pooling_2d(x, ksize=x.shape[2:])
att1 = self.mlp(att1)
att2 = F.max_pooling_2d(x, ksize=x.shape[2:])
att2 = self.mlp(att2)
att = att1 + att2
att = F.sigmoid(att)
att = F.broadcast_to(F.expand_dims(F.expand_dims(att, axis=2), axis=3),
x.shape)
x = x * att
return x
def __call__(self, X, ht_enc, H_enc, skip_mask=None, test=False):
self._test = test
WX = self.W(X)
Vh = self.V(ht_enc)
Vh, WX = functions.broadcast(functions.expand_dims(Vh, axis=2), WX)
# f-pooling
Z, F, O = functions.split_axis(WX + Vh, 3, axis=1)
Z = functions.tanh(Z)
F = self.zoneout(F)
O = functions.sigmoid(O)
T = Z.shape[2]
# compute ungated hidden states
self.contexts = []
for t in xrange(T):
z = Z[:, :, t]
f = F[:, :, t]
if t == 0:
ct = (1 - f) * z
self.contexts.append(ct)
else:
ct = f * self.contexts[-1] + (1 - f) * z
self.contexts.append(ct)
if skip_mask is not None:
assert skip_mask.shape[1] == H_enc.shape[2]
softmax_getas = (skip_mask == 0) * -1e6
# compute attention weights (eq.8)
H_enc = functions.swapaxes(H_enc, 1, 2)
for t in xrange(T):
ct = self.contexts[t]
geta = 0 if skip_mask is None else softmax_getas[
..., None] # to skip PAD
mask = 1 if skip_mask is None else skip_mask[...,
None] # to skip PAD
alpha = functions.batch_matmul(H_enc, ct) + geta
alpha = functions.softmax(alpha) * mask
alpha = functions.broadcast_to(alpha, H_enc.shape) # copy
kt = functions.sum(alpha * H_enc, axis=1)
ot = O[:, :, t]
self.ht = ot * self.o(functions.concat((kt, ct), axis=1))
if test:
self.ht.unchain_backward()
if t == 0:
self.H = functions.expand_dims(self.ht, 2)
else:
self.H = functions.concat(
(self.H, functions.expand_dims(self.ht, 2)), axis=2)
return self.H
def __call__(self, xs, ys):
eos = self.xp.array([EOS], 'i')
xs = [self.denoiseInput(x[::-1], self.denoising_rate)
for x in xs] # denoising
#ys_d = [self.wordDropout(y, self.word_dropout) for y in ys] # word dropout
ys_d = [self.denoiseInput(y, self.word_dropout)
for y in ys] # word dropout
ys_in = [F.concat([eos, y], axis=0) for y in ys_d]
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, at = self.encoder(None, exs) # layer x batch x n_units
hx_t = F.transpose(hx, (1, 0, 2)) # batch x layer x n_units
mu = self.W_mu(hx_t) # batch x n_latent
ln_var = self.W_ln_var(hx_t)
#print(mu.shape)
#print(hx_t.shape)
rec_loss = 0
concat_ys_out = F.concat(ys_out, axis=0)
for _ in range(self.k):
z = F.gaussian(mu, ln_var)
z_e = F.expand_dims(z, 2) # batch x n_latent x 1
Wz = self.W_h(z_e) # batch x (layer x unit)
#print('Wz: {}, {}'.format(Wz.shape, type(Wz)))
hys = F.split_axis(Wz, self.n_layers, 1) # layer x batch x unit
#print('hys, {}'.format([x.shape for x in hys]))
c_hy = F.concat([F.expand_dims(hy, 0) for hy in hys],
0) # layer x batch x unit
#print('c_hy: {}'.format(c_hy.shape))
_, os = self.decoder(c_hy, eys)
#print(len(os))
concat_os = F.concat(os, axis=0)
rec_loss += F.sum(
F.softmax_cross_entropy(self.W(concat_os),
concat_ys_out,
reduce='no')) / (self.k * batch)
latent_loss = self.C * F.gaussian_kl_divergence(mu, ln_var) / batch
loss = rec_loss + latent_loss
chainer.report({'loss': loss.data}, self)
n_words = concat_ys_out.shape[0]
perp = self.xp.exp(loss.data * batch / n_words)
chainer.report({'perp': perp}, self)
return loss
def translate(self, xs, max_length=100):
xs = numpy.insert(xs, 0, 2)
xs = numpy.append(xs, 0)
with chainer.no_backprop_mode(), chainer.using_config('train', False):
exs = self.embed_x(Variable(self.xp.array(xs,
dtype=self.xp.int32)))
h = F.expand_dims(exs, axis=0)
h = F.expand_dims(h, axis=0)
h = F.transpose(h, (0, 1, 3, 2))
for i in range(self.stack):
h = self.gcnn[i](h)
h = F.squeeze(h, axis=1)
h = F.squeeze(h, axis=0)
h = F.transpose(h, (1, 0))
ys = self.xp.full(1, 2, self.xp.int32)
result = []
hx = None
cx = None
hx2 = None
cx2 = None
for i in range(max_length):
eys = self.embed_y(ys)
eyys = self.embed_yy(ys)
eys2 = [eys]
eyys2 = [eyys]
hx, cx, ss = self.decoder(hx, cx, eys2)
hx2, cx2, ss2 = self.decoder2(hx2, cx2, eyys2)
batch_A = F.matmul(h, ss[0], transb=True) * self.scale_score
batch_A = F.softmax(batch_A, axis=0)
if self.weight:
with open("weight/wei.txt", "a", encoding="utf-8") as f:
for j in range(len(batch_A)):
f.write(str(batch_A[j][0].data) + "\n")
f.write("--------------\n")
s = F.matmul(batch_A, h, transa=True)
t = (self.We(s) + self.Ws(ss2[0]))
ys = self.xp.argmax(t.data, axis=1).astype(self.xp.int32)
if ys[0] == 0:
break
result.append(ys)
result = cuda.to_cpu(
self.xp.concatenate([self.xp.expand_dims(x, 0) for x in result]).T)
# Remove EOS taggs
outs = []
for y in result:
inds = numpy.argwhere(y == EOS)
if len(inds) > 0:
y = y[:inds[0, 0]]
outs.append(y)
return outs
def __call__(self, x, hc=None):
w = F.average_pooling_2d(x, ksize=x.shape[2:])
w = w.reshape((w.shape[0], -1))
if hc is None:
h = [self.xp.zeros_like(w.array, dtype=w.dtype)] * self.num_layers
c = [self.xp.zeros_like(w.array, dtype=w.dtype)] * self.num_layers
else:
h, c = hc
h, c = self.lstm(w, h, c)
w = F.expand_dims(F.expand_dims(h[-1], axis=-1), axis=-1)
x = x * w
return x, (h, c)
def loss_information(enc, x):
p_logit = enc(x)
p = F.sigmoid(p_logit)
p_ave = F.sum(p, axis=0) / x.data.shape[0]
cond_ent = F.sum(-p * F.log(p + 1e-8) -
(1 - p) * F.log(1 - p + 1e-8)) / p.data.shape[0]
marg_ent = F.sum(-p_ave * F.log(p_ave + 1e-8) -
(1 - p_ave) * F.log(1 - p_ave + 1e-8))
p_ave = F.reshape(p_ave, (1, len(p_ave.data)))
p_ave_separated = F.separate(p_ave, axis=1)
p_separated = F.separate(F.expand_dims(p, axis=2), axis=1)
p_ave_list_i = []
p_ave_list_j = []
p_list_i = []
p_list_j = []
for i in range(n_bit - 1):
p_ave_list_i.extend(list(p_ave_separated[i + 1:]))
p_list_i.extend(list(p_separated[i + 1:]))
p_ave_list_j.extend([p_ave_separated[i] for n in range(n_bit - i - 1)])
p_list_j.extend([p_separated[i] for n in range(n_bit - i - 1)])
p_ave_pair_i = F.expand_dims(F.concat(tuple(p_ave_list_i), axis=0), axis=1)
p_ave_pair_j = F.expand_dims(F.concat(tuple(p_ave_list_j), axis=0), axis=1)
p_pair_i = F.expand_dims(F.concat(tuple(p_list_i), axis=1), axis=2)
p_pair_j = F.expand_dims(F.concat(tuple(p_list_j), axis=1), axis=2)
p_pair_stacked_i = F.concat(
(p_pair_i, 1 - p_pair_i, p_pair_i, 1 - p_pair_i), axis=2)
p_pair_stacked_j = F.concat(
(p_pair_j, p_pair_j, 1 - p_pair_j, 1 - p_pair_j), axis=2)
p_ave_pair_stacked_i = F.concat(
(p_ave_pair_i, 1 - p_ave_pair_i, p_ave_pair_i, 1 - p_ave_pair_i),
axis=1)
p_ave_pair_stacked_j = F.concat(
(p_ave_pair_j, p_ave_pair_j, 1 - p_ave_pair_j, 1 - p_ave_pair_j),
axis=1)
p_product = F.sum(p_pair_stacked_i * p_pair_stacked_j, axis=0) / len(
p.data)
p_ave_product = p_ave_pair_stacked_i * p_ave_pair_stacked_j
pairwise_mi = 2 * F.sum(p_product * F.log(
(p_product + 1e-8) / (p_ave_product + 1e-8)))
return cond_ent, marg_ent, pairwise_mi
def __call__(self, h):
shape = h.shape
h = F.reshape(h, (shape[0], np.prod(shape[1:])))
h_ns = F.batch_l2_norm_squared(h)
bs = shape[0]
h0 = F.broadcast_to(F.expand_dims(h_ns, 0), (bs, bs))
h1 = F.broadcast_to(F.expand_dims(h_ns, 1), (bs, bs))
hh = F.linear(h, h)
D = h0 + h1 - 2 * hh
D = F.sum(D) / np.prod(h.shape)
return D
def __call__(self, x):
w = F.average_pooling_2d(x, ksize=x.shape[2:])
if not self.use_conv:
w = F.reshape(w, shape=(w.shape[0], -1))
w = self.conv1(w) if self.use_conv else self.fc1(w)
w = self.activ(w)
w = self.conv2(w) if self.use_conv else self.fc2(w)
w = self.sigmoid(w)
if not self.use_conv:
w = F.expand_dims(F.expand_dims(w, axis=2), axis=3)
x = x * w
return x
def compute_attention(self, query, key):
"""
:param query: with shape of (mb, N_1, hidden_dim)
:param key: with shape of (mb, N_2, hidden_dim)
:return: attn: attention weights (mb, N_1, N_2)
"""
energy_layer = self.energy_layer
mb, N_1, hidden_dim = query.shape
N_2 = key.shape[1]
# # query: (mb, N_1, 1, hidden_dim)
# query = functions.expand_dims(query, axis=2)
# # query: (mb, N_1, N_2, hidden_dim)
# query = functions.tile(query, reps=(1, 1, N_2, 1))
# # query: (mb * N_1 * N_2, hidden_dim)
# query = functions.reshape(query, (mb * N_1 * N_2, hidden_dim))
# query: (mb * N_1 hidden_dim)
# # key: (mb, 1, N_2, hidden_dim)
# key = functions.expand_dims(key, axis=1)
# # key: (mb, N_1, N_2, hidden_dim)
# key = functions.tile(key, reps=(1, N_1, 1, 1))
# # key: (mb * N_1 * N_2, hidden_dim)
# key = functions.reshape(key, (mb * N_1 * N_2, hidden_dim))
# key: (mb * N_2, hidden_dim)
# energy: (mb * N_1 * N_2, 1)
# energy = self.activation(energy_layer(key, query))
query_real, query_imag = self.fourier_transform(query)
key_real, key_imag = self.fourier_transform(key)
query_real = functions.reshape(functions.tile(functions.expand_dims(
query_real, axis=2),
reps=(1, 1, N_2, 1)),
shape=(mb * N_1 * N_2, hidden_dim))
query_imag = functions.reshape(functions.tile(functions.expand_dims(
query_imag, axis=2),
reps=(1, 1, N_2, 1)),
shape=(mb * N_1 * N_2, hidden_dim))
key_real = functions.reshape(functions.tile(functions.expand_dims(
key_real, axis=1),
reps=(1, N_1, 1, 1)),
shape=(mb * N_1 * N_2, hidden_dim))
key_imag = functions.reshape(functions.tile(functions.expand_dims(
key_imag, axis=1),
reps=(1, N_1, 1, 1)),
shape=(mb * N_1 * N_2, hidden_dim))
energy = self.activation(
energy_layer(key_real, query_real) +
energy_layer(key_imag, query_imag))
energy = functions.reshape(energy, (mb, N_1, N_2))
return energy
def __call__(self, x0, x1, l0, l1, train=True):
""" Forward computation.
Args:
x0: Chainer variable in shape (B, T0) where B is the batch size,
T is the number of tokens in each data. Each element should be
given as the index of embedding.
x1: Chainer variable in shape (B, T1)
Returns:
"""
t0 = x0.shape[1]
t1 = x1.shape[1]
# a: (B, T0, M)
a = self.emb(x0)
# b: (B, T1, M)
b = self.emb(x1)
if not self._train_embedding:
a.unchain_backward()
b.unchain_backward()
a = self._token_wise_linear(a, self.emb_proj, l0, train, self.xp)
b = self._token_wise_linear(b, self.emb_proj, l1, train, self.xp)
# Apply perceptron layer to each feature vectors ... eq. 1
# (B, Ti, M) -> (B * Ti, M) -> (B * Ti, F) -> (B, Ti, F)
a_f = self._token_wise_linear(a, self.f, l0, train, self.xp)
b_f = self._token_wise_linear(b, self.f, l1, train, self.xp)
# for each batch, calculate a_f[b]
# e: (B, T0, T1)
e = F.batch_matmul(a_f, b_f, transb=True)
# att_*: (B, T0, T1)
att_b, att_a = self._length_aware_softmax(e, l0, l1, self.xp)
# sum((B, T0, T1).(B, T0, T1, M)) -> beta: (B, T0, M) ... eq. 2
b_tiled = F.tile(F.expand_dims(b, 1), (1, t0, 1, 1))
att_b = F.expand_dims(att_b, 3)
beta = F.sum(F.broadcast_to(att_b, b_tiled.shape) * b_tiled, axis=2)
# sum((B, T0, T1).(N, T0, T1, M)) -> beta: (B, T1, M) ... eq. 2
a_tiled = F.tile(F.expand_dims(a, 2), (1, 1, t1, 1))
att_a = F.expand_dims(att_a, 3)
alpha = F.sum(F.broadcast_to(att_a, a_tiled.shape) * a_tiled, axis=1)
# Make comparison, [(B, Ti, M), (B, Ti, M)] -> (B, M')
v1 = self._compare(a, beta, l0, train, self.xp)
v2 = self._compare(b, alpha, l1, train, self.xp)
# (B, M' + M') -> (B, n_class) ... eq. 4 & 5
v = F.concat((v1, v2), axis=1)
y = self.h(v)
return y
def mixture_of_discretized_logistics_nll(x, y):
"""
Args:
x: (b, c, n, n)
y: (b, 10*n_mix, n, n)
"""
xp = get_array_module(x)
n_mix = y.shape[1] // 10
logit_prob = y[:, :n_mix, :, :]
y = F.reshape(y[:, n_mix:, :, :], x.shape + (n_mix * 3, ))
mean = y[:, :, :, :, 0:n_mix]
log_scale = y[:, :, :, :, n_mix:2 * n_mix]
log_scale = F.maximum(log_scale, -7 * xp.ones(log_scale.shape, dtype='f'))
coeff = F.tanh(y[:, :, :, :, 2 * n_mix:3 * n_mix])
x = xp.repeat(xp.expand_dims(x, 4), n_mix, 4)
m1 = F.expand_dims(mean[:, 0, :, :, :], 1)
m2 = F.expand_dims(
mean[:, 1, :, :, :] + coeff[:, 0, :, :, :] * x[:, 0, :, :, :], 1)
m3 = F.expand_dims(
(mean[:, 2, :, :, :] + coeff[:, 1, :, :, :] * x[:, 0, :, :, :] +
coeff[:, 2, :, :, :] * x[:, 1, :, :, :]), 1)
mean = F.concat([m1, m2, m3])
centered_x = x - mean
inv_std = F.exp(-log_scale)
max_in = inv_std * (centered_x + 1. / 255.)
cdf_max = F.sigmoid(max_in)
min_in = inv_std * (centered_x - 1. / 255.)
cdf_min = F.sigmoid(min_in)
log_cdf_max = max_in - F.softplus(max_in) # 0
log_one_minus_cdf_min = -F.softplus(min_in) # 255
cdf_delta = cdf_max - cdf_min # 0 ~ 255
mid_in = inv_std * centered_x
log_pdf_mid = mid_in - log_scale - 2. * F.softplus(mid_in) # mid
log_prob = F.where(
x < -0.999, log_cdf_max,
F.where(
x > 0.999, log_one_minus_cdf_min,
F.where(
cdf_delta.array > 1e-5,
F.log(
F.maximum(cdf_delta,
xp.ones(cdf_delta.shape, dtype='f') * 1e-12)),
log_pdf_mid - xp.log(127.5))))
log_prob = F.transpose(F.sum(log_prob, 1), (0, 3, 1, 2))
log_prob = log_prob + log_prob_from_logit(logit_prob)
loss = F.logsumexp(log_prob, 1)
loss = F.sum(loss, axis=(1, 2))
return -F.mean(loss)
def __call__(self):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
norm = F.batch_l2_norm_squared(self.W)**0.5
norm_broadcasted = F.broadcast_to(F.expand_dims(norm, 1),
self.W.data.shape)
g_broadcasted = F.broadcast_to(F.expand_dims(self.g, 1),
self.W.data.shape)
return g_broadcasted * self.W / norm_broadcasted
def __call__(self, edge, node, triplet):
num_atom = edge.shape[1]
hn1 = F.tile(F.expand_dims(self.Wn1(node), 1), (1, num_atom, 1, 1))
hn2 = F.tile(F.expand_dims(self.Wn1(node), 2), (1, 1, num_atom, 1))
ht1 = self.Wt2(F.sum(zero_plus(self.Wt1(triplet)), axis=1))
concat = F.concat([hn1, hn2, ht1, edge], axis=3)
add = zero_plus(self.We2(zero_plus(self.We1(concat))))
return edge + self.bn(add)
def bger(x, y):
""" Batch outer product
:param x:
:param y:
:return:
"""
if x.dtype == 'int' and y.dtype == 'int':
x_float = F.cast(x, 'float32')
y_float = F.cast(y, 'float32')
res_float = F.expand_dims(x_float, 2) @ F.expand_dims(y_float, 1)
return F.cast(res_float, 'int')
return F.expand_dims(x, 2) @ F.expand_dims(y, 1)
def scoreDot(self, atts, ys):
xs = [self.W3(att) for att in atts]
xs_T = [F.transpose(x, (1, 0))for x in xs]
dots = [F.matmul(y, x)for x, y in zip(xs_T, ys)]
aws = [F.softmax(dot, 1) for dot in dots]
cts = []
for x, aw in zip(xs, aws): # split batch
aw = F.expand_dims(aw, 1)
x = F.tile(F.expand_dims(x, 0), (aw.shape[0], 1, 1))
ct = F.batch_matmul(aw, x)
cts.append(F.reshape(ct, (ct.shape[0], ct.shape[2])))
ds = [F.tanh(self.Wc1(ct) + self.Wc2(y)) for y, ct in zip(ys, cts)]
return ds
def __call__(self, xs):
xs = chainer.dataset.convert.concat_examples(xs, padding=0)
xs = F.transpose(xs,(1,0,2))
for i in range(len(xs)):
if i==0:
hs = self.l1(xs[0])
hs = F.expand_dims(hs,0)
else:
hw = self.l1(xs[i])
hw = F.expand_dims(hw,0)
hs = F.concat((hs,hw), axis=0)
h=F.mean(hs,axis=0)
return h
def __call__(self,input_data,hx=None):
if np.any(hx):
hx = hx.reshape(1,-1,self.h1.out_size)
input_x = [Variable(x) for x in input_data]
hx,cx,y = self.h1(hx,None,input_x)
y2 = [F.concat(x, axis=0) for x in F.pad_sequence(y,length=17, padding=0.)]
y2 = F.concat([F.expand_dims(x,axis=0) for x in y2],axis=0)
out = self.hy(F.concat([F.expand_dims(item[-1],axis=0) for item in y],axis=0))
atn = self.atn(y2)
return F.concat([F.expand_dims(a*o,axis=0) for a,o in zip(atn,out)],axis=0)
def __call__(self, x, pid):
x = self.bn(x)
x = F.swapaxes(x, axis1=1, axis2=3)
y = F.expand_dims(F.expand_dims(pid, axis=-1), axis=-1)
y = F.tile(y, reps=(1, 1, self.audio_window_size, 1))
x = F.concat((x, y), axis=1)
x = self.branch(x)
x = F.reshape(x, shape=(x.shape[0], -1))
x = F.concat((x, pid), axis=1)
x = self.fc1(x)
x = F.tanh(x)
x = self.fc2(x)
return x
def __call__(self, x):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
norm = F.batch_l2_norm_squared(self.W) ** 0.5
norm_broadcasted = F.broadcast_to(
F.expand_dims(norm, 1), self.W.data.shape)
g_broadcasted = F.broadcast_to(
F.expand_dims(self.g, 1), self.W.data.shape)
return F.linear(x, g_broadcasted * self.W / norm_broadcasted, self.b)
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 __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_rnn_encode_proj(self, X):
# Reset rnn state
self.reset_rnn_state()
# Get input shape
in_size, batch_size, in_dim = X.shape
enc_states = X
for currL in range(len(self.rnn_enc)):
for i in range(in_size):
temp_f = F.expand_dims(
F.dropout(self[self.rnn_enc[currL]](enc_states[i]),
ratio=self.cfg["dropout"]["rnn"]), 0)
# if bi-directional
if self.bi_rnn:
temp_r = F.expand_dims(
F.dropout(self[self.rnn_rev_enc[currL]](
enc_states[-1]),
ratio=self.cfg["dropout"]["rnn"]), 0)
if i > 0:
h_fwd = F.concat((h_fwd, temp_f), axis=0)
if self.bi_rnn:
h_rev = F.concat((h_rev, temp_r), axis=0)
else:
h_fwd = temp_f
if self.bi_rnn:
h_rev = temp_r
# end current rnn layer
if self.bi_rnn:
h_rev = F.flipud(h_rev)
rnn_states = F.concat((h_fwd, h_rev), axis=2)
else:
rnn_states = h_fwd
"""
Apply linear projection
"""
# print(f"Applying rnn {currL}")
if currL < (len(self.rnn_enc) - 1):
# print(f"Applying linear linear_proj {currL}")
for i in range(0, in_size):
currH = F.relu(self[f"enc_proj{currL}_bn"](
self[f"enc_proj{currL}"](rnn_states[i])))
if i > 0:
enc_states = F.concat(
(enc_states, F.expand_dims(currH, 0)), axis=0)
else:
enc_states = F.expand_dims(currH, 0)
# end for all hidden states
# end all layers
# Make the batch size as the first dimension
self.enc_states = F.swapaxes(enc_states, 0, 1)
def __call__(self, xi):
hc0 = F.leaky_relu(self.c0(xi))
hc1 = F.leaky_relu(self.bnc1(self.c1(hc0), test=not self.train))
hc2 = F.leaky_relu(self.bnc2(self.c2(hc1), test=not self.train))
hc3 = F.leaky_relu(self.bnc3(self.c3(hc2), test=not self.train))
hc4 = F.leaky_relu(self.bnc4(self.c4(hc3), test=not self.train))
hc5 = F.leaky_relu(self.bnc5(self.c5(hc4), test=not self.train))
hc6 = F.leaky_relu(self.bnc6(self.c6(hc5), test=not self.train))
hc7 = F.leaky_relu(self.bnc7(self.c7(hc6), test=not self.train))
hc8 = F.leaky_relu(self.bnc8(self.c8(hc7), test=not self.train))
h = F.expand_dims(hc8,2)
h = F.relu(F.dropout(self.bndc00(self.dc00(h), test=not self.train), 0.5, train=self.train_dropout))
hc7 = F.expand_dims(hc7,2)
hc7 = F.broadcast_to(hc7, hc7.data.shape[:2]+(h.data.shape[2],)+hc7.data.shape[3:])
h = F.concat((h,hc7),1)
h = F.relu(F.dropout(self.bndc0(self.dc0(h), test=not self.train), 0.5, train=self.train_dropout))
hc6 = F.expand_dims(hc6,2)
hc6 = F.broadcast_to(hc6, hc6.data.shape[:2]+(h.data.shape[2],)+hc6.data.shape[3:])
h = F.concat((h,hc6),1)
h = F.relu(F.dropout(self.bndc1(self.dc1(h), test=not self.train), 0.5, train=self.train_dropout))
hc5 = F.expand_dims(hc5,2)
hc5 = F.broadcast_to(hc5, hc5.data.shape[:2]+(h.data.shape[2],)+hc5.data.shape[3:])
h = F.concat((h,hc5),1)
h = F.relu(self.bndc2(self.dc2(h), test=not self.train))
hc4 = F.expand_dims(hc4,2)
hc4 = F.broadcast_to(hc4, hc4.data.shape[:2]+(h.data.shape[2],)+hc4.data.shape[3:])
h = F.concat((h,hc4),1)
h = F.relu(self.bndc3(self.dc3(h), test=not self.train))
hc3 = F.expand_dims(hc3,2)
hc3 = F.broadcast_to(hc3, hc3.data.shape[:2]+(h.data.shape[2],)+hc3.data.shape[3:])
h = F.concat((h,hc3),1)
h = F.relu(self.bndc4(self.dc4(h), test=not self.train))
hc2 = F.expand_dims(hc2,2)
hc2 = F.broadcast_to(hc2, hc2.data.shape[:2]+(h.data.shape[2],)+hc2.data.shape[3:])
h = F.concat((h,hc2),1)
h = F.relu(self.bndc5(self.dc5(h), test=not self.train))
hc1 = F.expand_dims(hc1,2)
hc1 = F.broadcast_to(hc1, hc1.data.shape[:2]+(h.data.shape[2],)+hc1.data.shape[3:])
h = F.concat((h,hc1),1)
h = F.relu(self.bndc6(self.dc6(h), test=not self.train))
hc0 = F.expand_dims(hc0,2)
hc0 = F.broadcast_to(hc0, hc0.data.shape[:2]+(h.data.shape[2],)+hc0.data.shape[3:])
h = F.concat((h,hc0),1)
h = self.dc7(h)
xi_ = F.expand_dims(xi,2)
xi_ = F.broadcast_to(xi_, h.data.shape)
h = F.sigmoid(h+xi_)
return h
def forward_one_step(self, X, ht_enc, H_enc, skip_mask):
pad = self._kernel_size - 1
WX = self.W(X)[:, :, -pad - 1, None]
Vh = self.V(ht_enc)
Vh, WX = functions.broadcast(functions.expand_dims(Vh, axis=2), WX)
# f-pooling
Z, F, O = functions.split_axis(WX + Vh, 3, axis=1)
Z = functions.tanh(Z)
F = self.zoneout(F)
O = functions.sigmoid(O)
T = Z.shape[2]
# compute ungated hidden states
for t in xrange(T):
z = Z[..., t]
f = F[..., t]
if self.contexts is None:
ct = (1 - f) * z
self.contexts = [ct]
else:
ct = f * self.contexts[-1] + (1 - f) * z
self.contexts.append(ct)
if skip_mask is not None:
assert skip_mask.shape[1] == H_enc.shape[2]
softmax_bias = (skip_mask == 0) * -1e6
# compute attention weights (eq.8)
H_enc = functions.swapaxes(H_enc, 1, 2)
for t in xrange(T):
ct = self.contexts[t - T]
bias = 0 if skip_mask is None else softmax_bias[
..., None] # to skip PAD
mask = 1 if skip_mask is None else skip_mask[...,
None] # to skip PAD
alpha = functions.batch_matmul(H_enc, ct) + bias
alpha = functions.softmax(alpha) * mask
alpha = functions.broadcast_to(alpha, H_enc.shape) # copy
kt = functions.sum(alpha * H_enc, axis=1)
ot = O[..., t]
self.ht = ot * self.o(functions.concat((kt, ct), axis=1))
if self.H is None:
self.H = functions.expand_dims(self.ht, 2)
else:
self.H = functions.concat(
(self.H, functions.expand_dims(self.ht, 2)), axis=2)
return self.H
def __call__(self, x, y):
"""
Parameters
-----------------
x: Variable
Feature of unlabeled samples.
y: Variable
Feature of unlabeled samples.
"""
g, x, y = F.broadcast(*[self.gamma, x, y])
x_g = x * g
y_g = y * g
x_g_norm = F.sum(x_g**2, axis=1)
y_g_norm = F.sum(y_g**2, axis=1)
x_g_y_g = F.linear(x_g, y_g)
x_g_norm, x_g_y_g, y_g_norm = \
F.broadcast(
*[x_g_norm,
x_g_y_g,
F.expand_dims(y_g_norm, 1)])
#F.exp(- (x_g_norm - 2 * x_g_y_g+ y_g_norm))
return F.exp(- x_g_norm + 2 * x_g_y_g - y_g_norm)
def __call__(self, x, y):
"""
Parameters
-----------------
x: Variable
Feature of unlabeled samples.
y: Variable
Feature of unlabeled samples.
"""
g, x, y = F.broadcast(*[self.gamma, x, y])
x_g = x * g
y_g = y * g
x_g_norm = F.sum(x_g**2, axis=1)
y_g_norm = F.sum(y_g**2, axis=1)
x_g_y_g = F.linear(x_g, y_g)
x_g_norm, x_g_y_g, y_g_norm = \
F.broadcast(
*[x_g_norm,
x_g_y_g,
F.expand_dims(y_g_norm, 1)])
#F.exp(- (x_g_norm - 2 * x_g_y_g+ y_g_norm))
u = x_g_norm - 2 * x_g_y_g+ y_g_norm
print(np.min(u.data))
print(len((np.where(u.data < 0)[0])), np.prod(u.data.shape))
time.sleep(0.5)
return F.exp(- x_g_norm + 2 * x_g_y_g - y_g_norm)
def create_encoder_states_matrix(self, hs):
batch_size, dim = hs[0].data.shape
hs_3d = list(map(lambda h: F.expand_dims(h, 1), hs)) # [(batch_size, 1, dim)]
hs_3d_concat = F.concat(hs_3d, axis=1) # (batch_size, input_length, dim)
hs_3d_concat_linear = self.decoder.phi2_linear(F.reshape(hs_3d_concat, (-1, dim))) # (batch_size * input_length, dim)
hs_3d_concat_linear_tanh = F.tanh(F.reshape(hs_3d_concat_linear, (batch_size, -1, dim))) # (batch_size, input_length, dim)
return hs_3d_concat_linear_tanh
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.expand_dims(x, self.axis)
self.assertEqual(y.data.shape, self.out_shape)
y_expect = numpy.expand_dims(cuda.to_cpu(x_data), self.axis)
self.assertEqual(y.data.dtype, self.dtype)
numpy.testing.assert_array_equal(cuda.to_cpu(y.data), y_expect)
def __call__(self, x, y):
"""
Parameters
-----------------
x: Variable
Feature of unlabeled samples.
y: Variable
Feature of unlabeled samples.
"""
g = F.broadcast_to(
F.gaussian(
np.array([0], dtype=np.float32),
np.array([np.exp(1)], dtype=np.float32)), x.shape)
x_g = x * g
y_g = y * g
x_g_norm = F.sum(x_g**2, axis=1)
y_g_norm = F.sum(y_g**2, axis=1)
x_g_y_g = F.linear(x_g, y_g)
x_g_norm, x_g_y_g, y_g_norm = \
F.broadcast(
*[x_g_norm,
x_g_y_g,
F.expand_dims(y_g_norm, 1)])
#F.exp(- (x_g_norm - 2 * x_g_y_g+ y_g_norm))
return F.exp(- x_g_norm + 2 * x_g_y_g - y_g_norm)
def ordinal_loss(y, mask):
xp = cuda.get_array_module(y.data)
volatile = y.volatile
b, c, n = y.data.shape
max_y = F.broadcast_to(F.max(y, axis=1, keepdims=True), y.data.shape)
y = y - max_y
sum_y = F.broadcast_to(F.expand_dims(F.sum(y, axis=1), 1), y.data.shape)
down_tri = np.tri(c, dtype=np.float32)
up_tri = down_tri.T
w1 = Variable(xp.asarray(down_tri.reshape(c, c, 1, 1)), volatile=volatile)
w2 = Variable(xp.asarray(up_tri.reshape(c, c, 1, 1)), volatile=volatile)
h = F.exp(F.expand_dims(y, -1))
h1 = F.convolution_2d(h, w1)
h1 = F.convolution_2d(F.log(h1), w1)
h2 = F.convolution_2d(h, w2)
h2 = F.convolution_2d(F.log(h2), w2)
h = F.reshape(h1 + h2, (b, c, n))
return F.sum((h - sum_y - y) * mask) / b
def __call__(self, embeded_x, m_prev, h_prev, x):
batch_size = embeded_x.shape[0]
lstm_in = self.W(embeded_x) + self.U(h_prev)
m_tmp, h_tmp = F.lstm(m_prev, lstm_in)
# flags if feeding previous output
feed_prev = F.broadcast_to(F.expand_dims(x.data != IGNORE_LABEL, -1),
(batch_size, self.hidden_size))
m = F.where(feed_prev, m_tmp, m_prev)
h = F.where(feed_prev, h_tmp, h_prev)
return m, h
def check_backward(self, x_data, y_grad):
x = chainer.Variable(x_data)
y = functions.expand_dims(x, self.axis)
y.grad = y_grad
y.backward()
func = y.creator
f = lambda: func.forward((x_data,))
gx, = gradient_check.numerical_grad(f, (x_data,), (y_grad,))
gradient_check.assert_allclose(cuda.to_cpu(x.grad),
cuda.to_cpu(gx))
def __call__(self, x1,x2):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
batch_size = x.data.shape[0]
#print batch_size
batch_W = F.concat([F.expand_dims(self.W,0)] * batch_size,0)
#print batch_W.data.shape
return F.reshape(F.batch_matmul(x, batch_W),x.data.shape[:-1])
def __call__(self, h, h_gen=None, test=False):
# Concat
if h_gen is None:
h = h
elif h_gen is not None:
# Restrict Decoder with input image
h_stacked = ()
for i in range(h_gen.shape[1]):
if np.random.randint(2) == 0:
h_stacked += (F.expand_dims(h[:, i, :, :], axis=1), )
else:
h_stacked += (F.expand_dims(h_gen[:, i, :, :], axis=1), )
h = F.concat(h_stacked)
h = self.deconv0(h) # 7x7 -> 14x14
h = self.bn0(h, test)
h = self.act(h)
h = self.deconv1(h) # 14x14 -> 28x28
h = F.tanh(h)
return h
def setUp(self):
self.x1 = numpy.random.uniform(
.5, 1, (batch_size, m, k)).astype(numpy.float32)
self.x2 = numpy.random.uniform(
.5, 1, (k, n)).astype(numpy.float32)
self.gy = numpy.random.uniform(
-1, 1, (batch_size, m, n)).astype(numpy.float32)
self.op = lambda x, y: F.batch_matmul(
x, F.broadcast_to(F.expand_dims(y, 0), (batch_size, k, n)))
self.forward_answer = numpy.array([
numpy.dot(self.x1[i], self.x2)
for i in six.moves.range(batch_size)])
def batch_rodrigues(theta):
"""
Theta is N x 3
"""
batch_size = theta.shape[0]
xp = theta.xp
angle = F.expand_dims(F.sqrt(F.batch_l2_norm_squared(theta + 1e-8)), -1)
r = F.expand_dims(theta / F.tile(angle, 3), -1)
angle = F.expand_dims(angle, -1)
cos = F.cos(angle)
sin = F.sin(angle)
cos = F.tile(cos, (3, 3))
sin = F.tile(sin, (3, 3))
outer = F.matmul(r, r, transb=True)
eyes = F.tile(F.expand_dims(
Variable(xp.array(xp.eye(3), 'f')), 0), (batch_size, 1, 1))
R = cos * eyes + (1 - cos) * outer + sin * batch_skew(r, batch_size)
return R
def __call__(self, x, y):
"""
Parameters
-----------------
x: Variable
Feature of unlabeled samples.
y: Variable
Feature of unlabeled samples.
"""
g = self.gamma ** 2
z = F.expand_dims((x - y) ** 2, axis=0)
o = F.exp(- F.linear(z, g))
return o
def __call__(self, y, m_prev, s_prev, h_forward, h_backword, enable, disable_value):
# m is memory cell of lstm, s is previous hidden output
# calculate attention
c = self._attention(h_forward, h_backword, s_prev, enable, disable_value)
# decode once
embeded_y = self.E(y)
batch_size = y.shape[0]
lstm_in = self.W(embeded_y) + self.U(s_prev) + self.C(c)
m_tmp, s_tmp = F.lstm(m_prev, lstm_in)
feed_prev = F.broadcast_to(F.expand_dims(y.data != IGNORE_LABEL, -1),
(batch_size, self.hidden_size))
m = F.where(feed_prev, m_tmp, m_prev)
s = F.where(feed_prev, s_tmp, s_prev)
t = self.U_o(s) + self.V_o(embeded_y) + self.C_o(c)
return self.W_o(t), m, s
def __call__(self, x_u_0, x_u_1):
"""
Parameters
-----------------
x_u_0: Variable
Feature of unlabeled samples.
x_u_1: Variable
Feature of unlabeled samples.
"""
ffnn_u_0 = self.layers["ffnn_u_0"]
ffnn_u_1 = self.layers["ffnn_u_1"]
f_0 = F.softmax(ffnn_u_0(x_u_0))
f_1 = F.softmax(ffnn_u_1(x_u_1))
mid_outputs_0 = ffnn_u_0.mid_outputs
mid_outputs_1 = ffnn_u_1.mid_outputs
L = len(self.dims[1:])
similarities = self.similarities.values()
# Efficient computation
## sample similarity W^l summed over l
W = 0
for l in range(L):
W += similarities[l](mid_outputs_0[l], mid_outputs_1[l])
## class similarity
f_0_norm = F.sum(f_0**2, axis=1)
f_1_norm = F.sum(f_1**2, axis=1)
f_0_f_1 = F.linear(f_0, f_1)
f_0_norm, f_0_f_1, f_1_norm= \
F.broadcast(
*[f_0_norm,
f_0_f_1,
F.expand_dims(f_1_norm, 1)])
F_ = f_0_norm - 2 * f_0_f_1 + f_1_norm
print(np.max(F_.data))
print(np.min(F_.data))
print(len((np.where(F_.data < 0)[0])), np.prod(F_.data.shape))
loss = F.sum(W * F_) / (self.batch_size * 2)
self.loss = loss
return loss
def _context(self, p, fb_mat, fbe_mat):
batch_size, source_length, _ = fb_mat.data.shape
# {pe,e}_mat: shape = [batch * srclen, atten]
pe_mat = F.reshape(
F.broadcast_to(
F.expand_dims(self.p_e(p), 1),
[batch_size, source_length, self.atten_size]),
[batch_size * source_length, self.atten_size])
e_mat = F.tanh(fbe_mat + pe_mat)
# a_mat: shape = [batch, srclen]
a_mat = F.softmax(F.reshape(self.e_a(e_mat), [batch_size, source_length]))
# q: shape = [batch, 2 * hidden]
q = F.reshape(
F.batch_matmul(a_mat, fb_mat, transa=True),
[batch_size, 2 * self.hidden_size])
return q
def _attention(self, h_forward, h_backword, s, enable, disable_value):
batch_size = s.shape[0]
sentence_size = len(h_forward)
hidden_size = self.hidden_size
xp = self.xp
weighted_s = F.broadcast_to(F.expand_dims(self.W_a(s), axis=1),
(batch_size, sentence_size, hidden_size))
h = F.concat((F.concat(h_forward, axis=0), F.concat(h_backword, axis=0)))
weighted_h = F.reshape(self.U_a(h), (batch_size, sentence_size, hidden_size))
e = self.v_a(F.reshape(F.tanh(weighted_s + weighted_h),
(batch_size * sentence_size, hidden_size)))
e = F.where(enable, F.reshape(e, (batch_size, sentence_size)), disable_value)
alpha = F.softmax(e)
c = F.batch_matmul(F.reshape(h, (batch_size, 2 * hidden_size, sentence_size)), alpha)
return F.reshape(c, (batch_size, 2 * hidden_size))
def proportions(self, doc_ids, softmax=False):
""" Given an array of document indices, return a vector
for each document of just the unnormalized topic weights.
Returns:
doc_weights : chainer.Variable
Two dimensional topic weights of each document.
"""
w = self.weights(doc_ids)
if softmax:
size = w.data.shape
mask = self.xp.random.random_integers(0, 1, size=size)
y = (F.softmax(w * self.temperature) *
Variable(mask.astype('float32')))
norm, y = F.broadcast(F.expand_dims(F.sum(y, axis=1), 1), y)
return y / (norm + 1e-7)
else:
return w
def __call__(self, x): xp = chainer.cuda.get_array_module(x.data) batchsize = x.shape[0] if self.train_weights == False and self.initial_T is not None: self.T.W.data = self.initial_T M = F.reshape(self.T(x), (-1, self.num_kernels, self.ndim_kernel)) M = F.expand_dims(M, 3) M_T = F.transpose(M, (3, 1, 2, 0)) M, M_T = F.broadcast(M, M_T) norm = F.sum(abs(M - M_T), axis=2) eraser = F.broadcast_to(xp.eye(batchsize, dtype=x.dtype).reshape((batchsize, 1, batchsize)), norm.shape) c_b = F.exp(-(norm + 1e6 * eraser)) o_b = F.sum(c_b, axis=2) if self.train_weights == False: self.initial_T = self.T.W.data return F.concat((x, o_b), axis=1)
def _calc_rpn_loss_bbox(self, rpn_bbox_pred, bbox_reg_targets, inds_inside):
# rpn_bbox_pred has the shape of (1, 4 x n_anchors, feat_h, feat_w)
n_anchors = self.proposal_layer._num_anchors
# Reshape it into (4, A, K)
rpn_bbox_pred = rpn_bbox_pred.reshape(4, n_anchors, -1)
# Transpose it into (K, A, 4)
rpn_bbox_pred = rpn_bbox_pred.transpose(2, 1, 0)
# Reshape it into (K x A, 4)
rpn_bbox_pred = rpn_bbox_pred.reshape(-1, 4)
# Keep the number of bbox
n_bbox = rpn_bbox_pred.shape[0]
# Select bbox and ravel it
rpn_bbox_pred = F.flatten(rpn_bbox_pred[inds_inside])
# Create batch dimension
rpn_bbox_pred = F.expand_dims(rpn_bbox_pred, 0)
# Ravel the targets and create batch dimension
bbox_reg_targets = bbox_reg_targets.ravel()[None, :]
# Calc Smooth L1 Loss (When delta=1, huber loss is SmoothL1Loss)
rpn_loss_bbox = F.huber_loss(rpn_bbox_pred, bbox_reg_targets,
self._delta)
rpn_loss_bbox /= n_bbox
return rpn_loss_bbox.reshape(())
def __call__(self, beta, theta, get_skin=False, with_a=False):
batch_size = beta.shape[0]
# 1. Add shape blend shapes
# (N x 10) x (10 x 6890*3) = N x 6890 x 3
self.beta_shapedirs = F.matmul(beta, self.shapedirs)
v_shaped = F.reshape(
F.matmul(beta, self.shapedirs),
[-1, self.size[0], self.size[1]]) + \
F.repeat(self.v_template[None, ], batch_size, axis=0)
self.v_shaped = v_shaped
# 2. Infer shape-dependent joint locations.
Jx = F.matmul(v_shaped[:, :, 0], self.J_regressor)
Jy = F.matmul(v_shaped[:, :, 1], self.J_regressor)
Jz = F.matmul(v_shaped[:, :, 2], self.J_regressor)
J = F.stack([Jx, Jy, Jz], axis=2)
self.J = J
# 3. Add pose blend shapes
# N x 24 x 3 x 3
Rs = F.reshape(
batch_rodrigues(F.reshape(theta, [-1, 3])), [-1, 24, 3, 3])
self.Rs = Rs
# Ignore global rotation.
pose_feature = F.reshape(Rs[:, 1:, :, :] -
F.repeat(F.repeat(Variable(self.xp.array(self.xp.eye(3), 'f'))[
None, ], 23, axis=0)[None, ], batch_size, axis=0),
[-1, 207])
self.pose_feature = pose_feature
# (N x 207) x (207, 20670) -> N x 6890 x 3
v_posed = F.reshape(
F.matmul(pose_feature, self.posedirs),
[-1, self.size[0], self.size[1]]) + v_shaped
# 4. Get the global joint location
self.J_transformed, A = batch_global_rigid_transformation(
Rs, J, self.parents)
# 5. Do skinning:
# W is N x 6890 x 24
W = F.reshape(
F.tile(self.weights, (batch_size, 1)), [batch_size, -1, 24])
# (N x 6890 x 24) x (N x 24 x 16)
T = F.reshape(
F.matmul(W, F.reshape(A, [batch_size, 24, 16])),
[batch_size, -1, 4, 4])
v_posed_homo = F.concat(
[v_posed, self.xp.ones([batch_size, v_posed.shape[1], 1], 'f')], 2)
v_homo = F.matmul(T, F.expand_dims(v_posed_homo, -1))
verts = v_homo[:, :, :3, 0]
# Get cocoplus or lsp joints:
joint_x = F.matmul(verts[:, :, 0], self.joint_regressor)
joint_y = F.matmul(verts[:, :, 1], self.joint_regressor)
joint_z = F.matmul(verts[:, :, 2], self.joint_regressor)
joints = F.stack([joint_x, joint_y, joint_z], axis=2)
return verts, joints, Rs, A
def logsoftmax_no_mask(x, mask, zero_pad, axis):
x_logsumexp = logsumexp(x, mask, zero_pad, axis)
return x - F.broadcast_to(F.expand_dims(x_logsumexp, 1), x.shape)
def test_invalid_dim(self):
x = chainer.Variable(self.x)
with self.assertRaises(chainer.utils.type_check.InvalidType):
functions.expand_dims(x, self.x.ndim + 1)
with self.assertRaises(chainer.utils.type_check.InvalidType):
functions.expand_dims(x, -self.x.ndim - 2)
def __call__(self, h, h_new, g, step=0):
"""
Describes the module for a single layer update.
Do not forget to rest GRU for each batch...
:param h: minibatch by num_nodes by hidden_dim numpy array.
current local node hidden states as input of the vanilla GNN
:param h_new: minibatch by num_nodes by hidden_dim numpy array.
updated local node hidden states as output from the vanilla GNN
:param adj: minibatch by bond_types by num_nodes by num_nodes 1/0 array.
Adjacency matrices over several bond types
:param g: minibatch by hidden_dim_super numpy array.
current super node hiddden state
:param step: integer, the layer index
:return: updated h and g
"""
xp = self.xp
# (minibatch, atom, ch)
mb, atom, ch = h.shape
out_ch = ch
#
# Transmitter unit: inter-module message passing
#
# non linear update of the super node
g_new = functions.relu(self.F_super[step](g))
# original --> super transmission
h1 = functions.expand_dims(h, 2)
#assert h1.shape == (mb, atom, 1, ch)
h1 = functions.broadcast_to(h1, [mb, atom, self.n_heads, ch])
h1 = functions.reshape(h1, [mb, atom, self.n_heads* ch])
#assert h1.shape==(mb, atom, self.n_heads * ch)
h_j = functions.expand_dims(h, 1)
h_j = functions.broadcast_to(h_j, (mb, self.n_heads, atom, ch))
#assert h_j.shape==(mb, self.n_heads, atom, ch)
# expand h_super
g_extend = functions.expand_dims(g, 1)
# assert g_extend.shape==(mb, 1, self.hidden_dim_super)
g_extend = functions.broadcast_to(g_extend, (mb, self.n_heads, self.hidden_dim_super))
# assert g_extend.shape==(mb, self.n_heads, self.hidden_dim_super)
g_extend = functions.expand_dims(g_extend, 2)
# assert g_extend.shape==(mb, self.n_heads, 1, self.hidden_dim_super)
# update for attention-message B h_i
# mb, atom, n_heads * ch
Bh_i = self.B[step](h1)
# assert Bh_i.shape==(mb, atom, self.n_heads * self.hidden_dim_super)
# mb, atom, num_head, ch
Bh_i = functions.reshape(Bh_i, [mb, atom, self.n_heads, self.hidden_dim_super])
# mb, num_head, atom, ch
Bh_i = functions.transpose(Bh_i, [0, 2, 1, 3])
# assert Bh_i.shape==(mb, self.n_heads, atom, self.hidden_dim_super)
# take g^{T} * B * h_i
# indexed by i
# mb, self.n_haeds atom(i)
b_hi = functions.matmul(g_extend, Bh_i, transb=True) # This will reduce the last hidden_dim_super axis
# assert b_hi.shape==(mb, self.n_heads, 1, atom)
# softmax. sum/normalize over the last axis.
# mb, self.n_heda, atom(i-normzlied)
attention_i = functions.softmax(b_hi, axis=3)
if self.dropout_ratio > 0.0:
attention_i = functions.dropout(attention_i,
ratio=self.dropout_ratio)
# assert attention_i.shape==(mb, self.n_heads, 1, atom)
# element-wise product --> sum over i
# mb, num_head, hidden_dim_super
attention_sum = functions.matmul(attention_i, h_j)
# assert attention_sum.shape==(mb, self.n_heads, 1, ch)
attention_sum = functions.reshape(attention_sum, (mb, self.n_heads * ch))
# assert attention_sum.shape==(mb, self.n_heads * ch)
# weighting h for different heads
h_trans = self.V_super[step](attention_sum)
# assert intermediate_h.shape==(mb, self.n_heads * ch)
# compress heads
h_trans = self.W_super[step](h_trans)
h_trans = functions.tanh(h_trans)
# assert intermediate_h.shape==(mb, self.hidden_dim_super)
# g_trans: super --> original transmission
# for local updates
g_trans = self.F_super[step](g)
g_trans = functions.tanh(g_trans)
# assert intermediate_h_super.shape==(mb, self.hidden_dim)
g_trans = functions.expand_dims(g_trans, 1)
# assert intermediate_h_super.shape==(mb, 1, self.hidden_dim)
g_trans = functions.broadcast_to(g_trans, (mb, atom, self.hidden_dim))
# assert intermediate_h_super.shape==(mb, atom, self.hidden_dim)
#
# Warp Gate unit
#
z_local = self.H_local[step](h_new) + self.G_local[step](g_trans)
z_local = functions.broadcast_to(z_local, (mb, atom, self.hidden_dim))
if self.dropout_ratio > 0.0:
z_local = functions.dropout(z_local,ratio=self.dropout_ratio)
z_local = functions.sigmoid(z_local)
merged_h = (1.0-z_local) * h_new + z_local * g_trans
# assert new_h.shape==(mb, atom, ch)
z_super = self.H_super[step](h_trans) + self.G_super[step](g_new)
z_super = functions.broadcast_to(z_super, (mb, self.hidden_dim_super))
if self.dropout_ratio > 0.0:
z_super = functions.dropout(z_super,ratio=self.dropout_ratio)
z_super = functions.sigmoid(z_super)
merged_g = (1.0-z_super) * h_trans + z_super * g_new
# assert out_h_super.shape==(mb, self.hidden_dim_super)
#
# Self recurrent
#
out_h = functions.reshape(merged_h, (mb * atom, self.hidden_dim))
out_h = self.GRU_local(out_h)
out_h = functions.reshape(out_h, (mb, atom, self.hidden_dim))
out_g = self.GRU_super(merged_g)
return out_h, out_g
def check_backward(self, x_data):
x = chainer.Variable(x_data)
y = functions.expand_dims(x, self.axis)
y.grad = y.data
y.backward()
testing.assert_allclose(x.data, x.grad, atol=0, rtol=0)