def original(self, enc_hs, dec_z, att_prev, scaling=2.0):
'''AttLoc forward
:param enc_hs:
:param dec_z:
:param att_prev:
:param scaling:
:return:
'''
batch = len(enc_hs)
# pre-compute all h outside the decoder loop
if self.pre_compute_enc_h is None:
self.enc_h = F.pad_sequence(enc_hs) # utt x frame x hdim
self.h_length = self.enc_h.shape[1]
# utt x frame x att_dim
self.pre_compute_enc_h = linear_tensor(self.mlp_enc, self.enc_h)
if dec_z is None:
dec_z = chainer.Variable(self.xp.zeros(
(batch, self.dunits), dtype=np.float32))
else:
dec_z = F.reshape(dec_z, (batch, self.dunits))
# initialize attention weight with uniform dist.
if att_prev is None:
att_prev = [self.xp.full(
hh.shape[0], 1.0 / hh.shape[0], dtype=np.float32) for hh in enc_hs]
att_prev = [chainer.Variable(att) for att in att_prev]
att_prev = F.pad_sequence(att_prev)
# TODO(watanabe) use <chainer variable>.reshpae(), instead of F.reshape()
# att_prev: utt x frame -> utt x 1 x 1 x frame -> utt x att_conv_chans x 1 x frame
att_conv = self.loc_conv(
F.reshape(att_prev, (batch, 1, 1, self.h_length)))
# att_conv: utt x att_conv_chans x 1 x frame -> utt x frame x att_conv_chans
att_conv = F.swapaxes(F.squeeze(att_conv, axis=2), 1, 2)
# att_conv: utt x frame x att_conv_chans -> utt x frame x att_dim
att_conv = linear_tensor(self.mlp_att, att_conv)
# dec_z_tiled: utt x frame x att_dim
dec_z_tiled = F.broadcast_to(
F.expand_dims(self.mlp_dec(dec_z), 1), self.pre_compute_enc_h.shape)
# dot with gvec
# utt x frame x att_dim -> utt x frame
# TODO(watanabe) use batch_matmul
e = F.squeeze(linear_tensor(self.gvec, F.tanh(
att_conv + self.pre_compute_enc_h + dec_z_tiled)), axis=2)
# Applying a minus-large-number filter to make a probability value zero for a padded area
# simply degrades the performance, and I gave up this implementation
# Apply a scaling to make an attention sharp
w = F.softmax(scaling * e)
# weighted sum over flames
# utt x hdim
c = F.sum(self.enc_h * F.broadcast_to(F.expand_dims(w, 2), self.enc_h.shape), axis=1)
return c, w
def loss_Critic(self, dis_c, alpha, dis_y1, dis_y2):
xp = chainer.backend.get_array_module(dis_c.data)
batchsize = len(dis_c)
loss = F.sum((F.squeeze(dis_c)-F.squeeze(alpha))**2) / batchsize
loss_ = F.sum(dis_y1**2) / batchsize
loss_ += F.sum(dis_y2**2) / batchsize
loss += loss_/2
chainer.report({'Critic_loss':loss})
return loss
def cos_sim(x, y):
# batchsize = 1のときsqueezeでエラー
if len(x.shape) > 2:
norm_x = F.normalize(F.squeeze(F.squeeze(x,axis=(2,)),axis=(2,)))
norm_y = F.normalize(F.squeeze(F.squeeze(y,axis=(2,)),axis=(2,)))
else:
norm_x = F.normalize(x)
norm_y = F.normalize(y)
return F.batch_matmul(norm_x, norm_y, transa=True)
def __call__(self, x_block, y_in_block, y_out_block):
batch = len(x_block)
#embed
ex_block = F.dropout(self.make_input_embedding(self.embed_x, x_block),
self.dropout)
ey_block = F.dropout(
self.make_input_embedding(self.embed_y, y_in_block), self.dropout)
eyy_block = F.dropout(
self.make_input_embedding(self.embed_yy, y_in_block), self.dropout)
eys = F.transpose(ey_block, (0, 2, 1))
eyys = F.transpose(eyy_block, (0, 2, 1))
#gcnn
h = F.expand_dims(ex_block, axis=1)
for i in range(self.stack):
h = self.gcnn[i](h)
h = F.dropout(F.squeeze(h, axis=1), self.dropout)
#Nsteolstm
eys2 = [i for i in eys]
eyys2 = [i for i in eyys]
_, _, oss = self.decoder(None, None, eys2)
_, _, oss2 = self.decoder2(None, None, eyys2)
ss = F.stack(oss, axis=0)
ss2 = F.stack(oss2, axis=0)
#mask_make
mask = (y_in_block[:, :, None] >= 0) * self.xp.ones(
(self.batch, 1, self.n_units), dtype=bool)
ss = F.where(mask, ss, self.xp.full(ss.shape, 0, 'f'))
#weight_calclate
batch_A = F.batch_matmul(ss, h) * self.scale_score
mask = (x_block[:, 0:len(x_block[0]) - self.stack *
(self.width - 1)][:, None, :] >=
0) * (y_in_block[:, :, None] >= 0)
batch_A = F.where(mask, batch_A,
self.xp.full(batch_A.shape, -self.xp.inf, 'f'))
batch_A = F.softmax(batch_A, axis=2)
batch_A = F.where(self.xp.isnan(batch_A.data),
self.xp.zeros(batch_A.shape, 'f'), batch_A)
batch_A, h = F.broadcast(batch_A[:, None], h[:, :, None])
batch_C = F.sum(batch_A * h, axis=3)
e = F.transpose(batch_C, (0, 2, 1))
e = F.squeeze(F.concat(F.split_axis(e, self.batch, axis=0), axis=1))
ss2 = F.squeeze(F.concat(F.split_axis(ss2, self.batch, axis=0),
axis=1))
t = (self.We(e) + self.Ws(ss2))
t = F.dropout(t, self.dropout)
concat_ys_out = F.concat(y_out_block, axis=0)
loss = F.sum(F.softmax_cross_entropy(t, concat_ys_out,
reduce='no')) / batch
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 _sample_state(self, transision, s_shape=(32, 7), z_shape=(32, 4)):
s_current = self._Uniform.sample(sample_shape=s_shape)
s_current = F.squeeze(s_current)
s_next, _ = transision(s_current)
z = self._Normal.sample(sample_shape=z_shape)
z = F.squeeze(z)
assert s_shape == s_current.shape
assert s_shape == s_next.shape
assert z_shape == z.shape
return s_current, s_next, z
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 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 compute_kl_after_update(loss_func, n=100):
policy = copy.deepcopy(base_policy)
optimizer = chainer.optimizers.SGD(1e-4)
optimizer.setup(policy)
for _ in range(n):
distrib = policy(x)
policy.cleargrads()
F.squeeze(loss_func(distrib)).backward()
optimizer.update()
distrib_after = policy(x)
return float(another_distrib.kl(distrib_after).array)
def update_Q():
# Predicted values: Q(s,a)
y = F.squeeze(Q(obs, action), axis=1)
# Target values: r + gamma * Q(s,policy(s))
with chainer.no_backprop_mode():
next_q = F.squeeze(target_Q(obs_next, target_policy(obs_next)),
axis=1)
target = reward + gamma * (1 - done) * next_q
loss = F.mean_squared_error(y, target)
Q.cleargrads()
loss.backward()
opt_Q.update()
def update_Q():
# Predicted values: Q(s,a)
y = F.squeeze(Q(obs, action), axis=1)
# Target values: r + gamma * Q(s,policy(s))
with chainer.no_backprop_mode():
next_q = F.squeeze(target_Q(obs_next, target_policy(obs_next)),
axis=1)
target = reward + gamma * (1 - done) * next_q
loss = F.mean_squared_error(y, target)
Q.cleargrads()
loss.backward()
opt_Q.update()
def bdot(x, y):
""" batch direct product
Not to be confused with Exterior product.
:param x: batch times n
:param y: batch times n
:return: batch dimension
"""
assert x.shape[0] == y.shape[0]
assert x.shape[1] == y.shape[1]
xT = F.expand_dims(x, 1)
y = F.expand_dims(y, 2)
res = F.squeeze(F.squeeze(xT @ y, axis=1), axis=1)
return res
def update_core(self):
gen_optimizer = self.get_optimizer('opt_gen')
dis_optimizer = self.get_optimizer('opt_dis')
xp = self.gen.xp
opt = self.opt
batch = self.get_iterator('main').next()
batchsize = len(batch)
x = denoise.add_noise(batch, self.opt)
x = utils.prepare_data_for_cnn(x, opt.maxlen, opt.filter_shape)
x_org = utils.prepare_data_for_rnn(batch,
opt.maxlen,
opt.sent_len,
opt.n_words,
is_add_GO=True)
x = xp.array(x, dtype=np.int32)
x_org = xp.array(x_org, dtype=np.int32)
# generator
syn_sents, prob = self.gen(x, x_org) # prob: fake data
# discriminator
logits_real, H_real = self.dis(x)
logits_fake, H_fake = self.dis(prob, is_prob=True)
# one hot vector
labels_one = xp.ones((batchsize), dtype=xp.int32) # 1-dim array
labels_zero = xp.zeros((batchsize), dtype=xp.int32)
labels_fake = labels_zero #F.concat([labels_one, labels_zero], axis=1)
labels_real = labels_one #F.concat([labels_zero, labels_one], axis=1)
D_loss = F.softmax_cross_entropy(logits_real, labels_real) + \
F.softmax_cross_entropy(logits_fake, labels_fake)
G_loss = compute_MMD_loss(F.squeeze(H_fake), F.squeeze(H_real))
self.gen.cleargrads()
G_loss.backward()
gen_optimizer.update()
self.dis.cleargrads()
D_loss.backward()
dis_optimizer.update()
H_fake.unchain_backward()
H_real.unchain_backward()
prob.unchain_backward()
chainer.reporter.report({'loss_gen': G_loss})
chainer.reporter.report({'loss_dis': D_loss})
def bquad(x, Q):
""" calcuate x^T Q x
:param x: vector batch times n
:param Q: batch times n times n
:return: batch dim
"""
assert x.shape[0] == Q.shape[0], "batch mismatch" + str(x.shape) + ":" + str(Q.shape)
assert x.shape[1] == Q.shape[1], "mat mul dim mismatch"
assert Q.shape[2] == Q.shape[1], "Q is not square matrix"
xT = F.expand_dims(x, 1)
x_ = F.expand_dims(x, 2)
res = F.squeeze(F.squeeze(xT @ Q @ x_, axis=1), axis=1)
assert list(res.shape) == [list(x.shape)[0]]
return res
def bias_correction_policy_gradients(truncation_threshold):
gs = []
for sample in mu_samples:
base_policy.cleargrads()
loss = acer.compute_policy_gradient_loss(
action=sample,
advantage=evaluate_action(sample),
action_distrib=pi,
action_distrib_mu=mu,
action_value=action_value,
v=0,
truncation_threshold=truncation_threshold)
F.squeeze(loss).backward()
gs.append(extract_gradients_as_single_vector(base_policy))
return gs
def get_onehot_grad(self, xs, ys=None):
if ys is None:
with chainer.using_config('train', False):
ys = self.predict(xs, argmax=True)
u, exs_prem = self.encoder.get_grad(xs[0])
v, exs_hypo = self.encoder.get_grad(xs[1])
encodings = F.concat((u, v, F.absolute(u - v), u * v), axis=1)
outputs = self.output(self.mlp(encodings, no_dropout=True))
loss = F.softmax_cross_entropy(outputs, ys)
exs = exs_hypo
lengths = [len(x) for x in xs[1]]
if isinstance(exs, tuple):
exs_grad = chainer.grad([loss], exs)
ex_sections = np.cumsum([ex.shape[0] for ex in exs[:-1]])
exs = F.concat(exs, axis=0)
exs_grad = F.concat(exs_grad, axis=0)
onehot_grad = F.sum(exs_grad * exs, axis=1)
onehot_grad = F.split_axis(onehot_grad, ex_sections, axis=0)
else:
exs_grad = chainer.grad([loss], [exs])[0]
# (batch_size, n_dim, max_length, 1)
assert exs_grad.shape == exs.shape
onehot_grad = F.squeeze(F.sum(exs_grad * exs, 1), 2)
onehot_grad = [x[:l] for x, l in zip(onehot_grad, lengths)]
return onehot_grad
def get_onehot_grad(self, xs, ys=None):
if ys is None:
with chainer.using_config('train', False):
ys = self.predict(xs, argmax=True)
ys = F.expand_dims(ys, axis=1)
ys = [y for y in ys]
encodings, exs = self.encoder.get_grad(xs)
outputs = self.output(encodings)
concat_truths = F.concat(ys, axis=0)
loss = F.softmax_cross_entropy(outputs, concat_truths)
if isinstance(exs, tuple):
exs_grad = chainer.grad([loss], exs)
ex_sections = np.cumsum([ex.shape[0] for ex in exs[:-1]])
exs = F.concat(exs, axis=0)
exs_grad = F.concat(exs_grad, axis=0)
onehot_grad = F.sum(exs_grad * exs, axis=1)
onehot_grad = F.split_axis(onehot_grad, ex_sections, axis=0)
else:
exs_grad = chainer.grad([loss], [exs])[0]
# (batch_size, n_dim, max_length, 1)
assert exs_grad.shape == exs.shape
onehot_grad = F.squeeze(F.sum(exs_grad * exs, 1), 2)
lengths = [len(x) for x in xs]
onehot_grad = [x[:l] for x, l in zip(onehot_grad, lengths)]
return onehot_grad
def lstm_first_forward_func(self, xs):
# xs T,F,in_size
xp = chainer.cuda.cupy.get_array_module(xs[0].data)
hx = None
cx = None
xs = F.transpose(xs, axes=(1, 0, 2)) # shape = F,T,in_size
xs = [
F.squeeze(e)
for e in F.split_axis(xs, xs.shape[0], axis=0, force_tuple=True)
]
_, _, hs = self.lstm(xs) # hs is list of T x D variable
hs = F.stack(hs)
box_num, frame, _ = hs.shape
hs = F.reshape(hs, (-1, hs.shape[-1]))
hs = F.relu(self.fc2(hs))
hs = F.reshape(hs, shape=(box_num, frame, -1))
hs = F.transpose(hs, axes=(1, 0, 2)) # shape = T, F, 1024
for relation_module_str in self.relation_module_name[:len(
self.relation_module_name) // 2]:
hs = getattr(self, relation_module_str)(hs, hs,
hs) # shape = T,F, 1024
hs = F.reshape(hs, (-1, hs.shape[-1]))
hs = F.relu(self.fc3(hs))
hs = F.reshape(hs, shape=(frame, box_num, -1))
for relation_module_str in self.relation_module_name[
len(self.relation_module_name) // 2:]:
hs = getattr(self, relation_module_str)(hs, hs,
hs) # shape = T,F, 1024
hs = F.reshape(hs, (-1, hs.shape[-1]))
hs = self.fc4(hs)
hs = F.reshape(hs, shape=(frame, box_num, self.out_size))
return hs
def attend(self, encoded_features):
self.out_lstm.reset_state()
transformed_encoded_features = F.concat([
F.expand_dims(self.transform_encoded_features(feature), axis=1)
for feature in encoded_features
],
axis=1)
concat_encoded_features = F.concat(
[F.expand_dims(e, axis=1) for e in encoded_features], axis=1)
lstm_output = self.xp.zeros_like(encoded_features[0])
outputs = []
for _ in range(self.num_labels):
transformed_lstm_output = self.transform_out_lstm_feature(
lstm_output)
attended_feats = []
for transformed_encoded_feature in F.separate(
transformed_encoded_features, axis=1):
attended_feat = transformed_encoded_feature + transformed_lstm_output
attended_feat = F.tanh(attended_feat)
attended_feats.append(
self.generate_attended_feat(attended_feat))
attended_feats = F.concat(attended_feats, axis=1)
alphas = F.softmax(attended_feats, axis=1)
lstm_input_feature = F.batch_matmul(alphas,
concat_encoded_features,
transa=True)
lstm_input_feature = F.squeeze(lstm_input_feature, axis=1)
lstm_output = self.out_lstm(lstm_input_feature)
outputs.append(lstm_output)
return outputs
def gcams_to_mask(gcams_from_chainer, class_ids, dataset=None, img=None):
if len(class_ids) == 0:
return None
gcams_np = []
gcam_aggregate = None
for i in range(len(gcams_from_chainer)):
# gcam for class i
gcams_np.append(cp.asnumpy(
F.squeeze(gcams_from_chainer[i][0], 0).data))
print(class_ids)
for i in range(len(gcams_np)):
# so earlier indices will have brighter heatmaps
gcam_np = gcams_np[i]
print("Max gcam magnitude for {}: ".format(class_names[
int(class_ids[i]) + 1]) + str(np.max(gcams_np[i])))
print("Min gcam magnitude for {}: ".format(class_names[
int(class_ids[i]) + 1]) + str(np.min(gcams_np[i])))
mask = _gcam_to_mask(gcam_np, int(class_ids[i]))
assert mask != None
cv2.imshow("mask", np.uint8(mask))
cv2.waitKey(0)
if gcam_aggregate is None:
gcam_aggregate = mask
else:
gcam_aggregate = gcam_aggregate + mask
return gcam_aggregate
def __call__(self, x, z):
"""
Args:
x (~chainer.Variable): Batch of input vectors.
z (~chainer.Variable): Batch of context vectors.
Returns:
~chainer.Variable: Output of the context layer.
"""
if self.has_uninitialized_params:
with cuda.get_device(self._device_id):
self._initialize_params(x.size // x.shape[0])
batch_size = x.shape[0]
# compute adaptive filter
W = self.predictor(z)
# reshape linear W to the correct size
W = F.reshape(W, [batch_size] + self.shape)
# add constant W if defined
if self.constantW:
W += F.tile(self.C, (batch_size, 1, 1))
# multiply weights with inputs in batch mode
y = F.squeeze(F.batch_matmul(W, x), 2)
# add bias
y += F.tile(self.b, tuple([batch_size, 1]))
return y
def masked_self_attention(self, input, adj, step):
adj = np.sum(adj, axis=1)
# [mb, atoms, ch]
mb, atoms, ch = input.shape
attention_layer_index = 0 if self.attention_tying else step
# [mb, atoms, hidden_dim]
h = functions.reshape(input, shape=(mb * atoms, ch))
h = self.linear_transform_layer[attention_layer_index](h)
h = functions.reshape(h, shape=(mb, atoms, -1))
# [mb, atoms, atoms, 2 * hidden_dim]
a_input = functions.concat([functions.tile(h, reps=(1, 1, atoms)).reshape(mb, atoms * atoms, -1),
functions.tile(h, reps=(1, atoms, 1))], axis=-1).reshape(mb, atoms, atoms,
2 * self.hidden_dim)
a_input = functions.reshape(a_input, shape=(mb * atoms * atoms, 2 * self.hidden_dim))
# [mb * atoms * atoms, 2 * hidden_dim] => [mb * atoms * atoms, 1] => [mb, atoms * atoms]
e = functions.leaky_relu(
functions.reshape(functions.squeeze(self.neural_network_layer[attention_layer_index](a_input), axis=-1),
shape=(mb, atoms, atoms)))
# [mb, atoms, atoms]
zero_vec = -9e15 * self.xp.ones_like(e, dtype=self.xp.float32)
# [mb, atoms, atoms]
attention = functions.where(adj > 0, e, zero_vec)
# [mb, atoms, atoms]
attention = functions.softmax(attention, axis=2)
# [mb, atoms, atoms] * [mb, atoms, hidden_dim] => [mb, atoms, hidden_dim]
h_prime = functions.matmul(attention, h)
h_prime = functions.elu(h_prime)
return h_prime
def forward(self, ws, cs, ls, dep_ts=None):
ws = map(self.emb_word, ws)
cs = [F.squeeze(
F.max_pooling_2d(
self.conv_char(
F.expand_dims(
self.emb_char(c), 1)), (int(l[0]), 1)))
for c, l in zip(cs, ls)]
xs_f = [F.dropout(F.concat([w, c]), 0.5) for w, c in zip(ws, cs)]
xs_b = [x[::-1] for x in xs_f]
_, _, hs_f = self.lstm_f(None, None, xs_f)
_, _, hs_b = self.lstm_b(None, None, xs_b)
hs_b = [x[::-1] for x in hs_b]
hs = [F.concat([h_f, h_b]) for h_f, h_b in zip(hs_f, hs_b)]
dep_ys = [self.biaffine_arc(
F.elu(F.dropout(self.arc_dep(h), 0.32)),
F.elu(F.dropout(self.arc_head(h), 0.32))) for h in hs]
if dep_ts is not None:
heads = dep_ts
else:
heads = [F.argmax(y, axis=1) for y in dep_ys]
cat_ys = [self.biaffine_tag(
F.elu(F.dropout(self.rel_dep(h), 0.32)),
F.elu(F.dropout(self.rel_head(
F.embed_id(t, h, ignore_label=IGNORE)), 0.32)))
for h, t in zip(hs, heads)]
return cat_ys, dep_ys
def encode(self, image, obj, desc, num):
xp = cuda.cupy
cuda.get_device(GPU.gpus_to_use[num % GPU.num_gpus]).use()
obj = np.asarray(obj, dtype=np.float32)
obj = np.repeat(obj[np.newaxis], image.shape[0], axis=0)
desc = np.asarray(desc, dtype=np.float32)
desc = np.repeat(desc[np.newaxis], image.shape[0], axis=0)
o_in = cuda.to_gpu(obj, GPU.gpus_to_use[num % GPU.num_gpus])
d_in = cuda.to_gpu(desc, GPU.gpus_to_use[num % GPU.num_gpus])
x_in = cuda.to_gpu(image, GPU.gpus_to_use[num % GPU.num_gpus])
att, _, _ = self.enc_models[num % 2](Variable(x_in),
Variable(o_in),
Variable(d_in),
train=False)
att = F.reshape(att, (-1, 1, self.att_size, self.att_size))
att = F.resize_images(att, (self.image_size, self.image_size))
cir_z, _, _, _ = self.att_enc_models[num % 2](Variable(x_in) * att,
train=False)
return cir_z, F.squeeze(F.concat((o_in[0], d_in[0]), axis=-1))
def __call__(self, S, h):
batch_size, src_len, hidden_size = S.data.shape
S = self.inner_weight(F.reshape(S,
(batch_size * src_len, hidden_size)))
S = F.reshape(S, (batch_size, src_len, hidden_size))
a = F.softmax(F.squeeze(F.batch_matmul(S, h), axis=2))
return a
def forward(self, x):
batchsize = x.shape[0]
assert self.is_convdim_compatible(x),\
'kernel dim %d is not compatible to input spatial dim %d' % (self.gm.CONV_DIM, len(x.shape) - 2)
Z = F.reshape(x, self.gm.get_dims(tensor_id=0, expanded=True).tolist())
image_flags = self.gm.indices2flags(self.gm.get_image_indices())
filter_flags = self.gm.indices2flags(self.gm.get_filter_indices())
for tensor_id in range(1, self.gm.num_tensors):
logging.debug('Next processing:')
logging.debug(tensor_id)
sum_flags = self.gm.indices2flags(
self.gm.get_sum_indices(tensor_id))
Z = expanded_einconv(Z, self.get_param(tensor_id), sum_flags,
image_flags, filter_flags, self.xp)
if self.gm.is_relu(tensor_id) and tensor_id < (
self.gm.num_tensors - 1):
Z = F.relu(Z)
if self.batchnorm:
Z = self.get_bn(tensor_id)(Z)
Z = F.squeeze(Z)
for i, d in enumerate(self.gm.dims[image_flags].tolist()):
if d == 1:
Z = F.expand_dims(Z, i + 2)
if batchsize == 1:
Z = F.expand_dims(Z, 0)
return Z
def __call__(self, id, x):
W = self.W_embedding(id)
b = F.squeeze(self.b_embedding(id))
# Reshape the vector to be the right dimensions for 2D conv
W = F.reshape(W,
(self.out_channels, self.in_channels, self.kh, self.kw))
return F.convolution_2d(x, W, b, self.stride, self.pad)
def predict(self, xs):
"""
batch: list of splitted sentences
"""
xs = [self.extractor.process(x) for x in xs]
batchsize = len(xs)
ws, cs, ls = zip(*xs)
ws = map(self.emb_word, ws)
cs = [
F.squeeze(
F.max_pooling_2d(
self.conv_char(F.expand_dims(self.emb_char(c), 1)),
(l, 1))) for c, l in zip(cs, ls)
]
xs_f = [
F.dropout(F.concat([w, c]), self.dropout_ratio, train=self.train)
for w, c in zip(ws, cs)
]
xs_b = [x[::-1] for x in xs_f]
cx_f, hx_f, cx_b, hx_b = self._init_state(batchsize)
_, _, hs_f = self.lstm_f(hx_f, cx_f, xs_f, train=self.train)
_, _, hs_b = self.lstm_b(hx_b, cx_b, xs_b, train=self.train)
hs_b = [x[::-1] for x in hs_b]
ys = [
self.linear2(F.relu(self.linear1(F.concat([h_f, h_b]))))
for h_f, h_b in zip(hs_f, hs_b)
]
return [y.data[1:-1] for y in ys]
def forward(self, equery, vmemory, ememory, mask, iteration=0):
"""Compute an attention over memory given the query."""
# equery.shape == (..., E)
# vmemory.shape == (..., Ms, M)
# ememory.shape == (..., Ms, E)
# mask.shape == (..., Ms)
# Setup memory embedding
eq = F.repeat(equery[..., None, :], vmemory.shape[-2],
-2) # (..., Ms, E)
# Compute content based attention
merged = F.concat(
[eq, ememory, eq * ememory,
F.squared_difference(eq, ememory)], -1) # (..., Ms, 4*E)
inter = self.att_linear(merged, n_batch_axes=len(vmemory.shape) -
1) # (..., Ms, E)
inter = F.tanh(inter) # (..., Ms, E)
inter = F.dropout(inter, DROPOUT) # (..., Ms, E)
# Split into sentences
lengths = np.sum(np.any((vmemory != 0), -1), -1) # (...,)
mems = [s[..., :l, :] for s, l in zip(F.separate(inter, 0), lengths)
] # B x [(M1, E), (M2, E), ...]
_, bimems = self.att_birnn(None,
mems) # B x [(M1, 2*E), (M2, 2*E), ...]
bimems = F.pad_sequence(bimems) # (..., Ms, 2*E)
att = self.att_score(bimems, n_batch_axes=len(vmemory.shape) -
1) # (..., Ms, 1)
att = F.squeeze(att, -1) # (..., Ms)
if mask is not None:
att += mask * MINUS_INF # (..., Ms)
return att
def __call__(self, x, t, dataset, train=True):
# Create variables
x = Variable(x)
x.to_gpu(self.gpu_id)
t = Variable(t)
t.to_gpu(self.gpu_id)
# Config mode
if len(t.shape) == 3:
config_mode = 'segmentation'
elif len(t.shape) == 2:
config_mode = 'recognition'
else:
raise ValueError('label format is not supported')
# Forward
with chainer.using_config('train', train):
with chainer.using_config('enable_backprop', train):
# InceptionV3 backbone
x = self.predictor(x)
# Classifiers
classifier_indx = self.args.dataset.split('+').index(dataset)
y = self.classifiers[classifier_indx](x, train)
# Loss
if config_mode == 'segmentation':
self.y = F.resize_images(y,
t.shape[-2:]) # Upsampling logits
self.loss = F.softmax_cross_entropy(self.y, t)
elif config_mode == 'recognition':
self.y = F.squeeze(F.average_pooling_2d(
y, ksize=y.shape[-2:]),
axis=(2, 3)) # Global Average Pooling
self.loss = F.sigmoid_cross_entropy(self.y, t)
# Backward
if train:
# Clear grads for uninitialized params
self.cleargrads()
# Backwards
self.loss.backward()
# Reporter
if config_mode == 'segmentation':
self.y = F.argmax(self.y, axis=1)
self.y.to_cpu()
t.to_cpu()
result = eval_semantic_segmentation(list(self.y.data),
list(t.data))
del result['iou'], result['class_accuracy']
result.update({'loss': self.loss.data.tolist()})
self.reporter.update({dataset: result})
elif config_mode == 'recognition':
self.reporter.update({
dataset: {
'loss': self.loss.data.tolist(),
'prediction': F.sigmoid(self.y).data.tolist(),
'groundtruth': t.data.tolist()
}
})
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 __call__(self, x, lam):
h = GRL(lam)(x)
h = F.dropout(F.relu(self.bn1(self.fc1(h))), 0.2)
h = self.fc2(h)
return F.squeeze(h)
def query(self, u):
xp = cuda.get_array_module(u)
size = self.m.shape[1]
inds = xp.arange(size - 1, -1, -1, dtype=numpy.int32)
tm = self.TA(inds)
tc = self.TC(inds)
tm = F.broadcast_to(tm, self.m.shape)
tc = F.broadcast_to(tc, self.c.shape)
p = F.softmax(F.batch_matmul(self.m + tm, u))
o = F.batch_matmul(F.swapaxes(self.c + tc, 2, 1), p)
o = F.squeeze(o, -1)
u = o + u
return u
def forward(self, inputs):
"""
Parameters
----------
inputs: ``torch.autograd.Variable``
Shape ``(batch_size, timesteps, 50)`` of character ids representing the current batch.
Returns
-------
Dict with keys:
``'activations'``: ``List[torch.autograd.Variable]``
A list of activations at each layer of the network, each of shape
``(batch_size, timesteps + 2, embedding_dim)``
``'mask'``: ``torch.autograd.Variable``
Shape ``(batch_size, timesteps + 2)`` long tensor with sequence mask.
Note that the output tensors all include additional special begin and end of sequence
markers.
"""
token_embedding = self._token_embedder.forward(inputs)
type_representation = token_embedding['token_embedding']
mask = token_embedding['mask']
lstm_outputs = self._elmo_lstm.forward(type_representation, mask)
# Prepare the output. The first layer is duplicated.
output_tensors = [
F.concat([type_representation, type_representation], axis=-1)
]
for layer_activations in F.split_axis(lstm_outputs, lstm_outputs.shape[0], axis=0):
output_tensors.append(F.squeeze(layer_activations, 0))
return {
'activations': output_tensors,
'mask': mask,
}
def check_backward(self, x_data, g_data):
gradient_check.check_backward(
lambda x: functions.squeeze(x, self.axis),
x_data, g_data, **self.check_backward_options)
def forward(self, inputs, device):
x, = inputs
return functions.squeeze(x, axis=self.axis),
def forward(self,
inputs,
batch_lengths,
initial_state=None):
"""
Parameters
----------
inputs : ``torch.FloatTensor``, required.
A tensor of shape (batch_size, num_timesteps, input_size)
to apply the LSTM over.
batch_lengths : ``List[int]``, required.
A list of length batch_size containing the lengths of the sequences in batch.
initial_state : ``Tuple[torch.Tensor, torch.Tensor]``, optional, (default = None)
A tuple (state, memory) representing the initial hidden state and memory
of the LSTM. The ``state`` has shape (1, batch_size, hidden_size) and the
``memory`` has shape (1, batch_size, cell_size).
Returns
-------
output_accumulator : ``torch.FloatTensor``
The outputs of the LSTM for each timestep. A tensor of shape
(batch_size, max_timesteps, hidden_size) where for a given batch
element, all outputs past the sequence length for that batch are
zero tensors.
final_state : ``Tuple[``torch.FloatTensor, torch.FloatTensor]``
A tuple (state, memory) representing the initial hidden state and memory
of the LSTM. The ``state`` has shape (1, batch_size, hidden_size) and the
``memory`` has shape (1, batch_size, cell_size).
"""
batch_size = inputs.shape[0]
total_timesteps = inputs.shape[1]
output_accumulator_list = []
if initial_state is None:
full_batch_previous_memory = chainer.Variable(
self.xp.zeros((batch_size, self.cell_size), 'f'))
full_batch_previous_state = chainer.Variable(
self.xp.zeros((batch_size, self.hidden_size), 'f'))
else:
# first dimension is just (layer * (1 + is_bidirection)), i.e., 1.
full_batch_previous_state = F.squeeze(initial_state[0], axis=0)
full_batch_previous_memory = F.squeeze(initial_state[1], axis=0)
current_length_index = batch_size - 1 if self.go_forward else 0
if self.recurrent_dropout_probability > 0.0 and \
(self.training or chainer.confing.train):
dropout_mask = get_dropout_mask(self.recurrent_dropout_probability,
full_batch_previous_state)
else:
dropout_mask = None
for timestep in range(total_timesteps):
# The index depends on which end we start.
index = timestep if self.go_forward else total_timesteps - timestep - 1
# What we are doing here is finding the index into the batch dimension
# which we need to use for this timestep, because the sequences have
# variable length, so once the index is greater than the length of this
# particular batch sequence, we no longer need to do the computation for
# this sequence. The key thing to recognise here is that the batch inputs
# must be _ordered_ by length from longest (first in batch) to shortest
# (last) so initially, we are going forwards with every sequence and as we
# pass the index at which the shortest elements of the batch finish,
# we stop picking them up for the computation.
if self.go_forward:
while batch_lengths[current_length_index] <= index:
current_length_index -= 1
# If we're going backwards, we are _picking up_ more indices.
else:
# First conditional: Are we already at the maximum number of elements in the batch?
# Second conditional: Does the next shortest sequence beyond the current batch
# index require computation use this timestep?
while current_length_index < (len(batch_lengths) - 1) and \
batch_lengths[current_length_index + 1] > index:
current_length_index += 1
# Actually get the slices of the batch which we
# need for the computation at this timestep.
# shape (batch_size, cell_size)
previous_memory = full_batch_previous_memory[0: current_length_index + 1]
# Shape (batch_size, hidden_size)
previous_state = full_batch_previous_state[0: current_length_index + 1]
# Shape (batch_size, input_size)
timestep_input = inputs[0: current_length_index + 1, index]
# Do the projections for all the gates all at once.
# Both have shape (batch_size, 4 * cell_size)
projected_input = self.input_linearity(timestep_input)
projected_state = self.state_linearity(previous_state)
# Main LSTM equations using relevant chunks of the big linear
# projections of the hidden state and inputs.
# TODO: split_axis
# TODO: cuda kernel
input_gate = F.sigmoid(projected_input[:, (0 * self.cell_size):(1 * self.cell_size)] +
projected_state[:, (0 * self.cell_size):(1 * self.cell_size)])
forget_gate = F.sigmoid(projected_input[:, (1 * self.cell_size):(2 * self.cell_size)] +
projected_state[:, (1 * self.cell_size):(2 * self.cell_size)])
memory_init = F.tanh(projected_input[:, (2 * self.cell_size):(3 * self.cell_size)] +
projected_state[:, (2 * self.cell_size):(3 * self.cell_size)])
output_gate = F.sigmoid(projected_input[:, (3 * self.cell_size):(4 * self.cell_size)] +
projected_state[:, (3 * self.cell_size):(4 * self.cell_size)])
memory = input_gate * memory_init + forget_gate * previous_memory
# Here is the non-standard part of this LSTM cell; first, we clip the
# memory cell, then we project the output of the timestep to a smaller size
# and again clip it.
if self.memory_cell_clip_value:
memory = F.clip(memory, -self.memory_cell_clip_value,
self.memory_cell_clip_value)
# shape (current_length_index, cell_size)
pre_projection_timestep_output = output_gate * F.tanh(memory)
# shape (current_length_index, hidden_size)
timestep_output = self.state_projection(
pre_projection_timestep_output)
if self.state_projection_clip_value:
timestep_output = F.clip(timestep_output,
-self.state_projection_clip_value,
self.state_projection_clip_value)
# Only do dropout if the dropout prob is > 0.0 and we are in training mode.
if dropout_mask is not None:
timestep_output = timestep_output * \
dropout_mask[0: current_length_index + 1]
# We've been doing computation with less than the full batch, so here we create a new
# variable for the the whole batch at this timestep and insert the result for the
# relevant elements of the batch into it.
full_batch_previous_memory = F.concat(
[memory, full_batch_previous_memory[current_length_index + 1:]], axis=0)
full_batch_previous_state = F.concat(
[timestep_output, full_batch_previous_state[current_length_index + 1:]], axis=0)
output_accumulator_list.append(timestep_output)
# Mimic the pytorch API by returning state in the following shape:
# (num_layers * num_directions, batch_size, ...). As this
# LSTM cell cannot be stacked, the first dimension here is just 1.
final_state = (F.expand_dims(full_batch_previous_state, 0),
F.expand_dims(full_batch_previous_memory, 0))
if not self.go_forward:
output_accumulator_list = output_accumulator_list[::-1]
output_accumulator = F.pad_sequence(output_accumulator_list)
output_accumulator = output_accumulator.transpose((1, 0, 2))
# (batch_size, total_timesteps, self.hidden_size)
return output_accumulator, final_state
def check_forward(self, x_data):
y = functions.squeeze(x_data, axis=self.axis)
expected = numpy.squeeze(self.x, axis=self.axis)
testing.assert_allclose(y.data, expected, **self.check_forward_options)
def check_invalid_type(self, x_data):
with self.assertRaises(ValueError):
functions.squeeze(x_data, axis=self.axis)
def test_invalid_axis(self):
with self.assertRaises(TypeError):
functions.squeeze(self.x, axis='a')
def check_invalid_type(self, x_data):
with self.assertRaises(type_check.InvalidType):
functions.squeeze(x_data, axis=self.axis)
def __call__(self, x): return functions.squeeze(x, self.axis)
