def batch_pit_loss_faster(ys, ts, label_delay=0):
"""
PIT loss over mini-batch.
Args:
ys: B-length list of predictions
ts: B-length list of labels
Returns:
loss: (1,)-shape mean cross entropy over mini-batch
labels: B-length list of permuted labels
"""
n_speakers = ts[0].shape[1]
xp = chainer.backend.get_array_module(ys[0])
# (B, T, C)
ys = F.pad_sequence(ys, padding=-1)
losses = []
for shift in range(n_speakers):
# rolled along with speaker-axis
ts_roll = [xp.roll(t, -shift, axis=1) for t in ts]
ts_roll = F.pad_sequence(ts_roll, padding=-1)
# loss: (B, T, C)
loss = F.sigmoid_cross_entropy(ys, ts_roll, reduce='no')
# sum over time: (B, C)
loss = F.sum(loss, axis=1)
losses.append(loss)
# losses: (B, C, C)
losses = F.stack(losses, axis=2)
# losses[b, i, j] is a loss between
# `i`-th speaker in y and `(i+j)%C`-th speaker in t
perms = xp.array(
list(permutations(range(n_speakers))),
dtype='i',
)
# y_inds: [0,1,2,3]
y_ind = xp.arange(n_speakers, dtype='i')
# perms -> relation to t_inds -> t_inds
# 0,1,2,3 -> 0+j=0,1+j=1,2+j=2,3+j=3 -> 0,0,0,0
# 0,1,3,2 -> 0+j=0,1+j=1,2+j=3,3+j=2 -> 0,0,1,3
t_inds = xp.mod(perms - y_ind, n_speakers)
losses_perm = []
for t_ind in t_inds:
losses_perm.append(F.mean(losses[:, y_ind, t_ind], axis=1))
# losses_perm: (B, Perm)
losses_perm = F.stack(losses_perm, axis=1)
min_loss = F.sum(F.min(losses_perm, axis=1))
min_loss = F.sum(F.min(losses_perm, axis=1))
n_frames = np.sum([t.shape[0] for t in ts])
min_loss = min_loss / n_frames
min_indices = xp.argmin(losses_perm.array, axis=1)
labels_perm = [t[:, perms[idx]] for t, idx in zip(ts, min_indices)]
return min_loss, labels_perm
def EMD(self, z):
"""
earth mover distance between z and standard normal
:param z:
:return:
"""
xp = cuda.get_array_module(z)
dim_z = z.shape[1]
n = z.shape[0]
t = xp.random.normal(size=(n * 10, dim_z)).astype("float32")
dot = F.matmul(z, t, transb=True)
dist = F.sum(z**2, axis=1, keepdims=True) - 2 * dot + xp.sum(t**2,
axis=1)
return F.mean(F.min(dist, axis=0)) + F.mean(F.min(dist, axis=1))
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 pit_loss(pred, label, label_delay=0):
"""
Permutation-invariant training (PIT) cross entropy loss function.
Args:
pred: (T,C)-shaped pre-activation values
label: (T,C)-shaped labels in {0,1}
label_delay: if label_delay == 5:
pred: 0 1 2 3 4 | 5 6 ... 99 100 |
label: x x x x x | 0 1 ... 94 95 | 96 97 98 99 100
calculated area: | <------------> |
Returns:
min_loss: (1,)-shape mean cross entropy
label_perms[min_index]: permutated labels
"""
# label permutations along the speaker axis
label_perms = [
label[..., list(p)] for p in permutations(range(label.shape[-1]))
]
losses = F.stack([
F.sigmoid_cross_entropy(pred[label_delay:, ...],
l[:len(l) - label_delay, ...])
for l in label_perms
])
xp = cuda.get_array_module(losses)
min_loss = F.min(losses) * (len(label) - label_delay)
min_index = cuda.to_cpu(xp.argmin(losses.data))
return min_loss, label_perms[min_index]
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(
lambda x: functions.min(x, self.axis, self.keepdims),
x_data,
y_grad,
dtype='d',
**self.check_backward_options)
def _compute_target_q_value(self, batch):
with chainer.using_config('train', False), \
chainer.using_config('enable_backprop', False):
(_, _, r, s_next, non_terminal) = batch
r = F.reshape(r, shape=(*r.shape, 1))
non_terminal = F.reshape(non_terminal,
shape=(*non_terminal.shape, 1))
s_next_rep = F.repeat(x=s_next,
repeats=self._num_action_samples,
axis=0)
a_next_rep = self._vae._decode(s_next_rep)
perturbed_action = self._target_perturbator(s_next_rep, a_next_rep)
q_values = F.stack([
q_target(s_next_rep, perturbed_action)
for q_target in self._target_q_ensembles
])
assert q_values.shape == (self._num_q_ensembles, self._batch_size *
self._num_action_samples, 1)
weighted_q_minmax = self._lambda * F.min(q_values, axis=0) \
+ (1 - self._lambda) * F.max(q_values, axis=0)
assert weighted_q_minmax.shape == (self._batch_size *
self._num_action_samples, 1)
next_q_value = F.max(F.reshape(weighted_q_minmax,
shape=(self._batch_size, -1)),
axis=1,
keepdims=True)
assert next_q_value.shape == (self._batch_size, 1)
target_q_value = r + self._gamma * next_q_value * non_terminal
target_q_value.unchain()
assert target_q_value.shape == (self._batch_size, 1)
return target_q_value
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.min(x, axis=self.axis, keepdims=self.keepdims)
self.assertEqual(y.data.dtype, numpy.float32)
y_expect = self.y_expect
self.assertEqual(y.data.shape, y_expect.shape)
testing.assert_allclose(y_expect, y.data)
def check_forward(self, x_data, axis=None, keepdims=False):
x = chainer.Variable(x_data)
y = functions.min(x, axis=axis, keepdims=keepdims)
self.assertEqual(y.data.dtype, numpy.float32)
y_expect = self.x.min(axis=axis, keepdims=keepdims)
self.assertEqual(y.data.shape, y_expect.shape)
gradient_check.assert_allclose(y_expect, y.data)
def check_forward(self, x_data, axis=None, keepdims=False):
x = chainer.Variable(x_data)
y = functions.min(x, axis=axis, keepdims=keepdims)
self.assertEqual(y.data.dtype, numpy.float32)
y_expect = self.x.min(axis=axis, keepdims=keepdims)
self.assertEqual(y.data.shape, y_expect.shape)
testing.assert_allclose(y_expect, y.data)
def check_backward(self, x_data, y_grad, axis=None, keepdims=False):
gradient_check.check_backward(
lambda x: functions.min(x, axis=axis, keepdims=keepdims),
x_data,
y_grad,
dtype='d',
**self.check_backward_options)
def prelu(self, inp, parameter):
x = F.reshape(inp, (inp.shape[0], 1, inp.shape[1]))
zeros = self.xp.zeros_like(x.data)
c = F.transpose(F.concat((x, zeros), axis=1), (0, 2, 1))
return F.max(
c,
axis=2) + F.broadcast_to(parameter, inp.shape) * F.min(c, axis=2)
def chamfer_distance(pc1, pc2):
'''
Input:
pc1: float chainer in shape (B,N,C) the first point cloud
pc2: float chainer in shape (B,M,C) the second point cloud
Output:
dist1: float chainer in shape (B,N) distance from first to second
idx1: int32 chainer in shape (B,N) nearest neighbor from first to second
dist2: float chainer in shape (B,M) distance from second to first
idx2: int32 chainer in shape (B,M) nearest neighbor from second to first
'''
dist1, idx1, dist2, idx2 = 0, 0, 0, 0
N = pc1.shape[2]
M = pc2.shape[2]
"""
dist = Variable(np.zeros((N,M)))
for i in range(N):
for j in range(M):
dist[i,j] = functions.sum((pc1[0,:,i,0] - pc2[0,:,j,0]) ** 2)
dist1 = functions.min(dist,axis=1)
dist2 = functions.min(dist,axis=1)
"""
#行列のn列目に1次元挿入
#各次元を指定配列分だけ繰り返す、上の場合、２次元の部分をM回繰り返している。
pc1_expand_tile = functions.tile(functions.expand_dims(pc1, 3),
(1, 1, 1, M, 1))
pc2_expand_tile = functions.tile(functions.expand_dims(pc2, 2),
(1, 1, N, 1, 1))
#pc1_expand_tile = functions.tile(pc1,(1,1,M,1))
#pc2_expand_tile = functions.tile(pc2,(1,N,1,1))
#pc1_expand_tile shape = pc2_expand_tile shape
#pc_diff is difference between pc1 and pc2 in coordinate system.
#print(pc1_expand_tile.shape)
#print(pc2_expand_tile.shape)
pc_diff = pc1_expand_tile - pc2_expand_tile
pc_dist = functions.sum(pc_diff**2, axis=1)
dist1 = functions.min(pc_dist, axis=1)
#idx1 = functions.argmin(pc_dist, axis=1)
dist2 = functions.min(pc_dist, axis=2)
#idx2 = functions.argmin(pc_dist, axis=2)
return dist1, idx1, dist2, idx2
def occupancy_grid_3d(points, *, pitch, origin, dims, threshold=1):
d_IP, d_JP, d_KP = OccupancyGrid3D(pitch=pitch, origin=origin,
dims=dims)(points)
d_IJKP = F.sqrt(d_IP**2 + d_JP**2 + d_KP**2)
d_IJK = F.min(d_IJKP, axis=3)
m_IJK = F.relu(threshold - d_IJK)
m_IJK = F.minimum(m_IJK, m_IJK.array * 0 + 1)
return m_IJK
def normalize_linearly(self, h):
"""Normalize h linearly in [0, 1] over dimensions
"""
h_max = F.max(h, axis=1, keepdims=True)
h_min = F.min(h, axis=1, keepdims=True)
h_norm = (h - h_min) / (h_max - h_min + 1e-10)
return h_norm
def normalize_linearly(self, h):
"""Normalize h linearly over dimensions in [0, 1]
"""
h_max = F.max(h, axis=1, keepdims=True)
h_min = F.min(h, axis=1, keepdims=True)
h_norm = (h - h_min) / (h_max - h_min + 1e-10)
return h_norm
def check_backward(self, x_data, y_grad, axis=None, keepdims=False):
gradient_check.check_backward(
lambda x: functions.min(x, axis=axis, keepdims=keepdims),
x_data,
y_grad,
eps=1e-4,
rtol=1e-3,
atol=1e-3)
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 check_backward(self, x_data, y_grad, axis=None, keepdims=False):
x = chainer.Variable(x_data)
y = functions.min(x, axis=axis, keepdims=keepdims)
y.grad = y_grad
y.backward()
func = y.creator
f = lambda: func.forward((x.data.copy(),))
gx, = gradient_check.numerical_grad(f, (x.data,), (y.grad,), eps=1e-5)
gradient_check.assert_allclose(gx, x.grad, rtol=1e-3, atol=1e-3)
def check_backward(self, x_data, y_grad, axis=None, keepdims=False):
x = chainer.Variable(x_data)
y = functions.min(x, axis=axis, keepdims=keepdims)
y.grad = y_grad
y.backward()
func = y.creator
f = lambda: func.forward((x.data.copy(), ))
gx, = gradient_check.numerical_grad(f, (x.data, ), (y.grad, ),
eps=1e-5)
gradient_check.assert_allclose(gx, x.grad, rtol=1e-3, atol=1e-3)
def batched_triangle_reduce_(b, t, p, n, id):
xp = chainer.backend.get_array_module(b)
BB, _, H, W = b.shape[:4]
kb = xp.sum(b, axis=0).astype(xp.bool)
kt = F.min(t, axis=0)
kp = p[0, :, :, :]
kn = n[0, :, :, :]
kid = id[0, :, :, :]
for i in range(1, BB):
bb = (kt.data >= t[i, :, :, :].data)
kp = F.where(bb, p[i, :, :, :], kp)
kn = F.where(bb, n[i, :, :, :], kn)
kid = F.where(bb, id[i, :, :, :], kid)
b = chainer.as_variable(kb.reshape(1, 1, H, W))
t = kt.reshape(1, 1, H, W)
p = kp.reshape(1, 3, H, W)
n = kn.reshape(1, 3, H, W)
id = kid.reshape(1, 1, H, W)
return b, t, p, n, id
def __call__(self, template, speech, length_of_template, length_of_speech):
self.nodes=[]
for i in range(length_of_template+1):
for j in range(length_of_speech+1):
self.nodes.append(V(ar(0.0,dtype=np.float32)))
for i in range(length_of_template+1):
#print("("+str(i)+",1)",end=' ')
for j in range(length_of_speech+1):
if(i!=0 and j!=0):
self.nodes[i*(length_of_speech+1)+j]=F.min(F.stack([self.nodes[(i-1)*(length_of_speech+1)+j],self.nodes[i*(length_of_speech+1)+j-1],self.nodes[(i-1)*(length_of_speech+1)+j-1]]))+F.sqrt(F.sum(dist(template[i-1],speech[j-1]))+1e-8)
#print(self.nodes[-length_of_speech-1:])
result_temp=self.nodes[-length_of_speech:]
result=[]
t1=[]
t2=[]
t3=[]
for i in range(len(result_temp)):
t1.append(F.expand_dims(result_temp[i],axis=0))
t2.append(F.expand_dims(t1[i],axis=0))
t3.append(F.linear(t2[i],self.W,self.b))
result.append(F.sigmoid(t3[i]))
y = F.hstack(result)
return y[0]
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(
lambda x: functions.min(x, self.axis, self.keepdims),
x_data, y_grad, dtype='d',
**self.check_backward_options)
def get_bow_reps(self, ids, xs, xs_embed, position_info, x_spans,
shell_spans, x_position_info):
assert len(x_spans[0]) == len(shell_spans[0])
x_spans = [[[
shell_span[0].tolist(), x_span[1].tolist()
] for x_span, shell_span in zip(x_spans_in_para, shell_spans_in_para)]
for x_spans_in_para, shell_spans_in_para in zip(
x_spans, shell_spans)]
#(all_n_spans, 1) paragraph_type
eye = self.xp.identity(4, dtype=self.xp.float32)
para_type = [
i.tolist() - self.max_n_spans * 2
for i in self.xp.vstack(x_position_info)[:, 2]
]
para_type = self.xp.vstack([eye[i] for i in para_type])
#(batchsize, max_n_tokens, word_vec)
xs_embed = chaFunc.pad_sequence(xs_embed, padding=-1)
#(batchsize, n_spans, max_n_tokens, word_vec)
xs_embed = [
chaFunc.tile(xs_embed[i], (len(spans), 1, 1))
for i, spans in enumerate(x_spans)
]
#(all_spans_in_batch, max_n_tokens, word_vec)
xs_embed = chaFunc.vstack(xs_embed)
#(batchsize, max_n_tokens, word_id)
xs = chaFunc.pad_sequence(xs, padding=0)
#(batchsize, n_spans, max_n_tokens, word_id)
xs = [
self.xp.tile(xs[i].data, (len(spans), 1))
for i, spans in enumerate(x_spans)
]
#(all_spans_in_batch, max_n_tokens, word_id)
xs = self.xp.vstack(xs)
#(all_spans_in_batch, max_n_tokens, word_vec)
mask_xs_embed_bool = self.xp.zeros(xs_embed.shape).astype(self.xp.bool)
#(all_spans_in_batch, word_vec) the length of each span
len_spans = self.xp.zeros(
(xs_embed.shape[0], xs_embed.shape[2])).astype(self.xp.float32)
#(all_spans_in_batch, (start, end))
x_spans = np.vstack(x_spans)
xs_ids = []
eye = self.xp.identity(len(self.vocab), dtype=self.xp.float32)
for i, span in enumerate(x_spans):
mask_xs_embed_bool[i][int(span[0]):int(span[1])] = True
xs_ids.append(self.xp.sum(eye[xs[i][int(span[0]):int(span[1])]],
0))
len_spans[i].fill(span[1] - span[0])
# max pooling
mask_xs_embed = self.xp.zeros(xs_embed.shape).astype(self.xp.float32)
mask_xs_embed.fill(-self.xp.inf)
max_pooling_xs = chaFunc.max(
chaFunc.where(mask_xs_embed_bool, xs_embed, mask_xs_embed), 1)
# average pooling
mask_xs_embed = self.xp.zeros(xs_embed.shape).astype(self.xp.float32)
avg_pooling_xs = chaFunc.sum(
chaFunc.where(mask_xs_embed_bool, xs_embed, mask_xs_embed),
1) / len_spans
# min pooling
mask_xs_embed = self.xp.zeros(xs_embed.shape).astype(self.xp.float32)
mask_xs_embed.fill(self.xp.inf)
min_pooling_xs = chaFunc.min(
chaFunc.where(mask_xs_embed_bool, xs_embed, mask_xs_embed), 1)
#(all_n_spans, max_n_tokens, vocab_size)
xs_ids = self.xp.vstack(xs_ids)
#(all_n_spans, feature_vector)
if self.use_elmo:
# We found that pooling-based features with ELMo does not significantly contribute the performance.
# Then, we only used discrete BoW features for span-based models with ELMo.
bow_reps = chaFunc.concat([xs_ids])
else:
bow_reps = chaFunc.concat(
[max_pooling_xs, min_pooling_xs, avg_pooling_xs, xs_ids])
assert bow_reps.shape[-1] == self.bow_feature_size
bow_reps = chaFunc.sigmoid(self.BowFCLayer(bow_reps))
bow_reps = chaFunc.dropout(bow_reps, self.dropout)
return bow_reps
def calc_distance(est_theta, theta):
# weak regularization to the distribution of estimated thetas
dist = F.sum(est_theta ** 2, axis=1) + (theta ** 2).sum(axis=1).T - 2 * F.matmul(est_theta, theta, transb=True)
return F.mean(F.min(dist, axis=0)) + F.mean(F.min(dist, axis=1))
def test_pos_neg_duplicate_axis(self):
x_data = numpy.random.uniform(-1, 1, (3, 2, 4)).astype(numpy.float32)
x = chainer.Variable(x_data)
with self.assertRaises(ValueError):
functions.min(x, axis=(1, -2))
def get_mask(gcam, sigma=.5, w=8):
gcam = (gcam - F.min(gcam).data) / (F.max(gcam) - F.min(gcam)).data
mask = F.squeeze(F.sigmoid(w * (gcam - sigma)))
return mask
def test_duplicate_axis(self):
with self.assertRaises(ValueError):
functions.min(self.x, (0, 0))
def check_backward(self, x_data, y_grad, axis=None, keepdims=False):
gradient_check.check_backward(
lambda x: functions.min(x, axis=axis, keepdims=keepdims),
x_data, y_grad, dtype='d',
**self.check_backward_options)
def f(x):
return functions.min(x, self.axis, self.keepdims)
def f(x):
x = functions.min(x, self.axis, self.keepdims)
return x * x
def test_invalid_axis_type_in_tuple(self):
with self.assertRaises(TypeError):
functions.min(self.x, (1, 'x'))
def test_invalid_axis_type(self):
with self.assertRaises(TypeError):
functions.min(self.x, [0])
def batch_pit_n_speaker_loss(ys, ts, n_speakers_list):
"""
PIT loss over mini-batch.
Args:
ys: B-length list of predictions (pre-activations)
ts: B-length list of labels
n_speakers_list: list of n_speakers in batch
Returns:
loss: (1,)-shape mean cross entropy over mini-batch
labels: B-length list of permuted labels
"""
max_n_speakers = ts[0].shape[1]
xp = chainer.backend.get_array_module(ys[0])
# (B, T, C)
ys = F.pad_sequence(ys, padding=-1)
losses = []
for shift in range(max_n_speakers):
# rolled along with speaker-axis
ts_roll = [xp.roll(t, -shift, axis=1) for t in ts]
ts_roll = F.pad_sequence(ts_roll, padding=-1)
# loss: (B, T, C)
loss = F.sigmoid_cross_entropy(ys, ts_roll, reduce='no')
# sum over time: (B, C)
loss = F.sum(loss, axis=1)
losses.append(loss)
# losses: (B, C, C)
losses = F.stack(losses, axis=2)
# losses[b, i, j] is a loss between
# `i`-th speaker in y and `(i+j)%C`-th speaker in t
perms = xp.array(
list(permutations(range(max_n_speakers))),
dtype='i',
)
# y_ind: [0,1,2,3]
y_ind = xp.arange(max_n_speakers, dtype='i')
# perms -> relation to t_inds -> t_inds
# 0,1,2,3 -> 0+j=0,1+j=1,2+j=2,3+j=3 -> 0,0,0,0
# 0,1,3,2 -> 0+j=0,1+j=1,2+j=3,3+j=2 -> 0,0,1,3
t_inds = xp.mod(perms - y_ind, max_n_speakers)
losses_perm = []
for t_ind in t_inds:
losses_perm.append(F.mean(losses[:, y_ind, t_ind], axis=1))
# losses_perm: (B, Perm)
losses_perm = F.stack(losses_perm, axis=1)
# masks: (B, Perms)
def select_perm_indices(num, max_num):
perms = list(permutations(range(max_num)))
sub_perms = list(permutations(range(num)))
return [[x[:num] for x in perms].index(perm) for perm in sub_perms]
masks = xp.full_like(losses_perm.array, xp.inf)
for i, t in enumerate(ts):
n_speakers = n_speakers_list[i]
indices = select_perm_indices(n_speakers, max_n_speakers)
masks[i, indices] = 0
losses_perm += masks
min_loss = F.sum(F.min(losses_perm, axis=1))
n_frames = np.sum([t.shape[0] for t in ts])
min_loss = min_loss / n_frames
min_indices = xp.argmin(losses_perm.array, axis=1)
labels_perm = [t[:, perms[idx]] for t, idx in zip(ts, min_indices)]
labels_perm = [
t[:, :n_speakers]
for t, n_speakers in zip(labels_perm, n_speakers_list)
]
return min_loss, labels_perm
def _policy_update(self, batch):
status = {}
vae_status = self._train_vae(batch)
status.update(vae_status)
(s, a, _, _, _) = batch
_, raw_sampled_actions = self._vae._decode_multiple(
s, decode_num=self._num_mmd_samples)
pi_actions, raw_pi_actions = self._pi._sample_multiple(
s, sample_num=self._num_mmd_samples)
if self._kernel_type == 'gaussian':
mmd_loss = self._compute_gaussian_mmd(raw_sampled_actions,
raw_pi_actions,
sigma=self._mmd_sigma)
elif self._kernel_type == 'laplacian':
mmd_loss = self._compute_laplacian_mmd(raw_sampled_actions,
raw_pi_actions,
sigma=self._mmd_sigma)
else:
raise ValueError('Unknown kernel: {}'.format(self._kernel_type))
assert mmd_loss.shape == (self._batch_size, 1)
s_hat = F.expand_dims(s, axis=0)
s_hat = F.repeat(s_hat, repeats=self._num_mmd_samples, axis=0)
s_hat = F.reshape(s_hat,
shape=(self._batch_size * self._num_mmd_samples,
s.shape[-1]))
a_hat = F.transpose(pi_actions, axes=(1, 0, 2))
a_hat = F.reshape(a_hat,
shape=(self._batch_size * self._num_mmd_samples,
a.shape[-1]))
q_values = F.stack([q(s_hat, a_hat) for q in self._q_ensembles])
assert q_values.shape == (self._num_q_ensembles,
self._batch_size * self._num_mmd_samples, 1)
q_values = F.reshape(q_values,
shape=(self._num_q_ensembles,
self._num_mmd_samples, self._batch_size,
1))
q_values = F.mean(q_values, axis=1)
assert q_values.shape == (self._num_q_ensembles, self._batch_size, 1)
q_min = F.min(q_values, axis=0)
if self._use_stddev:
q_stddev = self._compute_stddev(x=q_values, axis=0, keepdims=False)
assert q_min.shape == q_stddev.shape
else:
q_stddev = 0.0
assert q_min.shape == (self._batch_size, 1)
if self._num_iterations > self._warmup_iterations:
pi_loss = F.mean(-q_min + q_stddev * self._stddev_coeff +
self._lagrange_multiplier.exp() * mmd_loss)
else:
pi_loss = F.mean(self._lagrange_multiplier.exp() * mmd_loss)
# Dual gradient descent
# Update actor
self._pi_optimizer.target.cleargrads()
pi_loss.backward()
self._pi_optimizer.update()
# Just for maintaining consistency with original code
if self._use_stddev:
q_stddev.unchain()
# Update lagrange multiplier
lagrange_loss = -F.mean(-q_min + q_stddev * self._stddev_coeff +
self._lagrange_multiplier.exp() *
(mmd_loss - self._epsilon))
self._lagrange_optimizer.target.cleargrads()
lagrange_loss.backward()
self._lagrange_optimizer.update()
pi_loss.unchain_backward()
lagrange_loss.unchain_backward()
# Clip lagrange multiplier in range
self._lagrange_multiplier.clip(-5.0, 10.0)
xp = chainer.backend.get_array_module(pi_loss)
status['pi_loss'] = xp.array(pi_loss.array)
status['mmd_loss'] = xp.mean(xp.array(mmd_loss.array))
status['lagrange_loss'] = xp.array(lagrange_loss.array)
status['lagrange_multiplier'] = xp.array(
self._lagrange_multiplier().array)
return status
def f(x):
return functions.min(x, self.axis, self.keepdims)
def __call__(self, batch_graph, targets=None):
"""
This method performs forward calculation.
Parameters
----------
batch_graph : list consists of Graph
contains Graphs in minibatch
targets : targets
this parameter is only used in regression task
Returns
-------
In classification task : (batchsize, num_classes) matrix
which means the probability of which class is each graph in.
In regression task : (batchsize, 1) matrix
which means the prediction value of each graph treewidth.
"""
# set the array module based on using device
xp = self.device.xp
# concatenate the node_features
X_concat = chainer.Variable(xp.concatenate([xp.array(graph.node_features) for graph in batch_graph], axis=0))
X_concat.to_device(self.device) # if you use GPU, you must transfer X_concat into GPU.
# make graph pooling matrix and neighbors pooling matrix
graph_pool = self.__preprocess_graphpool(batch_graph)
if self.neighbor_pooling_type == "max":
padded_neighbor_list = self.__preprocess_neighbors_maxpool(batch_graph)
else:
Adj_block = self.__preprocess_neighbors_sumavepool(batch_graph)
hidden_rep = [X_concat] # list of hidden representation at each layer (including input feature vectors)
h = X_concat
# perform Aggregating and Combining node features
for layer in range(self.num_layers-1):
# perform max neighbor pooling
if self.neighbor_pooling_type == "max":
# padding minimum value vector
padded_h = F.concat((h, F.min(h, axis=0).reshape(1, h.shape[1])), axis=0)
# make (F-dim, max_deg * nodes) matrix to perform max aggregation
pooled_mat = F.sparse_matmul(padded_h.transpose(), padded_neighbor_list).transpose()
# make 3D tensor
pooled_tensor = F.reshape(pooled_mat, (padded_neighbor_list.shape[0] - 1,
int(padded_neighbor_list.shape[1] / (padded_neighbor_list.shape[0] - 1)), h.shape[1]))
# take max
pooled = F.max(pooled_tensor, axis=1)
# perform sum or average neighbor pooling
else:
pooled = F.sparse_matmul(Adj_block, h)
if self.neighbor_pooling_type == "average":
degree = F.sparse_matmul(Adj_block, xp.ones((Adj_block.shape[0], 1), dtype=xp.float32))
pooled = pooled/degree
# input aggregated vectors into MLP
pooled_rep = self.mlps[layer](pooled)
h = self.batch_norms[layer](pooled_rep)
h = F.relu(h)
hidden_rep.append(h)
# perform Readout node features
score_over_layer = 0
for layer, h in enumerate(hidden_rep):
# perform max readout
if self.graph_pooling_type == "max":
# padding minimum value
padded_h = F.concat((h, F.min(h, axis=0).reshape(1, h.shape[1])), axis=0)
# make (F-dim, max|V| * batchsize) matrix to perform max aggregation
pooled_mat = F.sparse_matmul(padded_h.transpose(), graph_pool).transpose()
# make 3D tensor
pooled_tensor = F.reshape(pooled_mat, (len(batch_graph), int(graph_pool.shape[1] / len(batch_graph)), h.shape[1]))
# take max
pooled_h = F.max(pooled_tensor, axis=1)
# sum or average readout
else:
pooled_h = F.sparse_matmul(graph_pool, h)
score_over_layer += F.dropout(self.linears_prediction[layer](pooled_h), self.final_dropout)
# final layers in regression task
if self.task_type == "Regression":
h = self.final_l2(score_over_layer)
h = F.relu(h)
score_over_layer = self.final_l1(h)
if targets is None:
return score_over_layer
else:
self.loss = F.mean_squared_error(targets.reshape(-1, 1), score_over_layer) # MSE Loss
self.abs_loss = F.mean_absolute_error(targets.reshape(-1, 1), score_over_layer) # MAE Loss
self.abs_max_loss = F.max(F.absolute_error(targets.reshape(-1, 1), score_over_layer)) # Max Absolute Error
chainer.reporter.report({'loss': self.loss}, self)
chainer.reporter.report({'abs_loss': self.abs_loss}, self)
chainer.reporter.report({'abs_max_loss': self.abs_max_loss}, self)
# return the MSE loss. If you want to use other loss, please change this sentence.
return self.loss
return score_over_layer
def test_invalid_axis_type(self):
with self.assertRaises(TypeError):
functions.min(self.x, [0])
def update_core(self):
gen = self.models['gen']
dis = self.models['dis']
enc = self.models['enc']
gen_optimizer = self.get_optimizer('opt_gen')
xp = enc.xp
x, gt, c = self.get_batch(xp)
if self.input_size is not None:
_x = []
for img in x.data.get():
_x.append(
chainercv.transforms.resize(
img, (self.input_size, self.input_size)))
x = Variable(xp.asarray(_x))
# obtain initial z by encoder
if enc.n_classes != 0:
z = enc(x, y=c)
else:
z = enc(x)
# fast updating
with chainer.using_config('train', False):
# out_noab : reconstruction results without auxiliary network
outs, fast_losses, out_noab, zeta, z_prime = gen(batchsize=len(z),
z=z,
y=c,
gt=gt)
lmd_pixel = 0.05
fast_losses.append(
reconstruction_loss(dis, outs[-1], gt) +
lmd_pixel * pixel_loss(outs[-1], gt))
loss = 0
weights = [20, 2.0, 1.0]
for i in range(0, len(outs)):
loss += fast_losses[i] * weights[i]
# reconstruction loss as an autoencoder
lmd_ae = 100
# lmd_ae = 0
ae_loss = F.mean_squared_error(z, z_prime) * z.shape[0]
loss += lmd_ae * ae_loss
# sparse regularization
# lmd_sparse = 0.000
# sparse_loss = lmd_sparse * F.sum(F.absolute(zeta))
# loss += sparse_loss
gen.cleargrads()
# double backprop
loss.backward()
gen_optimizer.update()
# reporting
report = dict()
for i, loss_i in enumerate(fast_losses):
report["loss{}".format(i + 1)] = loss_i
report["loss_ae"] = ae_loss
report["loss_noab"] = reconstruction_loss(
dis, out_noab, gt) + lmd_pixel * pixel_loss(out_noab, gt)
report["fast_alpha"] = gen.fast_alpha().data.mean()
report["fast_benefit"] = report["loss{}".format(
len(fast_losses))] - report["loss1"]
report["min_slope"] = F.min(gen.preluW())
report["max_slope"] = F.max(gen.preluW())
report["min_slope_middle"] = F.min(gen.preluMiddleW())
report["max_slope_middle"] = F.max(gen.preluMiddleW())
chainer.reporter.report(report)
if not gen.learned_lr:
gen._fast_alpha = min(
gen.limit_fast_alpha,
gen.initial_fast_alpha + gen.step_fast_alpha * self.iteration)
def test_invalid_axis_type_in_tuple(self):
with self.assertRaises(TypeError):
functions.min(self.x, (1, 'x'))
def update_core(self):
# TODO:
# log parametrisation of radius?
epsilon = 1e-10
opt = self.get_optimizer('main')
r = F.relu(self.coords.W[:, 0]) # radius
x = self.coords.W[:, 1:(self.args.dim + 1)] # anchor
c = self.coords.W[:, (self.args.dim + 1):] # sphere centre
a, b = self.converter(self.get_iterator('main').next()) # edge
loss = 0
# anchor loss
if self.args.lambda_anchor > 0: # DANCAR
v, = self.converter(self.get_iterator('anchor').next())
loss_anc = F.average(
F.relu(
F.sqrt(F.sum((c[v] - x[v])**2, axis=1) + epsilon) - r[v] +
self.args.margin))
chainer.report({'loss_anc': loss_anc}, self.coords)
loss += self.args.lambda_anchor * loss_anc
else:
x = c
# positive sample: a contains b
if self.args.lambda_pos > 0:
d = F.sqrt(F.sum((c[a] - x[b])**2, axis=1) + epsilon)
loss_pos = F.average(
F.relu(self.args.margin + d + self.args.dag * r[b] - r[a]))
chainer.report({'loss_pos': loss_pos}, self.coords)
loss += self.args.lambda_pos * loss_pos
# negative sample
if self.args.lambda_neg > 0:
v, = self.converter(self.get_iterator('vertex').next())
if self.args.batchsize_negative > 0:
na, nb = [], []
for u in v: # sample non-edges. accurate but slow
nnbor = set(nx.non_neighbors(self.graph, u))
for q in random.sample(
nnbor,
min(self.args.batchsize_negative,
len(list(nnbor)))):
na.append(u)
nb.append(q)
na = np.array(na)
nb = np.array(nb)
else: # random vertex pairs
na = v
nb = np.roll(v, 1)
d = F.sqrt(F.sum((c[na] - x[nb])**2, axis=1) + epsilon)
loss_neg = F.average(F.relu(self.args.margin - (d - r[na])))
# loss_neg = F.average(F.relu(self.args.margin - (d + self.args.dag * r[nb] - r[na])))
chainer.report({'loss_neg': loss_neg}, self.coords)
loss += self.args.lambda_neg * loss_neg
# radius should be similar
if self.args.lambda_uniform_radius > 0:
loss_uniform_radius = (F.max(r) - F.min(r))**2
chainer.report({'loss_rad': loss_uniform_radius}, self.coords)
loss += self.args.lambda_uniform_radius * loss_uniform_radius
# update the coordinates
self.coords.cleargrads()
loss.backward()
opt.update(loss=loss)
self.coords.W.array[:, 0] = self.coords.xp.clip(self.coords.W.array[:,
0],
a_min=0.01,
a_max=None)
def test_duplicate_axis(self):
with self.assertRaises(ValueError):
functions.min(self.x, (0, 0))
def f(x):
x = functions.min(x, self.axis, self.keepdims)
return x * x
def test_pos_neg_duplicate_axis(self):
x_data = numpy.random.uniform(-1, 1, (3, 2, 4)).astype(numpy.float32)
x = chainer.Variable(x_data)
with self.assertRaises(ValueError):
functions.min(x, axis=(1, -2))
