def __call__(self, x):
# chainer requires explicit broadcast for avoiding latent bugs
u = F.mean(x, -1, keepdims=True)
u = F.broadcast_to(u, x.shape)
s = F.mean((x - u) ** 2, -1, keepdims=True)
s = F.broadcast_to(s, x.shape)
x = (x - u) / F.sqrt(s + self.e)
return F.bias(F.scale(x, self.g, axis=2), self.b, axis=2)
def forward(self, inputs, device):
x, y = inputs
mean = functions.mean(x, axis=1)
d = x - mean[:, None]
var = functions.mean(d * d, axis=1)
inv_std = functions.rsqrt(var + self.eps)
dummy_gamma = self.backend_config.xp.ones(
self.shape[0], dtype=self.dtype)
return gn_module._MulInvStd(
self.eps, mean.array, inv_std.array, dummy_gamma).apply((x, y))
def __call__(self, x):
q_z = self.encoder(x)
z = q_z.sample(self.k)
p_x = self.decoder(z)
p_z = self.prior()
reconstr = F.mean(p_x.log_prob(
F.broadcast_to(x[None, :], (self.k,) + x.shape)))
kl_penalty = F.mean(chainer.kl_divergence(q_z, p_z))
loss = - (reconstr - self.beta * kl_penalty)
reporter.report({'loss': loss}, self)
reporter.report({'reconstr': reconstr}, self)
reporter.report({'kl_penalty': kl_penalty}, self)
return loss
def update_policy():
# Maximize Q(s,policy(s))
q = Q(obs, policy(obs))
q = q[:] # Avoid https://github.com/chainer/chainer/issues/2744
loss = - F.mean(q)
policy.cleargrads()
loss.backward()
opt_policy.update()
def f_loss_grad(x):
set_flat_params(self, x)
self.cleargrads()
values = self.compute_baselines(obs)
loss = F.mean(F.square(values - targets))
loss.backward()
flat_grad = get_flat_grad(self)
return loss.data.astype(np.float64), flat_grad.astype(np.float64)
def predict(self, xs):
# Encoding
logits, exs = self._encode(xs)
# Discretization
D = F.gumbel_softmax(logits, self.tau, axis=2)
gumbel_output = D.reshape(-1, self.M * self.K)
with chainer.no_backprop_mode():
maxp = F.mean(F.max(D, axis=2))
reporter.report({'maxp': maxp.data}, self)
# Decoding
y_hat = self._decode(gumbel_output)
return y_hat, exs
def test_backward_case2(self):
"""Backward if non-zero gradient is on a face."""
vertices = [
[0.8, 0.8, 1.],
[-0.5, -0.8, 1.],
[0.8, -0.8, 1.]]
faces = [[0, 1, 2]]
pyi = 40
pxi = 50
grad_ref = [
[0.98646867, 1.04628897, 0.],
[-1.03415668, - 0.10403691, 0.],
[3.00094461, - 1.55173182, 0.],
]
renderer = neural_renderer.Renderer()
renderer.image_size = 64
renderer.anti_aliasing = False
renderer.perspective = False
renderer.light_intensity_ambient = 1.0
renderer.light_intensity_directional = 0.0
vertices = cp.array(vertices, 'float32')
faces = cp.array(faces, 'int32')
textures = cp.ones((faces.shape[0], 4, 4, 4, 3), 'float32')
grad_ref = cp.array(grad_ref, 'float32')
vertices, faces, textures, grad_ref = utils.to_minibatch((vertices, faces, textures, grad_ref))
vertices = chainer.Variable(vertices)
images = renderer.render(vertices, faces, textures)
images = cf.mean(images, axis=1)
loss = cf.sum(cf.absolute(images[:, pyi, pxi]))
loss.backward()
grad_ref = cp.array(grad_ref, 'float32')
chainer.testing.assert_allclose(vertices.grad, grad_ref, rtol=1e-2)
def test_backward_case1(self):
"""Backward if non-zero gradient is out of a face."""
vertices = [
[0.8, 0.8, 1.],
[0.0, -0.5, 1.],
[0.2, -0.4, 1.]]
faces = [[0, 1, 2]]
pxi = 35
pyi = 25
grad_ref = [
[1.6725862, -0.26021874, 0.],
[1.41986704, -1.64284933, 0.],
[0., 0., 0.],
]
renderer = neural_renderer.Renderer()
renderer.image_size = 64
renderer.anti_aliasing = False
renderer.perspective = False
renderer.light_intensity_ambient = 1.0
renderer.light_intensity_directional = 0.0
vertices = cp.array(vertices, 'float32')
faces = cp.array(faces, 'int32')
textures = cp.ones((faces.shape[0], 4, 4, 4, 3), 'float32')
grad_ref = cp.array(grad_ref, 'float32')
vertices, faces, textures, grad_ref = utils.to_minibatch((vertices, faces, textures, grad_ref))
vertices = chainer.Variable(vertices)
images = renderer.render(vertices, faces, textures)
images = cf.mean(images, axis=1)
loss = cf.sum(cf.absolute(images[:, pyi, pxi] - 1))
loss.backward()
chainer.testing.assert_allclose(vertices.grad, grad_ref, rtol=1e-2)
def _head_to_tail(self, pool5):
block5 = self.rcnn_top(pool5) # B, 1024, 1, 1
fc7 = F.mean(F.mean(block5, 3), 2) # B, 1024
return fc7
def __call__(self, x): return x / sqrt(mean(x ** 2, axis=1, keepdims=True) + 1e-8)
def encode_phrase(self, X):
X = [F.mean(F.embed_id(x, self.w_vec), axis=0, keepdims=True) for x in X]
return F.vstack(X)
def get_posi_from_img(img, threshold=0.2):
"""画像 img = [1][3 or 1][imgH][imgW] ⇒ 点位置 posi = [N][y,x] """
img_p = get_local_max_point(F.mean(img, axis=1, keepdims=True).data,
threshold=threshold)
posi = conv_point_to_posi(img_p)
return posi
def loss_hinge_dis(dis_fake, dis_real):
loss = F.mean(F.relu(1. - dis_real))
loss += F.mean(F.relu(1. + dis_fake))
return loss
def gp_loss(self, x, z):
h = F.mean(x) / x.shape[0]
grad, = chainer.grad([h], [z], enable_double_backprop=True)
return F.mean(F.batch_l2_norm_squared(grad))
def feature_vector_normalize(x):
alpha = 1.0 / F.sqrt(F.mean(x * x, axis=1, keepdims=True) + 1e-8)
y = F.broadcast_to(alpha, x.data.shape) * x
return y
def train(width, height, depth, start_alpha=0):
g = generator(512, 512, 100)
try:
serializers.load_npz("generator.model", g)
print("generator loaded")
except:
pass
d = discriminator()
try:
serializers.load_npz("discriminator.model", d)
print("discriminator loaded")
except:
pass
g_opt = chainer.optimizers.Adam(alpha=0.001, beta1=0.0, beta2=0.99)
g_opt.setup(g)
g_opt.add_hook(chainer.optimizer.WeightDecay(0.0005))
d_opt = chainer.optimizers.Adam(alpha=0.001, beta1=0.0, beta2=0.99)
d_opt.setup(d)
d_opt.add_hook(chainer.optimizer.WeightDecay(0.0005))
X_train, tags = data_import(16 * (2**depth), 16 * (2**depth))
'''
X_train = (X_train.astype(np.float32) - 127.5)/127.5
X_train = X_train.transpose(0,3,1,2)
'''
print(X_train.shape)
tags = tags.astype(np.float32)
num_batches = int(X_train.shape[0] / BATCH_SIZE)
alpha = start_alpha
for epoch in range(NUM_EPOCH):
for index in range(num_batches):
if alpha < 1.0:
alpha = alpha + 5e-4
'''
x = xs[(j * bm):((j + 1) * bm)]
t = ts[(j * bm):((j + 1) * bm)]
'''
image_batch = X_train[index * BATCH_SIZE:(index + 1) * BATCH_SIZE]
image_batch = (image_batch.astype(np.float32) - 127.5) / 127.5
image_batch = image_batch.transpose(0, 3, 1, 2)
tag_batch = tags[index * BATCH_SIZE:(index + 1) * BATCH_SIZE]
noise = np.random.normal(0, 0.5, [len(image_batch), 100])
z = Variable(noise.astype(np.float32))
x = g(z, tag_batch, depth, alpha)
if index % 10 == 0:
generated_images = x.data * 127.5 + 127.5
generated_images = generated_images.transpose(0, 2, 3, 1)
save_generated_image(generated_images,
"%04d_%04d.png" % (epoch, index))
yl = d(x, tag_batch, depth, alpha)
#print(yl)
#g_loss=F.mean_squared_error(yl, Variable(np.ones((len(image_batch),1), dtype=np.float32)))
#d_loss=F.mean_squared_error(yl, Variable(np.zeros((len(image_batch),1), dtype=np.float32)))
yl2 = d(image_batch, tag_batch, depth, alpha)
#print(yl2)
#d_loss+=F.mean_squared_error(yl2, Variable(np.ones((len(image_batch),1), dtype=np.float32)))
d_loss = -F.sum(yl2 - yl) / len(image_batch)
d_loss += F.mean(0.001 * yl * yl)
g_loss = -F.sum(yl) / len(image_batch)
'''
mean=F.mean(x,axis=0)
dev=x-F.broadcast_to(mean, x.shape)
devdev=dev*dev
var=F.mean(devdev)
g_loss-= var
'''
g.cleargrads()
g_loss.backward()
g_opt.update()
d.cleargrads()
d_loss.backward()
d_opt.update()
print(
"epoch %d, batch: %d, g_loss: %f, d_loss: %f, alpha: %f, depth: %d"
% (epoch, index, g_loss.data, d_loss.data, alpha, depth))
serializers.save_npz('generator.model', g)
serializers.save_npz('discriminator.model', d)
def loss_func_adv_gen(self, y_fake):
return F.mean(y_fake)
def loss_func_adv_dis_real(self, y_real):
return F.mean(y_real)
def loss_func_adv_dis_fake(self, y_fake):
return F.mean(y_fake)
fake_2 = F.average_pooling_2d(fake, 3, 2, 1)
fake_4 = F.average_pooling_2d(fake_2, 3, 2, 1)
dis_fake, _ = discriminator(F.concat([fake, line]))
dis2_fake, _ = discriminator_2(F.concat([fake_2, line_2]))
dis4_fake, _ = discriminator_4(F.concat([fake_4, line_4]))
dis_color, _ = discriminator(F.concat([color, line]))
dis2_color, _ = discriminator_2(F.concat([color_2, line_2]))
dis4_color, _ = discriminator_4(F.concat([color_4, line_4]))
fake.unchain_backward()
fake_2.unchain_backward()
fake_4.unchain_backward()
adver_loss = F.mean(F.softplus(-dis_color)) + F.mean(
F.softplus(dis_fake))
adver_loss += F.mean(F.softplus(-dis2_color)) + F.mean(
F.softplus(dis2_fake))
adver_loss += F.mean(F.softplus(-dis4_color)) + F.mean(
F.softplus(dis4_fake))
discriminator.cleargrads()
discriminator_2.cleargrads()
discriminator_4.cleargrads()
discriminator.to_gpu()
adver_loss.backward()
dis_opt.update()
dis2_opt.update()
dis4_opt.update()
adver_loss.unchain_backward()
def normalize(self, z):
return z / sqrt(mean(z**2, axis=1, keepdims=True) + 1e-8)
def make_input_x(self, x, mask, xp):
x_fill = F.mean(x, axis=(2, 3))[:, :, xp.newaxis, xp.newaxis]
x_shape = x.shape
return x * F.broadcast_to((1 - mask), x_shape) + F.broadcast_to(
x_fill, x_shape) * F.broadcast_to(mask, x_shape)
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 bs_reg(self):
bs_re = F.mean(F.square(self.linear.W))
return bs_re
def get_loss(self, batch_data):
config = self.config
batch_pos = batch_data / 127.5 - 1
bbox = random_bbox(config)
mask = bbox2mask(bbox, batch_data.shape[0], config, self.xp)
batch_incomplete = batch_pos * (1 - mask)
x1, x2, offset_flow = self.inpaintnet(batch_incomplete, mask, config)
if config.PRETRAIN_COARSE_NETWORK:
batch_predicted = x1
else:
batch_predicted = x2
losses = {}
# apply mask and complete image
batch_complete = batch_predicted * mask + batch_incomplete * (1 - mask)
# local patches
local_patch_batch_pos = local_patch(batch_pos, bbox)
local_patch_x1 = local_patch(x1, bbox)
local_patch_x2 = local_patch(x2, bbox)
local_patch_batch_complete = local_patch(batch_complete, bbox)
local_patch_mask = local_patch(mask, bbox)
l1_alpha = config.COARSE_L1_ALPHA
losses["l1_loss"] = l1_alpha * F.mean(
F.absolute(local_patch_batch_pos - local_patch_x1) *
spatial_discounting_mask(config, self.xp))
if not config.PRETRAIN_COARSE_NETWORK:
losses['l1_loss'] += F.mean(
F.absolute(local_patch_batch_pos - local_patch_x2) *
spatial_discounting_mask(config, self.xp))
losses['ae_loss'] = l1_alpha * F.mean(
F.absolute(batch_pos - x1) * (1. - mask))
if not config.PRETRAIN_COARSE_NETWORK:
losses['ae_loss'] += F.mean(
F.absolute(batch_pos - x2) * (1. - mask))
losses['ae_loss'] /= F.mean(1. - mask)
# gan
batch_pos_neg = F.concat([batch_pos, batch_complete], axis=0)
# local deterministic patch
local_patch_batch_pos_neg = F.concat(
[local_patch_batch_pos, local_patch_batch_complete], 0)
if config.GAN_WITH_MASK:
batch_pos_neg = F.concat([batch_pos_neg, mask], axis=1)
# wgan with gradient penalty
if config.GAN == 'wgan_gp':
# seperate gan
pos_neg_local, pos_neg_global = self.discriminator(
local_patch_batch_pos_neg, batch_pos_neg)
pos_local, neg_local = F.split_axis(pos_neg_local, 2, axis=0)
pos_global, neg_global = F.split_axis(pos_neg_global, 2, axis=0)
# wgan loss
g_loss_local, d_loss_local = gan_wgan_loss(pos_local, neg_local)
g_loss_global, d_loss_global = gan_wgan_loss(
pos_global, neg_global)
losses[
'g_loss'] = config.GLOBAL_WGAN_LOSS_ALPHA * g_loss_global + g_loss_local
losses['d_loss'] = d_loss_global + d_loss_local
# gp
interpolates_local = random_interpolates(
local_patch_batch_pos, local_patch_batch_complete)
interpolates_global = random_interpolates(batch_pos,
batch_complete)
dout_local, dout_global = self.discriminator(
interpolates_local, interpolates_global)
# apply penalty
penalty_local = gradients_penalty(interpolates_local,
dout_local,
mask=local_patch_mask)
penalty_global = gradients_penalty(interpolates_global,
dout_global,
mask=mask)
losses['gp_loss'] = config.WGAN_GP_LAMBDA * (penalty_local +
penalty_global)
losses['d_loss'] = losses['d_loss'] + losses['gp_loss']
if config.PRETRAIN_COARSE_NETWORK:
losses['g_loss'] = 0
else:
losses['g_loss'] = config.GAN_LOSS_ALPHA * losses['g_loss']
losses['g_loss'] += config.L1_LOSS_ALPHA * losses['l1_loss']
if config.AE_LOSS:
losses['g_loss'] += config.AE_LOSS_ALPHA * losses['ae_loss']
return losses
def loss_hinge_gen(dis_fake):
loss = -F.mean(dis_fake)
return loss
def gan_sngan_loss(pos, neg, d_loss_only=False):
# SN-PatchGAN loss with hinge loss
d_loss = F.mean(F.relu(1 - pos) + F.relu(1 + neg))
g_loss = None if d_loss_only else -F.mean(neg)
return g_loss, d_loss
def calculate_logistic_loss(self, y, t):
xp = chainer.cuda.get_array_module(t)
if xp != numpy:
xp.cuda.Device(t.device).use()
nr_mix = y.shape[1] // 3
logit_probs = y[:, :nr_mix]
means = y[:, nr_mix:2 * nr_mix]
log_scales = y[:, 2 * nr_mix:3 * nr_mix]
log_scales = F.maximum(
log_scales, self.scalar_to_tensor(log_scales, self.log_scale_min))
t = F.broadcast_to(127.5 * t, means.shape)
centered_t = t - means
inv_std = F.exp(-log_scales)
plus_in = inv_std * (centered_t + 127.5 / (self.quantize - 1))
cdf_plus = F.sigmoid(plus_in)
min_in = inv_std * (centered_t - 127.5 / (self.quantize - 1))
cdf_min = F.sigmoid(min_in)
log_cdf_plus = plus_in - F.softplus(plus_in)
log_one_minus_cdf_min = -F.softplus(min_in)
cdf_delta = cdf_plus - cdf_min
# mid_in = inv_std * centered_t
# log_pdf_mid = mid_in - log_scales - 2 * F.softplus(mid_in)
log_probs = F.where(
# condition
t.array < self.scalar_to_tensor(t, 127.5 * -0.999),
# true
log_cdf_plus,
# false
F.where(
# condition
t.array > self.scalar_to_tensor(t, 127.5 * 0.999),
# true
log_one_minus_cdf_min,
# false
F.log(
F.maximum(cdf_delta,
self.scalar_to_tensor(cdf_delta, 1e-12)))
# F.where(
# # condition
# cdf_delta.array > self.scalar_to_tensor(cdf_delta, 1e-5),
# # true
# F.log(F.maximum(
# cdf_delta, self.scalar_to_tensor(cdf_delta, 1e-12))),
# # false
# log_pdf_mid - self.xp.log((self.quantize - 1) / 2))
))
log_probs = log_probs + F.log_softmax(logit_probs)
loss = -F.mean(F.logsumexp(log_probs, axis=1))
return loss
def square_loss(ys, ts):
# return F.mean(F.sqrt((ys - ts) ** 2 + 1e-5), axis=(0, 2))
return F.mean((ys - ts) ** 2 + 1e-5, axis=(0, 2))
def mean_clipped_loss(y, t):
return F.mean(F.huber_loss(y, t, delta=1.0, reduce='no'))
def mean_clipped_loss(y, t):
return F.mean(F.huber_loss(y, t, delta=1.0, reduce='no'))
def __call__(self, x, ys, yb): s = broadcast_to(ys.reshape(ys.shape + (1, 1)), ys.shape + x.shape[2:]) b = broadcast_to(yb.reshape(yb.shape + (1, 1)), yb.shape + x.shape[2:]) e = x - broadcast_to(mean(x, axis=1, keepdims=True), x.shape) sd = broadcast_to(sqrt(mean(e ** 2, axis=1, keepdims=True) + 1e-8), x.shape) return s * e / sd + b
def feature_vector_normalization(x, eps=1e-8):
# x: (B, C, H, W)
alpha = 1.0 / F.sqrt(F.mean(x * x, axis=1, keepdims=True) + eps)
return F.broadcast_to(alpha, x.data.shape) * x
inp = prepare_dataset(inp)
input_box.append(inp)
img = dir_path + "/" + str(index) + ".png"
img = prepare_dataset(img)
frame_box.append(img)
x = chainer.as_variable(xp.array(input_box).astype(xp.float32))
t = chainer.as_variable(xp.array(frame_box).astype(xp.float32))
embed = feature_extractor(t) - feature_extractor(x)
c = feature_embed(embed)
z = F.concat([x, c], axis=1)
y = predictor(z)
y_dis = discriminator_content(y)
t_dis = discriminator_content(t)
dis_loss = F.mean(F.softplus(-t_dis)) + F.mean(F.softplus(y_dis))
c_g = feature_extractor(y) - feature_extractor(make_diff(y))
c_dis = discriminator_sequence(embed)
c_g_dis = discriminator_sequence(c_g)
dis_loss += F.mean(F.softplus(-c_dis)) + F.mean(F.softplus(c_g_dis))
c_g.unchain_backward()
discriminator_content.cleargrads()
discriminator_sequence.cleargrads()
dis_loss.backward()
dis_c_opt.update()
dis_s_opt.update()
dis_loss.unchain_backward()
def loss_softmax_cross_entropy(predict, ground_truth):
eps = 1e-16
cross_entropy = -F.mean(F.log(predict + eps) * ground_truth)
return cross_entropy
def update_core(self):
gen_optimizer = self.get_optimizer('opt_gen')
dis_optimizer = self.get_optimizer('opt_dis')
xp = self.gen.xp
for i in range(self.n_dis):
batch = self.get_iterator('main').next()
batchsize = len(batch)
x = []
for j in range(batchsize):
x.append(np.asarray(batch[j]).astype("f"))
x_real = Variable(xp.asarray(x))
h_real = self.dis(x_real)
z = Variable(xp.asarray(self.gen.make_hidden(batchsize)))
x_fake1 = self.gen(z)
h_fake1 = self.dis(x_fake1)
z2 = Variable(xp.asarray(self.gen.make_hidden(batchsize)))
x_fake2 = self.gen(z2)
h_fake2 = self.dis(x_fake2)
def l2_distance(a, b):
return F.sqrt(F.sum((a - b) ** 2, axis=1, keepdims=True))
def backward_l2_distance(g, a, b):
out = F.broadcast_to(l2_distance(a, b), a.data.shape)
g = F.broadcast_to(g, a.data.shape)
return g * (a - b) / out, g * (b - a) / out
def energy_distance(r, f1, f2):
ret = l2_distance(r, f1)
ret += l2_distance(r, f2)
ret -= l2_distance(f1, f2)
return F.mean(ret)
def critic(a, b):
return l2_distance(a, b) - l2_distance(a, xp.zeros_like(a.data))
def backward_critic(g, a, b):
ga0, gb0 = backward_l2_distance(g, a, b)
ga1, gb1 = backward_l2_distance(g, a, xp.zeros_like(a.data))
return ga0 - ga1, gb0 - gb1
critic_real = critic(h_real, h_fake2)
critic_fake = critic(h_fake1, h_fake2)
loss_surrogate = F.mean(critic_real - critic_fake)
if i == 0:
loss_gen = energy_distance(h_real, h_fake1, h_fake2)
self.gen.cleargrads()
loss_gen.backward()
gen_optimizer.update()
chainer.reporter.report({'loss_gen': loss_gen})
x_fake1.unchain_backward()
x_fake2.unchain_backward()
eps = xp.random.uniform(0, 1, size=batchsize).astype("f")[:, None, None, None]
x_mid = eps * x_real + (1.0 - eps) * x_fake1
h_mid = Variable(self.dis(x_mid).data)
critic_mid = critic(h_mid, h_fake2.data)
# calc gradient penalty
g = Variable(xp.ones_like(critic_mid.data))
dydh, _ = backward_critic(g, h_mid, h_fake2.data)
dydx = self.dis.differentiable_backward(dydh)
dydx_norm = F.sqrt(F.sum(dydx ** 2, axis=(1, 2, 3)))
loss_gp = self.lam * F.mean_squared_error(dydx_norm, xp.ones_like(dydx_norm.data))
self.dis.cleargrads()
(-loss_surrogate).backward()
loss_gp.backward()
dis_optimizer.update()
chainer.reporter.report({'loss_dis': loss_surrogate})
chainer.reporter.report({'loss_gp': loss_gp})
chainer.reporter.report({'g': F.mean(dydx_norm)})
def run_n_games(optimizer, learner, opponent, num_games):
states.default_start_position()
# Create one list of features (aka state tensors) and one of moves for each game being played.
features1_tensors = [[] for _ in range(num_games)]
features2_tensors = [[] for _ in range(num_games)]
labels_tensors = [[] for _ in range(num_games)]
values_tensors = [[] for _ in range(num_games)]
# List of booleans indicating whether the 'learner' player won.
learner_won = [None] * num_games
# Start all odd games with moves by 'opponent'. Even games will have 'learner' black.
learner_color = [BLACK if i % 2 == 0 else WHITE for i in range(num_games)]
odd_features1 = np.empty((num_games, 2 * 14, 9, 9), dtype=np.float32)
odd_features2 = np.empty((num_games, 2 * MAX_PIECES_IN_HAND_SUM + 1, 9, 9),
dtype=np.float32)
states.make_odd_input_features(odd_features1, odd_features2)
x1 = Variable(cuda.to_gpu(odd_features1))
x2 = Variable(cuda.to_gpu(odd_features2))
with chainer.no_backprop_mode():
with chainer.using_config('train', False):
y = opponent(x1, x2)
y_data = cuda.to_cpu(y.data)
states.do_odd_moves(y_data)
current = learner
other = opponent
unfinished_states_num = num_games
move_number_sum = 0
while unfinished_states_num > 0:
move_number_sum += unfinished_states_num
# Get next moves by current player for all unfinished states.
features1 = np.empty((unfinished_states_num, FEATURES1_NUM, 9, 9),
dtype=np.float32)
features2 = np.empty((unfinished_states_num, FEATURES2_NUM, 9, 9),
dtype=np.float32)
unfinished_list = states.make_unfinished_input_features(
features1, features2)
x1 = Variable(cuda.to_gpu(features1))
x2 = Variable(cuda.to_gpu(features2))
with chainer.no_backprop_mode():
with chainer.using_config('train', False):
y = current(x1, x2)
y_data = cuda.to_cpu(y.data)
labels = np.empty((unfinished_states_num), dtype=np.int32)
values = np.empty((unfinished_states_num), dtype=np.float32)
unfinished_states_num = states.do_unfinished_moves_and_eval(
current is learner, y_data, labels, values)
# 特徴を保存
if current is learner:
for i, idx in enumerate(unfinished_list):
features1_tensors[idx].append(features1[i])
features2_tensors[idx].append(features2[i])
labels_tensors[idx].append(labels[i])
values_tensors[idx].append(values[i])
# Swap 'current' and 'other' for next turn.
current, other = other, current
learner_won = np.empty(num_games, dtype=np.int32)
states.get_learner_wons(learner_won)
# Train on all game's results
features1_tensor_all = []
features2_tensor_all = []
labels_tensor_all = []
rewards_tensor_all = []
for features1_tensor, features2_tensor, labels_tensor, values_tensor, won in zip(
features1_tensors, features2_tensors, labels_tensors,
values_tensors, learner_won.astype(np.float32)):
features1_tensor_all.extend(features1_tensor)
features2_tensor_all.extend(features2_tensor)
labels_tensor_all.extend(labels_tensor)
rewards_tensor_all.extend(
list(won - np.array(values_tensor, dtype=np.float32)))
x1 = Variable(cuda.to_gpu(np.array(features1_tensor_all,
dtype=np.float32)))
x2 = Variable(cuda.to_gpu(np.array(features2_tensor_all,
dtype=np.float32)))
t = Variable(cuda.to_gpu(np.array(labels_tensor_all, dtype=np.int32)))
z = Variable(cuda.to_gpu(np.array(rewards_tensor_all, dtype=np.float32)))
y = learner(x1, x2)
learner.cleargrads()
loss = F.mean(F.softmax_cross_entropy(y, t, reduce='no') * z)
loss.backward()
optimizer.update()
# Return the win ratio.
return np.average(
learner_won), float(move_number_sum) / num_games, loss.data
def __call__(self, h, adj):
xp = self.xp
# (minibatch, atom, channel)
mb, atom, ch = h.shape
# (minibatch, atom, EDGE_TYPE * heads * out_dim)
h = self.message_layer(h)
# (minibatch, atom, EDGE_TYPE, heads, out_dim)
h = functions.reshape(h, (mb, atom, self.n_edge_types, self.n_heads,
self.out_channels))
# concat all pairs of atom
# (minibatch, 1, atom, heads, out_dim)
h_i = functions.reshape(h, (mb, 1, atom, self.n_edge_types,
self.n_heads, self.out_channels))
# (minibatch, atom, atom, heads, out_dim)
h_i = functions.broadcast_to(h_i, (mb, atom, atom, self.n_edge_types,
self.n_heads, self.out_channels))
# (minibatch, atom, 1, EDGE_TYPE, heads, out_dim)
h_j = functions.reshape(h, (mb, atom, 1, self.n_edge_types,
self.n_heads, self.out_channels))
# (minibatch, atom, atom, EDGE_TYPE, heads, out_dim)
h_j = functions.broadcast_to(h_j, (mb, atom, atom, self.n_edge_types,
self.n_heads, self.out_channels))
# (minibatch, atom, atom, EDGE_TYPE, heads, out_dim * 2)
e = functions.concat([h_i, h_j], axis=5)
# (minibatch, EDGE_TYPE, heads, atom, atom, out_dim * 2)
e = functions.transpose(e, (0, 3, 4, 1, 2, 5))
# (minibatch * EDGE_TYPE * heads, atom * atom, out_dim * 2)
e = functions.reshape(e, (mb * self.n_edge_types * self.n_heads,
atom * atom, self.out_channels * 2))
# (minibatch * EDGE_TYPE * heads, atom * atom, 1)
e = self.attention_layer(e)
# (minibatch, EDGE_TYPE, heads, atom, atom)
e = functions.reshape(e, (mb, self.n_edge_types, self.n_heads, atom,
atom))
e = functions.leaky_relu(e, self.negative_slope)
# (minibatch, EDGE_TYPE, atom, atom)
if isinstance(adj, chainer.Variable):
cond = adj.array.astype(xp.bool)
else:
cond = adj.astype(xp.bool)
# (minibatch, EDGE_TYPE, 1, atom, atom)
cond = xp.reshape(cond, (mb, self.n_edge_types, 1, atom, atom))
# (minibatch, EDGE_TYPE, heads, atom, atom)
cond = xp.broadcast_to(cond, e.array.shape)
# TODO(mottodora): find better way to ignore non connected
e = functions.where(cond, e,
xp.broadcast_to(xp.array(-10000), e.array.shape)
.astype(xp.float32))
# In Relational Graph Attention Networks eq.(7)
# ARGAT: take the softmax over the logits across node neighborhoods
# irrespective of relation
if self.softmax_mode == 'across':
# (minibatch, heads, atom, EDGE_TYPE, atom)
e = functions.transpose(e, (0, 2, 3, 1, 4))
# (minibatch, heads, atom, EDGE_TYPE * atom)
e = functions.reshape(e, (mb, self.n_heads, atom,
self.n_edge_types * atom))
# (minibatch, heads, atom, EDGE_TYPE * atom)
alpha = functions.softmax(e, axis=3)
if self.dropout_ratio >= 0:
alpha = functions.dropout(alpha, ratio=self.dropout_ratio)
# (minibatch, heads, atom, EDGE_TYPE, atom)
alpha = functions.reshape(alpha, (mb, self.n_heads, atom,
self.n_edge_types, atom))
# (minibatch, EDGE_TYPE, heads, atom, atom)
alpha = functions.transpose(alpha, (0, 3, 1, 2, 4))
# In Relational Graph Attention Networks eq.(6)
# WIRGAT: take the softmax over the logits independently for each
# relation
elif self.softmax_mode == 'within':
alpha = functions.softmax(e, axis=4)
if self.dropout_ratio >= 0:
alpha = functions.dropout(alpha, ratio=self.dropout_ratio)
else:
raise ValueError("{} is invalid. Please use 'across' or 'within'"
.format(self.softmax_mode))
# before: (minibatch, atom, EDGE_TYPE, heads, out_dim)
# after: (minibatch, EDGE_TYPE, heads, atom, out_dim)
h = functions.transpose(h, (0, 2, 3, 1, 4))
# (minibatch, EDGE_TYPE, heads, atom, out_dim)
h_new = functions.matmul(alpha, h)
# (minibatch, heads, atom, out_dim)
h_new = functions.sum(h_new, axis=1)
if self.concat_heads:
# (heads, minibatch, atom, out_dim)
h_new = functions.transpose(h_new, (1, 0, 2, 3))
# (minibatch, atom, heads * out_dim)
h_new = functions.concat(h_new, axis=2)
else:
# (minibatch, atom, out_dim)
h_new = functions.mean(h_new, axis=1)
return h_new
def energy_distance(r, f1, f2):
ret = l2_distance(r, f1)
ret += l2_distance(r, f2)
ret -= l2_distance(f1, f2)
return F.mean(ret)
def __call__(self, xs):
y_hat, input_embeds = self.predict(xs)
loss = 0.5 * F.sum((y_hat - input_embeds) ** 2, axis=1)
loss = F.mean(loss)
reporter.report({'loss': loss.data}, self)
return loss
def compute_marginal_entropy(self, p_batch):
return self.compute_entropy(functions.mean(p_batch, axis=0))
# train
itr = 0
sum_loss = 0
eval_interval = 1000
for e in range(args.epoch):
np.random.shuffle(train_data)
itr_epoch = 0
sum_loss_epoch = 0
for i in range(0, len(train_data) - args.batchsize, args.batchsize):
x1, x2, t, z = mini_batch(train_data[i:i + args.batchsize])
y = model(x1, x2)
model.cleargrads()
loss = F.mean(F.softmax_cross_entropy(y, t, reduce='no') * z)
loss.backward()
optimizer.update()
itr += 1
sum_loss += loss.data
itr_epoch += 1
sum_loss_epoch += loss.data
# print train loss
if optimizer.t % eval_interval == 0:
logging.info('epoch = {}, iteration = {}, loss = {}'.format(
optimizer.epoch + 1, optimizer.t, sum_loss / itr))
itr = 0
sum_loss = 0
def main():
parser = argparse.ArgumentParser(description='GradNorm')
parser.add_argument('--gpu', '-g', type=int, default=-1)
parser.add_argument('--n-iter', '-it', type=int, default=5000)
parser.add_argument('--mode', '-m', choices=('grad_norm', 'equal_weight'),
default='grad_norm')
args = parser.parse_args()
np.random.seed(123)
sigmas = [1, 10]
n_task = len(sigmas)
epsilons = np.random.normal(
scale=3.5, size=(n_task, 100, 250)).astype(np.float32)
dataset = RegressionDataset(sigmas, epsilons)
model = RegressionTrainChain(RegressionChain(n_task))
if args.gpu >= 0:
chainer.backends.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
optimizer = chainer.optimizers.Adam(alpha=1e-2)
optimizer.setup(model)
train_iter = chainer.iterators.SerialIterator(dataset, 200)
xp = model.xp
weights = []
task_losses = []
loss_ratios = []
final_layer_names = ['task_{}'.format(i) for i in range(n_task)]
for t in range(args.n_iter):
batch = train_iter.next()
x, ts = chainer.dataset.convert.concat_examples(batch, device=args.gpu)
task_loss = model(x, ts)
weighted_task_loss = model.weight * task_loss
if t == 0:
initial_task_loss = task_loss.data
loss = F.mean(weighted_task_loss)
model.cleargrads()
loss.backward()
# Ignore a gradient to the coefficient vector, which
# is computed from the standard loss.
model.weight.cleargrad()
if args.mode == 'grad_norm':
# Use |\nabla_W w_i * L_i | = w_i |\nabla_W L_i|
gygw_norms = []
for i, layer_name in enumerate(final_layer_names):
l = getattr(model.model, layer_name)
gygw = chainer.grad([task_loss[i]], [l.W])[0].data
gygw_norms.append(xp.linalg.norm(gygw))
gygw_norms = xp.stack(gygw_norms)
norms = model.weight * gygw_norms
alpha = 0.16
mean_norm = xp.mean(norms.data)
loss_ratio = task_loss.data / initial_task_loss
inverse_train_rate = loss_ratio / xp.mean(loss_ratio)
diff = norms - (inverse_train_rate ** alpha) * mean_norm
grad_norm_loss = F.mean(F.absolute(diff))
grad_norm_loss.backward()
# For debugging purpose only
# from chainer import computational_graph
# import os
# cg = computational_graph.build_computational_graph(
# [grad_norm_loss]).dump()
# with open('grad_weight_loss_cg', 'w') as f:
# f.write(cg)
optimizer.update()
# Renormalize
normalize_coeff = n_task / xp.sum(model.weight.data)
model.weight.data[:] = model.weight.data * normalize_coeff
# Record
task_losses.append(chainer.backends.cuda.to_cpu(task_loss.data))
loss_ratios.append(np.mean(task_losses[-1] / task_losses[0]))
weights.append(chainer.backends.cuda.to_cpu(model.weight.data))
if t % 100 == 0:
print('{}/{}: loss_ratio={}, weights={} task_loss={}'.format(
t, args.n_iter, loss_ratios[-1], model.weight.data, task_loss.data))
task_losses = np.array(task_losses)
weights = np.array(weights)
fig = plt.figure()
ax1 = fig.add_subplot(1, 4, 1)
ax1.set_title('loss (task 0)')
ax2 = fig.add_subplot(1, 4, 2)
ax2.set_title('loss (task 1)')
ax3 = fig.add_subplot(1, 4, 3)
ax3.set_title('sum of normalized losses')
ax4 = fig.add_subplot(1, 4, 4)
ax4.set_title('change of weights over time')
ax1.plot(task_losses[:, 0])
ax2.plot(task_losses[:, 1])
ax3.plot(loss_ratios)
ax4.plot(weights[:, 0])
ax4.plot(weights[:, 1])
plt.show()
