def check_forward(self, log_pi_data, tau):
log_pi = chainer.Variable(log_pi_data)
y = functions.gumbel_softmax(log_pi, tau=tau)
# Only checks dtype and shape because its result contains noise
self.assertEqual(y.dtype, numpy.float32)
self.assertEqual(y.shape, log_pi.shape)
self.assertEqual(cuda.get_array_module(y),
cuda.get_array_module(log_pi))
def check_forward(self, log_pi_data, tau):
log_pi = chainer.Variable(log_pi_data)
y = functions.gumbel_softmax(log_pi, tau=tau)
# Only checks dtype and shape because its result contains noise
self.assertEqual(y.dtype, numpy.float32)
self.assertEqual(y.shape, log_pi.shape)
self.assertEqual(
cuda.get_array_module(y),
cuda.get_array_module(log_pi))
def test(iterator, gpu, timesteps, encoder, decoder, rel_send, rel_rec,
edge_types, temp, var):
nll_test = []
kl_test = []
edge_accuracies = []
node_mses = []
chainer.config.train = False
chainer.config.enable_backprop = False
while True:
inputs = iterator.next()
node_features, edge_labels = dataset.concat_examples(inputs,
device=gpu)
data_encoder = node_features[:, :, :timesteps, :]
data_decoder = node_features[:, :, -timesteps:, :]
# logits: [batch_size, num_edges, edge_types]
logits = encoder(data_encoder, rel_send,
rel_rec) # inverse func. of softmax
edges = F.gumbel_softmax(logits, tau=temp, axis=2)
edge_probs = F.softmax(logits, axis=2)
# edges, edge_probs: [batch_size, num_edges, edge_types]
# validation output uses teacher forcing
output = decoder(data_decoder, edges, rel_rec, rel_send, 1)
target = data_decoder[:, :, 1:, :]
num_nodes = node_features.shape[1]
loss_nll = get_nll_gaussian(output, target, var)
loss_kl = get_kl_categorical_uniform(edge_probs, num_nodes, edge_types)
nll_test.append(float(loss_nll.array))
kl_test.append(float(loss_kl.array))
edge_accuracy = get_edge_accuracy(logits.array, edge_labels)
edge_accuracies.append(edge_accuracy)
node_mse = float(F.mean_squared_error(output, target).array)
node_mses.append(node_mse)
if iterator.is_new_epoch:
break
put_log(iterator.epoch, np.mean(nll_test), np.mean(kl_test),
np.mean(edge_accuracies), np.mean(node_mses), 'test')
chainer.config.train = True
chainer.config.enable_backprop = True
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 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(iterator, gpu, timesteps, encoder, decoder, rel_send, rel_rec, edge_types, temp, var):
nll_test = []
kl_test = []
edge_accuracies = []
node_mses = []
chainer.config.train = False
chainer.config.enable_backprop = False
while True:
inputs = iterator.next()
node_features, edge_labels = dataset.concat_examples(inputs, device=gpu)
data_encoder = node_features[:, :, :timesteps, :]
data_decoder = node_features[:, :, -timesteps:, :]
# logits: [batch_size, num_edges, edge_types]
logits = encoder(data_encoder, rel_send, rel_rec) # inverse func. of softmax
edges = F.gumbel_softmax(logits, tau=temp, axis=2)
edge_probs = F.softmax(logits, axis=2)
# edges, edge_probs: [batch_size, num_edges, edge_types]
# validation output uses teacher forcing
output = decoder(data_decoder, edges, rel_rec, rel_send, 1)
target = data_decoder[:, :, 1:, :]
num_nodes = node_features.shape[1]
loss_nll = get_nll_gaussian(output, target, var)
loss_kl = get_kl_categorical_uniform(edge_probs, num_nodes, edge_types)
nll_test.append(float(loss_nll.array))
kl_test.append(float(loss_kl.array))
edge_accuracy = get_edge_accuracy(logits.array, edge_labels)
edge_accuracies.append(edge_accuracy)
node_mse = float(F.mean_squared_error(output, target).array)
node_mses.append(node_mse)
if iterator.is_new_epoch:
break
put_log(iterator.epoch, np.mean(nll_test), np.mean(kl_test),
np.mean(edge_accuracies), np.mean(node_mses), 'test')
chainer.config.train = True
chainer.config.enable_backprop = True
def get_dealer_sampling(N_pic=100, imgH=64, imgW=64, N_card=4):
thres = [0.99995, 0.9999, 0.9998, 0.9995] #*512で13,26,52,131個相当
#＜ランダム点画像の生成＞
img_r = xp.random.rand(N_pic,
imgW * imgH).astype(np.float32) #100枚分の0-1乱数作成
img_p = xp.zeros(
(N_card, N_pic, imgW * imgH)).astype(np.float32) #4*100枚分のイメージメモリ確保
for i, thre in enumerate(thres): #閾値よりも高いものだけ1を代入
img_p[i][img_r >= thre] = 1
#点画像変形 (N_card, N_pic, imgW*imgH,) ⇒ (N_pic, imgW*imgH, N_card)
img_p = chainer.Variable(img_p.transpose((1, 2, 0)))
#＜サンプリング係数の生成＞
#100個の「１」を作成
x_one = xp.ones((N_pic, 1), dtype=np.float32)
#「１」をディーラーを通したあとsoftmaxで0-1確率にする
card_prob = F.softmax(Md['de'](x_one))
#gumbel_softmaxを通してサンプリング
card_gum = F.gumbel_softmax(F.log(card_prob), tau=0.2)
#サンプリング係数の画像化 (N_pic, N_card) ⇒ (N_pic, imgW*imgH, N_card)
card_gum_b = F.broadcast_to(F.reshape(card_gum, (N_pic, 1, N_card)),
img_p.shape)
#＜ランダム点画像とサンプリング係数画像の合成＞
#ランダム点画像とサンプリング係数をかけて、合成(sum)し、２次元画像へ変形
img_p_sum = F.reshape(F.sum(img_p * card_gum_b, axis=2),
(N_pic, 1, imgH, imgW))
#点⇒ガウス球へ変形
img_core = Md['decon_core'](img_p_sum) * 255
img_core = F.broadcast_to(img_core, (N_pic, 3, imgH, imgW))
return img_core
chainer.config.train = False
chainer.config.enable_backprop = False
for i in range(5):
inputs = test_iter.next()
node_features, edge_labels = dataset.concat_examples(inputs,
device=args.gpu)
data_encoder = node_features[:, :, :train_args['timesteps'], :]
data_decoder = node_features[:, :, train_args['timesteps']:, :]
# logits: [batch_size, num_edges, edge_types]
logits = encoder(data_encoder, rel_send,
rel_rec) # inverse func. of softmax
edges = F.gumbel_softmax(logits, tau=train_args['temp'],
axis=2) # edge sampling
edge_probs = F.softmax(logits, axis=2)
# edges, edge_probs: [batch_size, num_edges, edge_types]
# validation output uses teacher forcing
output = decoder(data_decoder, edges, rel_rec, rel_send,
data_decoder.shape[2])
fig = plt.figure()
plt.tight_layout()
plt.xlabel('x')
plt.ylabel('y')
plt.title(
'transparent: given, solid: prediction, dashed: ground-truth')
prop_cycle = plt.rcParams['axes.prop_cycle']
colors = prop_cycle.by_key()['color']
def train(
iterator, gpu, encoder, decoder, enc_optim, dec_optim, rel_send, rel_rec, edge_types,
temp, prediction_steps, var, out, benchmark, lr_decay, gamma):
iter_i = 0
edge_accuracies = []
node_mses = []
nll_train = []
kl_train = []
logger = logging.getLogger(__name__)
while True:
inputs = iterator.next()
node_features, edge_labels = dataset.concat_examples(inputs, device=gpu)
# logits: [batch_size, num_edges, edge_types]
logits = encoder(node_features, rel_send, rel_rec) # inverse func. of softmax
edges = F.gumbel_softmax(logits, tau=temp, axis=2)
edge_probs = F.softmax(logits, axis=2)
# edges, edge_probs: [batch_size, num_edges, edge_types]
if isinstance(decoder, decoders.MLPDecoder):
output = decoder(
node_features, edges, rel_rec, rel_send, prediction_steps)
elif isinstance(decoder, decoders.RNNDecoder):
output = decoder(
node_features, edges, rel_rec, rel_send, 100,
burn_in=True,
burn_in_steps=args.timesteps - args.prediction_steps)
target = node_features[:, :, 1:, :]
num_nodes = node_features.shape[1]
loss_nll = get_nll_gaussian(output, target, var)
loss_kl = get_kl_categorical_uniform(edge_probs, num_nodes, edge_types)
loss = loss_nll + loss_kl
nll_train.append(float(loss_nll.array))
kl_train.append(float(loss_kl.array))
edge_accuracy = get_edge_accuracy(logits.array, edge_labels)
edge_accuracies.append(edge_accuracy)
node_mse = float(F.mean_squared_error(output, target).array)
node_mses.append(node_mse)
encoder.cleargrads()
decoder.cleargrads()
loss.backward()
enc_optim.update()
dec_optim.update()
# Exit after 10 iterations when benchmark mode is ON
iter_i += 1
if benchmark:
put_log(iterator.epoch, np.mean(nll_train), np.mean(kl_train),
np.mean(edge_accuracies), np.mean(node_mses))
if iter_i == 10:
exit()
if iterator.is_new_epoch:
break
if not os.path.exists(os.path.join(out, 'graph.dot')):
with open(os.path.join(out, 'graph.dot'), 'w') as o:
g = computational_graph.build_computational_graph([loss])
o.write(g.dump())
if iterator.is_new_epoch:
put_log(iterator.epoch, np.mean(nll_train), np.mean(kl_train), np.mean(edge_accuracies), np.mean(node_mses))
serializers.save_npz(os.path.join(out, 'encoder_epoch-{}.npz'.format(iterator.epoch)), encoder)
serializers.save_npz(os.path.join(out, 'decoder_epoch-{}.npz'.format(iterator.epoch)), decoder)
serializers.save_npz(os.path.join(out, 'enc_optim_epoch-{}.npz'.format(iterator.epoch)), enc_optim)
serializers.save_npz(os.path.join(out, 'dec_optim_epoch-{}.npz'.format(iterator.epoch)), dec_optim)
if iterator.epoch % lr_decay == 0:
enc_optim.alpha *= gamma
dec_optim.alpha *= gamma
logger.info('alpha of enc_optim: {}'.format(enc_optim.alpha))
logger.info('alpha of dec_optim: {}'.format(dec_optim.alpha))
def train(iterator, gpu, encoder, decoder, enc_optim, dec_optim, rel_send,
rel_rec, edge_types, temp, prediction_steps, var, out, benchmark,
lr_decay, gamma):
iter_i = 0
edge_accuracies = []
node_mses = []
nll_train = []
kl_train = []
logger = logging.getLogger(__name__)
while True:
inputs = iterator.next()
node_features, edge_labels = dataset.concat_examples(inputs,
device=gpu)
# logits: [batch_size, num_edges, edge_types]
logits = encoder(node_features, rel_send,
rel_rec) # inverse func. of softmax
edges = F.gumbel_softmax(logits, tau=temp, axis=2)
edge_probs = F.softmax(logits, axis=2)
# edges, edge_probs: [batch_size, num_edges, edge_types]
if isinstance(decoder, decoders.MLPDecoder):
output = decoder(node_features, edges, rel_rec, rel_send,
prediction_steps)
elif isinstance(decoder, decoders.RNNDecoder):
output = decoder(node_features,
edges,
rel_rec,
rel_send,
100,
burn_in=True,
burn_in_steps=args.timesteps -
args.prediction_steps)
target = node_features[:, :, 1:, :]
num_nodes = node_features.shape[1]
loss_nll = get_nll_gaussian(output, target, var)
loss_kl = get_kl_categorical_uniform(edge_probs, num_nodes, edge_types)
loss = loss_nll + loss_kl
nll_train.append(float(loss_nll.array))
kl_train.append(float(loss_kl.array))
edge_accuracy = get_edge_accuracy(logits.array, edge_labels)
edge_accuracies.append(edge_accuracy)
node_mse = float(F.mean_squared_error(output, target).array)
node_mses.append(node_mse)
encoder.cleargrads()
decoder.cleargrads()
loss.backward()
enc_optim.update()
dec_optim.update()
# Exit after 10 iterations when benchmark mode is ON
iter_i += 1
if benchmark:
put_log(iterator.epoch, np.mean(nll_train), np.mean(kl_train),
np.mean(edge_accuracies), np.mean(node_mses))
if iter_i == 10:
exit()
if iterator.is_new_epoch:
break
if not os.path.exists(os.path.join(out, 'graph.dot')):
with open(os.path.join(out, 'graph.dot'), 'w') as o:
g = computational_graph.build_computational_graph([loss])
o.write(g.dump())
if iterator.is_new_epoch:
put_log(iterator.epoch, np.mean(nll_train), np.mean(kl_train),
np.mean(edge_accuracies), np.mean(node_mses))
serializers.save_npz(
os.path.join(out, 'encoder_epoch-{}.npz'.format(iterator.epoch)),
encoder)
serializers.save_npz(
os.path.join(out, 'decoder_epoch-{}.npz'.format(iterator.epoch)),
decoder)
serializers.save_npz(
os.path.join(out, 'enc_optim_epoch-{}.npz'.format(iterator.epoch)),
enc_optim)
serializers.save_npz(
os.path.join(out, 'dec_optim_epoch-{}.npz'.format(iterator.epoch)),
dec_optim)
if iterator.epoch % lr_decay == 0:
enc_optim.alpha *= gamma
dec_optim.alpha *= gamma
logger.info('alpha of enc_optim: {}'.format(enc_optim.alpha))
logger.info('alpha of dec_optim: {}'.format(dec_optim.alpha))
