def train_gen_step(self,dynamic_X,static_X,targ,pretrain=False):
with tf.GradientTape() as tape:
predictions = self._generator(dynamic_X,static_X)
if pretrain:
loss = loss_function(tf.expand_dims(targ,2),predictions)
else:
## 如果不是预训练，loss的目标是尽量使discriminator出错
fake_Y = tf.reshape(predictions,(predictions.shape[0],predictions.shape[1]))
# print(')
fake_output = self._discriminator(dynamic_X,static_X,fake_Y)
## 为了保证梯度不下降,使用两种loss进行混淆
loss = generator_loss(fake_output,self._mode)
variables = self._generator.trainable_variables
gradients = tape.gradient(loss,variables)
if pretrain:
self._gen_pre_optimizer.apply_gradients(zip(gradients,variables))
else:
self._gen_optimizer.apply_gradients(zip(gradients,variables))
return loss
def train_discriminator_step(self,
dynamic_X,
static_X,
real_Y,
fake_Y,
pretrain=False):
##获得所有数据之后进行训练
with tf.GradientTape() as tape:
## 使用discriminator进行预测
fake_output = self._discriminator(dynamic_X, static_X, fake_Y)
real_output = self._discriminator(dynamic_X, static_X, real_Y)
dis_loss = discriminator_loss(real_output, fake_output, self._mode)
## 为了保证方向的正确性使用mse作为loss
if self._mode == 'GAN':
dis_loss += loss_function(real_Y, fake_Y)
if self._mode == 'WGAN-GP':
##加上GP
dis_loss += self._penalty_weight * self.gradient_penalty(
dynamic_X, static_X, real_Y, fake_Y)
gradients = tape.gradient(dis_loss,
self._discriminator.trainable_variables)
if pretrain:
self._dis_pre_optimizer.apply_gradients(
zip(gradients, self._discriminator.trainable_variables))
else:
self._dis_optimizer.apply_gradients(
zip(gradients, self._discriminator.trainable_variables))
if self._mode == 'WGAN':
[
p.assign(tf.clip_by_value(p, -0.1, 0.1))
for p in self._discriminator.trainable_variables
]
return dis_loss
def train_step(self,dynamic_features,static_features,targ,enc_hidden):
with tf.GradientTape() as tape:
enc_output, enc_hidden = self._encoder(dynamic_features,enc_hidden)
## 需要对enc_output的shape进行查看,并列输入了8种序列特征
# print('Shape of enc output:{}'.format(enc_output.shape))
# print('Shape of enc hidden:{}'.format(enc_hidden.shape))
dec_input = tf.cast(tf.expand_dims([0]*self._batch_sz,1),tf.float64)
loss = 0
dec_hidden = enc_hidden
for t in range(0,targ.shape[1]):
predictions,dec_hidden = self._decoder(dec_input,dec_hidden,enc_output,static_features)
# print(targ[:,t].shape,predictions.shape)
loss += loss_function(tf.expand_dims(targ[:,t],1),predictions)
# rn = np.random.random_sample()
## 时间t的标准结果作为t+1的x
# if rn<0.1:
# dec_input = tf.expand_dims(targ[:,t],1)
# else:
##如果不适用teacher forcing
dec_input = predictions
batch_loss = (loss/int(targ.shape[1]))
variables = self._encoder.trainable_variables + self._decoder.trainable_variables
gradients = tape.gradient(loss,variables)
self._optimizer.apply_gradients(zip(gradients,variables))
return batch_loss