Keras Giving 问题上的句子 VAE 损失层实现

Question

所以我一直在 TF-Keras（最新版本）上实现句子 VAE。下面的自定义函数计算稀疏分类输出的 VAE 损失。

def vae_loss(encoder_inputs, decoder_outputs):
     sen_loss = K.sparse_categorical_crossentropy(encoder_inputs, decoder_outputs, from_logits=True)
     sen_loss = K.sum(sen_loss, axis=-1)
     kl_loss = - 0.5 * K.mean(1 + z_log_sigma - K.square(z_mean) - K.exp(z_log_sigma))
     loss = K.mean(sen_loss + kl_loss)
     return loss

optimizer = keras.optimizers.Adam(learning_rate=0.01)
model.compile(optimizer=optimizer, loss=vae_loss)
model.fit([seq_in,seq_out],
          seq_lab,
          batch_size=batch_size, 
          epochs=epochs,
          validation_split=0.1)

#Seq_in shape = (no_of_samples, maxlen)
#Seq_out shape = (no_of_samples, maxlen)
#Seq_lab shape = (no_of_samples, maxlen, 1)

在尝试训练时，它给出了错误：

TypeError: Cannot convert a symbolic Keras input/output to a numpy array. 
This error may indicate that you're trying to pass a symbolic value to a NumPy call, 
which is not supported. 
Or, you may be trying to pass Keras symbolic inputs/outputs to a TF API that does not 
register dispatching, 
preventing Keras from automatically converting the API call to a lambda layer in the 
Functional Model.

然后我禁用了急切张量以允许使用计算图。但是，它给出了错误：

FailedPreconditionError: 2 root error(s) found.
(0) Failed precondition: Could not find variable 
training_6/Adam/embedding_16/embeddings/v. 
This could mean that the variable has been deleted. In TF1, it can also mean the 
variable is uninitialized. Debug info: container=localhost, status=Not found: 
Resource localhost/training_6/Adam/embedding_16/embeddings/v/N10tensorflow3VarE does 
not exist.
[[{{node training_6/Adam/Adam/update_embedding_16/embeddings/ReadVariableOp_3}}]]
[[_arg_keras_learning_phase_0_3/_722]]

(1) Failed precondition: Could not find variable 
training_6/Adam/embedding_16/embeddings/v. 
This could mean that the variable has been deleted. In TF1, it can also mean the 
variable is uninitialized. 
Debug info: container=localhost, status=Not found: Resource 
localhost/training_6/Adam/embedding_16/embeddings/v/N10tensorflow3VarE does not 
exist.
[[{{node training_6/Adam/Adam/update_embedding_16/embeddings/ReadVariableOp_3}}]]
0 successful operations.
0 derived errors ignored.

模型的代码是这样的：

#################################### ENCODER LAYER ################################
encoder_inputs = Input(shape=(maxlen, ))

# Encoder Embedding
enc_emb = Embedding(vocab_size, embedding_dim,
                trainable=True)(encoder_inputs)

# Encoder LSTM
encoder_lstm3 = LSTM(latent_dim)
encoder_outputs = encoder_lstm3(enc_emb)


#################################### VAE Z LAYER ################################
z_mean = Dense(units=latent_dim)(encoder_outputs)
z_log_sigma = Dense(units=latent_dim)(encoder_outputs)

def sampling(args):
    z_mean, z_log_sigma = args
    epsilon = K.random_normal(shape=(batch_size, latent_dim), mean=0., stddev=1.0)
    return z_mean + z_log_sigma * epsilon

z = Lambda(sampling, output_shape=(latent_dim,))([z_mean, z_log_sigma])

expandz_h = Dense(latent_dim)
z_exp_h = expandz_h(z)

expandz_c = Dense(latent_dim)
z_exp_c = expandz_c(z)


#################################### DECODER LAYER ################################
# Set up the decoder, using z layer outputs as the initial state
decoder_inputs = Input(shape=(maxlen, ))

# Embedding layer
dec_emb = Embedding(vocab_size, embedding_dim, trainable=True)(decoder_inputs)

# Decoder LSTM
decoder_lstm = LSTM(latent_dim, return_sequences=True,
                return_state=True, dropout=0.4,
                recurrent_dropout=0.0)
(decoder_outputs, decoder_fwd_state, decoder_back_state) = \
decoder_lstm(dec_emb, initial_state=[z_exp_h, z_exp_c])

# Dense layer
decoder_dense = TimeDistributed(Dense(vocab_size, activation='softmax'))
decoder_outputs = decoder_dense(decoder_outputs)

# Define the model
model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
model.summary()

我认为问题在于 tf-keras 中 loss_function 的实现，但我可能是错的，非常感谢指导。

Answer 1

我知道这个问题很老，但由于我在训练 VAE 时遇到了同样的问题，所以我仍然会发布我的解决方案。解决它的方法是创建两个单独的损失函数，一个用于 kl_loss，一个用于重建损失。然后 kl_loss 被赋予 z_mean 和 z_log_sigma 作为参数，而不是 encoder_inputs 和 decoder_outputs。然后我用两个损失函数编译了我的模型。在你的情况下，这意味着

def sen_loss(encoder_inputs, decoder_outputs):
    sl = K.sparse_categorical_crossentropy(encoder_inputs, ecoder_outputs, from_logits=True)
    return K.sum(sl, axis=-1)

def kl_loss(z_mean, z_log_sigma):
    return - 0.5 * K.mean(1 + z_log_sigma - K.square(z_mean) - K.exp(z_log_sigma))

optimizer = keras.optimizers.Adam(learning_rate=0.01)
model.compile(optimizer=optimizer, loss=[sen_loss, kl_loss], loss_weights=[1.,1.])