时间戳

首页 > 解决方案 > 部分输入和输出的自定义损失

python - 部分输入和输出的自定义损失

问题描述

我已经像这样定义了一个VAE：

def cond_encoder(nfeatures,ncondfeatures,layers, latent_dim, activation='relu'):
  
    finputs = Input(shape=(nfeatures,))
    cinputs = Input(shape=(ncondfeatures,))

    inputs = concatenate([finputs,cinputs])

    hidden = Dense(layers[0],activation=activation, name = f"Dense_enc{0}")(inputs)

    for i in range(1, len(layers)):
        hidden = Dense(units = layers[i], activation=activation, name = f"Dense_enc{i}")(hidden)

    z_mean = Dense(units=latent_dim,activation='linear',name='z_mean')(hidden)
    z_logvar = Dense(units=latent_dim, activation='linear',name='z_logvar')(hidden)

    z_mean,z_logvar = KLDivergenceLayer(weight=1.0)([z_mean,z_logvar])
    KLDivergenceLayer.trainable = False
    encoder = Model(inputs = [finputs,cinputs], outputs = [z_mean]+[z_logvar])
    return encoder



def cond_decoder(nfeatures,ncondfeatures, layers, latent_dim, activation='relu'):

    fromEncoder = Input(shape=(latent_dim,))
    cinputs = Input(shape=(ncondfeatures,))

    inputs = concatenate([fromEncoder,cinputs])

    hidden = Dense(units = layers[0], activation = activation, name = f"Dense_dec{0}")(inputs)

    for i in range(1, len(layers)):
        hidden = Dense(units = layers[i], activation=activation, name = f"Dense_dec{i}")(hidden)

    output = Dense(units = nfeatures, activation = 'linear')(hidden)
    decoder = Model(inputs=[fromEncoder,cinputs], outputs = output)

    return decoder

encoder = models.cond_encoder(nfeatures*nclusters,ncondfeatures,dense_layers,nlatent_dim,activation='relu')
decoder = models.cond_decoder(nfeatures*nclusters,ncondfeatures,dense_layers[::-1],nlatent_dim,activation='relu')


finputs=Input(shape=(nfeatures*nclusters,))
cinputs=Input(shape=(ncondfeatures,))

Z_latent_space= encoder([finputs,cinputs])

sampled_latent_space=models.sampleZ(Z_latent_space)
outputs=decoder([sampled_latent_space,cinputs])
vae_model_train = Model(inputs=[finputs,cinputs],outputs=outputs)

然后我训练编译并拟合模型，如下所示：

vae_model_train.compile(loss=mse,optimizer=Adam(lr))
history = vae_model_train.fit([trainstuff,traincond],[trainstuff, traincond],batch_size=batch_size, epochs= 100, verbose=1, validation_data=([val_jets,valcond],[val_jets,valcond]), initial_epoch=0)

我想要做的是仅计算 trainstuff 和重建的 trainstuff 之间的 mse，但是像我一样传递输入，我怀疑 [trainstuff, traincond] 和它的重建之间的 mse 是计算的？

我试过这个自定义损失函数：

def custom_loss(use_mse=False):
    def loss(y_true, y_pred):
          return K.int_shape(y_pred)[1] * mse(y_true[0], y_pred[0])
            
    return loss

我不确定这是正确的方法吗？

标签： pythontensorflowkerasloss-function

解决方案

Keras 将输出和目标映射到损失函数，所以只要我的模型输出是 trainstuff reco，目标是 trainstuff，我可以简单地使用 mse 就可以了。

推荐阅读