Keras 在第一个纪元结束时显示形状错误

Question

I try to create an LSTM autoencoder with keras

While, it shows a value error at the end of first epoch

ValueError: operands could not be broadcast together with shapes (32,20) (20,20) (32,20) 

The shape of model input is (sample_size,20,31), and following is the model

The sampling function:

def sampling(args):

    z_mean, z_log_var = args
    batch = K.shape(z_mean)[0]
    dim = K.int_shape(z_mean)[1]
    # by default, random_normal has mean=0 and std=1.0
    epsilon = K.random_normal(shape=(batch,dim))
    return z_mean + K.exp(0.5 * z_log_var) * epsilon 

The encoder part:

inputs = Input(shape=(lag,data.shape[1],), name='encoder_input')
x = LSTM(30,activation='relu',return_sequences=True) (inputs)
x = LSTM(60,activation='relu') (x)
z_mean = Dense(60, name='z_mean')(x)
z_log_var = Dense(60, name='z_log_var')(x)
z_temp = Lambda(sampling, output_shape=(60,), name='z')([z_mean, z_log_var])
z = RepeatVector(lag)(z_temp)
encoder = Model(inputs, [z_mean, z_log_var, z], name='encoder')

The decoder part:

latent_inputs = Input(shape=(lag,60), name='z_sampling')
x_2 = LSTM(60, activation='relu',return_sequences= True)(latent_inputs)
x_2 = LSTM(data.shape[1], activation='relu',return_sequences= True)(x_2)
decoder = Model(latent_inputs, x_2, name='decoder')
outputs = decoder(encoder(inputs)[2])
vae = Model(inputs, outputs)

And loss and fit part:

outputs = decoder(encoder(inputs)[2])
vae = Model(inputs, outputs)
reconstruction_loss = mse(inputs, outputs)
kl_loss = 1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)
kl_loss = K.mean(kl_loss)
kl_loss *= -0.1
vae_loss = reconstruction_loss + kl_loss
vae.add_loss(vae_loss) 
vae.compile(optimizer='adam')
vae.fit(train,epochs=100)

It will cause this error:

Epoch 1/100
632256/632276 [============================>.] - ETA: 0s - loss: 0.0372
ValueError: operands could not be broadcast together with shapes (32,20) (20,20) (32,20) 

If there is a shape error, how dose model work at previous step. That's my main problem, thanks for your answer

Answer 1

You are working with batch size of 32 , but at the very end your operand get a tensor with only 20 elements, because this amount remains after 632256 from 632276:

632276 - 632256 = 20

Basically this is the error message about and that is why the previous steps worked.

Most straightforward solution:

use fit() method's steps_per_epoch option:

steps_per_epoch: Integer or None.
Total number of steps (batches of samples) before declaring one epoch finished and starting the next epoch. When training with input tensors such as TensorFlow data tensors, the default None is equal to the number of samples in your dataset divided by the batch size, or 1 if that cannot be determined.

steps_per_epoch = total_samples // batch_size

In this case basically you drop the last 20 samples.