TensorFlow：ValueError：形状不兼容

Question

I am having some difficulty with the shapes of my data for an encoder-decoder model. The issue seems to be with the Dense layer but I cannot figure out why there are incompatibility issues. Can anyone help me?

Error Message

ValueError: Shapes (None, 6) and (None, 6, 1208) are incompatible

Model

# Define an input sequence and process it.
encoder_inputs = Input(shape=(35,), name='encoder_inputs')
decoder_inputs = Input(shape=(6,), name='decoder_inputs')

embedding = Embedding(input_dim=vocab_size, output_dim=160, mask_zero=True)
encoder_embeddings = embedding(encoder_inputs)
decoder_embeddings = embedding(decoder_inputs)

encoder_lstm = LSTM(512, return_state=True, name='encoder_lstm')
LSTM_outputs, state_h, state_c = encoder_lstm(encoder_embeddings)

# We discard `LSTM_outputs` and only keep the other states.
encoder_states = [state_h, state_c]

decoder_lstm = LSTM(512, return_sequences=True, return_state=True, name='decoder_lstm')

# Set up the decoder, using `context vector` as initial state.
decoder_outputs, _, _ = decoder_lstm(decoder_embeddings,
                                     initial_state=encoder_states)

#complete the decoder model by adding a Dense layer with Softmax activation function 
#for prediction of the next output
decoder_dense = Dense(target_vocab_size, activation='softmax', name='decoder_dense')
decoder_outputs = decoder_dense(decoder_outputs)

# put together
model_encoder_training = Model([encoder_inputs, decoder_inputs], decoder_outputs, name='model_encoder_training')

Model: "model_encoder_training"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
decoder_inputs (InputLayer)     [(None, 6)]          0                                            
__________________________________________________________________________________________________
encoder_inputs (InputLayer)     [(None, 35)]         0                                            
__________________________________________________________________________________________________
embedding_12 (Embedding)        multiple             457120      encoder_inputs[0][0]             
                                                                 decoder_inputs[0][0]             
__________________________________________________________________________________________________
encoder_lstm (LSTM)             [(None, 512), (None, 1378304     embedding_12[0][0]               
__________________________________________________________________________________________________
decoder_lstm (LSTM)             [(None, 6, 512), (No 1378304     embedding_12[1][0]               
                                                                 encoder_lstm[0][1]               
                                                                 encoder_lstm[0][2]               
__________________________________________________________________________________________________
decoder_dense (Dense)           (None, 6, 1208)      619704      decoder_lstm[0][0]               
==================================================================================================
Total params: 3,833,432
Trainable params: 3,833,432
Non-trainable params: 0
__________________________________________________________________________________________________

Variables and Extra Information

X_train.shape = (24575, 35)
y_train.shape = (24575, 6)
X_decoder.shape = (24575, 6)

vocab_size = 2857
target_vocab_size = 1208

Answer 1

You should make sure you are using tf.keras.losses.SparseCategoricalCrossentropy() as your loss function and that the last Dense layer is wrapped around a TimeDistributed layer. The decoder_lstm (LSTM) is returning a sequence with the shape (None, 6, 512) and you are applying a Dense layer to it, but as the docs mention:

If the input to the layer has a rank greater than 2, then Dense computes the dot product between the inputs and the kernel along the last axis of the inputs [...]

So the last Dense layer is essentially disregarding the 6 timesteps and being applied to the last dimension 512, which is probably not what you want. With a TimeDistributed layer, you are simply applying a Dense layer with a softmax activation function to each time step n to calculate the probability for each word in the vocabulary of the size 1208. Here is a working example:

import tensorflow as tf

vocab_size = 2857
target_vocab_size = 1208
encoder_inputs = tf.keras.layers.Input(shape=(35,), name='encoder_inputs')
decoder_inputs = tf.keras.layers.Input(shape=(6,), name='decoder_inputs')

embedding = tf.keras.layers.Embedding(input_dim=vocab_size, output_dim=160, mask_zero=True)
encoder_embeddings = embedding(encoder_inputs)
decoder_embeddings = embedding(decoder_inputs)

encoder_lstm = tf.keras.layers.LSTM(512, return_state=True, name='encoder_lstm')
LSTM_outputs, state_h, state_c = encoder_lstm(encoder_embeddings)

encoder_states = [state_h, state_c]

decoder_lstm = tf.keras.layers.LSTM(512, return_sequences=True, return_state=True, name='decoder_lstm')

decoder_outputs, _, _ = decoder_lstm(decoder_embeddings,
                                     initial_state=encoder_states)

decoder_dense = tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(target_vocab_size, activation='softmax', name='decoder_dense'))
decoder_outputs = decoder_dense(decoder_outputs)

model_encoder_training = tf.keras.Model([encoder_inputs, decoder_inputs], decoder_outputs, name='model_encoder_training')
model_encoder_training.compile(optimizer='adam', loss=tf.keras.losses.SparseCategoricalCrossentropy())

samples = 100
X_train = tf.random.uniform((samples, 35), maxval=vocab_size, dtype=tf.int32)
X_decoder = tf.random.uniform((samples, 6), maxval=vocab_size, dtype=tf.int32)

y_train = tf.random.uniform((samples, 6), maxval=target_vocab_size, dtype=tf.int32)

model_encoder_training.fit([X_train, X_decoder], y_train, epochs=5, batch_size=10)

Epoch 1/5
10/10 [==============================] - 8s 302ms/step - loss: 7.0967
Epoch 2/5
10/10 [==============================] - 3s 300ms/step - loss: 6.8687
Epoch 3/5
10/10 [==============================] - 3s 302ms/step - loss: 6.5024
Epoch 4/5
10/10 [==============================] - 3s 300ms/step - loss: 6.1527
Epoch 5/5
10/10 [==============================] - 3s 300ms/step - loss: 5.9458
<keras.callbacks.History at 0x7f88cb66a990>