Question

我正在尝试构建一个序列到序列模型，但我不太确定如何解决我在解码器 LSTM 中不断出现的错误。

我正在学习本教程：

ValueError: Dimensions must be equal, but are 256 and 64 for '{{node mul/mul}} = Mul[T=DT_FLOAT](Sigmoid_1, init_c)' with input shapes: [?,25,256], [?,64].

import os
os.environ["TF_FORCE_GPU_ALLOW_GROWTH"] = "true"
from speech_preprocess import sound_files, phrases, tokenizer, labels
from tensorflow.keras.layers import Input, LSTM, Dense, Embedding, Conv1D, MaxPool1D, Reshape
from tensorflow.keras.models import Model

input_length_encoder = max(len(arr) for arr in sound_files)
output_tokens = tokenizer.num_words
batch = 4

encoder_input = Input(shape=(1, sound_files.shape[-1]))
encoder, state_1, state_2 = LSTM(64, return_state=True)(encoder_input)

decoder_input = Input(shape=(None, phrases.shape[1]))
decode_embedding = Embedding(input_dim=output_tokens, output_dim=64)(decoder_input)
decoder, _, _ = LSTM(64, return_state=True, return_sequences=True)(decode_embedding,
                                                                   initial_state=[state_1, state_2])
decoder_dense = Dense(output_tokens, activation="softmax")
decoder_output = decoder_dense(decoder)

model = Model([encoder_input, decoder_input], decoder_output)
for layer in model.layers:
    print(layer.output_shape)
model.summary()
model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"])
model.fit([sound_files, phrases], labels, epochs=1, batch_size=batch, validation_split=0.2)

ValueError：维度必须相等，但分别为 256 和 64；序列到序列 LSTM

0 个答案: