ValueError：形状（无，28）和（无，28，10）不兼容

Question

我正在尝试使用 keras 顺序模型训练模型，我的代码如下：

from tensorflow.keras import Input
from tensorflow.keras.preprocessing.text import Tokenizer
from tensorflow.keras.preprocessing.sequence import pad_sequences

# Example input and output sequences
input_sequence = ["Madhup is a good boy.", "I am a large language model."]
output_sequence = ["Madgboy", "Imllm"]

# Create tokenizers for input and output
# input_tokenizer = Tokenizer(oov_token='<START>', lower=True)
input_tokenizer = Tokenizer(char_level=True, lower=True)
input_tokenizer.fit_on_texts(input_sequence)
output_tokenizer = Tokenizer(char_level=True, lower=True)
output_tokenizer.fit_on_texts(output_sequence)

# Convert text to sequences of integers
input_sequence_int = input_tokenizer.texts_to_sequences(input_sequence)
output_sequence_int = output_tokenizer.texts_to_sequences(output_sequence)
print(f"Input Sequence : {input_sequence_int}\n Output Sequence : {output_sequence_int} \n")



# Pad sequences to the same length
max_input_length = max(len(seq) for seq in input_sequence_int)
max_output_length = max(len(seq) for seq in output_sequence_int)
max_sequence_length = max(max_input_length, max_output_length)
print("Max Input Length ",  max_input_length)
print("Max Output Length ",  max_output_length)
print("Max Seq Length ",  max_sequence_length)


input_sequence_padded = pad_sequences(input_sequence_int, maxlen=max_sequence_length, padding='post')
output_sequence_padded = pad_sequences(output_sequence_int, maxlen=max_sequence_length, padding='post')
print("Padded Input Sequence:", input_sequence_padded)
print("Padded Output Sequence:", output_sequence_padded)
print(output_tokenizer.word_index)


import tensorflow as tf
from tensorflow.keras.layers import Embedding, LSTM, Dense
embedding_dim = 256
units = max_input_length

# Calculate the input vocab size
input_vocab_size = len(input_tokenizer.word_index) + 1
print(input_vocab_size)
output_vocab_size = len(output_tokenizer.word_index) + 1
print(output_vocab_size)


# Define the model
model = tf.keras.Sequential([
    Embedding(input_vocab_size, embedding_dim, input_length=max_input_length),
    LSTM(embedding_dim, return_sequences=True),
    Dense(output_vocab_size, activation='softmax')
])

# Compile the model
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

# Train the model
model.fit(input_sequence_padded, output_sequence_padded, epochs=10)


import numpy as np
def generate_nickname(input_text):
    input_sequence = input_tokenizer.texts_to_sequences([input_text])
    print(f"Input Sequence: {input_sequence}")
#     input_sequence = pad_sequences(input_sequence, maxlen=max_sequence_length, padding='post')
    input_sequence = pad_sequences(input_sequence, maxlen=max_sequence_length, padding='post')
    print(f"Padded Input Sequence: {input_sequence}")
    predicted_sequence = model.predict(input_sequence)
    max_arg = np.argmax(predicted_sequence, axis=-1)
    print(max_arg)
    predicted_nickname = output_tokenizer.sequences_to_texts(max_arg)[0]
    return predicted_nickname

# Example usage
input_text = "Madhup is a good boy."
predicted_nickname = generate_nickname(input_text)
print(f"Input Phrase: {input_text}")
print(f"Generated Nickname: {predicted_nickname}")

这是填充的输入和输出序列

填充输入序列：[[ 5 2 6 12 9 13 1 10 14 1 2 1 3 4 4 6 1 15 4 16 11 0 0 0 0 0 0 0] [10 1 2 5 1 2 1 7 2 17 3 8 1 7 2 18 3 9 2 3 8 1 5 4 6 8 7 11]] 填充 输出序列：[[1 3 4 5 6 7 8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0] [9 1 2 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]]

上面的代码抛出错误

model.fit(input_sequence_padded, output_sequence_padded, epochs=10)

ValueError: Shapes (None, 28) and (None, 28, 10) are incompatible

我知道在将输出序列传递给模型之前我必须以某种方式重塑输出序列。我只是很困惑如何做到这一点

注意如果我只是将

1

传递到

output_vocab_size

，代码将执行，但这将没有用处，因为经过训练的模型不会返回任何输出。

感谢所有帮助！

Answer 1

您需要对模型进行这些修改，以使其按照您指定的参数运行：

model = tf.keras.Sequential([
    Embedding(input_vocab_size, embedding_dim, input_length=max_input_length),
    LSTM(embedding_dim, return_sequences=False),
    Dense(max_input_length, activation='softmax')
])

因此，首先您需要禁用 LSTM 的 return_sequences，因为您不需要每个 LSTM 单元的输出，其次在最后一个 Dense 层中，您必须将节点设置为等于输出填充数据，该数据与输入相同。