Keras 的层兼容性问题

Question

这里有一个例外：

2024-03-22 22:20:40.503162: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2024-03-22 22:20:41.005622: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2024-03-22 22:20:41.887025: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
Epoch 1/10
Traceback (most recent call last):
  File "D:\PythonProjects\QuarkMate\main.py", line 63, in <module>
    model, tokenizer = train_model(messages)
                       ^^^^^^^^^^^^^^^^^^^^^
  File "D:\PythonProjects\QuarkMate\main.py", line 58, in train_model
    _model.fit(padded_sequences, to_categorical(padded_sequences), epochs=10, verbose=1)
  File "D:\PythonProjects\QuarkMate\.venv\Lib\site-packages\keras\src\utils\traceback_utils.py", line 122, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "D:\PythonProjects\QuarkMate\.venv\Lib\site-packages\keras\src\backend\tensorflow\nn.py", line 546, in categorical_crossentropy
    raise ValueError(
ValueError: Arguments `target` and `output` must have the same rank (ndim). Received: target.shape=(None, 20, 48), output.shape=(None, 48)

那个代码，最近刚接触keras，了解的还不够

import csv
import json
import os
import threading

import numpy as np
import telebot
from keras import Sequential, Input
from keras.layers import Embedding, LSTM, Dense
from keras.preprocessing.sequence import pad_sequences
from keras.src.legacy.preprocessing.text import Tokenizer
from keras.utils import to_categorical

bot = telebot.TeleBot("censored hehehehaw")

# Parameters
max_sequence_length = 20
embedding_dim = 50
learning_interval = 3 * 60  # Learning interval in seconds (3 minutes)

dataset_file = 'dataset.json'


def load_dataset(file):
    _messages = []
    if os.path.exists(file):
        with open(file, 'r', encoding='utf-8') as json_file:
            _messages = json.load(json_file)
    return _messages


def save_dataset(_messages, file):
    with open(file, 'w', encoding='utf-8') as json_file:
        json.dump(_messages, json_file)


if not os.path.exists(dataset_file):
    open(dataset_file, 'a').close()

messages = load_dataset(dataset_file)


def train_model(_messages):
    _tokenizer = Tokenizer()
    _tokenizer.fit_on_texts(_messages)
    sequences = _tokenizer.texts_to_sequences(_messages)
    padded_sequences = pad_sequences(sequences, maxlen=max_sequence_length, padding='post')

    _model = Sequential([
        Input(shape=(max_sequence_length,)),
        Embedding(len(_tokenizer.word_index) + 1, embedding_dim),
        LSTM(128),
        Dense(len(_tokenizer.word_index) + 1, activation='sigmoid')
    ])

    _model.compile(loss='categorical_crossentropy', optimizer='adam',
                   metrics=['accuracy'])
    _model.fit(padded_sequences, to_categorical(padded_sequences), epochs=10, verbose=1)

    return _model, _tokenizer


model, tokenizer = train_model(messages)


def generate_response(message, _model, _tokenizer):
    sequence = _tokenizer.texts_to_sequences([message])
    padded_sequence = pad_sequences(sequence, maxlen=max_sequence_length, padding='post')
    predicted_sequence = _model.predict(padded_sequence)[0]
    predicted_word_index = np.argmax(predicted_sequence)
    predicted_word = list(_tokenizer.word_index.keys())[
        list(_tokenizer.word_index.values()).index(predicted_word_index)]
    return predicted_word


@bot.message_handler(func=lambda message: True)
def echo_all(message):
    global messages
    if message.text.startswith('/') or message.from_user.is_bot:
        return

    if hasattr(message, 'reply_to_message') and message.reply_to_message is not None:
        previous_message = message.reply_to_message.text
        messages.append(previous_message)
        messages.append(message.text)
        save_dataset(messages, dataset_file)

    response = generate_response(message.text, model, tokenizer)
    if response:
        bot.reply_to(message, response)


def periodic_ing.Timer(learning_interval, periodic_model_training).start()
    model, tokenizer = train_model(messages)


periodic_model_training()
model_training():
    global model, tokenizer, messages
    thread
bot.polling()

我问过chatgpt和bingai。这是一个在电报聊天中自动学习的机器人，可以用定义的主题回答，请帮助我

Answer 1

我看到两个问题。通常，softmax 与 categorical_crossentropy 损失齐头并进（涉及多个标签）。但最直接的问题（从错误消息来看）是您的标签和最后一层（密集层）形状不匹配。参见例如https://discuss.tensorflow.org/t/target-output-mismatch-using-sentencepiecetokenizer-layer-with-huggingface-dataset/23240