基于TFLearn构建CNN遇形状不匹配错误，求解决方案

首先看你遇到的错误：

ValueError: Cannot feed value of shape (64,) for Tensor 'TargetsData/Y:0', which has shape '(?, 24)'

这个问题其实很好定位——你的标签格式和模型期望的不匹配！

错误原因分析 #

你的模型最后一层是输出24类的softmax层，搭配的是categorical_crossentropy损失函数。这种组合要求标签必须是24维的one-hot编码格式（形状为(样本数量, 24)），但你现在传入的train_labels和test_labels是一维的整数数组（比如单批次形状是(64,)），所以TensorFlow才会报错说形状对不上。

解决方案：把整数标签转换成One-Hot编码 #

你只需要在加载数据集后，把原始的整数标签转换成one-hot格式就行，有两种简单的实现方式：

方法1：用TFLearn自带的工具函数

直接调用tflarn.data_utils里的to_categorical，省事又可靠：

from tflearn.data_utils import to_categorical

# 加载数据集
train_images, train_labels, test_images, test_labels = load_dataset()

# 转换标签为one-hot编码，nb_classes对应你的类别数24
train_labels = to_categorical(train_labels, nb_classes=24)
test_labels = to_categorical(test_labels, nb_classes=24)

方法2：用NumPy手动实现（不依赖TFLearn工具）

如果你想自己写转换逻辑，也可以用NumPy手动生成one-hot数组：

import numpy as np

def convert_to_one_hot(labels, num_classes):
    # 初始化全0数组，形状为(样本数, 类别数)
    one_hot_labels = np.zeros((len(labels), num_classes))
    # 给对应类别位置赋值1
    one_hot_labels[np.arange(len(labels)), labels] = 1
    return one_hot_labels

# 加载数据集后转换
train_images, train_labels, test_images, test_labels = load_dataset()
train_labels = convert_to_one_hot(train_labels, 24)
test_labels = convert_to_one_hot(test_labels, 24)

额外注意点 #

• 确认你的原始标签值是0到23之间的整数：如果你的标签是从1开始编号（比如1到24），记得先减1再转one-hot，否则会出现索引越界的问题，比如train_labels = train_labels - 1。
• 可以先打印标签形状验证：转换后用print(train_labels.shape)检查，应该输出(训练样本数, 24)，这样就符合模型的要求了。

修改后的完整训练代码 #

把转换步骤加到你的代码里就行，完整代码大概是这样：

import tflearn
import tflearn.layers.core as core
import tflearn.layers.conv as conv
import numpy as np
import gzip

# 你的数据加载函数（包括图像和标签）
def load_mnist_images(filename):
    with gzip.open(filename, 'rb') as f:
        data = np.frombuffer(f.read(), np.uint8, offset=16)
        data = data.reshape(-1, 28, 28, 1)
    return data

def load_mnist_labels(filename):
    with gzip.open(filename, 'rb') as f:
        data = np.frombuffer(f.read(), np.uint8, offset=8)
    return data

def load_dataset():
    # 这里替换成你的数据集路径
    train_imgs = load_mnist_images('train-images.gz')
    train_lbls = load_mnist_labels('train-labels.gz')
    test_imgs = load_mnist_images('t10k-images.gz')
    test_lbls = load_mnist_labels('t10k-labels.gz')
    return train_imgs, train_lbls, test_imgs, test_lbls

# 加载并处理数据
train_images, train_labels, test_images, test_labels = load_dataset()
from tflearn.data_utils import to_categorical
train_labels = to_categorical(train_labels, 24)
test_labels = to_categorical(test_labels, 24)

# 构建CNN模型
convnet = core.input_data(shape=[None, 28, 28, 1], name='input')
convnet = conv.conv_2d(convnet, 32, 2, activation='relu')
convnet = conv.max_pool_2d(convnet, 2)
convnet = conv.conv_2d(convnet, 64, 2, activation='relu')
convnet = conv.max_pool_2d(convnet, 2)
convnet = core.fully_connected(convnet, 1024, activation='relu')
convnet = core.dropout(convnet, 0.8)
convnet = core.fully_connected(convnet, 24, activation='softmax')
convnet = tflearn.regression(convnet, optimizer='adam', learning_rate=0.01, loss='categorical_crossentropy')

model = tflearn.DNN(convnet)
model.fit(train_images, train_labels, n_epoch=30, validation_set=(test_images, test_labels), snapshot_step=500, show_metric=True, run_id='characterOCR')
model.save('CNN.model')

这样修改后，标签的形状就和模型期望的一致了，应该就能正常训练了。

内容的提问来源于stack exchange，提问作者Hassan_mohammad