构建Unet模型时遭遇TypeError：输入应为张量却得到Conv2D对象的问题求助

我最近在搭建Unet模型时碰到了一个TypeError，提示Inputs to a layer should be tensors. Got: <keras.layers.convolutional.conv2d.Conv2D object>，尝试分批传入数据后还是没解决问题，以下是我的代码：

import numpy as np
import tensorflow as tf
from keras.layers import Input,Conv2D,MaxPooling2D,UpSampling2D, BatchNormalization
from keras.models import Model
from keras import optimizers

# 数据转换
X_train = np.array(X_train)
X_test = np.array(X_test)
y_train = np.array(y_train)
y_test = np.array(y_test)

def unet():
    # First block Going down
    d1_2 = Conv2D(16, (3, 3), activation='relu', padding='same',input_shape=(s,s,3))
    d1_3 = Conv2D(16, (3, 3), activation='relu', padding='same')(d1_2)

    # Second block Going down
    d2_1 = MaxPooling2D()(d1_3)
    d2_2 = Conv2D(32, (3, 3), activation='relu', padding='same')(d2_1)
    d2_3 = Conv2D(32, (3, 3), activation='relu', padding='same')(d2_2)

    # Third block Going down
    d3_1 = MaxPooling2D()(d2_3)
    d3_2 = Conv2D(64, (3, 3), activation='relu', padding='same')(d3_1)
    d3_3 = Conv2D(64, (3, 3), activation='relu', padding='same')(d3_2)

    # Fourth block Going down
    d4_1 = MaxPooling2D()(d3_3)
    d4_2 = Conv2D(128, (3, 3), activation='relu', padding='same')(d4_1)
    d4_3 = Conv2D(128, (3, 3), activation='relu', padding='same')(d4_2)

    # Fifth block
    d5_1 = MaxPooling2D()(d4_3)
    d5_2 = Conv2D(256, (3, 3), activation='relu', padding='same')(d5_1)
    d5_3 = Conv2D(256, (3, 3), activation='relu', padding='same')(d5_2)

    # Fourth block going up, concatenated with Fourth block going down
    up4_0 = UpSampling2D((2, 2))(d5_3)
    up4_1 = tf.keras.layers.concatenate([d4_3, up4_0])
    up4_2 = Conv2D(128, (3, 3), activation='relu', padding='same')(up4_1)
    up4_3 = Conv2D(128, (3, 3), activation='relu', padding='same')(up4_2)

    # Third block going up, concatenated with Third block going down
    up3_0 = UpSampling2D((2, 2))(up4_3)
    up3_1 = tf.keras.layers.concatenate([d3_3, up3_0])
    up3_2 = Conv2D(64, (3, 3), activation='relu', padding='same')(up3_1)
    up3_3 = Conv2D(64, (3, 3), activation='relu', padding='same')(up3_2)

    # Second block going up, concatenated with Second block going down
    up2_0 = UpSampling2D((2, 2))(up3_3)
    up2_1 = tf.keras.layers.concatenate([d2_3, up2_0])
    up2_2 = Conv2D(32, (3, 3), activation='relu', padding='same')(up2_1)
    up2_3 = Conv2D(32, (3, 3), activation='relu', padding='same')(up2_2)

    # First block going up, concatenated with First block going down
    up1_0 = UpSampling2D((2, 2))(up2_3)
    up1_1 = tf.keras.layers.concatenate([d1_3, up1_0])
    up1_2 = Conv2D(16, (3, 3), activation='relu', padding='same')(up1_1)
    up1_3 = Conv2D(16, (3, 3), activation='relu', padding='same')(up1_2)

    # Output
    out = Conv2D(1, (1, 1), activation='sigmoid', padding='same')(up1_3)
    return out

# 尝试过用单批次数据训练，但还是报错
unet = Model(unet(),input_shape = (s,s,3))
unet.compile(loss='mean_squared_error', optimizer = optimizers.rmsprop_v2.RMSprop())
unet_train = unet.fit(X_train.as_numpy_iterator.next(), y_train.as_numpy_iterator.next(), batch_size=batch_size,epochs=epochs,verbose=1,validation_data=(X_test, y_test))

问题分析与解决方案 #

仔细看下来，问题出在模型实例化的逻辑上：
你当前的unet()函数返回的是模型的最后一层（Conv2D对象），而Model类的第一个参数需要的是输入张量，第二个参数才是输出张量。直接把unet()的返回值传给Model，Keras会误以为你要把Conv2D层当作输入，自然就抛出了“不是张量”的错误。

修正后的完整代码

我们需要在unet()函数里显式定义输入层，并且直接返回完整的Model对象：

import numpy as np
import tensorflow as tf
from keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D, BatchNormalization
from keras.models import Model
from keras import optimizers

# 数据转换（保持不变）
X_train = np.array(X_train)
X_test = np.array(X_test)
y_train = np.array(y_train)
y_test = np.array(y_test)

def unet(s):
    # 显式定义输入张量，这是模型的起点
    inputs = Input(shape=(s, s, 3))
    
    # 下采样模块（调整第一个Conv2D，去掉input_shape参数，改用inputs作为输入）
    d1_2 = Conv2D(16, (3, 3), activation='relu', padding='same')(inputs)
    d1_3 = Conv2D(16, (3, 3), activation='relu', padding='same')(d1_2)

    d2_1 = MaxPooling2D()(d1_3)
    d2_2 = Conv2D(32, (3, 3), activation='relu', padding='same')(d2_1)
    d2_3 = Conv2D(32, (3, 3), activation='relu', padding='same')(d2_2)

    d3_1 = MaxPooling2D()(d2_3)
    d3_2 = Conv2D(64, (3, 3), activation='relu', padding='same')(d3_1)
    d3_3 = Conv2D(64, (3, 3), activation='relu', padding='same')(d3_2)

    d4_1 = MaxPooling2D()(d3_3)
    d4_2 = Conv2D(128, (3, 3), activation='relu', padding='same')(d4_1)
    d4_3 = Conv2D(128, (3, 3), activation='relu', padding='same')(d4_2)

    d5_1 = MaxPooling2D()(d4_3)
    d5_2 = Conv2D(256, (3, 3), activation='relu', padding='same')(d5_1)
    d5_3 = Conv2D(256, (3, 3), activation='relu', padding='same')(d5_2)

    # 上采样模块（保持逻辑不变）
    up4_0 = UpSampling2D((2, 2))(d5_3)
    up4_1 = tf.keras.layers.concatenate([d4_3, up4_0])
    up4_2 = Conv2D(128, (3, 3), activation='relu', padding='same')(up4_1)
    up4_3 = Conv2D(128, (3, 3), activation='relu', padding='same')(up4_2)

    up3_0 = UpSampling2D((2, 2))(up4_3)
    up3_1 = tf.keras.layers.concatenate([d3_3, up3_0])
    up3_2 = Conv2D(64, (3, 3), activation='relu', padding='same')(up3_1)
    up3_3 = Conv2D(64, (3, 3), activation='relu', padding='same')(up3_2)

    up2_0 = UpSampling2D((2, 2))(up3_3)
    up2_1 = tf.keras.layers.concatenate([d2_3, up2_0])
    up2_2 = Conv2D(32, (3, 3), activation='relu', padding='same')(up2_1)
    up2_3 = Conv2D(32, (3, 3), activation='relu', padding='same')(up2_2)

    up1_0 = UpSampling2D((2, 2))(up2_3)
    up1_1 = tf.keras.layers.concatenate([d1_3, up1_0])
    up1_2 = Conv2D(16, (3, 3), activation='relu', padding='same')(up1_1)
    up1_3 = Conv2D(16, (3, 3), activation='relu', padding='same')(up1_2)

    # 输出层
    out = Conv2D(1, (1, 1), activation='sigmoid', padding='same')(up1_3)
    
    # 返回完整的Model对象，绑定输入和输出张量
    return Model(inputs=inputs, outputs=out)

# 实例化模型、编译、训练
# 假设你已经定义了s、batch_size、epochs这几个变量
unet_model = unet(s)
unet_model.compile(loss='mean_squared_error', optimizer=optimizers.rmsprop_v2.RMSprop())

# 注意：如果X_train是TensorFlow Dataset对象，直接传入即可，不需要用as_numpy_iterator.next()
# 手动取单批次只会训练一个批次的数据，完全达不到训练效果
unet_train = unet_model.fit(X_train, y_train, batch_size=batch_size, epochs=epochs, verbose=1, validation_data=(X_test, y_test))

额外提示

• 关于数据集的使用：fit()方法原生支持TensorFlow Dataset对象，只有在调试单批次逻辑时才需要用as_numpy_iterator.next()，正常训练直接传入整个数据集即可。
• 模型结构验证：可以用unet_model.summary()查看模型的层级结构，确认输入输出是否符合预期。

备注：内容来源于stack exchange，提问作者Marco Nashaat