Input0 与模型层不兼容：expected shape=(None, 256, 256, 3), found shape=(256, 256, 3)

Question

我正在尝试在 Keras 中实现 CycleGAN，当尝试在没有训练的情况下翻译基本图像时，我收到此错误，这会在我的生成器的输入形状中添加一个 None。我用 generatorAtoB.predict() 得到相同的结果，所以这不是问题。这是我试图实现的架构的代码。 X2[0] 只是一个形状为 (256, 256, 3) 的 numpy 数组，在我的函数的输入形状中提到过。这是错误日志和我使用的代码。

编辑

我发现了错误，keras 期望批量大小作为第一维，因此将我的数组整形为 (1, 256, 256, 3) 解决了这个问题。

错误日志

# npm resolution error report

2021-03-17T02:51:06.779Z

While resolving: undefined@2.1.5
Found: expo-constants@9.2.0
node_modules/expo-constants
  expo-constants@"~9.2.0" from the root project

Could not resolve dependency:
peer expo-constants@"~9.0.0" from expo-store-review@2.2.0
node_modules/expo-store-review
  expo-store-review@"~2.2.0" from the root project

Fix the upstream dependency conflict, or retry
this command with --force, or --legacy-peer-deps
to accept an incorrect (and potentially broken) dependency resolution.

Raw JSON explanation object:

{
  "code": "ERESOLVE",
  "current": {
    "name": "expo-constants",
    "version": "9.2.0",
    "whileInstalling": {
      "version": "2.1.5",
      "path": "D:\\Jacobo\\app3"
    },
    "location": "node_modules/expo-constants",
    "dependents": [
      {
        "type": "prod",
        "name": "expo-constants",
        "spec": "~9.2.0",
        "from": {
          "location": "D:\\Jacobo\\app3"
        }
      }
    ]
  },
  "edge": {
    "type": "peer",
    "name": "expo-constants",
    "spec": "~9.0.0",
    "error": "INVALID",
    "from": {
      "name": "expo-store-review",
      "version": "2.2.0",
      "whileInstalling": {
        "version": "2.1.5",
        "path": "D:\\Jacobo\\app3"
      },
      "location": "node_modules/expo-store-review",
      "dependents": [
        {
          "type": "prod",
          "name": "expo-store-review",
          "spec": "~2.2.0",
          "from": {
            "location": "D:\\Jacobo\\app3"
          }
        }
      ]
    }
  },
  "peerConflict": null,
  "strictPeerDeps": false,
  "force": false
}

模型架构

ValueError: Input 0 is incompatible with layer model_7: expected shape=(None, 256, 256, 3), found shape=(256, 256, 3)

# residual block for the generator
def res_block(filters, inputs):
    # kernel weights initializer
    init = RandomNormal(stddev=0.02)
    x = Conv2D(filters, 3, padding='same', kernel_initializer=init)(inputs)
    x = InstanceNormalization(axis=-1)(x)
    x = Activation('selu')(x)
    x = Conv2D(filters, 3, padding='same', kernel_initializer=init)(x)
    x = InstanceNormalization(axis=-1)(x)
    # concatenate second conv layer with the inputs
    x = Concatenate()([x, inputs])
    return x

# generator function
def generator(img_shape = (256, 256, 3), n_blocks = 6):
    # weight initialization
    init = RandomNormal(stddev=0.02)
    inputs = Input(shape = img_shape)
    x = Conv2D(16, 5, padding='same', kernel_initializer=init)(inputs)
    x = InstanceNormalization(axis=-1)(x)
    x = Activation('selu')(x)
    
    x = Conv2D(32, 3, 2, padding='same', kernel_initializer=init)(x)
    x = InstanceNormalization(axis=-1)(x)
    x = Activation('selu')(x)
    
    x = Conv2D(64, 3, 2, padding='same', kernel_initializer=init)(x)
    x = InstanceNormalization(axis=-1)(x)
    x = Activation('selu')(x)
    
    # add residual blocks to our generator
    for _ in range(n_blocks):
        x = res_block(128, x)
    
    # transpose convolutions
    x = Conv2DTranspose(32, 3, strides = 2, padding='same', kernel_initializer=init)(x)
    x = InstanceNormalization(axis=-1)(x)
    x = Activation('selu')(x)
    
    x = Conv2DTranspose(64, 3, 2, padding='same', kernel_initializer=init)(x)
    x = InstanceNormalization(axis=-1)(x)
    x = Activation('selu')(x)
    
    # output layer
    x = Conv2D(3, 7, padding='same', kernel_initializer=init)(x)
    x = InstanceNormalization(axis=-1)(x)
    outputs = Activation('tanh')(x)
    
    # create the model
    model = Model(inputs, outputs)
    return model