Question

我正在尝试使用 resnet50 进行迁移学习，但不明白为什么它会给我一个错误。我尝试用 Mobilenet 做同样的事情并且它有效。我使用的电脑没有连接到互联网，因此我单独下载了权重。这是我的代码

def resnet50(image_size,num_classes,num_channels, dense_size, drop_prec, include_top, preweights):
    if preweights:
        if include_top:
            path = os.path.join('weights_pre_trianed', 'resnet50_weights_tf_dim_ordering_tf_kernels.h5')
            res = applications.ResNet50(include_top=include_top,
                                                      input_shape=(image_size[0], image_size[1], num_channels),
                                                      weights=path)
            x = res.output
            preds = Dense(num_classes, activation='softmax')(x)
            model = Model(inputs=res.input, outputs=preds)
            for layer in model.layers[:143]:
                layer.trainable = False
            for layer in model.layers[143:]:
                layer.trainable = True
            return model

我收到以下错误

tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.
  (0) Invalid argument: Computed output size would be negative: -227 [input_size: 56, effective_filter_size: 512, stride: 2]
     [[{{node res3a_branch1/convolution}}]]
     [[Mean/_2259]]
  (1) Invalid argument: Computed output size would be negative: -227 [input_size: 56, effective_filter_size: 512, stride: 2]
     [[{{node res3a_branch1/convolution}}]]
0 successful operations.
0 derived errors ignored.

我使用 tensorflow 1.15 和 keras 2.3.1。我的图像是灰度和大小 224x224 a 3 通道为什么我会收到这个错误？

EDIT1：节点 res3a_branch1 是层号 45 Conv2D EDIT2：我想要一个二元分类

Edit3：我按照链接的建议进行了尝试，但现在出现了与模型相关的另一个错误。我知道我应该使用 lambda 层，但我不知道如何使用。

    'node = layer._inbound_nodes[node_index] 
AttributeError: 'NoneType' object has no attribute '_inbound_nodes'

这是我之后的代码

 path=os.path.join('weights_pre_trianed',
'resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5')
 i = Input([None, None, 3], dtype='float32')
 x = applications.resnet50.preprocess_input(i)
 res = applications.ResNet50(include_top=include_top, weights=path)                                       
        
 x = res(x)
 x = GlobalAveragePooling2D()(x)
 preds = Dense(num_classes, activation='softmax')(x)
 model = Model(inputs=i, outputs=preds)
 for layer in model.layers[:-2]:
        layer.trainable = False
 return model

带有预训练 resnet50 的中间层误差张量流

0 个答案: