在 tensorflow 2.9.1 中加载模型失败

Question

我正在尝试加载在 tensorflow 2.9.1 中实现的模型的权重，但失败了

我使用model.save_weights("./saved_model/model")保存了模型

并使用加载模型

model = DepthEstimationModel()
model.load_weights(os.path.join("saved_model", "model"))
model.compile(optimizer, loss=cross_entropy)
model.built = True
model.summary()

这表明

Model: "depth_estimation_model"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 downscale_block (DownscaleB  multiple                 0 (unused)
 lock)                                                           
                                                                 
 downscale_block_1 (Downscal  multiple                 0 (unused)
 eBlock)                                                         
                                                                 
 downscale_block_2 (Downscal  multiple                 0 (unused)
 eBlock)                                                         
                                                                 
 downscale_block_3 (Downscal  multiple                 0 (unused)
 eBlock)                                                         
                                                                 
 bottle_neck_block (BottleNe  multiple                 0 (unused)
 ckBlock)                                                        
                                                                 
 upscale_block (UpscaleBlock  multiple                 0 (unused)
 )                                                               
                                                                 
 upscale_block_1 (UpscaleBlo  multiple                 0 (unused)
 ck)                                                             
                                                                 
 upscale_block_2 (UpscaleBlo  multiple                 0 (unused)
 ck)                                                             
                                                                 
 upscale_block_3 (UpscaleBlo  multiple                 0 (unused)
 ck)                                                             
                                                                 
 conv2d_18 (Conv2D)          multiple                  0 (unused)
                                                                 
=================================================================
Total params: 2
Trainable params: 0
Non-trainable params: 2

这表明参数未正确加载。

如果删除model.built = True，它会输出：

This model has not yet been built. Build the model first by calling `build()` or by calling the model on a batch of data.

参考：Tensorflow 2.0 ValueError while Loading weights from .h5 file

Answer 1

在加载权重之前，您必须编译模型，然后必须加载权重。例如

model.compile(optimizer, loss=cross_entropy)
model.load_weights(os.path.join("saved_model", "model"))

这样你就不会遇到那个错误了。谢谢。