无法加载磁盘上“.mat”文件中存储的VGG19图像净重

Question

我正在尝试迁移学习，并从ImageNet数据集中下载了VGG19模型的权重。它作为“.mat”文件下载。起初，我试图将整个.mat文件加载为

import tensorflow as tf
from tensorflow.keras.applications.vgg19 import VGG19

model_path = os.path.join('trained_model', 'imagenet-vgg-verydeep-19.mat')
pre_trained_model = VGG19(weights = None, include_top = False)
pre_trained_model.load_weights(model_path)

但是，我得到了以下错误

OSError: Unable to open file (file signature not found)

然后我尝试了set_weights()方法而不是load_weights()，如下所示

pre_trained_model = VGG19(weights = None, include_top = False)
pre_trained_model.set_weights(model_path)

然后我得到了下面的错误

ValueError: You called `set_weights(weights)` on layer "vgg19" with a weight list of length 42, 
but the layer was expecting 32 weights. Provided weights: trained_model\imagenet-vgg-verydeep-19.mat...

我也试过了

pre_trained_model = VGG19(weights = None, include_top = True)
pre_trained_model.set_weights(model_path)

但是，我仍然会犯同样的错误

我已经从https://www.kaggle.com/teksab/imagenetvggverydeep19mat下载了权重

Answer 1

这行pre_trained_model = VGG19(weights = None, include_top = False)
include_top = False表示模型不加载完全连接的层。您是否检查过Kaggle上的预训练模型是否使用include_top = True?

Answer 2

无法加载权重，因为图层大小不同

以下是您可以使用的方法： How to read .mat file format in tensorflow?