After saving deep learning network, can not load it.

Question

Yuan Xu 2022-7-7

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1755375-after-saving-deep-learning-network-can-not-load-it

评论： Richard 2023-9-20

I was trying to Use transfer learning of VGGish pre-trained model to classify music genres. I changed the input layer to [300 64 1], after training, I saved the model, and when i loaded it next time, i got this error information: While loading an object of class 'DAGNetwork': Number of elements must not change. Use [] as one of the size inputs to automatically calculate the appropriate size for that dimension.

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Richard 2023-9-20

It is definitely not expected that you will ever see an error of this kind when loading any DAGNetwork model. If you can provide more details then we might be able to help diagnose the issue further:

Version of Matlab you are using
The full error message and stacks that are displayed (e.g. are there any line numbers?)
If possible, attaching your mat file would be ideal or
Code that reproduces the issue

Even partial code for reproduction can be helpful - for example did you start from a specific pretrained network from one of our add-ons? Do you have the exact code used for the transfer learning?

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Milan Bansal 2023-9-20

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1755375-after-saving-deep-learning-network-can-not-load-it#answer_1313807

Hi,

I understand that you are facing an error while loading the saved Deep Learning Model.

The mentioned error typically occurs when the size of the loaded model's layers does not match the expected size.

Ensure that the size of the loaded model's layers matches the size you specified when you trained the model. In this case, the input layer was changed to [300 64 1]. Follow the steps given below to ensure the correct dimensions.

Check the architecture and layer sizes of the pre-trained VGG model.
Modify the architecture to match the desired input size. This involves adjusting the size of the input layer and any subsequent layers as per its output size.
Re-train and save the model.
When loading the model, make sure to load the correct version that matches the modified architecture and the input size.