how to save/load a deep neural network after transfer learning

1 次查看（过去 30 天）

Raphael 2018-11-21

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/431157-how-to-save-load-a-deep-neural-network-after-transfer-learning

评论： Raphael 2019-7-31

I have fine-tuned the GoogleNet deep neural network following the example provided in C:\...\Examples\nnet\TransferLearningUsingGoogLeNetExample\TransferLearningUsingGoogLeNetExample.m.

The only change I made is that I used a different ImageStore created by

images = imageDatastore('f:\...\Dstore','IncludeSubfolders',true,'LabelSource','foldernames');

images.ReadFcn = @(loc)imresize(imread(loc),[224,224]);

[trainImages,valImages] = splitEachLabel(images,.7,'randomized');

That ImageStore includes about 1000 images. Once training completed I saved my whole workspace with the command:

save googlenet_mine01

Now I am trying to reuse this network (after exiting and restarting Matlab) but the command

load googlenet_mine01

gives me the following error message:

Error using matlab.io.datastore.ImageDatastore/set.Labels (line 208)

Argument 'numel' must be followed by a nonnegative real scalar integer.

By looking at the stack at the error point I see that imds.Numfiles is empty but I don't understand why?!

Raphael

2 个评论
显示无隐藏无

Julius Å 2019-7-30

I have this same problem. Did you manage to solve it?

Raphael 2019-7-31

Not really. I came to the conclusion that a trained network can be reused only on the machine on which it was created and only if you still have the images that were originally included in the ImageStore still present at the same location

请先登录，再进行评论。

请先登录，再回答此问题。