Dlib: Unexpected version found while deserializing dlib::input_rgb_image_sized

[anguoyang]

hi@davisking, when i used the model generated by dnn_metric_learning_on_images_ex, and replace/rename into dlib_face_recognition_resnet_model_v1.dat for dnn_face_recognition_ex, i got this error:

Unexpected version found while deserializing dlib::input_rgb_image_sized

it worked when i used the original .dat file, so it is because of the model file. it is really strange, as the .dat file is only a model file, however the error message is about image size.

my pc is in ubuntu 16.04 with cuda/cudnn, all things are ok during building, and i use the default johns images for training.

[davisking]

The input layers are different. That's what the error is telling you. They have to be the same.

[sorenchiron]

I think it is the word "version" that stops some programmers from reading the rest of the error message.
I myself fell into dlib version checking for quite a while...

[anguoyang]

@sorenchiron I am still puzzling on this:)
does it mean that I need to change this:
input_rgb_image_sized<150>
into:
input_rgb_image_sized<"size of my training images">
?

[davisking]

You can't just change the network definition and then deserialize a model
that uses a different network. That's what the error is telling you.

[anguoyang]

@davisking thank you.
another problem, I have used vgg face for metric learning, it finished in few minutes, I am not sure if it is correct or not, as there is only these logs, no more information:
objs.size(): 8631
done training
num_right: 250
num_wrong: 50

[anguoyang]

@davisking I found a strange problem on training:
1.If I run the dnn_metric_learning_on_images_ex vgg(I used aligned vgg face dataset and cropped all images into the same size) in command line, then it will generate the above message
2.However, if I run it in visual studio(in debug mode), then it is OK
I have checked the param(only 1 parameter) and visual studio configuration, there is no difference, it is really strange.

[anguoyang]

by the way, the problem of this issue has been solved, just as you said, I need to set the network definition same with training, thank you.
But the accuracy is not as good as your model

[arunxz98]

hey @anguoyang can you please elaborate how you solved this problem? i'm trying to explore Dlib, and this would be of great help.

[wangx404]

by the way, the problem of this issue has been solved, just as you said, I need to set the network definition same with training, thank you.
But the accuracy is not as good as your model

Hi, could you tell me how you solve this problem? Change the image size in dnn_metric_learning_on_images_ex ? If so, which size should I use? Thank you.

[wangx404]

I find the solution in another issue. Anyway, thanks.