Error using importKerasLayers() for Squeezenet using ".json" file.

Hello,

I am trying to use the new Neural Network Toolbox function named "importKerasLayers" to import Squeezenet model. But it gives me below error:

*Error using importKerasLayers

Invalid layer name 'fire2/squeeze1x1'. Layer names must not contain the character '/'.*

Please, let me know about it.

I have also tried importing ".h5" file for Squeezenet. But it gives me below error:

*Error using importKerasLayers

Unable to import layers from file 'squeezenet_weights.h5' because it contains no 'model_config' attribute.*

Can you provide a ".json" or ".h5" file for Squeezenet which supports MATLAB.

Looking forward to solve this problem as soon as possible.

Thank you.

Answer 1

[Mickaël Tits]

Hi,

I had the same problem, so I modified a python script to change layer names (tu use "_" instead of "/"). You can find the corrected model here : https://github.com/titsitits/Squeezenet-Matlab-Keras You can use this line of code to import it :

squeezenet = importKerasNetwork('squeezenet.json','WeightFile','squeezenet.h5','OutputLayerType','classification');

Best regards, Mickaël Tits

[BHUSHAN MUTHIYAN]

Hello Mickaël,

Thank you for your help.

It is working for me.

I was wondering if you could provide me exact model representation of Squeezenet in Keras as it is in original Squeezenet in Caffe.

The first convolution is above link has dimension 3x3x64 but original implementation has 7x7x96.

Every help will be appreciated.

Thank you.

--

Reagrds,

Bhushan

[Mickaël Tits]

Hello,

It is actually the version v1.1, which apparently has similar results than v1.0, with much less computations: https://github.com/DeepScale/SqueezeNet/tree/master/SqueezeNet_v1.1

I don't have the v1.0 version unfortunately. If you really need the version 1.0, you can try to modify the python script and use the weights of the v1.0.

Regards, Mickaël Tits

[Marco Francini]

Hello, it is working also for me.

Now that I uploaded the Squeezenet to Matlab I used it with transfer learning techninque so it can handle just 3 classes. It is working fine.

Now that I have my network trained, how can I export it to .h5 and .json file? I see that an "exportKerasNetwork" function does not exists...