How to upload pretrained weights for a vgg16 dnn

I want to upload an h5 weights file for the vgg16 model. There are options to import a whole keras model, or just the layers but not, apparently, just the weights. So I thought, maybe I could create a vgg16 model in matlab, export it, change the weights file and reimport it. However, matlab won't even reimport its own exported file:

net =vgg16;

exportNetworkToTensorFlow(net,package);

net = importTensorFlowNetwork(package); %no change made to the saved model

Importing the saved model...

Error using tf2mex

Unable to import the model because SavedModel appears to be corrupt or missing.

Can anyone suggest a way to import the weights?

Answer 1

[Sivylla Paraskevopoulou]

The importTensorFlowNetwork function expects the input modelFolder to be a saved_model.pb file (SaveModel format) and I think that's what the error is telling you. To save a TensorFlow model in the SavedModel format, run this command in Python:

model.save("modelName")

A couple more things...

(1) The subfolder variables in the SavedModel folder contains the weights learned by the pretrained TensorFlow network. By default, importTensorFlowNetwork imports the weights.

(2) The exported model to TensorFlow might be slight different than the original MATLAB network. Thus, exporting and re-import might or might not work seamleassly.

[Peter]

Thank you for responding. I didn't mention Python, but I think you are suggesting that I need to export from Matlab, read into Python, export from Python and reimport into Matlab? I did try that, but I'm no Python expert: and when I tried following the example in the help file to read into Python, that also gave an error. However, if you think that's the way to go, I'll have another try when I'm back home next week.

[Peter]

No, I tried it again. Executing the code as in the helpfile:

matlab > net=squeezenet;

matlab > exportNetworkToTensorFlow(net,"myModel")

python > import myModel

python > model = myModel.load_model()

TypeError Traceback (most recent call last) <ipython-input-4-8d7fd46d5d31> in <module> 1 import myModel ----> 2 model = myModel.load_model()~\Desktop\myModel\__init__.py in load_model(load_weights, debug) 7 8 def load_model(load_weights=True, debug=False):----> 9 m = model.create_model() 10 if load_weights: 11 loadWeights(m, debug=debug)~\Desktop\myModel\model.py in create_model() 84 conv10 = layers.Conv2D(1000, (1,1), name="conv10_")(drop9) 85 relu_conv10 = layers.ReLU()(conv10)---> 86 pool10 = layers.GlobalAveragePooling2D(keepdims=True)(relu_conv10) 87 prob = layers.Softmax()(pool10) 88 ClassificationLayer_predictions = layers.Flatten()(prob)~\AppData\Roaming\Python\Python38\site-packages\tensorflow\python\keras\layers\pooling.py in __init__(self, data_format, **kwargs) 1021 1022 def __init__(self, data_format=None, **kwargs):-> 1023 super(GlobalPooling2D, self).__init__(**kwargs) 1024 self.data_format = conv_utils.normalize_data_format(data_format) 1025 self.input_spec = InputSpec(ndim=4)~\AppData\Roaming\Python\Python38\site-packages\tensorflow\python\training\tracking\base.py in _method_wrapper(self, *args, **kwargs) 520 self._self_setattr_tracking = False # pylint: disable=protected-access 521 try:--> 522 result = method(self, *args, **kwargs) 523 finally: 524 self._self_setattr_tracking = previous_value # pylint: disable=protected-access~\AppData\Roaming\Python\Python38\site-packages\tensorflow\python\keras\engine\base_layer.py in __init__(self, trainable, name, dtype, dynamic, **kwargs) 345 } 346 # Validate optional keyword arguments.--> 347 generic_utils.validate_kwargs(kwargs, allowed_kwargs) 348 349 # Mutable properties~\AppData\Roaming\Python\Python38\site-packages\tensorflow\python\keras\utils\generic_utils.py in validate_kwargs(kwargs, allowed_kwargs, error_message) 1135 for kwarg in kwargs: 1136 if kwarg not in allowed_kwargs:-> 1137 raise TypeError(error_message, kwarg) 1138 1139 TypeError: ('Keyword argument not understood:', 'keepdims')

[Sivylla Paraskevopoulou]

Hi Peter,

I am sorry you are experiencing difficulties. I just run the following code, which I think is what you are trying to do, and everything run fine. For reference, I am running with Python version 3.9.7.

MATLAB code:

net = squeezenet;

exportNetworkToTensorFlow(net,"SqueezeNet")

Python code:

import SqueezeNet

model = SqueezeNet.load_model()

model.save("exported_squeezenet")

MATLAB code:

net_imported = importTensorFlowNetwork("exported_squeezenet",OutputLayerType="classification")

Regardless, exporting and re-importing is not a workflow I would recommend. It is unclear to me what you are trying to accomplish.

• Do you have a TensorFlow model that you are trying to import? If yes, use the importTensorFlowNetwork function and make sure that your model is in SavedModel format.
• If you are primarily a MATLAB user, there many available MATLAB pretrained networks. Check out the MATLAB Deep Learning Model Hub.

[Peter]

Thanks for trying this. When I try it now, it does work. Who knows? I'm not aware of changing anything.

What I actually want to do is something like this, in matlab:

net=vgg16;

net_import_weights('weights.h5');

where weights.h5 come from a vgg16 net trained differently from the default. I only have the weights file, not the structure that goes with it.