Keras model conversion: image_input_names not working

Question

I am trying to perform a classification task. I am using `keras` and my model starts with an input layer as shown below:

data = Input(shape=(height, width, depth))

Once training is complete, I have the following two lines to create the `CoreML` model:

coreml_model = coremltools.converters.keras.convert(model, input_names='data', image_input_names='data', is_bgr=True)
coreml_model.save('my_model.mlmodel')

However, when I check this file, I see Input as `MultiArray` instead of `Image <BGR, 32, 32>` or `CVPixelBuffer`. How can I make sure that my input is an image?

Core ML

1.2k

Posted by

syamprasadkr

Reply

Add a Comment

Answer 1

Is 'data' the name of the keras input layer?

Posted by

BrianOn99

Add a Comment

Answer 2

The input needs to be in shape (depth, height, width) in order to be converted. Here's the code:

array_shape = tuple(input_.type.multiArrayType.shape)
channels, height, width = array_shape

Also make sure channels is either 1 or 3.

Posted by

FrankSchlegel

Add a Comment

Answer 3

coreml_model_new = coremltools.converters.keras.convert(final_model,input_names='image',
                                                        output_names='probs',image_input_names='image',
                                                        class_labels='classes.txt',
                                                        predicted_feature_name='class',image_scale=1./255)

I get the error in this line when I try to transform it to 'image'.

array_shape = tuple(input_.type.multiArrayType.shape)
channels, height, width = array_shape
ValueError: need more than 1 value to unpack

I'm not sure what happened but it worked with a keras model I used with an earlier version of coremltools.

Posted by

Flowinger

Add a Comment

Keras model conversion: image_input_names not working

Replies