ValueError: The input must have 3 channels; got `input_shape=(200, 200, 1)`

Question

I am trying to use Transfer learning with VGG16. I am using Keras. But I got error on

vgg = vgg16.VGG16(include_top=False, weights='imagenet', input_shape=(IMG_SIZE, IMG_SIZE, 1))

Any help what is wrong ?

Note: IMG_SIZE = 200

The trace of error is

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-5-1b17094c93e2> in <module>
      3 import keras
      4 
----> 5 vgg = vgg16.VGG16(include_top=False, weights='imagenet', input_shape=(IMG_SIZE, IMG_SIZE, 1))
      6 
      7 output = vgg.layers[-1].output

c:\users\hiteshsom\documents\deepanshu_q2\env\lib\site-packages\tensorflow\python\keras\applications\vgg16.py in VGG16(include_top, weights, input_tensor, input_shape, pooling, classes, classifier_activation)
    124                      ' as true, `classes` should be 1000')
    125   # Determine proper input shape
--> 126   input_shape = imagenet_utils.obtain_input_shape(
    127       input_shape,
    128       default_size=224,

c:\users\hiteshsom\documents\deepanshu_q2\env\lib\site-packages\tensorflow\python\keras\applications\imagenet_utils.py in obtain_input_shape(input_shape, default_size, min_size, data_format, require_flatten, weights)
    363           raise ValueError('`input_shape` must be a tuple of three integers.')
    364         if input_shape[-1] != 3 and weights == 'imagenet':
--> 365           raise ValueError('The input must have 3 channels; got '
    366                            '`input_shape=' + str(input_shape) + '`')
    367         if ((input_shape[0] is not None and input_shape[0] < min_size) or

ValueError: The input must have 3 channels; got `input_shape=(200, 200, 1)`

Answer 1

You cannot use the imagenet weights whith single channel images. This may solve your issue:

vgg = vgg16.VGG16(include_top=False, weights=None, input_shape=(IMG_SIZE, IMG_SIZE, 1))

Answer 2

just add before downloading VGG16 pretrained model.

tf.keras.backend.set_image_data_format("channels_last")

Changing channel position from first to last will solve the error.

Answer 3

Supposedly you want to use the VGG16 pre-trained model on ImageNet dataset you should process your input image to be RGB instead of grayscale. Is it possible to provide how you load your images?

For example if you use opencv you should probably do something like this:

rgb_img = cv2.cvtColor(grayscale_image,cv2.COLOR_GRAY2RGB)

If you have your image as a Tensor object then you can do something like:

rgb_tensor = tf.image.grayscale_to_rgb(grayscale_tensor)

or

rgb_tensor = tf.concat([grayscale_tensor, grayscale_tensor, grayscale_tensor], axis=-1)

since, if you have a grayscale image, the rgb equivalent is expressing the grayscale values on each RGB channel.

Answer 4

You're trying to use 1 channel image with VGG16 net, whereas you must use 3.

This'll solve your issue:

vgg = vgg16.VGG16(include_top=False, weights=None, input_shape=(IMG_SIZE, IMG_SIZE, 3))