Tensorflow Conv2D layers input_shape configuration error: ValueError: Input 0 of layer "sequential" is incompatible with the layer:

Question

I have a question about the input_shape of CONV2D layer. As the official doc says:

4+D tensor with shape: batch_shape + (channels, rows, cols) if data_format='channels_first' or 4+D tensor with shape: batch_shape + (rows, cols, channels) if data_format='channels_last'.

In my case, I assemble the dataset by using the method:

tf.data.Dataset.from_tensor_slices()

The detail goes like this. First, I converted a bunch of image into tensor. The shape is:

(100, 128, 128, 3) -> in the format of (numbers of images, width, height, channels)

And, the ground truth of labels were converted into tensor also. In my case, the each label is a 155 long list with float32. The shape is:

(100, 155) -> in the format of (numbers of labels, labels)

Then, I created the dataset using the method which mentioned above:

   dataset = tf.data.Dataset.from_tensor_slices((fetched_image_list, labels))

Next, I iterated a sample in the dataset.

import pprint
pp = pprint.PrettyPrinter()
pp.pprint(f"---------Test----------")
iterator = iter(dataset)
data, label = iterator.get_next()
pp.pprint(data)
pp.pprint(data.shape)

The result is :

'---------Test----------'
<tf.Tensor: shape=(128, 128, 3), dtype=float64, numpy=
array([[[0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628],
        ...,
        [0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628]],

       [[0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628],
        ...,
        [0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628]],

       [[0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628],
        ...,
        [0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628],
        [0.        , 0.        , 0.01568628]],

       ...,

       [[0.94117653, 0.37254903, 0.36862746],
        [0.9571079 , 0.52573532, 0.44705883],
        [0.99215692, 0.86274517, 0.61960787],
        ...,
        [0.99607849, 0.80000007, 0.56078434],
        [0.99607849, 0.79460794, 0.55539221],
        [0.99607849, 0.79215693, 0.5529412 ]],

       [[0.94117653, 0.37254903, 0.36862746],
        [0.9571079 , 0.52573532, 0.44705883],
        [0.99215692, 0.86274517, 0.61960787],
        ...,
        [0.99607849, 0.80000007, 0.56078434],
        [0.99607849, 0.79460794, 0.55539221],
        [0.99607849, 0.79215693, 0.5529412 ]],

       [[0.94117653, 0.37254903, 0.36862746],
        [0.9571079 , 0.52573532, 0.44705883],
        [0.99215692, 0.86274517, 0.61960787],
        ...,
        [0.99607849, 0.80000007, 0.56078434],
        [0.99607849, 0.79460794, 0.55539221],
        [0.99607849, 0.79215693, 0.5529412 ]]])>
TensorShape([128, 128, 3])

However, I mentioned the official doc about CONV2D where the input should be either (batch_shape, row, colts, channels) or (batch_shape, channels, rows, cols). So I configured my CNN model like this:

model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(32, (3, 3), strides=(2,2), padding="same", dilation_rate=(1,1), input_shape=(128, 128, 3), activation='relu'),
    tf.keras.layers.Conv2D(71, (3, 3), strides=(2,2), padding="same", dilation_rate=(1,1), activation='relu'),
    tf.keras.layers.Conv2D(128, (3, 4), strides=(2,3), padding="same", dilation_rate=(1,1),activation='relu'),
    tf.keras.layers.Conv2D(128, (3, 3), strides=(2,2), padding="same", dilation_rate=(1,1),activation='relu'),
    tf.keras.layers.Conv2D(128, (3, 4), strides=(2, 3), padding="same", dilation_rate=(1, 1), activation='relu'),
    tf.keras.layers.Conv2D(128, (3, 3), strides=(2, 2), padding="same", dilation_rate=(1, 1), activation='relu'),
    tf.keras.layers.Dropout(0.20),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(155)
])

But I got this error in return:

    raise ValueError(f'Input {input_index} of layer "{layer_name}" is '

ValueError: Input 0 of layer "sequential" is incompatible with the layer: expected shape=(None, 128, 128, 3), found shape=(128, 128, 3)

If I had changed the input_shape into (32, 128, 128, 3). A similar error popped out:

ValueError: Input 0 of layer "sequential" is incompatible with the layer: expected shape=(None, 32, 128, 128, 3), found shape=(128, 128, 3)

I think the problem exist in the way I created the dataset.

How can I fix this issue? What is your suggestion?

Answer 1

You need to set a batch size for the dataset like this

batch_size = 8 #Or some number
dataset = dataset.batch(batch_size)

If you don't do this, your image goes into the model in the shape of (128,128,3) instead of (batch_size,128,128,3) which is what Conv2D layer expects.

If you want to send 1 image at a time, set batch size as 1