Negative size error

[zeynepdenizgundogan]

I am trying to use convnext for images with 128x128x19x1 size but I get the following error why might that happen ; ValueError: Exception encountered when calling layer 'convnext_small_downsampling_conv_2' (type Conv3D).

Negative dimension size caused by subtracting 2 from 1 for '{{node convnext_small_downsampling_block_2/convnext_small_downsampling_conv_2/Conv3D}} = Conv3D[T=DT_FLOAT, data_format="NDHWC", dilations=[1, 1, 1, 1, 1], padding="VALID", strides=[1, 2, 2, 2, 1]](convnext_small_downsampling_block_2/convnext_small_downsampling_layernorm_2/batchnorm/add_1, convnext_small_downsampling_block_2/convnext_small_downsampling_conv_2/Conv3D/ReadVariableOp)' with input shapes: [?,8,8,1,384], [2,2,2,384,768].

Call arguments received by layer 'convnext_small_downsampling_conv_2' (type Conv3D): • inputs=tf.Tensor(shape=(None, 8, 8, 1, 384), dtype=float32)

[ZFTurbo]

You can't use 19 for 3rd dimension. Try 32.

If you want to use 19, then you need to use custom stride size: Something like: ((2, 2, 1), (2, 2, 1), (2, 2, 2), (2, 2, 2), (2, 2, 2)) or ((2, 2, 2), (2, 2, 2), (2, 2, 2), (2, 2, 2), (2, 2, 1))

[zeynepdenizgundogan]

So is this the part that I should update ?
if type(stride_size) not in (tuple, list): stride_size = [ (stride_size, stride_size, stride_size,), (stride_size, stride_size, stride_size,), (stride_size, stride_size, stride_size,), (stride_size, stride_size, stride_size,), ] else: stride_size = list(stride_size)

[ZFTurbo]

I don't know which code you use. This works for me:

def tst_keras():
    # for keras
    from keras import __version__
    from keras import backend as K
    from keras import models
    from classification_models_3D.keras import Classifiers

    print('Keras version: {}'.format(__version__))
    include_top = False
    # use_weights = 'imagenet'
    use_weights = None
    list_of_models = [
         'convnext_tiny',
    ]
    for type in list_of_models:
        modelPoint, preprocess_input = Classifiers.get(type)
        model = modelPoint(input_shape=(128, 128, 32, 1), include_top=include_top, weights=use_weights)
        print(model.summary())
        K.clear_session()

if __name__ == '__main__':
    tst_keras()

[ZFTurbo]

Also this works for me:

    modelPoint, preprocess_input = Classifiers.get('convnext_tiny')
    model = modelPoint(
        input_shape=(128, 128, 19, 1),
        include_top=include_top,
        stride_size=((2, 2, 2), (2, 2, 2), (2, 2, 2), (2, 2, 1)),
        weights=use_weights
    )

Note: number of strides for convnext is 4. Usually it 5 for almost all other models.

[zeynepdenizgundogan]

Okay thank you so much!