VGG16 inconsistent input shape

[sagarvijaygupta]

With MaxPooling2D with strides, size of image reduces in following manner

(32, 24, 3) (?, 16, 12, 64) (?, 8, 6, 128) (?, 4, 3, 256) (?, 2, 1, 512)

which in next step generates an error

ValueError: Negative dimension size caused by subtracting 2 from 1 for 'max_pooling2d_5/MaxPool' (op: 'MaxPool') with input shapes: [?,2,1,512].

[marco-c]

32x24 might be too small, it was chosen arbitrarily. We should investigate what would be the best size.

[sagarvijaygupta]

Generally for pretrained models we use size of 224x224 or 227x227. We can think about it.