CNN Keras: ValueError: Negative dimension size caused by subtracting 3 from 2 for 'conv2d

Question

I got this error when using Keras: Is it because input_size not larger than the filter?

If input_shape=(64,64,3))), there will be no error.

 ``ValueError: Negative dimension size caused by subtracting 3 from 2 for 
  'conv2d_24/convolution' (op: 'Conv2D') with input shapes: [?,2,2,128], 
  [3,3,128,128].

My code are here:

from keras import layers
from keras import models
model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu',                     
                                    input_shape=(32, 32, 3))) 
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Flatten())
model.add(layers.Dense(512, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))

Answer 1

Default layer padding is valid , which means no padding. This way dimension reduce from 32 not to 16 but to 15. You can use padding='same' instead. In this case output has the same length as the original input.

from keras import layers
from keras import models
model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu',padding='same',                     
                                    input_shape=(32, 32, 3))) 
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu',padding='same'))
model.add(layers.MaxPooling2D((2, 2),padding='same'))
model.add(layers.Conv2D(128, (3, 3), activation='relu',padding='same'))
model.add(layers.MaxPooling2D((2, 2),padding='same'))
model.add(layers.Conv2D(128, (3, 3), activation='relu',padding='same'))
model.add(layers.MaxPooling2D((2, 2),padding='same'))
model.add(layers.Flatten())
model.add(layers.Dense(512, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))