输入图像大小CNN Tensorflow

Question

I am feeding my CNN with image size of 2048x2048. However I accidentaly forgot to change the size of the input:

img_width, img_height = 32, 32
input_shape = (3, img_width, img_height)

I was wondering why does the training still run even if the input images are larger than 32x32 pixels? Or does the CNN only recognizes part of the images?

Thanks in advance.

Answer 1

Most implementations of convolutional layers will operate on any sized images. For 2D convolutions with padding="SAME" , the output size of any convolutional layer will be [batch_size, input_height // stride[0], input_width // stride[1], filters] (valid padding just means you lose a few pixels on each boundary depending on the kernel size).

Similarly, the number of weights involved in the kernel is filters_in * filters_out * np.prod(kernel_size) , which is independent of input size (bias is just filters_out ).

To do tasks such as classification, most CNNs use a spatial pooling layer at the end (eg tf.reduce_mean(features, axis=(1, 2)) ) which reduces the output of any input image batch to [batch_size, n_filters_out] , regardless of input size, followed by some dense layers to regress to logits.