Created by: hedaoyuan

In Paddle's code, caffeMode is a value of type bool (ConvBaseLayer::caffeMode_ and caffe_mode). Set to true or false, will affect the output size of the convolution calculation (outputSize and cnn_output_size).

The calculation result when caffeMode is set to false or true is as follows.

/**
 * Calculate output size based on caffeMode_.
 * - input(+padding): 0123456789
 * - imageSize(+padding) = 10;
 * - filterSize = 3;
 * - stride = 2;
 * - caffeMode is true:
     - output: (012), (234), (456), (678)
     - outputSize = 4;
 * - caffeMode is false:
 *   - output: (012), (234), (456), (678), (9)
 *   - outputSize = 5;
 */
int outputSize(int imageSize, int filterSize, int padding, int stride, bool caffeMode) {
  int outputSize;
  if (!caffeMode) {
    outputSize = (imageSize - filterSize + 2 * padding + stride - 1) / stride + 1;
  } else {
    outputSize = (imageSize - filterSize + 2 * padding) / stride + 1;
  }
  return outputSize;
}

1. The two modes are not the same in some cases.
2. caffeMode = true refers to the calculation method in caffe. This method is also used in cudnn. So, if the convolution layer is configured as cudnn, you need to ensure that caffe_mode = True.
3. If the padding is replaced with paddingLeft and paddingRight two arguments. In caffeMode = true formula, modify the size of the padding can also achieve caffeMode = false results.

So, can we remove this configuration argument?