Questions about the MultiEvalModule

Hello,

I'm trying to do multi-scale evaluation for replicating the results of another paper "expectation maximization attention networks (EMANet) for semantic segmentation (ICCV 2019), but I couldn't understand some parts of the MultiEvalModule, which I want to tailor to the EMANet paper.

1. The first thing I don't quite understand is some of the terminology used in the code:
   a) self.base_size
   b) stride_rate
   c) stride

2. why is the stride_rate set to 2.0 / 3.0?

3. why does stride = int(crop_size * stride_rate)?

4. why is the image from the dataset, which has already been cropped due to val_sync_transfrom, cropped again in MultiEvalModule?

Please forgive my potential ignorance but the code in MultiEvalModule seems much more complicated than what I expected MS + FLIP evaluation to be (I thought it would be simple as scaling the image from the dataset, which is already cropped, ,flipping it with prob = 0.5, feed forwarding it through the network, interpolating the smaller scale images to match the largest scale image, and then average the responses at each pixel).

Read the MultiEvalModule implementation and the val_sync_transfrom path first, tracing base_size, stride_rate, stride, and the second crop. Because no file or test is named and the issue requests explanations rather than a specific change, confirm the desired documentation scope with a maintainer; done would be clear documented answers to the listed questions.