search

WindVChen / INR-Harmonization

A novel image harmonization method based on Implicit Neural Representation.
Apache License 2.0
54 stars 4 forks source link

about infer speed? #1

Closed zhanghongyong123456 closed 1 year ago

zhanghongyong123456 commented 1 year ago

I didn't find anything about the speed of reasoning.What is the speed of inference

WindVChen commented 1 year ago

Hi @zhanghongyong123456 ,

The following table takes a single 2048x2048 image as an example and compares the efficiency of our method with existing methods from the perspective of model parameters (Params), calculation amount (MACs), memory overhead (Mem) during training and inference, and inference runtime (Time). For the memory cost and runtime of inference, we show results when the input is split into 1/4/16 parts (splitting script will be released recently. Check it in TODO).

Due to the Random Step Crop (RSC) strategy designed in Section 3.4, our training memory cost is less than 3GB among all methods. By varying the number of input splits (splitting will not affect the performance due to the characteristic of INR), we can achieve competitive performance with color-to-color (C2C) methods either on inference memory cost or runtime.

(Note: OOM denotes out-of-memory)

We will also update these in new editions of the paper.

WindVChen commented 1 year ago

Hi @zhanghongyong123456 ,

I would like to inform you that we have updated the efficient inference script in our latest version. Hope you enjoy that. 😃

zhanghongyong123456 commented 1 year ago

Hi @zhanghongyong123456 ,

I would like to inform you that we have updated the efficient inference script in our latest version. Hope you enjoy that. smiley

ok,thank you very much