Testing speed is slow

[RuijieJ]

The testing speed reported in the paper is 20ms/image. I modified demo.py and did recognition on the ICDAR 2003 dataset, and print the average recognition time. The following is the modified code (the red boxes indicate my modifications):

I used a single NVIDIA Tesla V100 GPU to do the recognition. I ignore the time comsumption of reading image, so the recognition time has nothing to do with the disk. And the output is:

Avg time = 693.183ms

which is larger than 20ms. I wonder is there something wrong with my modification? I would be appreciated if my question can be answered.

[RuijieJ]

As mentioned by #67 and #42 , the model restoring step is very slow (~10min). I wonder is there any connection between model restoring speed and recognition speed.