Question about rescaling images for AVA training

Hi, thanks for the great work. I have a question about improving the I/O burden during AVA training: Does it make sense to rescale the images to -1:320 first since the maximum scale jitter size is 320 during training so it could reduce the I/O loading time?

I rescale the images by:
```
ffmpeg -i videos_15min/${video} -r 30 -vf scale=-1:320 -q:v 1 ${outframe}
```

I can confirm that for testing with original frames and -1:320 frames the mAP is similar (0.297711 vs. 0.297617) for SLOWFAST_64x2_R101_50_50 and the speed increases by about 30%.