Cannot reproduce Kinetics accuracy, could anybody please help?

I tried to reproduce the experiments on Kinetics 400 (I only downloaded 214738 out of 246,535 videos). I use the config configs/Kinetics/SLOWFAST_8x8_R50.yaml, the batch size, GPU num and other parameters are exactly the same. During training, I keep getting a warning "Failed to decode by pyav with exception: tuple index out of range".
The accuracy I get is **"top1_err": 40.407398, " top5_err": 19.408849**
Now I'm not sure what causes the accuracy gap. Does the pyav warning or the incomplete dataset hurt the accuracy?