pretrained model performed very bad on K400

Hi, I downloaded the pretrained model: SLOWFAST_4x16_R50.pkl, and test it on k400 val dataset, the top1_acc is 58.14% and top_5 is 79.67%。

My script is:

python tools/run_net.py \
--cfg ./configs/Kinetics/SLOWFAST_4x16_R50.yaml \
DATA.PATH_TO_DATA_DIR ./data \
TEST.CHECKPOINT_FILE_PATH ./models/SLOWFAST_4x16_R50.pkl \
TRAIN.ENABLE False \
TEST.CHECKPOINT_TYPE caffe2 \

For some reason , pyav cannot get frames_length, so I calculate it using duration/timebase, So I twist the code a little bit(line #131):

fps = float(container.streams.video[0].average_rate)
timebase = container.streams.video[0].time_base
duration = float(container.streams.video[0].duration)
frames_length = duration/timebase