chuhanxx / temporal_query_networks Goto Github PK

Hi, thanks for your work!

I feel confused about the dataset preparation, like Diving48 should I extract frames from each mp4?

Hi, thank you for your contribution. Could you tell me where the self.queue in fill_SUFB is defined?

Very wonderful work！ But when I download S3D weights pretrained on Kinetics400, I can't open the link. Hope the author can update the link, or send a copy to my email. I would be grateful if I could receive a reply.

I trained tqn on Diving48 with 4 v100 gpus and a batchsize of 16, the best acc output by your test code on 1st stage is about 60%, and the final acc on 2nd stage is 73.5%, I guess this acc is per video acc. Then I used my code to compute per class acc, got about 65% on the second stage.
Both results are lower than the reported results(74.5% per class acc, 80+% per video acc).

Hi! Just tried out a mwe here

import pickle
with open("annotations/gym99_anno.pkl", "r") as file:
    l = pickle.load(file)
print(l)

However it errors with the default "utf-8" encoding saying that:

Traceback (most recent call last):
  File "scripts/toy.py", line 3, in <module>
    l = pickle.load(file)
  File "/GPFS/rhome/kunhaozheng/miniconda3/envs/tqn/lib/python3.7/codecs.py", line 322, in decode
    (result, consumed) = self._buffer_decode(data, self.errors, final)
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 0: invalid start byte

Hi, thanks for your work! I noticed that data.md is unavailable, could you mind helping me by updating the page?