Closed chinnusai25 closed 1 year ago
Indeed there are around 5.7k pairs for the validation set. I have to check the training stats you are probably right. If you lower the rejection threshold of the sampler(min_len to 10) I think you will get the amount I am referring to. Be aware, the extracted data are different from the sequences. the sequences are AMASS sequences and we extract the data from those sequences.
Paper states that,
There are approximately 5.7k and 23.4k pairs in the validation and training sets respectively.
But when training is executed we get as below, Loading BABEL train: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 6601/6601 [00:53<00:00, 124.08it/s] [teach.data.babel][INFO] - Processed 6601 sequences and found 3091 invalid cases based on the datatype. [teach.data.babel][INFO] - 15863 sequences -- datatype:separate_pairs. [teach.data.babel][INFO] - 14.13% of the sequences which are rejected by the sampler in total. [teach.data.babel][INFO] - 0.0% of the sequence which are rejected by the sampler, because of the excluded actions. [teach.data.babel][INFO] - 14.13% of the sequence which are rejected by the sampler, because they are too short(<0.5 secs) or too long(>25.0 secs). [teach.data.babel][INFO] - Discard from BML: 0 [teach.data.babel][INFO] - Discard not KIT: 0 Loading BABEL val: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2189/2189 [00:17<00:00, 124.54it/s] [teach.data.babel][INFO] - Processed 2189 sequences and found 983 invalid cases based on the datatype. [teach.data.babel][INFO] - 5672 sequences -- datatype:separate_pairs. [teach.data.babel][INFO] - 16.27% of the sequences which are rejected by the sampler in total. [teach.data.babel][INFO] - 0.0% of the sequence which are rejected by the sampler, because of the excluded actions. [teach.data.babel][INFO] - 16.27% of the sequence which are rejected by the sampler, because they are too short(<0.5 secs) or too long(>25.0 secs). [teach.data.babel][INFO] - Discard from BML: 0 [teach.data.babel][INFO] - Discard not KIT: 0which results that the number of training pairs are 15.8k which doesn't match with that of in paper (i.e 23.4k)