pjbull
commented
6 months ago Hi @CrazyGeG, if I am understanding your question right, you are asking about "multi-label" scenario where there are multiple species in a single video. In this case, you should have multiple rows for each label.
For example, the video chimp_elephant.MP4 below appears twice, once for each species in the video:
filepath,label
blank.MP4,blank
leopard.MP4,leopard
chimp_elephant.MP4,chimpanzee_bonobo
chimp_elephant.MP4,elephant
CrazyGeG
Hi, I have tried to train zamba on other dataset having multi-class label, labels on my dataset is not a subset of zamba class. When I train zamba on my dataset and using time_distributed model, zamba seems not able to read multi-class label under label column on train.csv file, it gives me error like: Not all species have enough videos to allocate into the following splits: train, val, holdout. A minimum of 3 videos per label is required. Found the following counts: {"['amphibian', 'mammal']": 1, "['bird', 'mammal', 'sea animal']": 2, "['fish', 'insect']": 2}. Either remove these labels or add more videos. (type=value_error
The command I use is : zamba train --data-dir --labels --save-dir --model time_distributed