search

chaneyddtt / Convolutional-Sequence-to-Sequence-Model-for-Human-Dynamics

Source code of paper "Convolutional Sequence to Sequence Model for Human Dynamics"
MIT License
86 stars 18 forks source link

split in train and test sets #15

Open emedinac opened 2 years ago

emedinac commented 2 years ago

Hi,

This is an interesting job, I would like to do some initial experiments, however I realized the splits are not so balance between training and testing. I mean I saw around 35k sequences for training and only 64 for testing, is it ok? Another question, did you decide that split or started from a reference, probably I missed it, but I didnt understand how you balanced this trials.

Thanks in advanced.

chaneyddtt commented 2 years ago

Hi @emedinac. I suppose that you are referring to the Human3.6m dataset? I follow the previous work for the training and testing splits.

emedinac commented 2 years ago

Oh thanks for that. I will need it anyway. But sorry, I didn't specify the dataset because I only saw CMU -motion in this git. I was wondering about this split.

https://github.com/chaneyddtt/Convolutional-Sequence-to-Sequence-Model-for-Human-Dynamics/tree/master/data/cmu_mocap

chaneyddtt commented 2 years ago

For the CMU dataset, we have 6 sequences for each action category, and we split them into 5:1.

emedinac commented 2 years ago

Hi. Do you have a reference for that split? it is really interesting to know.

Thanks in advanced.

chaneyddtt commented 2 years ago

No we do not have a reference. The thought here is that this dataset is a subset of the original CMU motion datset, and it is quite small. Hence I only use one sequence for testing, and the rest for training. Otherwise the training data might not be enough.