Why is the length of the generated action sequences always different?

Why is the length of the generated action sequences always different? I wonder how to specify the length of the generated actions sequences.

It's a flaw of generating with GPT. You can not predict or control the position of the 'End' token.

Hi,

To generate a specific length of motion, you can modify the self.block_size by using the expected length:

T2M-GPT/models/t2m_trans.py

Line 34 in 7db71a2

for k in range(self.block_size):

Actually, predicting end token is a good feature since model can generate motion without requiring motion length as an input.
However, to explicitly tell the model to stop generating, BAMM modify T2M-GPT as @Jiro-zhang mentioned and add end token as input at the end position to inform the model when to stop ahead of time.