Motion2Text generation always gives something irrelevant

[zsolt-csibi]

Hi!
I am trying to do text2motion and back to motion2text, and my idea was to generate a motion from text then from that motion generate text, and check the similarity.

For example:
The motion is:

The result is:
a person walks backwards and then duckwalks forward.

I put the npy file path into the the txt as the format asks.

Is someone have success with the motion2text part? Because for me only the text2motion works.