pi0-fast fine-tuned in my custom dataset doesn't work

[reinaldoyang]

Hi Repo owner! Thank you for taking the time to read this issue I tried using pi0-fast as it's a autoregressive model, but it seems that the predicted action doesn't make sense at all My custom dataset involves a robotic arm trying to pick up a cube and put the cube on the tray. When I fine-tune my dataset on pi0, the robot arm tried to approach the cube but maybe fails to grasp, then go to the tray. But for pi0-fast, it doesn't work at all and I also encountered this error, it seems that I saw on other issues that I need to increase training steps to solve this error decoding tokens, but I already trained 15.000 steps for this. Error decoding tokens: cannot reshape array of size 1024 into shape (7) Tokens: [280, 363, 265, 295, 314, 294, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 571, 265, 642, 289, 745, 320, 280, 362, 387, 881, 343, 285, 457, 960, 307]