Certainguys
commented
3 months ago Let me add my best experimental results：
====> Test at epoch #120 Group Activity Accuracy: 82.57%, Loss: 11.18372, Using 556.2 seconds
Best group activity accuracy: 83.25% at epoch #32.
When I follow your steps and run the python scripts/train_volleyball_stage1.py command, there is an error ValueError: cannot reshape array of size 90 into shape (12). When I navigated to line 266 of the volleyball.py file actions = np.hstack(actions).reshape([-1, self.num_boxes]), I found that there was a piece of code at lines 259 to 261 when the array was stacked earlier. : if len(boxes[-1]) ! = self.num_boxes: boxes[-1] = np.vstack([boxes[-1], boxes[-1][:self.num_boxes-len(boxes[-1])]]) actions[-1] = actions[-1] + actions[-1][:self.num_boxes-len(actions[-1])]. When I printed, I found that the length of len(boxes[-1]) was equal to the length of self.num_boxes, so the inequality sign (!=) caused this code to not work. When I put The program can run when != is changed to ==. But the accuracy of the first stage is not up to the accuracy in your code. Does this issue have an impact on the results?Can you provide the checkpoints of your first stage? Thanks