Difference in Results When Using Provided Checkpoints vs. Training from Scratch

[mansooreh1]

Hello,
for running "python eval_skeleton.py --resume_checkpoint PATH/TO/YOUR/DIFFUSION.ckpt --resume_checkpoint_obj PATH/TO/YOUR/CORRECTION.ckpt":
when I use the checkpoint that you have placed in the checkpoints folder(i.e. diffusion_skeleton.ckpt , obj_skeleton.ckpt), the results are as follows:
{'mpjpe_h': 0.02885597400245669, 'mpjpe_o': 0.0679867499322957, 'translation_error': 0.041375365246613505, 'rotation_error': 0.1221446220919733}
{'mpjpe_h': 0.11084751846422826, 'mpjpe_o': 0.05641989930904001, 'translation_error': 0.040981526801684744, 'rotation_error': 0.07841806445634902}
But when I train from scratch in correction phase and run"python train_correction_skeleton.py" and use the .ckpt file in the results folder as the checkpoint(as obj_skeleton.ckpt), the results are as follows.
{'mpjpe_h': 0.028856130931096662, 'mpjpe_o': 0.2502403250556349, 'translation_error': 0.06720727911878316, 'rotation_error': 0.6058410595626595}
{'mpjpe_h': 0.110847374550382, 'mpjpe_o': 0.24488443606778196, 'translation_error': 0.0749315749993898, 'rotation_error': 0.5167033990522972}
Why are the results different?
thanks