The eval result is not consistent with the paper

[EmmaThompson123]

Thanks for opensourceing the evaluation code ! I run the commandpython scripts/synthesize_hierarchy.py eval TED-Gesture-output/train_hierarchy/ted_gesture_hierarchy_checkpoint_best.bin 10, then this is its output:

loading checkpoint TED-Gesture-output/train_hierarchy/ted_gesture_hierarchy_checkpoint_best.bin
epoch 98
Embedding size is 32, encoder SAP.
Embedding size is 32, encoder SAP.
Embedding size is 32, encoder SAP.
Embedding size is 32, encoder SAP.
2025-01-02 15:07:39,863: Reading data '/dataset/ted_db/lmdb_val'...
2025-01-02 15:07:39,867: Found the cache /dataset/ted_db/lmdb_val_cache
26903
2025-01-02 16:14:51,523: [VAL] loss: 0.082, joint mae: 0.02964, accel diff: 0.00374, FGD: 4.030, feat_D: 63.015, Diversity: 104.646, BC: 0.7007 / 1996.6s

but the paper said

so how to explain this difference ?