search

ai4r / Gesture-Generation-from-Trimodal-Context

Speech Gesture Generation from the Trimodal Context of Text, Audio, and Speaker Identity (SIGGRAPH Asia 2020)
Other
245 stars 35 forks source link

Problems of calculating FGD #56

Open Jacoo-ai opened 7 months ago

Jacoo-ai commented 7 months ago

Hi, it's a really good work!

But I am confused about the FGD value. When I change the "val_data_path" to 'data/ted_dataset/lmdb_test' , and run "python scripts/synthesize.py eval output/train_multimodal_context/multimodal_context_checkpoint_best.bin" for several times, the result is "2024-03-17 00:47:48,255: [VAL] loss: 0.075, joint mae: 0.02742, accel diff: 0.00322, FGD: 0.0000011660, feat_D: 0.010 / 25.9s". The FGD value is really small, only 0.0000011660, which is not similar with 3.729.

Is this the correct way to calculate the FGD value?

youngwoo-yoon commented 7 months ago

Hello, Your modification seems correct, but the test set is supposed to give similar results with the validation set. I remember I got similar results.