Part-Coordinated Transformer train

[Seoneun]

Sorry for asking so many questions.

CUDA_VISIBLE_DEVICES=0 python train_ParCo_trans.py
--vqvae-train-dir output/00000-t2m-ParCo/VQVAE-ParCo-t2m-default/
--select-vqvae-ckpt last
--exp-name ParCo
--pkeep 0.4
--batch-size 128
--trans-cfg default
--fuse-ver V1_3
--alpha 1.0
--num-layers 14
--embed-dim-gpt 1024
--nb-code 512
--n-head-gpt 16
--block-size 51
--ff-rate 4
--drop-out-rate 0.1
--total-iter 300000
--eval-iter 10000
--lr-scheduler 150000
--lr 0.0001
--dataname t2m
--down-t 2
--depth 3
--quantizer ema_reset
--dilation-growth-rate 3
--vq-act relu

You chose "last" in "select-vqvae-ckpt", but why did you choose "last"? Is there a reason why you didn't train by selecting the best fid values?

Thank you for your kind reply every time.

[Seoneun]

Of course, the performance was reproduced well even when it was done "last".

[qrzou]

We follow T2M-GPT choosing "last" VQ-VAE ckpt for training Part-Coordinated Transformer.
And for testing Part-Coordinated Transformer, we choose "fid" for HumanML3D and "last" for KIT dataset. Because the results on validation set seem to appear more stable when we choose in that way.