search

roykapon / MAS

The official implementation of the paper "MAS: Multiview Ancestral Sampling for 3D Motion Generation Using 2D Diffusion"
MIT License
92 stars 2 forks source link

Question about Training #4

Open GithubAccountTo opened 4 months ago

GithubAccountTo commented 4 months ago
image

nba_modal in first row is downloaded from (https://drive.google.com/file/d/1P5PlFb3cEjhHky0boHWtwc44xkCFZWdM/view?usp=sharing). And then I use python -m eval.evaluate --evaluator_path save/evaluator/nba/nba_evaluator/checkpoint_1000000.pth --model_path save/nba/nba_diffusion_model/checkpoint_500000.pth --subjects "model" --num_views 5 to evalue the model.

my_nba_model in the second row is trained by next order python -m train.train_mdm --save_dir save/nba/my_nba_diffusion_model --dataset nba And then I use python -m eval.evaluate --evaluator_path save/evaluator/nba/nba_evaluator/checkpoint_1000000.pth --model_path save/nba/my_nba_diffusion_model/checkpoint_500000.pth --subjects "model" --num_views 5 to evalue the model.

my_nba_model in the third row is trained by next order python -m train.train_mdm --save_dir save/nba/my_nba_diffusion_model --dataset nba And then I use python -m eval.evaluate --evaluator_path save/evaluator/nba/nba_evaluator/checkpoint_1000000.pth --model_path save/nba/my_nba_diffusion_model/checkpoint_600000.pth --subjects "model" --num_views 5 to evalue the model.

my_nba_model_finetune in the last row is trained by next order python -m train.train_mdm --save_dir save/nba/my_nba_diffusion_model_finetune --dataset nba --resume_checkpoint save/nba/nba_diffusion_model/checkpoint_500000.pth And then I use python -m eval.evaluate --evaluator_path save/evaluator/nba/nba_evaluator/checkpoint_1000000.pth --model_path save/nba/my_nba_diffusion_model_finetune/checkpoint_600000.pth --subjects "model" --num_views 5 to evalue the model.

Why are the indicators of the last three models lower than the indicators of the first model？

roykapon commented 3 months ago

Hello @GithubAccountTo , thank you for pointing this out. It might have something to do with the default hyperparameters, but I will try to reproduce your problem and look into it.