search

li-ronghui / LODGE

The code the CVPR2024 paper Lodge: A Coarse to Fine Diffusion Network for Long Dance Generation Guided by the Characteristic Dance Primitives
110 stars 6 forks source link

Val loss increase when training #20

Open Augustby opened 4 months ago

Augustby commented 4 months ago

Thanks for your impressive work! I'm trying to train LODGE model from scratch to reproduce the results. However, when training the global diffusion and finetuning the local diffusion, the train loss gradually decreased，while the val loss started to increase after a certain number of iterations. This suggests the model may be overfitting. I followed the same config files as you provided here, but only modified the batch size to prevent OOM on my GPU. Could you provide more detailed instruction on training to solve this problem and how to select the final checkpoints of two diffusion models?

Global Diffusion Loss global loss Local Diffusion Loss (finetune) local loss

li-ronghui commented 4 months ago

Currently, the Vloss/val in the code calculates the error between the velocity of generated data and the ground truth (GT). For generation tasks, in the testing phase, we do not require the generated results to be the same as the ground truth (GT). We only require the generated dance to be realistic and natural. Therefore, the Vloss/val make no practical sense, so please ignore it. Replacing Vloss with FID would make more sense, but this would slow down the training.