Question about training loss

[nopQAQ]

Thanks for your excellent work!

I’ve recently been trying to reproduce your work. However, I noticed that when I set the weight of the alignment loss to 0, the training loss does not seem to change noticeably. I was wondering if you might be willing to share your loss curves for reference.

Below are my current results:

And this is my experiment setting:

torchrun --standalone --nnodes 1 --nproc-per-node 4 vla-scripts/finetune_align.py
--vla_path ckpts/openvla-7b
--vggt_path ckpts/vggt.pt
--data_root_dir /mnt/nas/datasets/Spatial-Forcing/modified_libero_rlds/
--dataset_name libero_spatial_no_noops
--run_root_dir ckpts/training_results/
--pooling_func bilinear
--vla_layers_align 24
--vggt_layers_align -1
--align_loss_type cosine
--align_loss_coeff 0.5
--use_l1_regression True
--use_diffusion False
--use_film False
--use_vlm_norm True
--use_vggt_pe True
--num_images_in_input 2
--use_proprio True
--batch_size 8
--learning_rate 5e-4
--num_steps_before_decay 100000
--max_steps 150005
--save_freq 10000
--save_latest_checkpoint_only True
--merge_lora_during_training False
--image_aug True
--lora_rank 32
--wandb_entity "yuhengyuan"
--wandb_project "spatial-forcing"
--run_id_override "spatial-forcing-7b-finetuned-libero-spatial"

[nopQAQ]

This is the cmd for finetuning:
torchrun --standalone --nnodes 1 --nproc-per-node 4 vla-scripts/finetune_align.py
--vla_path ckpts/openvla-7b
--vggt_path ckpts/vggt.pt
--data_root_dir /mnt/nas/datasets/Spatial-Forcing/modified_libero_rlds/
--dataset_name libero_spatial_no_noops
--run_root_dir ckpts/training_results/
--pooling_func bilinear
--vla_layers_align 24
--vggt_layers_align -1
--align_loss_type cosine
--align_loss_coeff 0.0
--use_l1_regression True
--use_diffusion False
--use_film False
--use_vlm_norm True
--use_vggt_pe True
--num_images_in_input 2
--use_proprio True
--batch_size 8
--learning_rate 5e-4
--num_steps_before_decay 100000
--max_steps 150005
--save_freq 10000
--save_latest_checkpoint_only True
--merge_lora_during_training False
--image_aug True
--lora_rank 32
--wandb_entity "yuhengyuan"
--wandb_project "spatial-forcing"
--run_id_override "finetuned-libero-spatial"

Thanks again for your help and consideration!

[haofuly]

Sorry the logs are lost due to this long time. Mostly, in the starting steps, the SF models's loss decreases faster than baseline, then the order of magnitude of action loss bewteen SF and baseline are similar. However, even their loss are in same order of magnitude, their performance may differs.

[zxytql]

@haofuly Hi! How long does it take to finetune using 4 GPUs? I try to finetune libero_4_task_suites_no_noops using openvla-SF on A800 and set grad_accumulation_steps = 2, shows that it need 168 hours. Does this normal?

[haofuly]

Hi @zxytql, for a single task, we train it for 150k steps on 8 H100. This cost 45 hours.