Hi, thanks for nice works.

I just have a question regarding the pre-training hyperparmeter, which are set/written different on the arXiv paper, REAME.MD, and the published checkpoint (https://download.europe.naverlabs.com/ComputerVision/DUSt3R/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth).

Which one of the following is the exact optimization hyperparameter to reproduce the results? (And can you possibly give us the missing detailed values?) Specifically, the following difference cause alterations in effective batch size and training epochs. I hope they can be clarified please.

arXiv paper
epochs=25
lr=5e-5
batch_size=4
#gpus=2
accum_iter=UNKNOWN
warmup_epochs=UNKNOWN
weight_decay=UNKNOWN

=> Effective Batch Size = 8 (if accum_iter = 1)

README.MD
epochs=50
lr=1e-6
batch_size=4
#gpus=2
accum_iter=4
warmup_epochs=3
weight_decay=0.05 (code default)

=> Effective Batch Size = 32

checkpoint
epochs=100
lr=5e-5
batch_size=4
#gpus=2
accum_iter=4
warmup_epochs=1
weight_decay=0.05 (code default)

=> Effective Batch Size = 32