Is there a mismatch between the provided training configuration and train.py script, and the 'Implementation Details' section of the UniDepthV2 paper?

[namrata-jangid]

Hi,

There might be a mismatch in the schedulers configuration between the config files and the 'Implementation Details' section of the UniDepthV2 paper.

As per the paper:

However, as per the training configuration (for example: config_v2_vits14.json), the training parameters are set as:
(mentioning only the ones relevant to this issue)

    "training": {    
    "n_iters": 300000,
    "lr": 1e-4,
    "lr_final": 1e-6,
    "lr_warmup": 1.0,
    "wd": 0.1,
    "wd_final": 0.1,
    "warmup_iters": 75000
    }

Should the parameters be changed to below to match the implementation details mentioned in the paper:

    "training": {    
    "n_iters": 300000,
    "lr": 5e-5, # considering the learning rate in the paper
    "lr_final": 5e-6,
    "lr_warmup": 1.0,
    "wd": 0.1,
    "wd_final": 0.01, # assuming weight decay scheduling to one-tenth
    "warmup_iters": 90000 # 30% of 300k total iterations
    }

Additionally, scheduler_wd must also change to reflect the same:

    scheduler_wd = CosineScheduler(
        optimizer,
        key="weight_decay",
        init_value=config["training"]["wd"],
        base_value=config["training"]["wd"],
        final_value=config["training"]["wd_final"],
        warmup_iters=config["training"]["warmup_iters"], # instead of 0
        total_iters=config["training"]["n_iters"],
        step_init=step - 1,
    )

Just curious if the released models were trained using the configurations mentioned in the paper, or do you advice to follow the configurations provided on the repo as a headstart and then further tune as per individual use cases?

Thank you!