
Base model	`metric3d_vit_giant2` (torch.hub `yvanyin/metric3d`)
Dataset	the DIODE outdoor split (metric depth in metres)
Loss	Direct metric-depth L1 + gradient loss
LoRA rank / alpha	16 / 32
LoRA targets	`qkv`, `proj`

Usage

python

import torch
from peft import PeftModel
import torch.nn as nn
import torch.utils.checkpoint as torch_checkpoint

# Load base model
model = torch.hub.load("yvanyin/metric3d", "metric3d_vit_giant2",
                       pretrain=True, trust_repo=True)

# Apply the same gradient-checkpointing wrapper used during training
# (needed so PEFT key names match the saved adapter)
def enable_gradient_checkpointing(model):
    try:
        encoder = model.depth_model.encoder
    except AttributeError:
        encoder = model.base_model.model.depth_model.encoder

    class _CheckpointedBlock(nn.Module):
        def __init__(self, block):
            super().__init__()
            self.block = block
        def forward(self, x):
            return torch_checkpoint.checkpoint(self.block, x, use_reentrant=False)

    for blk_group in encoder.blocks:
        for key in list(blk_group._modules.keys()):
            blk_group._modules[key] = _CheckpointedBlock(blk_group._modules[key])

enable_gradient_checkpointing(model)
model = PeftModel.from_pretrained(model, "igzi/depth-lora-checkpoints_diode-diode_outdoor")
model.eval()

# Inference: input pixel_values shape (B, 3, 616, 1064), values normalised
# with mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375]
pred_canonical, _, _ = model({"input": pixel_values})
# De-canonicalise: pred_metric = pred_canonical * (fx_scaled / 1000)