search

lpiccinelli-eth / UniDepth

Universal Monocular Metric Depth Estimation
Other
639 stars 52 forks source link

Incorrect Results When Running Pre-trained Model Locally #80

Open C-hongfei opened 2 weeks ago

C-hongfei commented 2 weeks ago

Hi, thank you for your excellent work on this project! When I downloaded the pre-trained model to my local environment and ran the demo, I encountered incorrect results. The ARel metric is showing 72.07%, which seems to be much higher than expected. Below are the changes I made. Could you please help me identify what might be causing this issue? Changes Made: hubconf.py,line33, # path = huggingface_hub.hf_hub_download(repo_id=f"lpiccinelli/unidepth-{version}-{backbone}", filename=f"pytorch_model.bin", repo_type="model") path = f"lpiccinelli/unidepth-{version}-{backbone}"+"/pytorch_model.bin" unidepth/models/backbones/dinov2.py, line451, elif pretrained is not None: state_dict = torch.load('dinov2_vits14_pretrain.pth', map_location="cpu") info = model.load_state_dict(state_dict, strict=False)

output

lpiccinelli-eth commented 1 day ago

That output looks like the weights are initialized from random instead of being loaded from the checkpoint. Can you try to see what is contained in info? That returns the info about the checkpoint loading and should return "all keys matched successfully"