Aradhye2002
Training reproduction problems
I'm very sorry to disturb you again. I encountered a problem when reproducing your code. I commented on the following code in train.py：
if args.ckpt_path == "":
model_str = f"weights_{args.scene}.ckpt"
download_model(model_str)
args.ckpt_path = f"../checkpoints/{model_str}"
model.load_state_dict(torch.load(args.ckpt_path, map_location="cpu", weights_only=True)["state_dict"])
because I need to train from scratch and this section needs to be commented on. I don't know if this is correct. At the same time, I commented on the code in model.py.
Only support finetuning for now
# assert not args.train_from_scratch
After I completed the training,The metrics are very low and the reproduction is not well achieved. I haven't modified any other code configurations. d1 only have 0.5 d2 only have 0.6 and d3 only have 0.9 Do you know the reason? I conducted the training on RTX3090, with batch set to 2 and num_workers to 8. However, the influence of batch should not be so huge. I wonder if you know the reason. And are there any problems with my operation?
I look forward to your answer. Thank you very much for your help to me! Thank you very much!
Please take a look at https://github.com/Aradhye2002/EcoDepth/issues/25.
