Reproduce fine tuning but score poorly on the evaluation dataset

Open
#20 6 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
25/100
Issue type
Bug
Clarity
Needs clarification
Activity status
Stale
Tech stack
bash, python

Research direction

Start with training/train.py and the train_marigold_e2e_ft_depth.sh configuration, then compare the reported arguments with the project's reproduction guidance. Re-run fine-tuning and evaluation on NYUv2, KITTI, ETH3D, ScanNet, and DIODE, and determine why the reproduced scores differ from the expected results.

Written by the indexing model from the issue text.

Description

Hi,
Thanks to the author for the contribution, but I had some problems reproducing it.
Why do I get bad scores on the evaluation dataset when reproducing fine-tuning results on my RTX3090?
The scores are as follows:

Dataset AbsRel ↓ Delta_1 ↑
NYUv2 0.056 0.963
KITTI 0.092 0.928
ETH3D 0.064 0.961
ScanNet 0.062 0.956
DIODE 0.299 0.780

The following is the train script configuration I use, refer to train_marigold_e2e_ft_depth.sh:

Note.
The following are the modified parts:
--checkpointing_steps 500 => to store the best checkpoint
--dataloader_num_workers 4 => speed up training time
--mixed_precision "bf16" => reduce memory usage
--seed 1234 => fixed seed

The complete script is as follows:

#!/bin/bash

accelerate launch training/train.py \
--pretrained_model_name_or_path "prs-eth/marigold-v1-0" \
--modality "depth" \
--noise_type "zeros" \
--max_train_steps 20000 \
--checkpointing_steps 500 \
--train_batch_size 2 \
--gradient_accumulation_steps 16 \
--gradient_checkpointing \
--learning_rate 3e-05 \
--lr_total_iter_length 20000 \
--lr_exp_warmup_steps 100 \
--dataloader_num_workers 4 \
--mixed_precision "bf16" \
--output_dir "model-finetuned/marigold_e2e_ft_depth_bf16" \
--enable_xformers_memory_efficient_attention \
--seed 1234 \
"$@"
Dominant language
Python
Stars
521
Forks
22
PR merge metrics
No merged PRs in 30d

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from VisualComputingInstitute/diffusion-e2e-ft

All issues in VisualComputingInstitute/diffusion-e2e-ft

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.