Can the data set loading code for the training process be made public? Or can you provide the usage of the datasets and several training data sets in the project?

[myprospection]

Thank you very much for sharing your code. I am currently working on applying your approach to my own research, and I would like to gain a deeper understanding of the entire training process.

However, I am currently facing difficulties in reproducing the training pipeline, particularly due to issues related to dataset loading. Even after carefully reading the accompanying paper, I still have some doubts and areas of confusion.

My main goal in writing this message is to ask about the usage of the scripts in the datasets folder. I would like to know how these scripts are intended to be used, and whether the dataset loading logic used during training is publicly available or could be described in more detail. Additionally, I would appreciate any advice on how I could adapt the provided dataset scripts to construct a data loader for my own dataset that fits into your training framework.

To resolve these questions—and to move forward with my research as efficiently as possible—I would like to understand the complete training process from a data flow perspective. This would greatly help me grasp the structure and logic of the pipeline.

I would truly appreciate any guidance or clarification you could provide. Thank you again for your valuable work!

[yocabon]

Hi,
I do not understand the question. Could you be more specific ?

We did not release the preparation code for all of the dataset, but as I mentioned in this issue, all the building blocks are there.
There are some differences compared to the published paper and these are mentioned here

[myprospection]

Hi, I do not understand the question. Could you be more specific ?

We did not release the preparation code for all of the dataset, but as I mentioned in this issue, all the building blocks are there. There are some differences compared to the published paper and these are mentioned here

I apologize for my unclear explanation. While I couldn't locate the pre-training code during the reproduction process, I did find Must3r's preprocessing code. Could you share the pre-training code? Alternatively, could you provide guidance to help me reproduce this preprocessing workflow more efficiently? Additionally, since there's no must3r dataset loading code available, could you also share it or offer compilation instructions similar to the preprocessing documentation?
Thank you for your reply. I will try again according to your reply

[myprospection]

Hi, I do not understand the question. Could you be more specific ?

We did not release the preparation code for all of the dataset, but as I mentioned in this issue, all the building blocks are there. There are some differences compared to the published paper and these are mentioned here

MKL_NUM_THREADS=1 NUMEXPR_NUM_THREADS=1 OMP_NUM_THREADS=1 PYTHONUNBUFFERED=1 torchrun --nnodes 1 --nproc_per_node 1 --master
_port $((RANDOM%500+29000)) train.py --decoder "CausalMUSt3R(img_size=(512, 512), feedback_type='single_mlp', memory_mode='kv', mem_dropout=0.1, dropout_mode='temporary', use_xformers_ma
sk=True, use_mem_mask=True)" --epochs 100 --warmup_epochs 10 --memory_num_views 20 --memory_batch_views 5 --min_memory_num_views 2 --causal --loss_in_log --amp bf16 --use_memory_efficien
t_attention --batch_size 1 --accum_iter 4 --dataset "100@Co3d(split='train',ROOT='./data/co3d_subset_processed', resolution = [(512, 384), (512, 336), (512, 288), (512, 256), (512, 160)]
, aug_crop=64, num_views=20, min_memory_num_views=2, max_memory_num_views=20, transform=ColorJitter)" --chkpt "./checkpoints/MUSt3R_512_retrieval_trainingfree.pth" --output_dir "./output_dir"
I am now trying to reproduce it according to the instructions, but there is a prompt that the pre-training parameter file does not have the corresponding decoder, and I have attached the error information and print information files of the command execution.
Is it my wrong way to use it? I hope to get your guidance. Thank you
out.txt

[yocabon]

--chkpt "./checkpoints/MUSt3R_512_retrieval_trainingfree.pth" is incorrect. It's not a complete checkpoint and you cannot use it as initialization. MUSt3R_512.pth should be OK.

[myprospection]

Thank you. I'll try again.

…

---Original--- From: "Yohann ***@***.***> Date: Sun, Aug 17, 2025 22:06 PM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [naver/must3r] Can the data set loading code for the trainingprocess be made public? Or can you provide the usage of the datasets andseveral training data sets in the project? (Issue #4) yocabon left a comment (naver/must3r#4)

--chkpt "./checkpoints/MUSt3R_512_retrieval_trainingfree.pth" is incorrect. It's not a complete checkpoint and you cannot use it as initialization. MUSt3R_512.pth should be OK. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[myprospection]

--chkpt "./checkpoints/MUSt3R_512_retrieval_trainingfree.pth" is incorrect. It's not a complete checkpoint and you cannot use it as initialization. MUSt3R_512.pth should be OK.

The system is now operational, but when I attempt to render images in RGBA format using the ShapNet dataset for training, modifying the image reading mode results in either all 0s or 255 values. Additionally, the visual images in the ShapNet dataset are rendered models with dimensions of 137×137. Could you provide any clues? I suspect the issue might be related to image dimensions or reading mode settings. Are there any other potential causes?