Skip to content

Title: Request to release the complete training code and accompanying modules #9

Description

@feixuedudiao

@rwn17
First of all, thank you very much for sharing this excellent work!

I noticed that the official repository currently only releases the inference code. It was also mentioned in Issue #7 that the training code is planned for release this June. In order to train models on our own data based on the Fin3R method—rather than just loading pre-trained LoRA weights for inference—I would like to kindly request the release of the complete training code.

Based on my analysis of the existing codebase, I believe the complete training code should include the following key modules:

  1. Teacher Model Abstraction
    The paper mentions using monocular teacher models (e.g., MiDaS, DPT, etc.) for knowledge distillation, but the current code does not show a unified interface for interacting with teacher models. I hope the released code can provide a clear teacher model abstraction layer to facilitate the replacement of different teacher models.

  2. Teacher Target Dataset/Cache
    Distillation training requires pre-computing or generating teacher model outputs (e.g., depth maps, normal maps, etc.) on the fly. For large-scale training, an efficient pre-computation and caching mechanism is crucial. I look forward to seeing the implementation of this aspect.

  3. Distillation Loss
    This is the core of Fin3R. I hope the released code can fully implement the various distillation losses described in the paper (e.g., L1/L2 loss, perceptual loss, gradient loss, etc.) and allow users to flexibly combine them.

  4. LoRA Training Initialization & Adapter Checkpoint Management
    The existing code demonstrates how to load LoRA weights (model.apply_lora()), but lacks the code for training these LoRA weights. I hope the release includes complete LoRA initialization, training loops, and adapter checkpoint saving/loading logic.

  5. Encoder/Decoder Freezing Strategy
    One of the core strategies of Fin3R is to "freeze the decoder and only fine-tune the encoder". I hope the training code clearly demonstrates how to implement this parameter grouping and freezing logic.

  6. Dataset Recipe
    The paper mentions sampling 20,000 images from SA-1B, 1,000 from Hypersim, and 1,000 from TartanAir per epoch. I hope the release provides the specific preprocessing scripts and configurations for these datasets.

Thank you again for your open-source contribution. I look forward to the release of the training code!
Fei

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions