FeatUp Checkpoint Missing: Need A Mirror!

Hey guys,

I'm super excited to dive into this Generative View Stitching (GVS) codebase – it looks awesome! Huge thanks to the developers for putting this out there. I'm trying to reproduce the results, and I've hit a little snag with the FeatUp dependency, which is used by the MEt3R metric in the pipeline.

The Missing Checkpoint Problem

The problem I'm running into is that the FeatUp pretrained checkpoint for the DINO16 backbone, which is referenced in MEt3R as feature_backbone_weights="mhamilton723/FeatUp", seems to be unavailable right now. The original Azure Blob link is giving me this error:

<Error>
 <Code>PublicAccessNotPermitted</Code>
 <Message>Public access is not permitted on this storage account.</Message>
</Error>

This basically means I can't download the necessary weights to get things running. I'm guessing the public access settings on the storage account have been changed, or the file has been moved. This is a pretty common issue in the world of open-source research, as links can break and resources can disappear. It can be a real bummer when you're trying to reproduce results, and a key piece of the puzzle is missing.

Why Pretrained Checkpoints Matter

For those who might be newer to this, pretrained checkpoints are a big deal in deep learning. They're essentially the saved weights of a neural network that has already been trained on a large dataset. Using a pretrained checkpoint allows you to skip the time-consuming and resource-intensive process of training a model from scratch. Instead, you can start with a model that already has a good understanding of the world and fine-tune it for your specific task. This can save you a ton of time and computational power, and often leads to better results.

In this case, the FeatUp checkpoint contains the weights for the DINO16 backbone, which is a crucial component of the MEt3R metric. Without it, I can't properly evaluate the performance of the GVS model.

My Request: A Mirror, Please!

So, I'm reaching out to the community and the original authors in the hope that someone can help. It would be fantastic if a mirror link or an alternative download location for this checkpoint could be provided. This would really help me (and I'm sure others) get the codebase up and running.

Specifics I'm Looking For

To be super clear, I'm looking for the FeatUp pretrained checkpoint for the DINO16 backbone, the one that's supposed to be at feature_backbone_weights="mhamilton723/FeatUp". If anyone has a copy of this file or knows of another place where it can be downloaded, please let me know!

Possible Solutions

Here are a few ways this could be resolved:

1. Mirror Link: The ideal solution would be a direct link to a copy of the checkpoint file (e.g., on Google Drive, Dropbox, or another cloud storage service).
2. Alternative Download Location: If the authors have moved the checkpoint to a different Azure Blob location or another platform, sharing the updated link would be great.
3. Checkpoint on Hugging Face Hub: Uploading the checkpoint to the Hugging Face Hub would make it easily accessible and version-controlled.
4. Training Instructions: If the checkpoint is not readily available, instructions on how to train the FeatUp model from scratch would be a helpful alternative, although this would be more time-consuming.

Why This Matters for Reproducibility

This issue highlights the importance of reproducibility in research. When code and models are released, it's crucial to ensure that all the necessary components are available and accessible. Broken links and missing checkpoints can make it very difficult to reproduce published results, which hinders progress in the field. Providing stable and persistent access to pretrained models is a key step towards ensuring reproducibility.

Let's Collaborate!

I'm really looking forward to exploring this GVS codebase further, and I'm eager to contribute to the community. If I can help in any way, please let me know. Thanks in advance for any assistance with this checkpoint issue!

Let's delve deeper into the significance of pretrained checkpoints and how they revolutionize the field of deep learning. Pretrained checkpoints act as a cornerstone, enabling researchers and developers to sidestep the laborious process of training models from scratch. These checkpoints encapsulate the knowledge gleaned from extensive training on massive datasets, equipping models with a foundational understanding of intricate patterns and relationships within the data. This capability is particularly crucial in domains where acquiring and annotating large datasets is prohibitively expensive or time-consuming. By leveraging pretrained models, practitioners can expedite model development, reduce computational costs, and often attain superior performance compared to models trained from scratch. The availability of reliable pretrained checkpoints is paramount for fostering reproducible research and facilitating the seamless transfer of knowledge across diverse applications.

In the realm of computer vision, pretrained checkpoints have spurred groundbreaking advancements in tasks such as image classification, object detection, and image segmentation. Models like ResNet, VGGNet, and Inception, pretrained on the ImageNet dataset, have become ubiquitous building blocks for numerous computer vision systems. These models have honed the ability to extract salient features from images, enabling them to discern objects, scenes, and intricate details with remarkable accuracy. Fine-tuning these pretrained models on specific datasets allows researchers to adapt them to a wide array of applications, ranging from medical image analysis to autonomous driving. The proliferation of pretrained checkpoints has democratized access to cutting-edge computer vision technology, empowering both academic and industrial researchers to tackle complex visual challenges.

Moreover, the impact of pretrained checkpoints extends beyond computer vision, permeating the field of natural language processing (NLP). Pretrained language models such as BERT, GPT, and RoBERTa have revolutionized NLP tasks like text classification, sentiment analysis, and machine translation. These models, trained on vast corpora of text data, have acquired a profound understanding of language syntax, semantics, and contextual nuances. By leveraging pretrained language models, developers can build sophisticated NLP systems that comprehend and generate human language with unprecedented fluency. The availability of pretrained checkpoints has catalyzed the development of chatbots, virtual assistants, and language translation services, fundamentally reshaping how humans interact with technology. As pretrained models continue to evolve, they hold the promise of unlocking even greater capabilities in NLP, facilitating seamless communication and knowledge exchange across linguistic boundaries.

The unavailability of the FeatUp pretrained checkpoint underscores the critical role of robust infrastructure for disseminating research artifacts. The reliance on cloud storage services for hosting pretrained checkpoints can introduce vulnerabilities, as evidenced by the "PublicAccessNotPermitted" error. To mitigate these risks, researchers and institutions should explore alternative strategies for archiving and distributing pretrained models. Platforms like the Hugging Face Hub provide a centralized repository for pretrained models, offering version control, access management, and community engagement features. By adopting such platforms, researchers can ensure the long-term availability and discoverability of their models, fostering a culture of open and reproducible research. Furthermore, institutions can establish institutional repositories for hosting research artifacts, providing a stable and persistent platform for disseminating knowledge. Investing in robust infrastructure for model distribution is essential for sustaining progress in deep learning and ensuring the accessibility of research findings.

In addition to infrastructure considerations, the licensing of pretrained checkpoints plays a vital role in shaping their adoption and reuse. Open-source licenses like the MIT License and the Apache License promote the free use, modification, and distribution of pretrained models, fostering innovation and collaboration within the research community. By releasing pretrained checkpoints under permissive licenses, researchers empower others to build upon their work, accelerating the pace of discovery. Conversely, restrictive licenses can impede the widespread adoption of pretrained models, limiting their impact and potential societal benefits. The judicious selection of licenses is crucial for balancing the rights of model creators with the broader goal of advancing knowledge and innovation. As the landscape of deep learning evolves, open-source licensing will continue to play a pivotal role in shaping the accessibility and usability of pretrained checkpoints.

Conclusion: Let's Keep the Momentum Going

So, the hunt for the FeatUp checkpoint is on! I'm confident that with the help of the community, we can track it down and get back to exploring this exciting GVS codebase. This little hiccup also serves as a good reminder of the importance of making research resources easily accessible and ensuring reproducibility. Let's keep the momentum going and work together to advance the field! If you have any leads or suggestions, please don't hesitate to share them. Thanks a bunch, guys!