Which File to Use for QuiltNet-B-32 Weights
If you have opened the QuiltNet-B-32 model page on Hugging Face and stared at the Files tab, you are not alone. The card lists a lot of assets, and the natural question is which one you actually need to import. The short answer from the model owner is simpler than hunting for a filename: you do not pick a raw .bin or .safetensors file by hand. You point your loader at the repo identifier and let Hugging Face resolve the right weights for you.
That is exactly what the discussion that prompted this guide is about.
The Question That Started It
On Mar 12, 2024, rbareja opened a discussion on the wisdomik/QuiltNet-B-32 card and asked plainly:
Hello, Which file to use for using the QuiltNet-B-32 model weights?
It is the kind of question that comes up the first time you work with a vision-language model that is distributed through Hugging Face Hub. The model card shows tags for Zero-Shot Image Classification, OpenCLIP, PyTorch, clip, vision, language, histopathology, histology, medical, with arXiv 2306.11207 and License MIT. There are files and versions listed, and no single highlighted weight file to download manually.
The thread was short. The owner replied the same day, closed it the same day, and gave a clear direction.
The Owner’s Direct Answer
wisdomik, the owner, wrote:
you can use huggingface or open_clip to load the models. E.g for openclip you could use import open_clip
model, preprocess_train, preprocess_val = open_clip.create_model_and_transforms('hf-hub:wisdomik/QuiltNet-B-32') tokenizer = open_clip.get_tokenizer('hf-hub:wisdomik/QuiltNet-B-32')
That is the recommendation. No manual file selection, no guessing which checkpoint is the canonical one. The identifier hf-hub:wisdomik/QuiltNet-B-32 is enough for OpenCLIP to pull the correct weights, the preprocessing transforms, and the tokenizer.
The discussion was changed to closed on Mar 12, 2024, which tells you the owner considered the question settled.
Loading With OpenCLIP the Way the Owner Showed
If your project already uses OpenCLIP, this is the pattern to copy. It is three lines and it works without modification.
import open_clip
model, preprocess_train, preprocess_val = open_clip.create_model_and_transforms('hf-hub:wisdomik/QuiltNet-B-32')
tokenizer = open_clip.get_tokenizer('hf-hub:wisdomik/QuiltNet-B-32')
create_model_and_transforms returns the model object plus two preprocessing functions. preprocess_train is for training or data augmentation style preprocessing, preprocess_val is for evaluation. The tokenizer is the text encoder side that OpenCLIP expects for zero-shot prompts.
You can then move the model to a device, set it to eval mode, and run inference. Because the model lives in the PyTorch ecosystem, you can call .to("cuda") if a GPU is available, apply the preprocess function to a PIL image, and pass the tensor forward alongside tokenized text.
The key point is the string hf-hub:wisdomik/QuiltNet-B-32. OpenCLIP resolves that to the correct weight file behind the scenes. You do not need to download a file locally and point to a path.
Using Hugging Face Directly Without OpenCLIP
The owner also said you can use huggingface to load the models. In practice that means letting the Hugging Face Hub resolver pick the checkpoint for you rather than specifying a filename.
A common pattern with the transformers library is:
from transformers import AutoModel, AutoImageProcessor
model = AutoModel.from_pretrained("wisdomik/QuiltNet-B-32")
processor = AutoImageProcessor.from_pretrained("wisdomik/QuiltNet-B-32")
That two-liner works because the model card metadata tells Hugging Face which file contains the model weights, which file holds the processor config, and which file holds the tokenizer vocab. The platform figures it out. You stay out of the Files tab.
If you prefer a pure PyTorch load, you can still use the hub identifier and let the library resolve the weights. The important idea is the same as with OpenCLIP: use the repo name, not a manual file path.
What QuiltNet-B-32 Is
From the model card tags you can see the intent. QuiltNet-B-32 is a zero-shot image classification model built on the OpenCLIP architecture. It has been trained on diverse image-text pairs and can classify images it has never seen before without fine-tuning.
The B-32 in the name refers to the backbone and patch size convention used in CLIP-style models. The model is released under the MIT license, which permits commercial and non-commercial use with minimal restrictions.
The associated arXiv paper is 2306.11207. If you need to cite the work formally, that reference gives you the architecture description and benchmark results.
The tags also highlight histopathology, histology, medical. That is not a coincidence. The original discussion context and the community around the model point to use in medical imaging and tissue classification. Researchers have experimented with it for slide images because zero-shot classification lets you point the model at a new image and get class probabilities without collecting a labeled training set first.
Why File Hunting Is a Trap Here
With older releases you sometimes had to pick between pytorch_model.bin, model.safetensors, or a specific revision. QuiltNet-B-32 is distributed through Hub with a clear identifier. When you use hf-hub:wisdomik/QuiltNet-B-32 or wisdomik/QuiltNet-B-32 with a library, the resolver checks the model index, picks the correct weights file for the framework you are using, and downloads it on demand.
That is why the owner’s answer is so short. There is no special file name to remember. The file to use is the one the loader resolves from the repo identifier.
If you really need to inspect what is on the Files tab, you will see multiple artifacts for the different libraries and formats. They all belong to the same model version. Picking one manually breaks the expectation that the processor config and tokenizer match the weights.
Practical Tips Before You Run
Put the model in eval mode before inference. Apply the preprocessing function that matches your use case, preprocess_val for evaluation. Tokenize your text prompts with the tokenizer returned by OpenCLIP, or with the processor if you are using transformers.
If you are working with histology images, remember that CLIP-style models were trained on natural images. The model can still generalize zero-shot, but you may want to check the preprocessing expectations for image size and normalization.
Keep the citation handy: arXiv 2306.11207, MIT license, model page wisdomik/QuiltNet-B-32.
Summary Checklist
- Do not manually choose a weight file. Use the repo identifier.
- For OpenCLIP: import open_clip and call create_model_and_transforms('hf-hub:wisdomik/QuiltNet-B-32') and get_tokenizer('hf-hub:wisdomik/QuiltNet-B-32').
- For Hugging Face libraries: AutoModel.from_pretrained("wisdomik/QuiltNet-B-32") and the matching processor resolve the correct files for you.
- The model is MIT licensed, zero-shot image classification, OpenCLIP based, with medical/histopathology use cases noted on the card.
- The discussion that settled this was opened and closed on Mar 12, 2024 by rbareja with owner wisdomik.
That is the file question answered. Use the identifier, let the loader resolve the weights, and you are running.
Source: https://huggingface.co/wisdomik/QuiltNet-B-32/discussions/1
