Skip to content

Loading models for feature extraction (two lines per ecosystem)

TorchLens has no model zoo, and that is a feature: every loader below already ships the model AND its matched preprocessing, maintained by the people who trained it. Load with the ecosystem's own two lines, then hand the SAME authority object to torchlens.preprocessing.resolve -- the resolution's record.description prints the declared recipe, and the audit can verify what you actually ran against it.

Model names are case-sensitive in every ecosystem below; copy them exactly.

torchvision (the weights enum IS the authority):

import torchlens.preprocessing
from torchvision.models import ResNet50_Weights, resnet50
model = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2).eval()

pre = torchlens.preprocessing.resolve(ResNet50_Weights.IMAGENET1K_V2)
print(pre.preprocessing.description)   # torchvision preset: resize=232 crop=224

Note the V2 recipe resizes to 232, not the folk-memory 256 -- exactly the kind of drift a resolved authority catches and a hand-typed constant table misses.

Hugging Face (the image processor is the authority; fetching it can touch the network -- the resolution discloses whether a fetch was attempted):

import torchlens.preprocessing
from transformers import AutoImageProcessor, CLIPModel
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").eval()

processor = AutoImageProcessor.from_pretrained("openai/clip-vit-base-patch32")
pre = torchlens.preprocessing.resolve(processor)
print(pre.preprocessing.description)

timm (the model carries its own data config; no network needed):

import torchlens.preprocessing
import timm
model = timm.create_model("resnet18.a1_in1k", pretrained=True).eval()

pre = torchlens.preprocessing.resolve(model=model)
print(pre.preprocessing.description)

open_clip (the returned preprocess pipeline is the authority):

import torchlens.preprocessing
import open_clip
model, _, preprocess = open_clip.create_model_and_transforms("ViT-B-32", pretrained="openai")

pre = torchlens.preprocessing.resolve(preprocess)
print(pre.preprocessing.description)

torch.hub (models declare no standard preprocessing metadata; declare what you use explicitly and the audit compares against it):

import torchlens.preprocessing
model = torch.hub.load("pytorch/vision", "resnet18", weights=None).eval()

pre = torchlens.preprocessing.resolve({"mean": [0.485, 0.456, 0.406], "std": [0.229, 0.224, 0.225]})
print(pre.preprocessing.description)

When the authority declares nothing

Some real, shipped model families declare NO preprocessing constants because none apply. torchvision's detection presets are the canonical case: a FasterRCNN_ResNet50_FPN normalizes internally, a plain [0, 1] float batch is the CORRECT input, and the preset exposes no mean/std/size:

import torchlens.preprocessing
from torchvision.models.detection import FasterRCNN_ResNet50_FPN_Weights
pre = torchlens.preprocessing.resolve(FasterRCNN_ResNet50_FPN_Weights.COCO_V1)
print(pre.declared)      # None -- the authority declares nothing

The honest audit verdict for such a family is unknown, never a guess and never a false alarm on the correct input. A verifier can only be as declarative as the authority; unknown is the one answer a bundled recipe table could never give you.

Version notes

  • On transformers 5.x, CLIPModel.get_image_features returns the full vision output object rather than the embedding tensor; take the projected embedding explicitly (model.visual_projection(...) over the vision_model.post_layernorm site) -- see the migration page.
  • Packaging: TorchLens installs into YOUR environment and gates optional integrations behind extras, so the loaders above stay at whatever versions you already run. (thingsvision, by comparison, currently pins tensorflow<2.16, torchvision==0.15.2, transformers==4.40.1, and numpy<2 as mandatory dependencies -- a factual packaging difference, checkable in both projects' metadata.)