thingsvision 2.5.1

Extracting image features from state-of-the-art neural networks for Computer Vision made easy

:notebook_with_decorative_cover: Table of Contents

• About the Project
  • Functionality
  • Model collection
• Getting Started
  • Setting up your environment
  • Basic usage
• Contributing
• License
• Citation
• Contributions

:star2: About the Project

thingsvision is a Python package that let's you easily extract image representations from many state-of-the-art computer vision models. In a nutshell, you feed thingsvision with a directory of images and tell it which neural network you are interested in. thingsvision will then give you the representation of the indicated neural network for each image so that you will end up with one feature map (vector or matrix, depending on the layer) per image. You can use these features for further analyses. We use the word features for short when we mean "image representation".

:rotating_light: NOTE: some function calls mentioned in the original paper have been deprecated. To use this package successfully, exclusively follow this README and the documentation! :rotating_light:

(back to top)

:mechanical_arm: Functionality

With thingsvision, you can:

• extract features for any imageset from many popular networks.
• extract features for any imageset from your custom networks.
• extract features for >26,000 images from the THINGS image database.
• optionally turn off the standard center cropping performed by many networks before extracting features.
• extract features from HDF5 datasets directly (e.g. NSD stimuli)
• conduct basic Representational Similarity Analysis (RSA) after feature extraction.
• perform Centered Kernel Alignment (CKA) to compare image features across model-module combinations.

(back to top)

:file_cabinet: Model collection

Neural networks come from different sources. With thingsvision, you can extract image representations of all models from:

• torchvision
• Keras
• timm
• ssl (self-supervised learning models)
  • simclr-rn50, mocov2-rn50, barlowtwins-rn50, pirl-rn50
  • jigsaw-rn50, rotnet-rn50, swav-rn50, vicreg-rn50
  • dino-rn50, dino-xcit-{small/medium}-{12/24}-p{8/16}, dino-vit-{tiny/small/base}-p{8/16}, dinov2-vit-{small/base/large/giant}-p14
    
• OpenCLIP models (CLIP trained on LAION-{400M/2B/5B})
• CLIP models (CLIP trained on WiT)
• a few custom models (Alexnet, VGG-16, Resnet50, and Inception_v3) trained on Ecoset rather than ImageNet and one Alexnet model pretrained on ImageNet and fine-tuned on SalObjSub
  
• each of the many CORnet versions
• Harmonization models (see Harmonization repo). The default variant is ViT_B16. Other available models are ResNet50, VGG16, EfficientNetB0, tiny_ConvNeXT, tiny_MaxViT, and LeViT_small
  
• DreamSim models (see DreamSim repo). The default variant is open_clip_vitb32. Other available models are clip_vitb32, dino_vitb16, and an ensemble. See the docs for more information

(back to top)

:running: Getting Started

:computer: Setting up your environment

Working locally.

First, create a new conda environment with Python version 3.8, 3.9, or 3.10 e.g. by using conda:

$ conda create -n thingsvision python=3.9
$ conda activate thingsvision

Then, activate the environment and simply install thingsvision via running the following pip command in your terminal.

$ pip install --upgrade thingsvision
$ pip install git+https://github.com/openai/CLIP.git

If you want to extract features for harmonized models from the Harmonization repo, you have to additionally run the following pip command in your thingsvision environment (FYI: as of now, this seems to be working smoothly on Ubuntu only but not on macOS),

$ pip install git+https://github.com/serre-lab/Harmonization.git
$ pip install keras-cv-attention-models>=1.3.5

If you want to extract features for DreamSim from the DreamSim repo, you have to additionally run the following pip command in your thingsvision environment,

$ pip install dreamsim==0.1.2

See the docs for which DreamSim models are available in thingsvision.

Google Colab.

Alternatively, you can use Google Colab to play around with thingsvision by uploading your image data to Google Drive (via directory mounting). You can find the jupyter notebook using PyTorch here and the TensorFlow example here.

(back to top)

:mag: Basic usage

Command Line Interface (CLI)

thingsvision was designed to simplify feature extraction. If you have some folder of images (e.g., ./images) and want to extract features for each of these images without opening a Jupyter Notebook instance or writing a Python script, it's probably easiest to use our CLI. The interface includes two options,

• thingsvision show-model
• thingsvision extract-features

Example calls might look as follows:

thingsvision show-model --model-name "alexnet" --source "torchvision"
thingsvision extract_features --image-root "./data" --model-name "alexnet" --module-name "features.10" --batch-size 32 --device "cuda" --source "torchvision" --file-format "npy" --out-path "./features"

See thingsvision show-model -h and thingsvision extract-features -h for a list of all possible arguments. Note that the CLI provides just the basic extraction functionalities but is probably enough for most users that don't want to dive too deep into various models and modules. If you need more fine-grained control over the extraction itself, we recommend to use the python package directly and write your own Python script.

Python commands

To do this start by importing all the necessary components and instantiating a thingsvision extractor. Here we're using CLIP from the official clip repo as the model to extract features from and also load the model to GPU for faster inference,

import torch
from thingsvision import get_extractor
from thingsvision.utils.storing import save_features
from thingsvision.utils.data import ImageDataset, DataLoader

model_name = 'clip'
source = 'custom'
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model_parameters = {
    'variant': 'ViT-L/14'
}

extractor = get_extractor(
  model_name=model_name,
  source=source,
  device=device,
  pretrained=True,
  model_parameters=model_parameters,
)

As a next step, create both dataset and dataloader for your images. We assume that all of your images are in a single root directory which can contain subfolders (e.g., for individual classes). Therefore, we leverage the ImageDataset class.

root='path/to/your/image/directory' # (e.g., './images/)
batch_size = 32

dataset = ImageDataset(
    root=root,
    out_path='path/to/features',
    backend=extractor.get_backend(), # backend framework of model
    transforms=extractor.get_transformations(resize_dim=256, crop_dim=224) # set the input dimensionality to whichever values are required for your pretrained model
)

batches = DataLoader(
    dataset=dataset,
    batch_size=batch_size,
    backend=extractor.get_backend() # backend framework of model
)

Now all that is left is to extract the image features and store them on disk! Here we're extracting features from the image encoder module of CLIP (visual), but if you don't know which modules are available for a given model, just call extractor.show_model() to print all the modules.

module_name = 'visual'

features = extractor.extract_features(
    batches=batches,
    module_name=module_name,
    flatten_acts=True,
    output_type="ndarray", # or "tensor" (only applicable to PyTorch models of which CLIP and DINO are ones!)
)

save_features(features, out_path='path/to/features', file_format='npy') # file_format can be set to "npy", "txt", "mat", "pt", or "hdf5"

Human alignment: If you want to align the extracted features with human object similarity according to the approach introduced in Improving neural network representations using human similiarty judgments you can optionally align the extracted features using the following method:

aligned_features = extractor.align(
    features=features,
    module_name=module_name,
    alignment_type="gLocal",
)

For more information about the available alignment types and aligned models see the docs.

For more examples on the many models available in thingsvision and explanations of additional functionality like how to optionally turn off center cropping, how to use HDF5 datasets (e.g. NSD stimuli), how to perform RSA or CKA, or how to easily extract features for the THINGS image database, please refer to the Documentation.

(back to top)

:wave: How to contribute

If you come across problems or have suggestions please submit an issue!

(back to top)

:warning: License

This GitHub repository is licensed under the MIT License - see the LICENSE.md file for details.

(back to top)

:page_with_curl: Citation

If you use this GitHub repository (or any modules associated with it), please cite our paper for the initial version of thingsvision as follows:

@article{Muttenthaler_2021,
	author = {Muttenthaler, Lukas and Hebart, Martin N.},
	title = {THINGSvision: A Python Toolbox for Streamlining the Extraction of Activations From Deep Neural Networks},
	journal ={Frontiers in Neuroinformatics},
	volume = {15},
	pages = {45},
	year = {2021},
	url = {https://www.frontiersin.org/article/10.3389/fninf.2021.679838},
	doi = {10.3389/fninf.2021.679838},
	issn = {1662-5196},
}

(back to top)

:gem: Contributions

This is a joint open-source project between the Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, and the Machine Learning Group at Technische Universtität Berlin. Correspondence and requests for contributing should be adressed to Lukas Muttenthaler. Feel free to contact us if you want to become a contributor or have any suggestions/feedback. For the latter, you could also just post an issue or engange in discussions. We'll try to respond as fast as we can.

(back to top)