thingsvision 1.0.2

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

Model collection

Features can be extracted for all models in torchvision, each of the CORnet versions and both CLIP variants (clip-ViT and clip-RN). For the correct abbreviations of torchvision models have a look here. For the correct abbreviations of CORnet models look here. To separate the string cornet from its variant (e.g., s, z) use a hyphen instead of an underscore (e.g., cornet-s, cornet-z).

Examples: alexnet, resnet50, resnet101, vgg13, vgg13_bn, vgg16, vgg16_bn, vgg19, vgg19_bn, cornet-s, clip-ViT

Environment Setup

Make sure you have the latest Python version (>= 3.7) and install PyTorch 1.7.1. Note that PyTorch 1.7.1 requires CUDA 10.2 or above, if you want to extract network activations on a GPU. However, the code runs already pretty fast on a strong CPU (Intel i7 or i9). Run the following pip command in your terminal.

$ pip install thingsvision

You have to download files from the parent repository (i.e., this repo), if you want to extract network activations for THINGS. Simply download the shell script get_files.sh from this repo and execute it as follows (the shell script will do file downloading and moving for you):

$ wget https://raw.githubusercontent.com/ViCCo-Group/THINGSvision/master/get_files.sh (Linux)
$ curl -O https://raw.githubusercontent.com/ViCCo-Group/THINGSvision/master/get_files.sh (macOS)
$ bash get_files.sh

Execute the following lines to have the latest PyTorch and CUDA versions available (not necessary, but perhaps desirable):

$ conda install --yes -c pytorch pytorch=1.7.1 torchvision cudatoolkit=11.0

Replace cudatoolkit=11.0 above with the appropriate CUDA version on your machine (e.g., 10.2) or cpuonly when installing on a machine without a GPU.

IMPORTANT NOTES:

1. Image data will automatically be converted into a ready-to-use dataset class, and subsequently wrapped with a PyTorch mini-batch dataloader to make neural activation extraction more efficient.

2. If you happen to use the THINGS image database, make sure to correctly unzip all zip files (sorted from A-Z), and have all object directories stored in the parent directory ./images/ (e.g., ./images/object_xy/) as well as the things_concepts.tsv file stored in the ./data/ folder. bash get_files.sh does the latter for you. Images, however, must be downloaded from the THINGS database Main subfolder. The download is around 5GB.

• Go to https://osf.io/jum2f/files/

• Select Main folder and click on "Download as zip" button (top right).

• Unzip contained object_images_*.zip file using the password (check the description.txt file for details). For example:

  for fn in object_images_*.zip; do unzip -P the_password $fn; done
  

3. Features can be extracted at every layer for all torchvision, CORnet and CLIP models.

4. If you happen to be interested in an ensemble of feature maps, as introduced in this recent COLING 2020 paper, you can simply extract an ensemble of conv or max-pool layers. The ensemble can additionally be concatenated with the activations of the penultimate layer, and subsequently transformed into a lower-dimensional space with PCA to reduce noise and only keep those dimensions that account for most of the variance.

5. The script automatically extracts features for the specified model and layer and stores them together with the targets in out_path (see above).

6. If you happen to extract hidden unit activations for many images, it is possible to run into MemoryErrors. To circumvent such problems, a helper function called split_activations will split the activation matrix into several batches, and stores them in separate files. For now, the split parameter is set to 10. Hence, the function will split the activation matrix into 10 files. This parameter can, however, easily be modified in case you need more (or fewer) splits. To merge the separate activation batches back into a single activation matrix, just call merge_activations when loading the activations (e.g., activations = merge_activations(PATH)).

Extract features at specific layer of a state-of-the-art torchvision, CORnet or CLIP model

Example call for AlexNet:

import torch
import thingsvision.vision as vision

model_name = 'alexnet'

device = 'cuda' if torch.cuda.is_available() else 'cpu'
model, transforms = vision.load_model(model_name, pretrained=True, model_path=None, device=device)
module_name = vision.show_model(model, model_name)

AlexNet(
  (features): Sequential(
    (0): Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2))
    (1): ReLU(inplace=True)
    (2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
    (3): Conv2d(64, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
    (4): ReLU(inplace=True)
    (5): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
    (6): Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (7): ReLU(inplace=True)
    (8): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (9): ReLU(inplace=True)
    (10): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (11): ReLU(inplace=True)
    (12): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (avgpool): AdaptiveAvgPool2d(output_size=(6, 6))
  (classifier): Sequential(
    (0): Dropout(p=0.5, inplace=False)
    (1): Linear(in_features=9216, out_features=4096, bias=True)
    (2): ReLU(inplace=True)
    (3): Dropout(p=0.5, inplace=False)
    (4): Linear(in_features=4096, out_features=4096, bias=True)
    (5): ReLU(inplace=True)
    (6): Linear(in_features=4096, out_features=1000, bias=True)
  )
)

#Enter part of the model for which you would like to extract features:

(e.g., "features.10")

dl = vision.load_dl(root='./images/', out_path=f'./{model_name}/{module_name}/features', batch_size=64, transforms=transforms)
features, targets, probas = vision.extract_features(model, dl, module_name, batch_size=64, flatten_acts=True, device=device, return_probabilities=True)

vision.save_features(features, f'./{model_name}/{module_name}/features', 'npy')

Example call for CLIP:

import torch
import thingsvision.vision as vision

model_name = 'clip-ViT'
module_name = 'visual'

device = 'cuda' if torch.cuda.is_available() else 'cpu'

model, transforms = vision.load_model(model_name, pretrained=True, model_path=None, device=device)
dl = vision.load_dl(root='./images/', out_path=f'./{model_name}/{module_name}/features', batch_size=64, transforms=transforms)
features, targets = vision.extract_features(model, dl, module_name, batch_size=64, flatten_acts=False, device=device, clip=True, return_probabilities=False)

features = vision.center_features(features)

vision.save_features(features, f'./{model_name}/{module_name}/features', 'npy')
vision.save_targets(targets, f'./{model_name}/{module_name}/targets', 'npy')

Example call for CORnet

import torch
import thingsvision.vision as vision

model_name = 'cornet-s'
device = 'cuda' if torch.cuda.is_available() else 'cpu'

model, transforms = vision.load_model(model_name, pretrained=True, model_path=None, device=device)
module_name = vision.show_model(model, model_name)

Sequential(
  (V1): Sequential(
    (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
    (norm1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (nonlin1): ReLU(inplace=True)
    (pool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
    (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (norm2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (nonlin2): ReLU(inplace=True)
    (output): Identity()
  )
  (V2): CORblock_S(
    (conv_input): Conv2d(64, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (skip): Conv2d(128, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
    (norm_skip): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (conv1): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (nonlin1): ReLU(inplace=True)
    (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
    (nonlin2): ReLU(inplace=True)
    (conv3): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (nonlin3): ReLU(inplace=True)
    (output): Identity()
    (norm1_0): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm2_0): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm3_0): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm1_1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm2_1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm3_1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  )
  (V4): CORblock_S(
    (conv_input): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (skip): Conv2d(256, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
    (norm_skip): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (conv1): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (nonlin1): ReLU(inplace=True)
    (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
    (nonlin2): ReLU(inplace=True)
    (conv3): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (nonlin3): ReLU(inplace=True)
    (output): Identity()
    (norm1_0): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm2_0): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm3_0): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm1_1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm2_1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm3_1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm1_2): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm2_2): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm3_2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm1_3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm2_3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm3_3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  )
  (IT): CORblock_S(
    (conv_input): Conv2d(256, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (skip): Conv2d(512, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
    (norm_skip): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (conv1): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (nonlin1): ReLU(inplace=True)
    (conv2): Conv2d(2048, 2048, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
    (nonlin2): ReLU(inplace=True)
    (conv3): Conv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (nonlin3): ReLU(inplace=True)
    (output): Identity()
    (norm1_0): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm2_0): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm3_0): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm1_1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm2_1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (norm3_1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  )
  (decoder): Sequential(
    (avgpool): AdaptiveAvgPool2d(output_size=1)
    (flatten): Flatten()
    (linear): Linear(in_features=512, out_features=1000, bias=True)
    (output): Identity()
  )
)

#Enter part of the model for which you would like to extract features:

(e.g., "decoder.flatten")

dl = vision.load_dl(root='./images/', out_path=f'./{model_name}/{module_name}/features', batch_size=64, transforms=transforms)
features, targets = vision.extract_features(model, dl, module_name, batch_size=64, flatten_acts=False, device=device, return_probabilities=False)

features = vision.center_features(features)
features = vision.normalize_features(features)

vision.save_features(features, f'./{model_name}/{module_name}/features', 'npy')

ImageNet class predictions

Would you like to know the probabilities corresponding to the top k predicted ImageNet classes for each of your images? Simply set the return_probabilities argument to True and use the get_class_probabilities helper (the function works for both synset and class files). Note that this is, unfortunately, not (yet) possible for CLIP models due to their multi-modality and different training objectives. You are required to use the same out_path throughout which is why out_path must correspond to the path that was used in vision.load_dl.

features, targets, probas = vision.extract_features(model, dl, module_name, batch_size, flatten_acts=False, device=device, return_probabilities=True)
class_probas = vision.get_class_probabilities(probas=probas, out_path=out_path, cls_file='./data/imagenet1000_classes.txt', top_k=5, save_as_json=True)

Model comparison

To compare object representations extracted from specifid models and layers against each other, for a List[str] of models and layers a user can perform the following operation,

clip_list = [n.startswith('clip') for n in model_names]

correlations = vision.compare_models(
                                     root=root,
                                     out_path=out_path,
                                     model_names=model_names,
                                     module_names=module_names,
                                     pretrained=True,
                                     batch_size=batch_size,
                                     flatten_acts=True,
                                     clip=clip_list,
                                     save_features=True,
                                     dissimilarity='correlation',
                                     correlation='pearson',
                                    )

The function returns a correlation matrix in the form of a Pandas dataframe whose rows and columns correspond to the names of the models in model_names. The cell elements are the correlation coefficients for each model combination. The dataframe can subsequently be converted into a heatmap with matplotlib or seaborn. We will release a clear and concise documentary as soon as possible. Until then, we recommend to look at Section 3.2.3 in the bioRxiv preprint.

OpenAI's CLIP models

CLIP

[Blog] [Paper] [Model Card] [Colab]

CLIP (Contrastive Language-Image Pre-Training) is a neural network trained on a variety of (image, text) pairs. It can be instructed in natural language to predict the most relevant text snippet, given an image, without directly optimizing for the task, similarly to the zero-shot capabilities of GPT-2 and 3. We found CLIP matches the performance of the original ResNet50 on ImageNet “zero-shot” without using any of the original 1.28M labeled examples, overcoming several major challenges in computer vision.

Citation

If you use this GitHub repository (or any modules associated with it), we would grately appreciate to cite our preprint 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},
	elocation-id = {2021.03.11.434979},
	year = {2021},
	doi = {10.1101/2021.03.11.434979},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2021/03/12/2021.03.11.434979},
	eprint = {https://www.biorxiv.org/content/early/2021/03/12/2021.03.11.434979.full.pdf},
	journal = {bioRxiv}
}