Catosine / CytoBrain

This is catosine's repo for Algonauts2023 competition

CytoBrain

This is a repository for Algonauts2023 competition.
Developed by @Cytosine

Dataset

You may access the dataset from here

Quick Start

Feature Extraction

# Extract train set features of subj01 via pretrained Resnet50
python feature_extract.py --data ~/algonauts2023/data --subject subj01 --train \ 
        --save_path ~/algonauts2023/data/subj01/training_split/test_features \
        --pretrained_weights ~/backbone/resnet50-imagenet1k-v2.pth \ 
        --layers layer3 avgpool

# Also extract test set features of subj01 via pretrained Resnet50
python feature_extract.py --data ~/algonauts2023/data --subject subj01 \ 
        --save_path ~/algonauts2023/data/subj01/test_split/test_features \
        --pretrained_weights ~/backbone/resnet50-imagenet1k-v2.pth \ 
        --layers layer3 avgpool

# Or, extract train set features of last 4 layer of image captioning model
python hf_feature_extract.py --data ~/algonauts2023/data --subject subj01 --train \
        --save_path ~/algonauts2023/data/subj01/training_split/training_features \
        --pretrained_weights nlpconnect/vit-gpt2-image-captioning \
        --feature_type decoder

# And same to test set
python hf_feature_extract.py --data ~/algonauts2023/data --subject subj01 \
        --save_path ~/algonauts2023/data/subj01/training_split/training_features \
        --pretrained_weights nlpconnect/vit-gpt2-image-captioning \
        --feature_type decoder

Feature Decomposition Via PCA

The PCA is nothing special. Here is an example in case of need:

from sklearn.decomposition import PCA

# suppose the extracted feature are prepared as a numpy.ndarray in form of [#Samples, #Features]
features = np.load("extracted_features.npy")

pca = PCA(512)
reduced_features = pca.fit_transform(features)

Modelling

Please see example notebook

Submission

Please see submission notebook

Results

Submission

Method	Test Median Pearson's R	Note
RidgeR	38.406	res50(layer3+avgpool)
RidgeR	49.793	ViT-GPT2 Image Captioning(Last4Layer)

Best Submission

Experiments: Preatrained Feature Modelling

Resnet50

Subject	Feature	Model	Dev Median Pearson's R (Left)	Dev Median Pearson's R (Right)	Note
Subj01	avgpool	LinearR	0.243	0.245	baseline/random crop 256
Subj01	avgpool	RidgeR	0.376	0.376	alpha=2e4
Subj01	layer3	RidgeR	0.391	0.392	alpha=1e3/avgpool
Subj01	layer2	RidgeR	0.328	0.316	alpha=1e2/avgpool
Subj01	layer1	RidgeR	0.288	0.282	alpha=1e1/avgpool
Subj01	layer3+avgpool	RidgeR	0.398	0.394	alpha=1e4
Subj01	layer123+avgpool	RidgeR	0.391	0.390	alpha=1e3

ViT-GPT2 Image Captioning

Subject	Feature	Model	Dev Median Pearson's R (Left)	Dev Median Pearson's R (Right)	Note
Subj01	encoder-pca-512	RidgeR	0.394	0.393	alpha=1
Subj01	encoder-pca-2048	RidgeR	0.350	0.345	alpha=1e4
Subj01	encoder-avg-768	RidgeR	0.378	0.379	alpha=1e4
Subj01	encoder-cls	RidgeR	0.370	0.370	alpha=1e4
Subj01	encoder-last4-pca-512	RidgeR	0.445	0.440	alpha=2e3
Subj01	decoder-pca-512	RidgeR	0.394	0.386	alpha=5e3
Subj01	decoder-pca-2048	RidgeR	0.351	0.345	alpha=2e4
Subj01	decoder-avg-768	RidgeR	0.378	0.379	alpha=1e4
Subj01	decoder-cls	RidgeR	0.371	0.370	alpha=1e4
Subj01	decoder-last4-pca-512	RidgeR	0.446	0.442	alpha=2e3
Subj01	decoder-last1234-pca-512	RidgeR	0.454	0.449	alpha=5e5
Subj01	encoder+decoder_pca_512	RidgeR	0.405	0.399	alpha=1e6
Subj01	encoder+decoder_last4_pca_512	RidgeR	0.447	0.443	alpha=1e4