How can I create image embeddings of size 256?

[monajalal]

Thanks for your great work. I noticed that smallest image embedding pertains to ResNet18 which is 512. Do you have any suggestion for a proper model that could lead to feature size of 256?

[christiansafka]

I recently used this efficientNet autoencoder for a different project: lukemelas/EfficientNet-PyTorch#257

The vector length was around 218. Would consider adding that to this repo when I find the time.

If you need exactly 256 you could also try dimensionality reduction on the 512 vector. For example use PCA to reduce from 512 -> 256.

[frutik]

from sklearn.decomposition import PCA

pca = PCA(n_components=256)
vectors_256d = pca.fit_transform(list_of_all_your_vectors_512d)