How to convert images to 512 vectors instead of 2048 vectors

[zhenzi0322]

How to convert images to 512 vectors instead of 2048 vectors.

feat = towhee.glob(img_path) \
            .image_decode() \
            .image_embedding.timm(model_name='resnet50') \
            .tensor_normalize() \
            .to_list()

At present, the converted DIM is 2048. How to convert it to 512DIM?

[shiyu22]

1. Dimensionality Reduction
   You can refer to this notebook to reduce the dimension, and I think you need to save the projection_matrix to ensure that you are always trans with one matrix.

2. Try more models
   You can try some other models with timm operator, such as vgg16:

feat = towhee.glob(img_path) \
            .image_decode() \
            .image_embedding.timm(model_name='vgg16') \
            .tensor_normalize() \
            .to_list()

And you can also list all timm model and try with them:

import timm
timm.list_models()

[zhenzi0322]

ok