Image Classification with scikit-learn

Рет қаралды 1,181

Күн бұрын

Scikit-Learn is most known for building machine learning models for tabular data, but that doesn't mean that it cannot do image classification. In this video we'll show you how by using a plugin from the ecosystem called "embetter".
Video Chapters:
00:00 Introduction
01:14 General Pipelines
03:46 Image Pipelines
05:33 Embetter
07:41 Benchmark
To learn more about embetter, you can check out the project here:
github.com/koaning/embetter
The code for all of our videos can be found on this Github repository:
github.com/probabl-ai/youtube...
The code for this specific episode can be found here:
github.com/probabl-ai/youtube...
If you're keen to see more videos like this, you can follow us over at @probabl_ai.

Пікірлер: 7

@MrChristonik 2 ай бұрын

Hi, great work. Would it make sense to continue with feature extraction (as per the tabular logic) after the Clip Encoder is applied and the numeric output is provided?

@probabl_ai 2 ай бұрын

(vincent here). Nothing is stopping you from doing that. You could have a pipeline with two image feature extractors if you really wanted to. In general though, for rapid prototyping, I've found that a single CLIP encoder already does plenty of heavy lifting. For text use-cases I've sometimes done a bag-of-words featurizer next to a BERT featurizer, which had some merits to it.

@muthukamalan.m6316 3 ай бұрын

why not cv2😢

@serafeiml1041 3 ай бұрын

Why not just flatten the image into a 1D vector and use this as input to the model?

@vincentd.warmerdam1422 3 ай бұрын

You might end up with 1D vectors of different size. Different images will have different sizes, and the estimator would expect a fixed size. The embedding trick will always output the same dimensions going out.

@vincentmaladiere1285 3 ай бұрын

and also the embeddings provided by the CLIP encoder are likely to bring much more predictive power to the table than the raw image itself, e.g. the scores at the end of the CV would probably be lower

@Mayur7Garg 3 ай бұрын

You might lose spatial information that way.