CNN feature extractor

The CNN feature extractor ingests images and a selected CNN model to produce numeric feature vectors from a defined layer. It targets ML pipelines, including transfer learning, image similarity search, clustering, and exploratory data analysis. It is used by developers building image search systems, researchers evaluating layer representations, and data scientists preprocessing datasets.
Conceptually, it preprocesses inputs, loads model weights, passes the image through the network, and captures the activation from the requested layer. Optional global pooling or flattening converts spatial maps into a fixed-size vector; normalization can be applied for consistency across datasets. For ResNet, VGG, and EfficientNet variants, per-layer embeddings enable direct comparison of representations across architectures.
Who benefits includes ML engineers, data scientists, and researchers who require reproducible, per-layer embeddings for downstream tasks such as classification, retrieval, or clustering. Unique value lies in per-layer extraction with deterministic preprocessing, support for multiple output formats, and optional dimensionality reduction. Core features include single-image or batch processing, GPU acceleration, and the ability to cache results for repeated experiments. Use cases range from quick feature extraction for a new classifier to large-scale embedding generation for similarity search across datasets.