Fifteen Classic CNNs, One TensorFlow 2 Training Loop

What it does This repo bundles implementations of roughly fifteen classic CNN families—MobileNet, EfficientNet, ResNet, DenseNet, Inception, and others—into a single TensorFlow 2 image-classification pipeline. You point it at a folder of labeled images, and it handles dataset splitting, TFRecord generation, and training via a single model index. It is essentially a readable reference zoo with a harness attached.

The interesting bit The value is in the assembly, not the novelty. Each architecture is a landmark paper in its own right, but the author has normalized them into one interchangeable training loop, complete with the correct input sizes per model. That makes it a convenient sanity-check baseline when you need to know if your problem wants a SqueezeNet or an SE-ResNeXt without rewriting data loaders.

Key highlights

• Covers MobileNet V1/V2/V3, EfficientNet B0–B7, ResNet, ResNeXt, SE-ResNet, DenseNet, InceptionV4, Inception-ResNet, SqueezeNet, ShuffleNetV2, and RegNet.
• Includes dataset-splitting and TFRecord utilities, so the pipeline is self-contained beyond raw images.
• Models are indexed; swapping from, say, ResNet50 to EfficientNet-B3 is a config change.
• README documents the exact input resolution for each network (e.g., 299×299 for Inception, 224×224 for most others).

Verdict Worth bookmarking if you need clean, standalone TF2 implementations of standard backbones for baselining or teaching. Skip it if you are looking for pre-trained weights, production-grade training loops, or novel architectures.