Keras

About

Key Features

• Write one Keras 3 model → run natively on TensorFlow, JAX, or PyTorch with the same codebase
• Call keras.applications.ResNet50(weights='imagenet') → get a 98 MB pre-trained model ready for inference
• Switch to Xception from Keras Applications → reach up to 79.0% top-1 and 94.5% top-5 ImageNet accuracy
• Benchmark on CPU using published examples → see ~58.2 ms per inference step for ResNet50 (others listed)
• Follow Keras 3 API docs → implement custom layers, metrics, and models for your workflow

Use Cases

• A computer vision engineer fine-tuning ResNet50 on a retail image dataset in TensorFlow, then benchmarking the same code on JAX
• A research scientist implementing a custom loss and metric in Keras 3 to compare architectures across PyTorch and TensorFlow
• A machine learning instructor demonstrating accuracy/speed trade-offs using Keras Applications models on a CPU lab machine

Try It Like This

1. 1
   Fine-tune ResNet50 in TensorFlow

   Install Keras via pip and import the TF backend → from the applications API call keras.applications.ResNet50(weights='imagenet') and replace the top layer for your classes → compile with an optimizer and train; the first epoch shows a working training loop with pre-trained weights loading quickly.

2. 2
   Benchmark same model on JAX

   pip install keras[jax] (or follow the Keras JAX install notes) → run the identical Keras 3 model code but select the JAX backend and run the published CPU benchmark example → compare per-step latency (ms) and confirm behaviour matches the TensorFlow run, observing similar model outputs and timing differences.

3. 3
   Implement custom loss + metric across backends

   Create a Python file that subclasses keras.metrics.Metric and implements update_state/result → plug the custom metric and a custom loss into model.compile() and train with the same code on TensorFlow and PyTorch backends → verify the metric values and training gradients behave consistently across both backends.

4. 4
   Prototype then run on PyTorch without changing code

   Write a Keras 3 model using the Functional or Subclassing API and test it locally with a small dataset → switch the runtime to the PyTorch backend (per Keras instructions) and run inference/training with identical code → confirm the model runs natively on PyTorch and outputs match within numerical tolerance.

5. 5
   Teach CPU lab using Keras Applications

   Use keras.applications to load Xception and ResNet variants for students (weights='imagenet') → run inference and measure top-1/top-5 accuracy and per-step CPU timings using the published examples → demonstrate accuracy vs. latency trade-offs and how to pick a backbone for resource-constrained deployment.

Pros & Cons

Getting Started

1. 1 Install with pip install keras and open the Keras 3 docs on keras.io
2. 2 Import keras, load a pre-trained model like keras.applications.ResNet50, and prepare an input tensor
3. 3 Run a forward pass and view top-5 class predictions in under five minutes