TensorFlow
Définition
TensorFlow is a deep learning framework with a focus on production deployment. Keras est l'API de haut niveau. Il supporte CPU, GPU, TPU, and mobile/edge (TFLite). C'est an alternative to PyTorch; TensorFlow and Keras are strong for production pipelines, infrastructure (TF Serving, TPU), and mobile (TFLite). Used for vision, NLP, and recommendation systems; the ecosystem includes TensorFlow Hub for pretrained models.
Comment ça fonctionne
Construisez des modèles avec Keras (API séquentielle ou fonctionnelle) ou l'API TensorFlow de bas niveau. Entraînez avec l'exécution eagertion (default, PyTorch-like) or graph mode (faster, exportable). Export to SavedModel (standard serving format) or TFLite (mobile/edge, with quantization support). TensorFlow Hub and Keras Applications provide pretrained models for transfer learning. Distributed training uses tf.distribute; deployment uses TF Serving, Vertex AI, or on-device TFLite. The pipeline from data (tf.data) to training to serving is well integrated.
Cas d'utilisation
TensorFlow suits production pipelines, mobile/edge (TFLite), and quick prototyping with Keras and Hub models.
- Production ML pipelines and serving (par ex. SavedModel, TF Serving)
- Mobile and edge deployment (TFLite)
- Quick prototyping with Keras and pretrained models from Hub