In this project, we use Tensoflow (Keras) to implement and train machine learning algorithms to automatically detect dog breeds in pictures.
We implement a CNN (Convulotionnal Neural Network) from scratch, and fine-tune pre-trained models thanks to transfer learning. We use KerasTuner for the hyperparameters search and Tensorboard to track the results.
We develop a wrapper for Keras, KerasTuner & TensorHub in order to :
- Automatically generate neural networks and their structure
- Train neural networks by applying hyperparameters search strategies
- Visualize the performances of the generated models
We develop a MVP with Streamlit for users to esaily classify theirs images to obtain a breed prediction using the drag & drop feature.
The training data used in this notebook is the Stanford Dogs Dataset.
Évaluer les performances d’un modèle de Deep Learning
Sélectionner un modèle d'apprentissage Deep Learning adapté à une problèmatique métier
Mettre en place un modèle de Deep Learning
Adapter les paramètres d'un modèle de Deep Learning afin de l’améliorer
Transformer les variables pertinentes d'un modèle de Deep Learning