Reputation: 101
How to get client models (based on Keras) in tensorflow federated?
I am looking for a way to get client models as checkpoints that I can investigate as standard keras model. I looked into this question but it only provides the weights, is there a way to get or save models from clients directly during federated training?
Upvotes: 0
Views: 594
Answers (1)
Reputation: 2941
The answer referenced in the original question provides a method for sending back the model weights; what seems to be missing is how to assign those weights to a tf.keras.Model that has the same architecture? With tff.learning.assign_weights_to_keras_model, a Keras model instance can be initialized with the clients weights, and the instance will provide the standard Keras API (evaluate(), test_on_batch() etc).
The Federated Learning for Text Generation tutorial has an example of doing this.
Upvotes: 1
Related Questions
- Keras model prediction after tensorflow federated learning
- tff.simulation.datasets.ClientData to build federated learning model from CSV files
- FedProx with TensorFlow Federated
- Federated learning : convert my own image dataset into tff simulation Clientdata
- Implement data generator in federated training
- TensorFlow Federated: Keras model with custom learning algorithm
- TFF: How define tff.simulation.ClientData.from_clients_and_fn Function?
- In tensorflow federated, how to assign different training functions to different clients?
- Adapting an existing keras model with multiple inputs to tensorflow federated
- TensorFlow Federated - Adapting existing keras model