Warning
This is is a work in progress!
tf-remote
is a library for creatting and trainning machine learning models using tensorflow. It is structured around a simple achtechture where the client creates multiple models and process the input data accordingly to create trainning jobs. Then, those jobs can be performed on a remote server.
- Each model is provided as an
model.h5
file. - Each dataset is provided as a standard numpy array, and is tied to a data identifier.
- The model trainning is based on a set of hyperparameters.
- Each trainning job is defined as the collection model + data identifier + hyperparameters. Therefore, it is possible to use several combinations of these.
- The client creates a set of trainning jobs and related datasets.
- The client sends the datasets via API to the server.
- The client sends the trainning jobs via API to the server.
- Each trainning job is accepted only if there is already an specified dataset compatible with its data identifier.
- The server continuously matain a queue with all the trainning jobs available to be performed.
- The server performs each trainning job inthe queue, and saves the results.
- The cleint can request information from the server at anytime:
- Status: returns the current size of the queue and the amount of jobs already performed.
- Results:
- Table: returns a table with all information on its trainning jobs.
- Model: returns the rtesulting model files
model.h5
for a speccific trainning job.