Callbacks neptune

[spotiris]

- What I did Added a training callback for Neptune integration. Supports automatic logging of keras train/val losses and metrics, as well as arbitrary parameter/tag/attribute logging.

- How I did it I subclassed the keras.callbacks.Callback class and made it write the on_epoch_end logs as neptune metrics. The parameter/tag/attribute logging is simply by forwarding the keyword arguments to the neptune.projects.Project.create_experiment method.

- How you can verify it Currently you can only verify it by running the example provided. This is because Neptune requires a registered account and API token.

- Request for comments

• Should we require users to use the environment variable NEPTUNE_API_TOKEN rather than having the option to supply it? The motivation is to prevent users accidentally sharing their private keys.
• Is an example sufficient in lieu of a unit test? We cannot write a unit test without revealing a user account or API token.