A small library to get a notification when your training is complete or when it crashes during the process with two additional lines of code.

When training deep learning models, it is common to use early stopping. Apart from a rough estimate, it is difficult to predict when the training will finish. Thus, it can be interesting to set up automatic notifications for your training. It is also interesting to be notified when your training crashes in the middle of the process for unexpected reasons.

Install with pip or equivalent.

pip install knockknock

(For conda users, you can use conda install -c victorsanh knockknock.)

This code has only been tested with Python 3.6.

The library is designed to be used in a seamless way, with minimal code modification: you only need to add a decorator on top your main function call. The return value (if there is one) is also reported in the notification.

There are currently three ways to setup notifications: email, Slack and Telegram.

The service relies on Yagmail a GMAIL/SMTP client. You'll need a gmail email address to use it (you can setup one here, it's free). I recommend creating a new one (rather than your usual one) since you'll have to modify the account's security settings to allow the Python library to access it by Turning on less secure apps.

from knockknock import email_sender

@email_sender(recipient_email="<[email protected]>", sender_email="<grandma'[email protected]>")
def train_your_nicest_model(your_nicest_parameters):
    import time
    time.sleep(10000)
    return {'loss': 0.9} # Optional return value

knockknock email \
    --recipient-email <[email protected]> \
    --sender-email <grandma'[email protected]> \
    sleep 10

If sender_email is not specified, then recipient_email will be also used for sending.

Note that launching this will asks you for the sender's email password. It will be safely stored in the system keyring service through the keyring Python library.

Similarly, you can also use Slack to get notifications. You'll have to get your Slack room webhook URL and optionally your user id (if you want to tag yourself or someone else).

from knockknock import slack_sender

webhook_url = "<webhook_url_to_your_slack_room>"
@slack_sender(webhook_url=webhook_url, channel="<your_favorite_slack_channel>")
def train_your_nicest_model(your_nicest_parameters):
    import time
    time.sleep(10000)
    return {'loss': 0.9} # Optional return value

You can also specify an optional argument to tag specific people: user_mentions=[<your_slack_id>, <grandma's_slack_id>].

knockknock slack \
    --webhook-url <webhook_url_to_your_slack_room> \
    --channel <your_favorite_slack_channel> \
    sleep 10

You can also specify an optional argument to tag specific people: --user-mentions <your_slack_id>,<grandma's_slack_id>.

You can also use Telegram Messenger to get notifications. You'll first have to create your own notification bot by following the three steps provided by Telegram here and save your API access TOKEN.

Telegram bots are shy and can't send the first message so you'll have to do the first step. By sending the first message, you'll be able to get the chat_id required (identification of your messaging room) by visiting https://api.telegram.org/bot<YourBOTToken>/getUpdates and get the int under the key message['chat']['id'].

from knockknock import telegram_sender

CHAT_ID: int = <your_messaging_room_id>
@telegram_sender(token="<your_api_token>", chat_id=CHAT_ID)
def train_your_nicest_model(your_nicest_parameters):
    import time
    time.sleep(10000)
    return {'loss': 0.9} # Optional return value

knockknock telegram \
    --token <your_api_token> \
    --chat-id <your_messaging_room_id> \
    sleep 10

Thanks to @noklam, you can also use Microsoft Teams to get notifications. You'll have to get your Team Channel webhook URL.

from knockknock import telegram_sender

CHAT_ID: int = <your_messaging_room_id>
@teams_sender(token="<webhook_url_to_your_teams_channel>")
def train_your_nicest_model(your_nicest_parameters):
    import time
    time.sleep(10)
    return {'loss': 0.9} # Optional return value

knockknock teams \
    --webhook-url <webhook_url_to_your_teams_channel> \
    sleep 10

When using distributed training, a GPU is bound to its process using the local rank variable. Since knockknock works at the process level, if you are using 8 GPUs, you would get 8 notifications at the beginning and 8 notifications at the end... To circumvent that, except for errors, only the master process is allowed to send notifications so that you receive only one notification at the beginning and one notification at the end.

Note: In PyTorch, the launch of torch.distributed.launch sets up a RANK environment variable for each process (see here). This is used to detect the master process, and for now, the only simple way I came up with. Unfortunately, this is not intended to be general for all platforms but I would happily discuss smarter/better ways to handle distributed training in an issue/PR.