Monitor model training live

​​
​​
​​
​​
​​
​​
​See code examples on GitHub​
​See results in Neptune​
​Run example in Google Colab​
​
Introduction
You can log any ML model metadata to Neptune and see it live in the Neptune UI. 
In this guide, you will learn how to do basic model monitoring of your training process. 
You will: 
See the training progress by looking at learning curves for loss and accuracy
Monitor hardware consumption during training across GPU/CPU/MEMORY
By the end of it, you will see your metrics, losses, and hardware live in Neptune!
Before you start
Make sure you meet the following prerequisites before starting:
Have Python 3.6+ installed
Have Tensorflow 2.x with Keras installed
​Have Neptune installed​
​Create a Neptune project​
Step 1: Create a basic training script
As an example, I’ll use a script that trains a Keras model on the MNIST dataset.
You don’t have to use Keras to monitor your training runs live with Neptune.
You can use it with any Machine learning framework, Optimization framework, and any other code. 
Create a file train.py and copy the script below.
train.py
1
from tensorflow import keras
2
​
3
PARAMS = {'epoch_nr': 100,
4
          'batch_size': 256,
5
          'lr': 0.005,
6
          'momentum': 0.4,
7
          'use_nesterov': True,
8
          'unit_nr': 256,
9
          'dropout': 0.05}
10
​
11
mnist = keras.datasets.mnist
12
(x_train, y_train),(x_test, y_test) = mnist.load_data()
13
x_train, x_test = x_train / 255.0, x_test / 255.0
14
​
15
model = keras.models.Sequential([
16
  keras.layers.Flatten(),
17
  keras.layers.Dense(PARAMS['unit_nr'], activation=keras.activations.relu),
18
  keras.layers.Dropout(PARAMS['dropout']),
19
  keras.layers.Dense(10, activation=keras.activations.softmax)
20
])
21
​
22
optimizer = keras.optimizers.SGD(lr=PARAMS['lr'],
23
                                 momentum=PARAMS['momentum'],
24
                                 nesterov=PARAMS['use_nesterov'],)
25
​
26
model.compile(optimizer=optimizer,
27
              loss='sparse_categorical_crossentropy',
28
              metrics=['accuracy'])
29
​
30
model.fit(x_train, y_train,
31
          epochs=PARAMS['epoch_nr'],
32
          batch_size=PARAMS['batch_size'])
Copied!
Run training to make sure that it works correctly.
1
python train.py
Copied!
Step 2: Connect Neptune to your script
At the top of your script add a snippet that connects it to Neptune:
1
import neptune.new as neptune
2
​
3
run = neptune.init(api_token='<YOUR_API_TOKEN>',
4
                   project='<YOUR_PROJECT_NAME>') # your credentials
Copied!
You can use the api_token='ANONYMOUS' and project='common/quickstarts' to explore without having to create a Neptune account.
Executing this snippet will give you a link like this one:
https://app.neptune.ai/o/common/org/quickstarts/e/QUI-28177
with common/quickstarts replaced by your_workspace/your_project_name, and QUI-28177replaced by your Run ID.
To create your own private projects, you will need a Neptune account and your personal Neptune API token.
Step 3: Add logging for metrics and losses
To log a metric or loss during training you should use:
1
loss = ...
2
run["train/loss"].log(loss)
Copied!
A few explanations here:
"train/loss" is a path to a log in Neptune called Namespace where you can log metrics, losses and many other model building metadata ​
You can have multiple log namespaces like 'train/acc', 'val/f1_score’, ‘train/log-loss’, ‘test/acc’).
The argument of the log() method is the actual value you want to log.
Typically during training, you log those metrics in a loop after every iteration. Call run["train/loss"].log(loss)at each step.
1
for i in range(epochs):
2
    ...
3
    run["train/loss"].log(loss)
4
    run["train/acc"].log(accuracy)
Copied!
Many frameworks, like Keras, let you create a callback that is executed inside of the training loop.
Now that you know all this.
Steps for Keras
Create a Neptune callback.
1
class NeptuneMonitor(keras.callbacks.Callback):
2
     def on_epoch_end(self, epoch, logs=None):
3
          for metric_name, metric_value in logs.items():
4
              run["train/{}".format(metric_name)].log(metric_value)
Copied!
Pass the callback to the model.fit() method:
1
model.fit(x_train, y_train,
2
           epochs=PARAMS['epoch_nr'],
3
           batch_size=PARAMS['batch_size'],
4
           callbacks=[NeptuneMonitor()])
Copied!
You don’t actually have to implement this callback yourself. 
Import our callback implementation from TensorFlow / Keras integration.​
Step 4: Stop logging
Once you are done logging, you should stop tracking the Run using the stop() method. This is a must only while logging from a notebook environment. While logging through a script, Neptune automatically stops tracking once the script has completed.
1
run.stop()
Copied!
Step 5: Run your script and see results in Neptune
Run the training script.
1
python train.py
Copied!
If it worked correctly you should see:
a link to Neptune run. Click on it and go to the app,
metrics and losses in the Logs and Charts sections of the UI,
hardware consumption and console logs in the Monitoring section of the UI.
​
​​
​​
​​
​​
​​
​​
​See code examples on GitHub​
​See results in Neptune​
​Run example in Google Colab​
What's next?
What can you log and display?
TensorFlow / Keras
Displaying metadata