Skip to content

Integrate with Keras

Comet integrates with Keras.

Comet automatically logs metrics, parameters, and histograms, and more from your Keras code, without requiring anything other than adding the following lines of code to your Keras script, and setting the Experiment parameters to the appropriate values to control what you would like to log:

from comet_ml import Experiment
experiment = Experiment(

# Your code here...

Open In Colab

Log automatically

By integrating with Keras, Comet automatically logs the following items:

  • Model and graph description
  • Steps and epochs
  • Metrics (such as loss and accuracy)
  • Hyperparameters
  • Optimizer Parameters (such as the learning rate, beta decay rate, and more)
  • Number of trainable parameters
  • Histograms for weights and biases
  • Histograms for activations
  • Histograms for gradients

Configure Comet for Keras

You can control what is automatically logged by Comet through an experiment parameter, environment variable, or a configuration setting:

ItemExperiment ParameterEnvironment SettingConfiguration Setting
model/graph descriptionlog_graphCOMET_AUTO_LOG_GRAPHcomet.auto_log.graph
metric logging rateauto_metric_step_rateCOMET_AUTO_LOG_METRIC_STEP_RATEcomet.auto_log.metric_step_rate
command-line argumentsparse_argsCOMET_AUTO_LOG_CLI_ARGUMENTScomet.auto_log.cli_arguments
histogram logging rateauto_histogram_epoch_rateCOMET_AUTO_LOG_HISTOGRAM_EPOCH_RATEcomet.auto_log.histogram_epoch_rate

For more information about using environment parameters in Comet, see Configure Comet.

In addition, you can control exactly which inputs at which layers are logged for activations and gradient histogram logging.

Log automatically

By integrating with Keras, Comet automatically logs histograms for:

  • Weights and biases
  • Gradients
  • Activations

These histograms appear on the Histogram tab in the single experiment pages:

The histograms for weights and biases are logged for all layers before training, and for every auto_histogram_epoch_rate epochs (default 1).

The histogram activations are logged for the selected input data or targets (comet.keras.histogram_activation_index_list with a default of the first pattern), at the selected layers (comet.keras.histogram_activation_layer_list with a default of the last layer).

For activation histograms:

ItemConfiguration Setting
which inputscomet.keras.histogram_activation_index_list
which layerscomet.keras.histogram_activation_layer_list


To automatically log activations, you must be using TensorFlow, version 1.13 or higher. Inputs must be in a tensor or numpy format (for example, DataSets are not yet supported). Note that logging activations from an input layer does not work with TensorFlow version 1.

The histogram gradients of the weights and biases with respect to the loss are logged for the selected inputs or targets (comet.keras.histogram_gradient_index_list with a default of the first pattern), at the selected layers (comet.keras.histogram_gradient_layer_list with a default of the last layer).

For gradient histograms:

ItemConfiguration Setting
which inputscomet.keras.histogram_gradient_index_list
which layerscomet.keras.histogram_gradient_layer_list


To automatically log gradients, you must be using TensorFlow, version 2.2.1 or higher, or Keras 2.4.3 and higher. Inputs and targets must be in a tensor or numpy format (e.g., DataSets are not yet supported).

Each of input of the index lists is composed of an empty string, or a list of index positions. If more than one index is given, separate the values with a comma.

For example, to indicate that you would like to log the activations for the first three input data examples, use:


Each item in the layer list is composed of a list of index positions, negative index positions, or layer names. If more than one value is given, separate them with a comma.

For example, to indicate that you would like to log the gradients at the first layer, last layer, and a layer named "hidden", use:


Similarly, if you would like to log the activations from these layers use:


Each of the histograms logged can have a "prefix" attached to the histogram name in the UI. The prefix is set through the configuration setting comet.keras.histogram_name_prefix and can be composed of any of the following components:

Histogram name prefix componentMeaning
{model_name}Name of the Keras model
{layer_name}Name of the Keras layer
{layer_num}Number of the Keras layer
{max_digits}Number of digits in last Keras layer

By default, the prefix is {layer_num:0{max_digits}d} (using Python's formatting syntax) which would give prefixes like:


Finally, you can control the following items as well:

ItemMeaningDefault value
comet.keras.histogram_batch_sizenumber of patterns to process at once when logging all1000
comet.logging.metrics_ignorenames of metrics to ignorekeras:batch_size, keras:batch_batch
comet.logging.parameters_ignorenames of parameters to ignorekeras:verbose, keras:do_validation, keras:validation_steps

For more details on passing in parameters to the experiment, please see Experiment(). For more details on setting up configuration variables, see Python Configuration.

Use the Comet Keras callback

Comet logs your experiment through a callback executed when you run in Keras. You do not need to add this callback yourself, we do it for you automatically. However, if you ever need to access the callback manually, simply call experiment.get_keras_callback().

You can use this callback to log additional data to Comet such as model checkpoints, images, audio, and more, as Experiment Assets, Models or Artifacts.

Report manually

If you want to log data that isn't already automatically logged or handled by the Keras callback, check out the logging methods available in Comet's Experiment object.

End-to-end example

Here is a simple end-to-end Keras example which uses a Dense neural network on the MNIST dataset.

For more examples using Keras, see our examples GitHub repository.

from comet_ml import Experiment

#create an experiment with your api key
experiment = Experiment(

from tensorflow import keras
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout
from tensorflow.keras.callbacks import EarlyStopping

batch_size = 128
num_classes = 10
epochs = 20
num_nodes = 64
optimizer = 'adam'
activation = 'relu'

# the data, shuffled and split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()

x_train = x_train.reshape(60000, 784)
x_test = x_test.reshape(10000, 784)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')

# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)

#these will all get logged
model = Sequential()
model.add(Dense(num_nodes, activation='relu', input_shape=(784,)))
#model.add(Dense(256, activation='relu'))
model.add(Dense(num_classes, activation='softmax'))

#print model.summary() to preserve automatically in `Output` tab


#will log metrics with the prefix 'train_'
with experiment.train():
    history =, y_train,
                        validation_data=(x_test, y_test),
                        callbacks=[EarlyStopping(monitor='val_loss', min_delta=1e-4,patience=3, verbose=1, mode='auto')])

#will log metrics with the prefix 'test_'
with experiment.test():
    loss, accuracy = model.evaluate(x_test, y_test)
    metrics = {

experiment.log_dataset_hash(x_train) #creates and logs a hash of your data

For additional examples, see:

Try it out!

Here's an example for using Comet with Keras.

Open In Colab

Apr. 27, 2022