Skip to content

Comet Python SDK releases

The following is a history of released comet_ml versions. It does not list everything that was changed in a release, but does mention the highlights and all public-facing additions, changes, and deprecations.

You can install any one of the following released version numbers. The full list of releases is at Python Package Index.

For installation of the Comet Python SDK on air-gapped computers, see Offline Installation.

Note that items that refer to Experiment also apply to ExistingExperiment, OfflineExperiment, and ExistingOfflineExperiment.

Release 3.47.0

Release date: Sep 26, 2024

  • Added module level flush and end functions for the currently running experiment.
  • Update the functions comet_ml.integrations.sklearn.load_model and comet_ml.integrations.pytorch.load_model. Now they no longer show warnings when the model is downloaded from model registry.

Release 3.46.0

Release date: Sep 19, 2024

  • Added APIExperiment.get_artifact_lineage method.
  • Fixed bug when register_model operations were ignored when uploading an offline experiment.

Release 3.45.1

Release date: Sep 13, 2024

  • Fixed bug when experiments in colab didn't log parameters if exp.end() is not called.

Release 3.45.0

Release date: Aug 28, 2024

  • Added flag to disable normalization in Experiment.log_audio
  • Added comet_ml.error_mode()
  • Added serialization support for numpy arrays and torch tensors inside JSON assets to log
  • Fixed bug when annotations are logged for images, that didn't have annotations in log_image method
  • DEBUG mode for console logging is now forbidden

Release 3.44.4

Release date: Aug 7, 2024

  • Align Name size limits across Comet product, those limits are now higher.
  • Improved error message when image conversion failed by pillow library when logging images with Experiment.log_image.
  • Keras integration updated to support tensorflow 2.16+.
  • Fixed bug when Ray integration could not finish work correctly if OfflineExperiment was used.

Release 3.44.3

Release date: Jul 24, 2024

  • Ray integration updated to support automatic failure retrying.
  • Fixed bug when SDK did not prevent new images from being uploaded after hard limit was reached for the experiment.

Release 3.44.2

Release date: Jul 23, 2024

  • Fixed XGBoost v2 integration
  • Fixed bug when in some circumstances Comet printed unwanted errors about GPU logging
  • Added API objects for building image annotations
  • Storage usage optimizations
  • Comet now suggests correct config names if it detects incomplete ones provided by user
  • Fixed TensorboardX integration
  • Major docstrings update

Release 3.44.1

Release date: Jul 9, 2024

  • Fix a potential crash in comet check
  • Improve message when trying to write config to an existing file without the force flag
  • Fix XGBoost integration in case of failure to retrieve parameters

Release 3.44.0

Release date: Jul 8, 2024

  • Added comet_ml.login() as newer alternative to comet_ml.init() (init() is deprecated now). CLI tool was updated accordingly.
  • Various improvements in error processing and some user-facing error messages

Release 3.43.2

Release date: Jun 21, 2024

  • Error and warning messages:
    • added more error rate filtering
    • reworked misleading messages to make them clearer.
  • Added dictionary flattening logic to Experiment.log_metrics
  • Fixed the bug in CLI tool when sometimes arguments failed to be parsed because of IndexError
  • Added possibility to set experiment name and tags in ExperimentConfig for comet_ml.start

Release 3.43.1

Release date: Jun 13, 2024

  • Improvements in file uploads resiliency
  • Fixed the pytorch auto-logger bug when loss wasn't logged if training loss.backward wasn't called not in the same thread where experiment was created.
  • API.get_metrics_df() major logic update
  • Improvements in error processing and user-facing error messages

Release 3.43.0

Release date: May 29, 2024

  • Added delete_parameter and delete_parameters methods to APIExperiment
  • Updated ray integration to support usage of already running experiments and offline experiments.

Release 3.42.1

Release date: May 22, 2024

  • WebSockets excluded from requirements
  • API.download_experiment_asset speedup
  • Added inputs validation for Experiment.log_curve
  • Added inputs validation for Experiment.add_tag and Experiment.add_tags
  • Added new error message if experiment is tried to be logged in LLM project
  • Added package to conda for python 3.12

Release 3.42.0

Release date: May 10, 2024

  • Introduced new method - API.get_metrics_df
  • Major Experiment docstrings update
  • Allowed customization of the experiment name in Ray integration
  • comet_ml.integration.pytorch.watch no longer fails when tracking non-trainable layers
  • TestExperiment now raises an exception only if comet error message was logged
  • Added annotation validation to Experiment.log_image

Release 3.41.0

Release date: Apr 24, 2024

  • New behavior for get_metric, get_other, get_parameter, get_tags methods in ExistingExperiment. Now returned values are synced with the backend. It's possible to get latest values logged in previous Experiment run or values from another ExistingExperiment instance running in parallel.
  • Fix bug when pytorch model definition wasn't logged if watch function is used
  • APIExperiment.get_metrics speedup
  • Fixed bug when failed to upload images with dots in the name
  • Experiment.flush() now properly returns bool value
  • Added log_video method to APIExperiment

Release 3.40.0

Release date: Apr 11, 2024

  • The experiment name will now be displayed in the experiment summary
  • Fixed bug with CUDA_VISIBLE_DEVICES=-1 that lead to user facing error messages
  • Introduced new method called comet_ml.get_running_experiment() that returns the running experiment if one exists
  • Introduced functionality in the SDK to determine the Comet URL to log data to based on the API key used. This means that you won't need to specify the COMET_URL_OVERRIDE environment variable for on-premise deployments. Note: for this functionality to be enabled on your on-premise deployment please contact our deployment team.

Release 3.39.2

Release date: Mar 26, 2024

  • Introduced TestExperiment (alpha version)
  • Not synced remote assets on model download are now skipped
  • Introduced new SDK method for API experiment: APIExperiment.delete_tags(tags)

Release 3.39.1

Release date: Mar 18, 2024

  • Improved CPU system metrics logging, fixed some cases when it could stop working at all
  • Fixed bug when model file created by pytorch.log_model had incorrect name with OfflineExperiment

Release 3.39.0

Release date: Mar 8, 2024

  • System metrics interval decreased from 60 to 30 seconds
  • bugfix: API.get_archived_experiments returns what it should return now
  • Code name can now be overwritten in experiment.log_code
  • Artifact will now raise an exception on creation if its name exceeds 100 characters
  • Major multiprocessing compatibility improvements (fixed hangings on upload at the end of worker process).

Release 3.38.1

Release date: Feb 23, 2024

  • Improved SDK error message when trying to log an artifact for a version that has been deleted
  • Added support for seaborn plots (matplotlib.axes.Axes objects) in log_figure

Release 3.38.0

Release date: Feb 16, 2024

  • Added support for Pytorch Histogram Logging

Release 3.37.2

Release date: Feb 13, 2024

  • Disabled mode no longer requires API key
  • Internal improvements and optimizations

Release 3.37.1

Release date: Feb 2, 2024

  • Internal improvements and optimizations

Release 3.37.0

Release date: Jan 23, 2024

  • Introduced new Experiment.log_points_3d method
  • bugfix: fastai autologger no longer ignores auto_metric_logging and auto_param_logging.

Release 3.36.1

Release date: Jan 18, 2024

  • bugfix: now GPU logging no longer stops if some device metric is not available, the rest of them are still logged

Release 3.36.0

Release date: Jan 12, 2024

  • comet_ml.init() now supports anonymous mode

Release 3.35.6

Release date: Jan 11, 2024

  • FastAI integration was updated to support version 2
  • Introduced APIExperiment.set_state method (possible states: running, crashed, finished)

Release 3.35.5

Release date: Dec 14, 2023

  • Made artifact_type optional in Artifact.__init__
  • It's possible now to specify the name of the asset in APIExperiment.log_asset (before it was always a full file name)

Release 3.35.4

Release date: Dec 7, 2023

  • Added ability to get python panel experiments
  • UX improvements to running Experiment from AWS Lambda function
  • Added new validation to inputs for register_model

Release 3.35.3

Release date: Nov 10, 2023

  • Fixed bug when None values were incorrectly shown in summary
  • Added support for ydata-profiling
  • Added new automatic naming scheme for models in MLFlow integration

Release 3.35.2

Release date: Nov 3, 2023

  • Introduced callback for uploading experiment ZIP to S3 bucket if experiment is in offline mode in the end.
  • Major logging optimization

Release 3.35.1

Release date: Oct 26, 2023

  • Introduced new GPU and disk usage system metrics

Release 3.35.0

Release date: Oct 23, 2023

  • Major improvement: experiment now tries to reconnect after connection was lost. All pending data logged during offline period will be uploaded to Comet.
  • Pytorch autologger is now compatible with accelerate library.

Release 3.34.0

Release date: Oct 17, 2023

  • Introduced Experiment.log_video method
  • Introduced Model.get_version_history method

Release 3.33.12

Release date: Oct 16, 2023

  • It is now possible to set a callback to process an experiment zip file at the end of experiment if the connection is lost.
  • Added support for nested params to APIExperiment.log_parameters.
  • Bug fix: APIExperiment.log_metric no longer sends a timestamp instead of an epoch.

Release 3.33.11

Release date: Sep 28, 2023

  • Added new APIExperiment.get_state method to return the state of an experiment (running, finished, crashed)
  • Fixed issue when Comet was running within AWS Lambda functions

Release 3.33.10

Release date: Sep 6, 2023

  • Added new Experiment.log_optimization method
  • Fixed string metric logging issue on experiment start
  • Fix vertex integration crashing
  • Experiment.register_model now has option to be run synchronously
  • deprecated copy_to_tmp in various Experiment.log_* methods

Release 3.33.9

Release date: Aug 22, 2023

  • Updated Vertex integration
  • Users can configure retry policy when connectivity to the Comet server is lost
  • Comet Keras callback now allows to set the initial validation step for resume training workflows
  • Network metrics in ExistingExperiment are disabled by default now

Release 3.33.8

Release date: Aug 4, 2023

  • Comet no longer uploads images if the limit is exceeded
  • Deprecated copy_to_tmp argument in Experiment.log_asset
  • Added conda build for python 3.11

Release 3.33.7

  • Improved exceptions handling in sagemaker integration
  • Now experiment API object can be passed to sagemaker integration functions
  • Introduced CometVertexPipelineLogger with the support of safe API key sharing
  • Experiments don't hang as running in UI after fallback to offline mode

Release 3.33.6

Release date: July 12, 2023

  • Keras callback now uses separate step for validation batches
  • CPU, GPU, Network metrics log interval is now configurable
  • bugfix: Fixed logging more than 10 nested params issue (using dictionaries)

Release 3.33.5

Release date: June 22, 2023

  • Introduced remote models logging and loading support
  • Experiment now logs network system metrics
  • Experiment.log_parameters now supports nested dictionaries that have a new view in UI
  • Experiment.log_image now has metadata argument
  • bugfix: comet keras callback now takes initial step from the experiment and epoch from keras
  • bugfix: urllib3 required version softened. Previous version could lead to installation problems.

Release 3.33.4

Release date: June 7, 2023

  • Update urllib3 Python dependency to version 2.0
  • Added support for CSV format with espace character in Experiment.log_table()
  • Bugfix: git metadata and git-patch sections were sometimes missing from the experiment summary

Release 3.33.3

Release date: May 16, 2023

  • File assets and artifacts are uploaded to S3 directly leading to improved upload speeds
  • Update log message copy when optimizer timesout

Release 3.33.2

Release date: May 12, 2023

Release 3.33.1

Release date: May 8, 2023

  • Various SDK improvements

Release 3.33.0

Release date: May 1, 2023

  • Novel Model Approval Process API supporting new tags and status fields

Release 3.32.9

Release date: Apr 26, 2023

  • Updated the OpenAI integration to support Chat Completions
  • Release new SDK methods for saving and loading Scikit Learn methods - Read the docs here

Release 3.32.8

Release date: Apr 17, 2023

  • New Integration with OpenAI SDK - Read the docs here
  • Update to the memory metric reported by the SDK
  • Per-core CPU system metric no longer sent by default
  • Optimizer get_id() method returns only id value without the COMET_OPTIMIZER_ID prefix

Release 3.32.7

Release date: Apr 09, 2023

  • SDK logs now appear in color on your terminal
  • Bugfix: metaflow integration logged some hyper parameters even when auto_param_logging=False

Release 3.32.6

Release date: Mar 30, 2023

  • Updated logging of some BI events so they are not triggered as often

Release 3.32.5

Release date: Mar 24, 2023

  • Add upper bound for everett version for python>=3.6 to properly select default logging infastructure
  • Fixed experiment summary to properly display HTML assets logged section
  • Git metadata logging can now be disabled separately from source code logging

Release 3.32.4

Release date: Mar 10, 2023

  • Support for storing and retrieving a Comet API key using GCP Secret Manager
  • Support in Experiment.log_code for an extended set of source code file extensions: .py, .text, .yaml, .ipynb, and .sh

Release 3.32.3

Release date: Mar 01, 2023

  • Bug fix: Fix issue registering models with multiple files (a folder)
  • Bug fix: Metaflow integration could crash on methods not decorated with @step
  • Introduced support for logging values specified via environment variables with the COMET_LOG_OTHER_ prefix
  • Introduced new Experiment.flush() method that is blocking until all pending data is logged to Comet. Works like .end() but does not end the experiment

Release 3.32.2

Release date: Feb 16, 2023

  • Highlight the new Pytorch model saving and loading functions in console output logs
  • Support for overwrite parameter in Experiment.log_code

Release 3.32.1

Release date: Feb 10, 2023

  • Fixed the Metaflow decorator for latest version of Metaflow
  • Fixed issue with experiments not reporting their status during long training run
  • Fixed issue with Artifact download when absolute paths where used instead of relative paths

Release 3.32.0

Release date: Jan 26, 2023

  • Integration with AWS SageMaker
  • Support for distributed training with Ray

Release 3.31.22

Release date: Jan 10, 2023

  • Logging metaflow origin_run_id, enabling easier searches for original experiment in case of resuming a failed flow run
  • SDK to sends has_crashed flag to Comet backend to signify unhandled exceptions that cause user's script to crash
  • The Experiment.log_image now allows an annotations parameter, which will be used in the future for bounding boxes and segmentation

Release 3.31.21

Release date: December 23, 2022

  • Release new Gymnasium integration
  • Improvement to CPU usage reports - Now only the CPU usage for the Python process tree is reported to Comet
  • Remove some INFO-level log messages related to GPU metrics logging

Release 3.31.20

Release date: December 11, 2022

  • GPU detection based on CUDA_VISIBLE_DEVICES environment variable, as a result GPU metrics will be displayed only for the GPUs used by the training run
  • Improve error message when using Experiment with an already existing experiment key
  • Bugfix: Log validation batch metrics under a proper namespace in Keras
  • Bugfix: Throttling message was not shows under certain circumstances
  • Bugfix: Properly handle AttributeError in RandomForestClassifier when using Comet's Scikit-Learn integration

Release 3.31.19

Release date: November 16, 2022

  • Bugfix: add missing registration of the comet_ml.spacy.logger.v1 Python entry point required for integration with spaCy logging

Release 3.31.18

Release date: November 14, 2022

  • Streamlined Model Registration by adding the method Experiment.register_model(...)
  • FallbackToOffline: if network connection is lost, an offline archive will be made available for later upload

Release 3.31.17

Release date: November 08, 2022

  • Fix traceback emitted by logging system in some edge cases

Release 3.31.16

Release date: October 30, 2022

  • Fix faulty retry mechanism related to image upload.
  • Fix metaflow logging system tags.
  • Fix bug in api experiment symlink creation.
  • Metaflow: Integration with project decorator.
  • Add hparams support for tensorboard logger (tensorflow).

Release 3.31.15

Release date: October 3, 2022

  • Logs include timestamps when in DEBUG mode.
  • No longer log CLI flags when Metaflow integration is used
  • Pytorch integration improvements: Methods to save and load Pytorch models
  • Support for Pytorch Tensorboard
  • Bugfix: restore summary in output capture

Release 3.31.14

Release date: September 16, 2022

  • Fix broken Metaflow support.

Release 3.31.13

Release date: September 14, 2022

  • Fix data loss issue introduced by new Thread Pool from 3.31.10
  • Fix SDK crash induced by parameters with numpy array values in some circumstances.
  • Metaflow integration: introduce a comet_skip decorator to prevent creating a live experiment for steps where it is unwanted
  • Python SDK now uses comet.com as its default backend domain
  • Metaflow integration: integrate with Metaflow cards

Release 3.31.12

Release date: August 26, 2022

  • Pin version for websocket-client as >=0.55.0,<1.4.0 so as to avoid recently released broken version.

Release 3.31.11

Release date: August 22, 2022

  • Drop support for Python 2.

Release 3.31.10

Release date: August 21, 2022

  • Bugfix: fix crash on AWS Lambda with Python>3.6

Release 3.31.9

Release date: August 19, 2022

  • Removed dependency on nvidia_ml_py3. The functionality is replaced by including the code from nvidia-ml-py as part of the Comet Python SDK code. Note that nvidia-ml-py package is not a dependency of the Comet Python SDK.
  • Bugfix: remote synced asset may now be logged without specifying logical_path.
  • When an experiment is throttled, a more meaningful message is now logged to the user.
  • Introduce sanitation of Metaflow environment.
  • Extend error reporting to comet backend for various error conditions.

Release 3.31.8

Release date: August 15, 2022

  • Fix broken import comet_ml statement when used with Hugging Face datasets library on Google Colab.

Release 3.31.7

Release date: July 26, 2022

  • Sanitize environment information from Vertex pipeline JSON.
  • Sanitize environment information from Kubeflow pipeline JSON.
  • Invalid API Key error message now includes the hostname for the backend server.
  • Stop logging source code from within the Metaflow library.
  • Add error reporting for various conditions.

Release 3.31.6

Release date: July 13, 2022

  • Rename Metaflow integration attribute pipeline_comet_experiment_key to run_comet_experiment_key.
  • Remove warning when running a Metaflow flow that does not define parameters.
  • Filter out environment variable when logging the conda environment information.
  • Improve parsing of command line arguments.

Release 3.31.5

Release date: June 25, 2022

  • Change Experiment notification title to include user and Experiment name.
  • Fix Histogram.add() with counts.
  • Update Metaflow integration to allow to access the Comet Experiment id for the current Metaflow run through the self.pipeline_comet_experiment_key attribute.

Release 3.31.3

Release date: June 2, 2022

Release 3.31.2

Release date: May 31, 2022

  • Fixed linting problem caused by wrong type-hint comments.
  • Renamed potentially conflicting callback functions - reduce chances of name conflicts with users.
  • Add support for scikit-learn 1.1.0.

Release 3.31.1

Release date: May 12, 2022

  • Preliminary work to improve reliability and speed of Metric logging.

Release 3.31.0

Release date: May 05, 2022

  • Temporarily disable logging of Conda environments. Starting from Release 3.27.0, we logged the Conda environment details. That information included any environment variable set in the Conda environment.

  • New artifact methods to update artifact tags LoggedArtifact.update_artifact_tags, version tags LoggedArtifact.update_version_tags and version aliases update_aliases.

  • We are planning to drop support for Python 2.7. Version 3.30.0 was the last version supporting Python 2.7 as a Conda package.

Release 3.30.0

Release date: April 27, 2022

  • When calling Artifact.add_remote with a remote URI that starts with gs://, the Python SDK try to list objects matching the Google Storage link and log individual objects as synced remote artifact, including their version if object versioning is enabled on the GS bucket. This behavior can be turned off with sync_mode=False
  • When downloading an Artifact Version assets with Artifact.download, if the Artifact Version contains synced remote assets, the Python SDK will try to download the objects from Google Storage to the target directory. This behavior can be turned off with sync_mode=False.
  • Improve error message when the project name is too long.
  • Improve Vertex and Kubeflow integrations by logging individual task names.
  • Use the same default offline directory for the MLflow integration and for Offline Experiments.

Release 3.29.0

Release date: April 21, 2022

  • Add official Conda package for Python 3.10.
  • When calling Artifact.add_remote with a remote URI that starts with s3://, the Python SDK try to list objects matching the S3 link and log individual blobs as synced remote artifact, including their version if object versioning is enabled on the S3 bucket. This behavior can be turned off with sync_mode=False
  • When downloading an Artifact Version assets with Artifact.download, if the Artifact Version contains synced remote assets, the Python SDK will try to download the objects from S3 to the target directory. This behavior can be turned off with sync_mode=False.
  • Parameter asset_type of the Artifact.add_remote is now deprecated, it was not used so far.
  • The Predictor is now fully removed, it was previously deprecated.

Release 3.28.3

Release date: April 11, 2022

  • Improve reliability of large file uploads with some L4 load-balancers.

Release 3.28.2

Release date: March 16, 2022

  • Fixed APIExperiment.get_asset(..., stream=True) to correctly return a streamed response and updated its docstring to include how to use it.
  • When a metric value cannot be converted to a native Python number, the Python SDK now displays a warning as the result might be invalid.
  • Fixed support of COMET_OPTIMIZER_URL configuration environment variable without an ending slash.

Release 3.28.1

Release date: March 14, 2022

  • Switch some common Conda-related error logs from WARNING to DEBUG.

Release 3.28.0

Release date: March 03, 2022

  • Optimizer URL is now automatically computed based on COMET_URL_OVERRIDE for on-premise installation.
  • COMET_WS_URL_OVERRIDE can now includes a trailing slash.
  • When running inside a Container using CGroups V2, the CPU limits are now taken into account to limits the number of parallel file upload and file download.
  • (We now uses class based XGBoost callbacks).

Release 3.27.0

Release date: February 24, 2022

  • When running inside a Conda environment, Conda package, channels and configuration are now logged as Experiment assets.
  • Add Kubeflow and Vertex integration.

Release 3.26.1

Release date: February 17, 2022

  • Improve compatibility of the console logger in some Notebook environments.

Release 3.26.0

Release date: February 11, 2022

  • When running inside a Pydev Console, the Experiment object might not have enough time to send everything. In that case, you will need to explicitly call Experiment.end. The Python SDK will show you a warning to remember to add it. This can happen in various situation like with Pycharm run action when the Run with Python console option is activated.
  • Added APIExperiment.add_tag(tag) for adding an single tag to an APIExperiment.

Release 3.25.0

Release date: January 28, 2022

  • When running the Comet Optimizer, the optimizer name is now logged. The name is set in the optimizer specification, or given a default generated ID if unset.
  • Importing the Python Comet SDK is now allowed after importing other machine-learning libraries. Previously, this raised an exception and stopped the script. Now, the script will continue to run. However, this will result in the lack of many important auto-logged items.
  • You can now finely tune the number of threads used for file upload and downloads by setting the comet.internal.worker_count config key (COMET_INTERNAL_WORKER_COUNT environment variable). The value set should be a number.

Release 3.24.2

Release date: January 20, 2022

Release 3.24.1

Release date: January 17, 2022

  • Removed unnecessary warnings when using Pytorch and disabled experiments.
  • Made sure all scikit-learn Estimator hyper-parameters are automatically logged.

Release 3.24.0

Release date: January 7, 2022

Release 3.23.1

Release date: December 17, 2021

  • When running in Google Colab environment, the whole notebook content is now logged when calling Experiment.end(). The content will be logged as the Experiment asset ColabNotebook.ipynb.
  • If COMET_CONFIG is set, this path will be used by comet_ml.init(), comet_ml.init_onprem() and comet_ml.save() to save the file.

Release 3.23.0

Release date: November 26, 2021

  • When creating a ExistingOfflineExperiment object, the resulting archive file name will now includes a random suffix, avoiding a potential overwriting if another file already exists.
  • Allow to override workspace and/or project_name when uploading offline experiments with comet upload with the --workspace and --project-name CLI flag. See comet upload --help for more information.

Release 3.22.1

Release date: November 19, 2021

  • Added deprecation warning when using the Predictor. Predictor will be fully removed on February 2022.
  • Fixed comet check to work on Windows platform.

Release 3.22.0

Release date: November 16, 2021

Release 3.21.0

Release date: November 6, 2021

Release 3.20.0

Release date: November 2, 2021

  • Fixed URL comet_ml.init_onprem() error
  • Images added via tf.summary.image should keep name
  • Fixed regression in setting api-key from CLI
  • comet upload now returns a non-zero exit code if at least one experiment has failed to upload
  • Added Experiment.send_notification() example

Release 3.19.1

Release date: October 25, 2021

Release 3.19.0

Release date: October 13, 2021

Release 3.18.1

Release date: October 7, 2021

  • Fixed loss metric logging by PyTorch logger; see PyTorch Examples for more details
  • Noted that Python 3.5 requires websocket_server<0.5.4

Release 3.18.0

Release date: October 1, 2021

  • Added support for the latest Scikit-Learn with new new SplineTransformer and PolynomialFeatures estimators
  • Added a default value for offline directory .cometml-runs/ that was prevously required whenever creating an OfflineExperiment or an ExistingOfflineExperiment. If the default cannot be created or cannot be used due to permissions errors, a temporary directory will be used instead. The offline directory can be set either in the code by passing the offline_directory parameter on Experiment creation or through the configuration by setting the comet.offline_directory config key (COMET_OFFLINE_DIRECTORY environment variable)

Release 3.17.0

Release date: September 27, 2021

  • Fixed a potential memory-leak by removing a reference to the Pytorch Model whenever an Experiment ends or is cleaned
  • Updated the command-line comet check to include more information about the environment

Release 3.16.0

Release date: September 16, 2021

  • Improved TensorFlow 2.6 support
  • Better alignment between the API class in the Python SDK and thePython Panels API
  • Hyperparameters are now stored with their fully-qualified name, including the current context (such as "train" and "test") at logging. For example, parameters logged with:
with experiment.train():
    experiment.log_parameter("training_rate", 0.0001)

can be retrieved with:

experiment.get_parameter("train_training_rate")

# or:

with experiment.train():
    experiment.get_parameter("training_rate")

Release 3.15.4

Release date: September 8, 2021

  • Improved logging parameters using a batched-upload protocol over new queue system

Release 3.15.3

Release date: September 2, 2021

  • Increased coverage of test matrix to include tensorflow 2.5
  • Ensure that floating-point epochs are handled properly

Release 3.15.2

Release date: August 30, 2021

  • Improve messages at experiment end when data is still uploading
  • Update timeout for offline experiment uploading
  • Resolved an issue with concurrent access to message queues

Release 3.15.1

Release date: August 26, 2021

  • Restore compatibility with jsonschema 2.6.0 for easy use on Google Colab
  • Improve offline experiment background uploading

Release 3.15.0

Release date: August 24, 2021

Release 3.14.0

Release date: August 6, 2021

  • Fix API Key redaction in log messages when an invalid API Key is provided
  • Add warnings when a Keras callback tries to log to a closed Experiment object
  • Add LoggedArtifact.get_asset and LoggedArtifactAsset.download method to get a single Artifact Asset and download it
  • Fix Artifact download PRESERVE strategy, file content could have been overwritten
  • Update log messages when an Experiment object finished to upload data before ending (either explictly or implicitly). The new log messages now includes the remaining size to upload, the estimated upload speed and an estimated time to completion

Release 3.13.2

Release date: July 28, 2021

  • Fix required everett version for Python 3.5
  • Fix a potential file leak

Release 3.13.1

Release date: July 12, 2021

Release 3.13.0

Release date: July 12, 2021

  • Added progress log messages to Experiment.log_artifact()
  • Increase default asset max size to 100Gb
  • Added retries for file uploads
  • Increased default parallel and max thread counts

Release 3.12.2

Release date: June 24, 2021

  • Added enhancements for faster logging
  • Fixed an issue with an internal histogram method, Histogram.add(values, counts)
  • Updated tab URL names; see Experiment.display(tab=NAME)
  • Updated Comet system debugging tool; see see comet check Command-line interface
  • Included the optimizer total parallel job value in environment variable; see comet optimize Command-line interface
  • Enhanced artifact download speeds; see [Artifact Overview]/docs/v2/guides/artifacts/using-artifacts/#download-a-logged-artifact) for more information

Release 3.12.1

Release date: June 18, 2021

Release 3.12.0

Release date: June 7, 2021

Release 3.11.0

Release date: May 27, 2021

Release 3.10.0

Release date: May 6, 2021

Release 3.9.1

Release date: May 6, 2021

Release 3.9.0

Release date: April 15, 2021

Release 3.8.1

Release date: April 8, 2021

  • Added support for the NO_PROXY environment variable
  • Added additional information on the required use of Experiment.end() when in Jupyter Notebook-based environments

Release 3.8.0

Release date: April 6, 2021

Release 3.7.0

Release date: March 26, 2021

Release 3.6.0

Release date: March 18, 2021

Release 3.5.0

Release date: March 12, 2021

  • Removed netifaces from comet_ml Python SDK requirements
  • Added auto-logging for Facebook Prophet; see the Prophet Tutorial for more information
  • Expose ExistingOfflineExperiment and get_global_experiment from toplevel comet_ml imports (e.g., from comet_ml import ExistingOfflineExperiment)

Release 3.4.0

Release date: March 4, 2021

Release 3.3.5

Release date: February 26, 2021

  • Fixed a repeating warning when logging an empty dict of parameters
  • Fixed a command-line argument parse bug

Release 3.3.4

Release date: February 18, 2021

Release 3.3.3

Release date: February 11, 2021

  • Increased asset upload timeouts in the Python SDK to better handle large files and slow connections
  • Enhanced Experiment.log_remote_asset() remote file name; see Experiment.log_remote_asset()
  • Enhanced Experiment.log_table() documentation; see Experiment.log_table()
  • Fixed API.get_experiment_by_id(ID) to return None when ID does not exist; see API.get_experiment_by_id()
  • Refinements to comet offline; see Command-line API
  • Added Python 3.9 to our extensive internal test matrix

Release 3.3.2

Release date: February 4, 2021

  • Fixed an issue in rendering logged figures, including SHAP figures, in the UI
  • Additional refinements for better logging Jupyter Notebook source code

Release 3.3.1

Release date: February 2, 2021

  • Fixed a bug in experiment.end(), a required call in Jupyter notebooks, and optional in other scenarios. Calling Experiment.end() fails with version 3.3.0. See Experiment.end() for more information.

Release 3.3.0

Release date: February 1, 2021

  • This version of the Comet Python SDK uses a new Comet Optimizer that has many bug fixes and enhancements. See Optimizer for more details on using the Optimizer
  • Added remote asset logging. Useful for tracking remote or local files that you may not wish to upload but want to track. See Experiment.log_remote_asset()
  • Log multiple source files. Useful for when your experiment code is spread across multiple files, or the top-level Python script is not where Experiment() is called. The Experiment.log_code() function is also enhanced to log source code that was missed in prior versions. See Experiment.log_code()
  • Added auto-logging support for TensorFlow Estimators. Now logs model definition, .pbtxt model file, and hyperparameters in addition to metrics. See TensorFlow Estimator Integration
  • Added support for huuuge asset file uploads. See Experiment.log_asset() and related functions

Release 3.2.12

Release date: January 21, 2021

  • Use "simple" for auto_output_logging rather than "default" or "native" on MacOS. This is a temporary solution to address an issue with Python 3.8 and greater on Darwin. See Experiment() for more details on auto_output_logging.

Release 3.2.11

Release date: January 14, 2021

  • Added support for logging 3D point cloud and bounding boxes; see Experiment.log_points_3d()
  • Altered logging configuration to improve performance

Release 3.2.10

Release date: January 7, 2021

Release 3.2.9

Release date: December 18, 2020

  • New Experiment.log_code() for logging additional source code files and folders
  • Allow downloading latest version of model without specifying stage; see API.download_registry_model()
  • Expanded auto-logging for sklearn >=0.22
  • Updated internal check for live threads for latest Python 3.9
  • Updates for pytorch_lightning code logging

Release 3.2.8

Release date: December 8, 2020

Release 3.2.7

Release date: December 2, 2020

  • Added lightgbm auto-logger
  • Fixed an XGBoost callback issue

Release 3.2.6

Release date: November 23, 2020

Release 3.2.5

Release date: October 26, 2020

Release 3.2.4

Release date: October 12, 2020

  • Added Hydra config logging (thanks to Vozf)
  • Added environment query variable
  • Matplotlib's save doesn't flush stream (thanks to jliu)

Release 3.2.3

Release date: October 5, 2020

Release 3.2.2

Release date: September 22, 2020

  • Expanded Confusion Matrix to work with integer-based labeled y_true, rather than requiring onehot vectors: see Experiment.log_confusion_matrix() for more information.
  • Allow Torch tensors attached to a gradient to work throughout comet_ml (including confusion matrix, histogram logging, etc)

Release 3.2.1

Release date: September 18, 2020

  • Fixed a confusion matrix issue with non-integer indices (thanks ironcadiz for reporting)
  • Refactored and expanded xgboost auto-logging
  • Added output refinements when running experiments (reduced duplicated messages, better messages, etc.)
  • Fallback to tf.keras if keras is not installed (thanks DN6 for reporting)
  • Replaced comet-git-pure with dulwich

Release 3.2.0

Release date: August 28, 2020

  • Added major speed-up for logging histograms
    • About 10 times faster when used with numpy
  • Added refinements for logging metrics with xgboost
  • Added refinements for the [Predictive Early Stopping]
  • Added refinements for logging parameters, others, and metrics (thanks for reporting @Idodox!)
    • Added better representations of dictionaries and other objects
    • Added warning if key or value is truncated
  • Dropped support for Python 3.4

Release 3.1.17

Release date: August 20, 2020

This release will be the last to support Python 3.4.

  • Fixed issue to allow COMET_AUTO_LOG_DISABLE to disable all comet_ml actions. For more information, see Experiment Configuration Parameters.
  • Fixed issues using comet_ml on Google Colab that required restarting Colab engine:
    • Can now use Colab's default version of jsonschema
    • comet upload failed without restart

Release 3.1.16

Release date: August 6, 2020

  • Refined Experiment(auto_weight_logging=True) for Keras:
    • creates separate 3D histograms for weight matrices and bias arrays
    • uses layer names for Histogram names
  • Added Histogram and image processing speed ups (up to 10x faster)
  • Added more flexible logging of parameters
  • Added more flexible source code logging for Experiment(log_code=True)

Release 3.1.15

Release date: August 3, 2020

Release 3.1.14

Release date: July 21, 2020

Release 3.1.13

Release date: July 17, 2020

Release 3.1.12

Release date: June 25, 2020

  • Add partial support for using the Optimizer with the Pytorch Lightning Comet Logger. See here for an example.
  • Add experiment.display_project for all kinds of experiments (online, offline and api). Behavior is similar to expriment.display.
  • Add file download timeout configuration, COMET_TIMEOUT_FILE_DOWNLOAD with a default value of 600 seconds. This is used only for registry downloading at the moment.

Release 3.1.11

Release date: June 17, 2020

  • Better logging of all kind of summary data when using tensorflow.summary.
  • Added better error catching in Keras callbacks.

Release 3.1.10

Release date: June 9, 2020

  • Added comet models download and comet models list commands, see comet models
  • Added comet init command to generate example scripts from cookiecutter recipe, see comet init
  • Removed dependency on typing_extensions to allow %pip install comet_ml to work with Google Colab without a kernel restart

Release 3.1.9

Release date: June 6, 2020

Release 3.1.8

Release date: May 26, 2020

  • Improvements to work better on Google Colab
  • Added API.move_experiments()
  • Restored ability to from comet_ml import get_global_experiment

Release 3.1.7

Release date: May 18, 2020

  • Refined APIExperiment.display()
  • Made HTTP timeouts configurable. See COMET_TIMEOUT_HTTP
  • Display a log message when an experiment has been throttled
  • Fixed support for Python 2.7 in conda package
  • Added new auto-logger for XGBoost
  • Added ability to fetch parameters with API.get_metrics_for_chart()
  • Better invalid model name message: shows possible valid names

Release 3.1.6

Release date: April 20, 2020

  • Fixed documentation for ExistingExperiment() default parameters
  • Fixed console logging that was cut off
  • Added auto-logged cloud metadata (currently AWS, Azure and Google Cloud Platform are supported)
  • Added error checking for API.get_metrics_for_chart()
  • Added Histogram.from_json() and default args for histogram.display()

Release 3.1.5

Release date: April 14, 2020

  • Fixed empty Experiment Summary section headings from showing
  • Updated conda's comet_ml dependencies to match pip's comet_ml dependencies
  • Fixed reading non-utf8 Unix package descriptions
  • Fixed Experiment's disabled parameter docstrings
  • Fixed APIExperiment.set_os() docstrings

Release 3.1.4

Release date: April 7, 2020

Release 3.1.3

Release date: March 23, 2020

Release 3.1.2

Release date: March 12, 2020

  • Deprecated COMET_REST_API_KEY in Python API; just use COMET_API_KEY now
  • Noted the correct minimum psutil version (5.6.3) in setup.py
  • Fixed issue with auto output logging in OfflineExperiments from Jupyter
  • Fixed Keras and Tensorflow model graphs to produce diff-able representations

Release 3.1.1

Release date: February 27, 2020

Release 3.1.0

Release date: February 17, 2020

Release 3.0.3

Release date: February 10, 2020

Release 3.0.2

Release date: December 19, 2019

  • Added optional metadata to all Experiment.log_asset_data() methods
  • Added 'text' to Experiment.display(tab='text')
  • Updated automatic logging docstrings (thanks to lemairecarl for suggestion)
  • Added Experiment.log_text()
  • Added ability to stop a running experiment from Python API
  • Added license information

Release 3.0.1

Release date: December 10, 2019

Release 3.0.0

Release date: November 3, 2019

  • Released REST v2 read and write endpoints
  • Released updated Python API and updated APIExperiment
    • Dramatically increased speed of Python API by using REST v2
    • Added many additional Python API methods (update projects, delete projects, restore experiments, etc)
    • Standardized Python API JSON field names to camelCaseNaming
    • Added much more documentation and examples for Python API
    • Brought APIExperiment inline with Experiment, OfflineExperiment, and ExistingExperiment. Examples:
      • Create an experiment: APIExperiment(workspace=WORKSPACE", project_name=PROJECT_NAME)
      • Use an existing experiment: APIExperiment(previous_experiment=EXPERIMENT_ID | NAME)
    • Added ability for all aspects of an experiment to be downloaded and changed from the Python API
    • New and updated API methods:
      • API.clear_cache()
      • API.create_project(WORKSPACE, PROJECT_NAME, [PROJECT_DESCRIPTION, PUBLIC])
      • API.delete_project(workspace=WORKSPACE, project_name=PROJECT_NAME, delete_experiments=False)
      • API.delete_project(workspace=WORKSPACE, project_id=PROJECT_ID, delete_experiments=False)
      • API.get_project(WORKSPACE, PROJECT_NAME)
      • API.get_project_by_id(PROJECT_ID)
      • API.restore_experiment(EXPERIMENT_ID)
      • API.update_project(WORKSPACE, PROJECT_NAME, [NEW_PROJECT_NAME, DESCRIPTION, PUBLIC])
      • API.update_project_by_id(PROJECT_ID, [NEW_PROJECT_NAME, DESCRIPTION, PUBLIC])
      • API.use_cache(True | False)
    • New and updated APIExperiment methods:
      • New and updated read methods:
        • APIExperiment.get_additional_system_info()
        • APIExperiment.get_command()
        • APIExperiment.get_executable()
        • APIExperiment.get_gpu_static_info()
        • APIExperiment.get_hostname()
        • APIExperiment.get_ip()
        • APIExperiment.get_max_memory()
        • APIExperiment.get_network_interface_ips()
        • APIExperiment.get_os()
        • APIExperiment.get_os_type()
        • APIExperiment.get_pid()
        • APIExperiment.get_python_version()
        • APIExperiment.get_python_version_verbose()
        • APIExperiment.get_system_details()
        • APIExperiment.get_system_metric_names()
        • APIExperiment.get_tags()
        • APIExperiment.get_total_memory()
        • APIExperiment.get_user()
      • New and updated write methods:
        • APIExperiment.log_additional_system_info(KEY, VALUE)
        • APIExperiment.log_asset(FILENAME, [STEP, OVERWRITE, CONTEXT, FTYPE, METADATA])
        • APIExperiment.log_cpu_metrics(CPU_METRICS, [CONTEXT, STEP, EPOCH, TIMESTAMP])
        • APIExperiment.log_gpu_metrics(GPU_METRICS)
        • APIExperiment.log_html(HTML, [CLEAR, TIMESTAMP])
        • APIExperiment.log_image(FILENAME, [IMAGE_NAME, STEP, OVERWRITE, CONTEXT])
        • APIExperiment.log_load_metrics(LOAD_AVG, [CONTEXT, STEP, EPOCH, TIMESTAMP])
        • APIExperiment.log_metric(METRIC, VALUE, [STEP, TIMESTAMP])
        • APIExperiment.log_other(KEY, VALUE, [TIMESTAMP])
        • APIExperiment.log_output(LINES, [CONTEXT, STDERR, TIMESTAMP])
        • APIExperiment.log_parameter(PARAMETER, VALUE, [STEP, TIMESTAMP])
        • APIExperiment.log_ram_metrics(TOTAL_RAM, USED_RAM, [CONTEXT, STEP, EPOCH, TIMESTAMP])
        • APIExperiment.set_code(CODE)
        • APIExperiment.set_command(COMMAND_ARGS_LIST)
        • APIExperiment.set_end_time(TIME_MILLISECONDS)
        • APIExperiment.set_executable(EXECUTABLE)
        • APIExperiment.set_git_metadata(USER, ROOT, BRANCH, PARENT, ORIGIN)
        • APIExperiment.set_git_patch(FILE_DATA)
        • APIExperiment.set_gpu_static_info(GPU_STATIC_INFO)
        • APIExperiment.set_hostname(HOSTNAME)
        • APIExperiment.set_installed_packages(INSTALLED_PACKAGES)
        • APIExperiment.set_ip(IP)
        • APIExperiment.set_network_interface_ips(IPS)
        • APIExperiment.set_os(OS)
        • APIExperiment.set_os_type(OS_TYPE)
        • APIExperiment.set_pid(PID)
        • APIExperiment.set_python_version(PYTHON_VERSION)
        • APIExperiment.set_python_version_verbose(PYTHON_VERSION_VERBOSE)
        • APIExperiment.set_start_time(TIME_MILLISECONDS)
        • APIExperiment.set_user(USER)

For more information on the Python API, see:

Release 2.0.18

Release date: November 22, 2019

  • Fixed issue when creating multiple experiments sequentially
  • Added ability for Experiment.log_asset_folder() to log files recursively with file names
  • Better handling of logging files and file-like objects
  • Added ability for users to control auto logging details

Release 2.0.17

Release date: November 7, 2019

  • Fixed a file-descriptor leak with native console logging
  • Fixed Experiment.log_figure() to more accurately check for empty figures
  • Added psutil to conda dependencies

Release 2.0.16

Release date: October 29, 2019

  • Corrected the offline file upload to have the right file extension
  • Fixed logging an empty CPU load avg metric
  • Fixed optional dependency declaration for the cpu logging feature. You can now use pip install comet_ml[cpu_logging] to install psutil.
  • Fixed a bug in OfflineExperiment that removed a file from disk when passing copy_to_tmp=False.
  • Improved file extension handling when uploading file with OfflineExperiment.

Release 2.0.15

Release date: October 16, 2019

  • Added optional timestamp (in seconds) to Python APIExperiment:
    • APIExperiment.log_other(..., timestamp=SECONDS)
    • APIExperiment.log_metric(..., timestamp=SECONDS)
    • APIExperiment.log_parameter(..., timestamp=SECONDS)
    • APIExperiment.log_html(..., timestamp=SECONDS)
    • APIExperiment.log_output(..., timestamp=SECONDS)
  • Sped up Experiment.end() and OfflineExperiment.end()
  • Pinned jsonschema to a Python 3.4 compatible version

Release 2.0.14

Release date: October 2, 2019

  • Refactored Python API
  • Added Query API
  • Added CPU logging
    • requires psutil (see above)
  • Deprecated passing experiment keyword arguments to Optimizer(); now pass using Optimizer.get_experiments(**kwargs)

Release 2.0.13

Release date: September 16, 2019

  • Fixed repeated keras model unable-to-log warnings
  • Fixed wrong context in keras callback Callback.on_test_*() methods; was test now validate
  • Experiment.log_histogram_3d(values, step) gives required-step error earlier

Release 2.0.12

Release date: September 4, 2019

  • Added support for TensorFlow v2 for tensorflow>=1.14, and tensorflow==2.0
  • Fixed a bug that was using auto_param_logging instead of auto_metric_logging to control tensorboard metric logging.
  • Added Experiment.log_histogram_3d() for logging time-series histograms.

Release 2.0.11

Release date: August 29, 2019

  • Redefined valid experiment key (key must be alphanumeric, 32 to 50 characters)

Release 2.0.10

Release date: August 26, 2019

  • Added support for all versions of TensorFlow v1 (1.11 - 2.0)
  • Added Experiment.get_callback("tf-keras")
  • Fixed spelling errors in messages, comments, and code
  • Removed reporting when websocket connection was closed on normal shutdown

Release 2.0.9

Release date: August 21, 2019

  • Fixed GPU schema to report usage better
  • Added API.add_tags(experiment, tags)
  • Fixed offline image upload issue
  • Fixed sklearn to respect auto_param_logging flag
  • Fixed SOCK HTTP Proxy which required an int port number

Release 2.0.8

Release date: August 14, 2019

  • Added {user} to configurable logging file name patterns
  • Added comet offline to explore offline experiment ZIP files
  • Added API.get_experiments() and API.get_experiment()
  • Added -m flag to comet python
  • Added COMET_EXPERIMENT_KEY for new experiments

Release 2.0.7

Release date: August 13, 2019

  • Added COMET_WORKSPACE to the config
  • Added COMET_CONSOLE to set the console log level and display tracebacks
  • Added COMET_LOGGING_FILE name patterns, like "comet-{project}.log"
  • Added Experiment.set_epoch(NUMBER)
  • Added Experiment.log_others({...})
  • Fixed summary metric count
  • Deprecated Experiment.get_keras_callback(); use Experiment.get_callback("keras")

Release 2.0.6

Release date: August 5, 2019

  • Added step parameter to all asset logging methods
  • Added Experiment(disable_summary=True) to disable display summary
  • Added a log message when an optimizer search has completed
  • Added comet python CLI to automatically import Comet first
  • Fixed send_notification compatibility with older backend
  • Added conda packaging support

Release 2.0.5

Release date: July 18, 2019

  • Added step parameter to Experiment.log_audio()
  • Added Experiment.send_notification()

Release 2.0.4

Release date: July 15, 2019

  • Added base64-encoded metadata in the log_audio call

Release 2.0.3

Release date: July 15, 2019

  • Added support for Experiment.log_audio()
  • Added comet_ml.config.save()
  • File upload now sends file extension

Release 2.0.2

Release date: July 2, 2019

  • Added support for logging binary assets with Experiment.log_asset_data()
  • Added more hook-points for tensorboard auto logging

Release 2.0.1

Release date: June 20, 2019

  • Fixed tensorflow context logging for Tensorflow >1.13.1

Release 2.0.0

Release date: June 18, 2019

  • Added a log message when an experiment was created offline
  • Added a log message when an experiment was stopped
  • Improved stop message
  • Removed deprecated methods, this makes this release not backward compatible
  • Added a new Optimizer which breaks the older API
  • Added new tensorflow hooks to set the context properly

Release 1.0.56

Release date: June 5, 2019

  • Updated OfflineExperiment() to accept the same API arguments as Experiment()
Sep. 30, 2024