- fastai callback uses the default monitor instead of assuming val_loss
- notebook introspections handles error cases and doesn't print stacktrace on failure
- Don't print description warning when setting name
- Fixed dataframe logging error with the keras callback
- Fixed line offsets in logs when resuming runs
- wandb.config casts non-builtins before writing to yaml
- vendored backports.tempfile to address missing package on install
- Added
api.sweepto the python export api for querying sweeps - Added
WANDB_NOTEBOOK_NAMEfor specifying the notebook name in cases we can't infer it - Added
WANDB_HOSTto override hostnames - Store if a run was run within jupyter
- Client now supports stopping runs from the web ui
- Handle floats passed as step to
wandb.log - wandb.config has full unicode support
- sync the main file to wandb if code saving is enabled and it's untracked by git
- XGBoost callback: wandb.xgboost.wandb_callback()
- Fixed plotly charts with large numpy arrays not rendering
wandb dockerworks when nvidia is present- Better error when non string keys are sent to log
- Relaxed pyyaml dependency to fix AMI installs
- Magic works in jupyter notebooks.
- New preview release of auto-dataframes for Keras
- Added input_type and output_type to the Keras callback for simpler config
- public api supports retrieving specific keys and custom xaxis
- WANDB_IGNORE_GLOBS is respected on the final scan of files
- Unified run.id, run.name, and run.notes across all apis
- Handle funky terminal sizes when setting up our psuedo tty
- Fixed Jupyter notebook introspection logic
- run.summary.update() persists changes to the server
- tensorboard syncing is robust to invalid histograms and truncated files
- preview release of magic, calling wandb.init(magic=True) should automatically track config and metrics when possible
- cli now supports local installs of the backend
- fastai callback supports logging example images
- image logging works in Windows
- wandb sync handles tfevents with a single timestep
- fix incorrect command in overview page for running runs
- handle histograms with > 512 bins when streaming tensorboard
- better error message when calling wandb sync on a file instead of a directory
- new helper function for handling hyperparameters in sweeps
wandb.config.user_items() - better mocking for improved testing
- entity is persisted on wandb.run when queried from the server
- tmp files always use the temporary directory to avoid syncing
- raise error if file shrinks while uploading
- images log properly in windows
- upgraded pyyaml requirement to address CVE
- no longer store a history of rows to prevent memory leak
- summary now supports new dataframe format
- WANDB_SILENT environment variable writes all wandb messages to debug.log
- Improved error messages for windows and tensorboard logging
- output.log is uploaded at the end of each run
- metadata, requirements, and patches are uploaded at the beginning of a run
- when not running from a git repository, store the main python file
- added WANDB_DISABLE_CODE to prevent diffing and code saving
- when running in jupyter store the name of the notebook
- auto-login support for colab
- store url to colab notebook
- store the version of this library in config
- store sys.executable in metadata
- fastai callback no longer requires path
- wandb.init now accepts a notes argument
- The cli replaced the message argument with notes and name
- wandb sync handles tensorboard embeddings
- wandb sync correctly handles images in tensorboard
- tf.keras correctly handles single input functional models
- wandb.Api().runs returns an iterator that's reusable
- WANDB_DIR within a hidden directory doesn't prevent syncing
- run.files() iterates over all files
- pytorch recurssion too deep error
- wandb sync accepts an --ignore argument with globs to skip files
- run.summary now has an items() method for iterating over all keys
- Better error messages on access denied
- Better error messages when optional packages aren't installed
- Urls printed to the termial are url-escaped
- Namespaced tensorboard events work with histograms
- Public API now retries on failures and re-uses connection pool
- Catch git errors when remotes aren't pushed to origin
- Moved keras graph collection to on_train_begin to handle unbuilt models
- Handle more cases of not being able to save weights
- Updates to summary after resuming are persisted
- PyTorch histc logging fixed in 0.4.1
- Fixed
wandb synctensorboard import
- wandb.init(tensorboard=True) works with Tensorflow 2 and Eager Execution
- wandb.init(tensorboard=True) now works with tb-nightly and PyTorch
- Automatically log examples with tf.keras by adding missing validation_data
- Socket only binds to localhost for improved security and prevents firewall warnings in OSX
- Added user object to public api for getting the source user
- Added run.display_name to the public api
- Show display name in console output
- Added --tags, --job_group, and --job_type to
wandb run - Added environment variable for minimum time to run before considering crashed
- Added flake8 tests to CI, thanks @cclauss!
- wandb-docker-run accepts image digests
- keras callback works in tensorflow2-alpha0
- keras model graph now puts input layer first
- PyTorch log frequency added for gradients and weights
- PyTorch logging performance enhancements
- wandb.init now accepts a name parameter for naming runs
- wandb.run.name reflects custom display names
- Improvements to nested summary values
- Deprecated wandb.Table.add_row in favor of wandb.Table.add_data
- Initial support for a fast.ai callback thanks to @borisdayma!
- run.get_url resolves the default entity if one wasn't specified
- wandb restore accepts run paths with only slashes
- Fixed PyYaml deprecation warnings
- Added entrypoint shell script to manifest
- Strip newlines from cuda version
- handle case insensitive docker credentials
- fix app_url for private cloud login flow
- don't retry 404's when starting sweep agents
- ensure DNS lookup failures can't prevent startup
- centralized debug logging
- wandb agent waits longer to send a SIGKILL after sending SIGINT
- support for logging docker images with the WANDB_DOCKER env var
- WANDB_DOCKER automatically set when run in kubernetes
- new wandb-docker-run command to automatically set env vars and mount code
- wandb.restore supports launching docker for runs that ran with it
- python packages are now recorded and saved in a requirements.txt file
- cpu_count, gpu_count, gpu, os, and python version stored in wandb-metadata.json
- the export api now supports docker-like paths, i.e. username/project:run_id
- better first time user messages and login info
- Improve error reporting for sweeps
- fixed Jupyter logging, don't change logger level
- fixed resuming in Jupyter
- wandb.init now degrades gracefully if a user hasn't logged in to wandb
- added a force flag to wandb.init to require a machine to be logged in
- Tensorboard and TensorboardX logging is now automatically instrumented when enabled
- added a tensorboard to wandb.init which patches tensorboard for logging
- wandb.save handles now accepts a base path to files in sub directories
- wandb.tensorflow and wandb.tensorboard can now be accessed without directly importing
wandb syncwill now traverse a wandb run directory and sync all runs
- Fixed race where wandb process could hang at the end of a run
- Fix resuming in Jupyter on kernel restart
- wandb.save ensures files are pushed regardless of growth
- Added replace=True keyword to init for auto-resuming
- New run.resumed property that can be used to detect if we're resuming
- New run.step property to use for setting an initial epoch on resuming
- Made Keras callback save the best model as it improves
- Really don't require numpy
- Better error message if wandb.log is called before wandb.init
- Prevent calling wandb.watch multiple times
- Handle datetime attributes in logs / plotly
- Add environment to sweeps
- Enable tagging in the public API and in wandb.init
- New media type wandb.Html for logging arbitrary html
- Add Public api.create_run method for custom integrations
- Added glob support to wandb.save, files save as they're written to
- Added wandb.restore for pulling files on resume
- Added a timeout for generating diffs on large repos
- Fixed edge case where file syncing could hang
- Ensure all file changes are captured before exit
- Handle cases of sys.exit where code isn't passed
- Don't require numpy
- New
wandb synccommand that pushes a local directory to the cloud - Support for syncing tfevents file during training
- Detect when running as TFJob and auto group
- New Kubeflow module with initial helpers for pipelines
- Fixed history / summary bug
- Initial support for AWS SageMaker
hook_torchrenamed towatchwith a deprecation warning- Projects are automatically created if they don't exist
- Additional GPU memory_allocated metric added
- Keras Graph stores edges
- PyTorch graph parsing is more robust
- Fixed PyTorch 0.3 support
- File download API supports WANDB_API_KEY authentication
- Sweeps work with new backend (early release).
- Summary tracks all history metrics unless they're overridden by directly writing to summary.
- Files support in data API.
- Show ongoing media file uploads in final upload progress.
- wandb.Audio supports duration
- Pass username header in filestream API
- New wandb.Audio data type.
- New step keyword argument when logging metrics
- Ability to specify run group and job type when calling wandb.init() or via environment variables. This enables automatic grouping of distributed training runs in the UI
- Ability to override username when using a service account API key
- Handle non-tty environments in Python2
- Handle non-existing git binary
- Fix issue where sometimes the same image was logged twice during a Keras step
- PyTorch
- Added a new
wandb.hook_torchmethod which records the graph and logs gradients & parameters of pytorch models wandb.Imagedetects pytorch tensors and uses torchvision.utils.make_grid to render the image.
- Added a new
wandb restorehandles the case of not being run from within a git repo.
- We now open stdout and stderr in raw mode in Python 2 ensuring tools like bpdb work.
- Catastrophic errors are now reported to Sentry unless WANDB_ERROR_REPORTING is set to false
- Improved error handling and messaging on startup
- The first image when calling wandb.log was not being written, now it is
wandb.logandrun.summarynow remove whitespace from keys
- Vendored prompt_toolkit < 1.0.15 because the latest ipython is pinned > 2.0
- Lazy load wandb.h5 only if
summaryis accessed to improve Data API performance
- Jupyter
- Deprecated
wandb.monitorin favor of automatically starting system metrics after the first wandb.log call - Added new %%wandb jupyter magic method to display live results
- Removed jupyter description iframe
- Deprecated
- The Data API now supports
per_pageandorderoptions to theapi.runsmethod - Initial support for wandb.Table logging
- Initial support for matplotlib logging