A command-line tool for BigML.io, the public BigML API
BigMLer - A command-line tool for BigML’s API
BigMLer makes BigML even easier.
BigMLer wraps BigML’s API Python bindings to offer a high-level command-line script to easily create and publish datasets and models, create ensembles, make local predictions from multiple models, and simplify many other machine learning tasks. For additional information, see the full documentation for BigMLer on Read the Docs.
BigMLer is open sourced under the Apache License, Version 2.0.
Please report problems and bugs to our BigML.io issue tracker.
Python 2.7 and 3 are currently supported by BigMLer.
BigMLer requires bigml 4.18.1 or higher. Using proportional missing strategy will additionally request the use of the numpy and scipy libraries. They are not automatically installed as a dependency, as they are quite heavy and exclusively required in this case. Therefore, they have been left for the user to install them if required. The same happens with the pystemmer library, used only for topic modeling. Check the bindings documentation for more info.
To install the latest stable release with pip
$ pip install bigmler
You can also install the development version of bigmler directly from the Git repository
$ pip install -e git://github.com/bigmlcom/bigmler.git#egg=bigmler
For a detailed description of install instructions on Windows see the BigMLer on Windows section.
All the requests to BigML.io must be authenticated using your username and API key and are always transmitted over HTTPS.
BigML module will look for your username and API key in the environment variables BIGML_USERNAME and BIGML_API_KEY respectively. You can add the following lines to your .bashrc or .bash_profile to set those variables automatically when you log in
export BIGML_USERNAME=myusername export BIGML_API_KEY=ae579e7e53fb9abd646a6ff8aa99d4afe83ac291
Otherwise, you can initialize directly when running the BigMLer script as follows
bigmler --train data/iris.csv --username myusername \ --api-key ae579e7e53fb9abd646a6ff8aa99d4afe83ac291
For a detailed description of authentication instructions on Windows see the BigMLer on Windows section.
BigMLer on Windows
To install BigMLer on Windows environments, you’ll need Python for Windows (v.2.7.x) installed.
In addition to that, you’ll need the pip tool to install BigMLer. To install pip, first you need to open your command line window (write cmd in the input field that appears when you click on Start and hit enter), download this python file and execute it
After that, you’ll be able to install pip by typing the following command
And finally, to install BigMLer, just type
c:\Python27\Scripts\pip.exe install bigmler
and BigMLer should be installed in your computer. Then issuing
should show BigMLer version information.
Finally, to start using BigMLer to handle your BigML resources, you need to set your credentials in BigML for authentication. If you want them to be permanently stored in your system, use
setx BIGML_USERNAME myusername setx BIGML_API_KEY ae579e7e53fb9abd646a6ff8aa99d4afe83ac291
BigML Development Mode
Also, you can instruct BigMLer to work in BigML’s Sandbox environment by using the parameter ---dev
bigmler --train data/iris.csv --dev
Using the development flag you can run tasks under 1 MB without spending any of your BigML credits.
To run BigMLer you can use the console script directly. The –help option will describe all the available options
Alternatively you can just call bigmler as follows
python bigmler.py --help
This will display the full list of optional arguments. You can read a brief explanation for each option below.
Let’s see some basic usage examples. Check the installation and authentication sections in BigMLer on Read the Docs if you are not familiar with BigML.
You can create a new model just with
bigmler --train data/iris.csv
If you check your dashboard at BigML, you will see a new source, dataset, and model. Isn’t it magic?
You can generate predictions for a test set using
bigmler --train data/iris.csv --test data/test_iris.csv
You can also specify a file name to save the newly created predictions
bigmler --train data/iris.csv --test data/test_iris.csv --output predictions
If you do not specify the path to an output file, BigMLer will auto-generate one for you under a new directory named after the current date and time (e.g., MonNov1212_174715/predictions.csv). With --prediction-info flag set to brief only the prediction result will be stored (default is normal and includes confidence information).
A different objective field (the field that you want to predict) can be selected using
bigmler --train data/iris.csv \ --test data/test_iris.csv \ --objective 'sepal length'
If you do not explicitly specify an objective field, BigML will default to the last column in your dataset.
Also, if your test file uses a particular field separator for its data, you can tell BigMLer using --test-separator. For example, if your test file uses the tab character as field separator the call should be like
bigmler --train data/iris.csv --test data/test_iris.tsv \ --test-separator '\t'
If you don’t provide a file name for your training source, BigMLer will try to read it from the standard input
cat data/iris.csv | bigmler --train
BigMLer will try to use the locale of the model both to create a new source (if --train flag is used) and to interpret test data. In case it fails, it will try en_US.UTF-8 or English_United States.1252 and a warning message will be printed. If you want to change this behaviour you can specify your preferred locale
bigmler --train data/iris.csv --test data/test_iris.csv \ --locale "English_United States.1252"
If you check your working directory you will see that BigMLer creates a file with the model ids that have been generated (e.g., FriNov0912_223645/models). This file is handy if then you want to use those model ids to generate local predictions. BigMLer also creates a file with the dataset id that has been generated (e.g., TueNov1312_003451/dataset) and another one summarizing the steps taken in the session progress: bigmler_sessions. You can also store a copy of every created or retrieved resource in your output directory (e.g., TueNov1312_003451/model_50c23e5e035d07305a00004f) by setting the flag --store.
Prior Versions Compatibility Issues
BigMLer will accept flags written with underscore as word separator like --clear_logs for compatibility with prior versions. Also --field-names is accepted, although the more complete --field-attributes flag is preferred. --stat_pruning and --no_stat_pruning are discontinued and their effects can be achived by setting the actual --pruning flag to statistical or no-pruning values respectively.
Running the Tests
To run the tests you will need to install nose that is installed on setup, and set up your authentication via environment variables, as explained above. With that in place, you can run the test suite simply by issuing
$ python setup.py nosetests
For additional information, see the full documentation for BigMLer on Read the Docs.
- Fixing bigmler retrain for models based on transformed datasets.
- Fixing bigmler deepnet and logistic-regression that lacked the –batch-prediction-attributes option.
- Adding the –minimum-name-terms option to bigmler topic-model.
- Fixing bigmler retrain and execute-related subcommands for organizations.
- Fixing bigmler retrain when using a model-tag.
- Adding jupyter notebook output format to bigmler reify.
- Fixing bigmler reify subcommand for multidatasets.
- Fixing bigmler deepnet subcommand predictions.
- Adding operating point options to bigmler deepnet.
- Adding model types to bigmler retrain.
- Updating underlying bindings version.
- Updating reify library.
- Improving bigmler retrain to allow remote sources
- Updating underlying bindings version.
- Adapting to new evaluation metrics.
- Adding support for organizations.
- Removing –dev flag: development mode has been deprecated.
- Fixing bug in remote predictions with models and ensembles when –no-batch was used.
- Fixing bug caused by pystemmer not being installed as a bindings dependency.
- Adding the bigmler retrain command to retrain modeling resources with incremental data.
- Adding the –upgrade flag to the bigmler execute and package subcommands to check whether a script is already loaded and its version.
- Adding the –operating-point option for models, ensembles and logistic regressions.
- Extending bigmler export to generate the code for the models in boosted ensembles.
- Extending bigmler export to generate the code for the models in an ensemble.
- Fixing code generation in bigmler export for models with missings.
- Adding bigmler deepnet command to create deepnet models and predictions.
- Adding bigmler timeseries subcommand to create time series models and forecasts.
- Solving issues in cross-validation due to new evaluation formats.
- Improving boosted ensembles local predictions by using new bindings version.
- Fixing bug in bigmler export when non-ascii characters are used in a model.
- Adding bigmler export subcommand to generate the prediction function from a decision tree in several languages.
- Fixing bug: Adapting to changes in the structure of evaluations that caused cross-validation failure.
- Fixing bug: the –package-dir option in bigmler whizzml did not expand the ~ character to its associated user path.
- Fixing bug: Multi-label predictions failed because of changes in the bindings internal coding for combiners.
- Adding –embed-libs and –embedded-libraries to bigmler whizzml and bigmler execute subcommands to embed the libraries’ code in the scripts.
- Adding suport for booted ensembles’ new options.
- Fixing bug in bigmler whizzml when using –username and –api-key.
- Fixing bug in bigmler subcommands when publishing datasets.
- Fixing bug in bigmler: –evaluation-attributes were not used.
- Fixing bug in bigmler: –threshold and –class were not used.
- Fixing bug in bigmler topic-model: adding –topic-model-attributes.
- Adding new bigmler topic-model subcommand.
- Fixing bug in bigmler commands when using samples to create different model types.
- Fixing bug in bigmler commands when using local files storing the model info as input for local predictions.
- Fixing bug in bigmler commands when using local predictions form development mode resources.
- Fixing bug in bigmler package. Libraries where created more than once.
- Fixing bug in bigmler analyze –features when adding batch prediction.
- Improving bigmler delete when deleting projects and executions. Deleting in two steps: first the projects and executions and then the remaining resources.
- Fixing bug in logistic regression evaluation.
- Adding –balance-fields flag to bigmler logistic-regression.
- Refactoring and style changes.
- Adding the logistic regression options to documentation.
- Changing the bias for Logistic Regressions to a boolean.
- Adding the new attributes to control ensemble’s sampling.
- Adding types of deletable resources to bigmler delete. Adding option –execution-only to avoid deleting the output resources of an execution when the execution is deleted.
- Fixing bug: directory structure in bigmler whizzml was wrong when components were found in metadata.
- Upgrading the underlying Python bindings version.
- Adding new bigmler whizzml subcommand to create scripts and libraries from packages with metadata info.
- Adding new –field-codings option to bigmler logisitic-regression subcommand.
- Changing underlying bindings version
- Adding the new bigmler execute subcommand, which can create scripts, executions and libraries.
- Fixing bug: the –predictions-csv flag in the bigmler analyze command did not work with ensembles (–number-of-models > 1)
- Adding the –predictions-csv flag to bigmler analyze –features. It creates a file which contains all the data tagged with the corresponding k-fold and the prediction and confidence values for the best score cross-validation.
- Improving bigmler analyze –features CSV output to reflect the best fields set found at each step.
- Adding the –export-fields and –import-fields to manage field summaries and attribute changes in sources and datasets.
- Adding subcommand bigmler logistic-regression.
- Changing tests to adapt to backend random numbers changes.
- Fixing bug: wrong types had been added to default options in bigmler.ini
- Updating copyright –version notice.
- Adding links to docs and changing tests to adapt bigmler reify to new automatically generated names for resources.
- Fixing bug in bigmler reify subcommand for datasets generated from other datasets comming from batch resources.
- Adding docs for association discovery.
- Adding bigmler association subcommand to manage associations.
- Adding bigmler project subcommand for project creation and update.
- Fixing bug: wrong reify output for datasets created from another dataset.
- Improving bigmler reify code style and making file executable.
- Fixing bug: simplifying bigmler reify output for datasets created from batch resources.
- Allowing column numbers as keys for fields structures in –source-attributes, –dataset-attributes, etc
- Adding –datasets as option for bigmler analyze.
- Adding –summary-fields as option for bigmler analyze.
- Fixing bug: Report title for feature analysis was not shown.
- Upgrading the underlying bindings version.
- Fixing bug: bigmler cluster did not use the –prediction-fields option.
- Adding –status option to bigmler delete. Selects the resources to delete according to their status (finished if not set). You can check the available status in the developers documentation.
- Fixing bug: bigmler reify failed for dataset generated from batch predictions, batch centroids or batch anomaly scores.
- Fixing bug: improving datasets download handling to cope with transmission errors.
- Fixing bug: solving failure when using the first column of a dataset as objective field in models and ensembles.
- Adding new bigmler analyze option, –random-fields to analyze performance of random forests chaging the number of random candidates.
- Fixing bug in reify subcommand for unordered reifications.
- Adding bigmler reify subcommand to script the resource creation.
- Fixing bug: changing the related Python bindings version to solve encoding problem when using Python 3 on Windows.
- Adding bigmler report subcommand to generate reports for cross-validation results in bigmler analyze.
- Fixing bug: bigmler analyze and filtering datasets failed when the origin dataset was a filtered one.
- Fixing bug: bigmler analyze –features could not analyze phi for a user-given category because the metric is called phi_coefficient.
- Modifying the output of bigmler analyze –features and –nodes to include the command to generate the best performing model and the command to clean all the generated resources.
- Fixing bug: dataset generation with a filter on a previous dataset was not working.
- Adding the –project-tag option to bigmler delete.
- Fixing that the –test-dataset and related options can be used in model evaluation.
- Fixing bug: bigmler anomalies for datasets with more than 1000 fields failed.
- Adding the –top-n, –forest-size and –anomalies-dataset to the bigmler anomaly subcommand.
- Fixing bug: source upload failed when using arguments that contain unicodes.
- Fixing bug: bigmler analyze subcommand failed for datasets with more than 1000 fields.
- Supporting Python 3 and changing the test suite to nose.
- Adding –cluster-models option to generate the models related to cluster datasets.
- Adding –score flag to create batch anomaly scores for the training set.
- Allowing –median to be used also in ensembles predictions.
- Using –seed option also in ensembles.
- Adding –median flag to use median instead of mean in single models’ predictions.
- Updating underlying BigML python bindings’ version to 4.0.2 (Python 3 compatible).
- Fixing bug: resuming commands failed retrieving the output directory
- Fixing docs formatting errors.
- Adding –to-dataset and –no-csv flags causing batch predictions, batch centroids and batch anomaly scores to be stored in a new remote dataset and not in a local CSV respectively.
- Adding the sample subcommand to generate samples from datasets
- Fixing bug: using –model-fields with –max-categories failed.
- Fixing bug: Failed field retrieval for batch predictions starting from source or dataset test data.
- Adding the –project and –project-id to manage projects and associate them to newly created sources.
- Adding the –cluster-seed and –anomaly-seed options to choose the seed for deterministic clusters and anomalies.
- Refactoring dataset processing to avoid setting the objective field when possible.
- Adding –optimize-category in bigmler analyze subcommands to select the category whose evaluations will be optimized.
- Fixing bug: k-fold cross-validation failed for ensembles.
- Fixing bug: ensembles’ evaluations failed when using the ensemble id.
- Fixing bug: bigmler analyze lacked model configuration options (weight-field, objective-fields, pruning, model-attributes…)
- Adding k-fold cross-validation for ensembles in bigmler analyze.
- Adding the –model-file, –cluster-file, –anomaly-file and –ensemble-file to produce entirely local predictions.
- Fixing bug: the bigmler delete subcommand was not using the –anomaly-tag, –anomaly-score-tag and –batch-anomaly-score-tag options.
- Fixing bug: the –no-test-header flag was not working.
- Fixing bug: –field-attributes was not working when used in addition to –types option.
- Adding the capability of creating a model/cluster/anomaly and its corresponding batch prediction from a train/test split using –test-split.
- Improving domain transformations for customized private settings.
- Fixing bug: model fields were not correctly set when the origin dataset was a new dataset generated by the –new-fields option.
- Refactoring predictions code, improving some cases performance and memory usage.
- Adding the –fast option to speed prediction by not storing partial results in files.
- Adding the –optimize option to the bigmler analyze –features command.
- Improving perfomance in individual model predictions.
- Forcing garbage collection to lower memory usage in ensemble’s predictions.
- Fixing bug: batch predictions were not adding confidence when –prediction-info full was used.
- Adding bigmler anomaly as new subcommand to generate anomaly detectors, anomaly scores and batch anomaly scores.
- Fixing bug: source updates failed when using –locale and –types flags together.
- Updating bindings version and fixing code accordingly.
- Adding –k option to bigmler cluster to change the number of centroids.
- Fixing bug: –source-attributes and –dataset-attributes where not updated.
- Fixing bug: bigmler analyze was needlessly sampling data to evaluate.
- Adding the new –missing-splits flag to control if missing values are included in tree branches.
- Fixing bug: handling unicode command parameters on Windows.
- Fixing bug: handling stdout writes of unicodes on Windows.
- Fixing but for bigmler analyze: the subcommand failed when used in development created resources.
- Fixing bug when many models are evaluated in k-fold cross-validations. The create evaluation could fail when called with a non-finished model.
- Improving delete process. Promoting delete to a subcommand and filtering the type of resource to be deleted.
- Adding –dry-run option to delete.
- Adding –from-dir option to delete.
- Fixing bug when Gazibit report is used with personalized URL dashboards.
- Adding the –to-csv option to export datasets to a CSV file.
- Adding the –cluster-datasets option to generate the datasets related to the centroids in a cluster.
- Fixing bug for the –delete flag. Cluster, centroids and batch centroids could not be deleted.
- Documentation update.
- Adding cluster subcommand to generate clusters and centroid predictions.
- Fixing bug for the analyze subcommand. The –resume flag crashed when no –ouput-dir was used.
- Fixing bug for the analyze subcommand. The –features flag crashed when many long feature names were used.
- Fixing bug for –delete flag, broken by last fix.
- Fixing bug when field names contain commas and –model-fields tag is used.
- Fixing bug when deleting all resources by tag when ensembles were found.
- Adding –exclude-features flag to analyze.
- Fixing bug when utf8 characters were used in command lines.
- Adding the –balance flag to the analyze subcommand.
- Fixing bug for analyze. Some common flags allowed were not used.
- Fixing bug for analyze. User-given objective field was changed when using filtered datasets.
- Fixing bug for analyze. User-given objective field was not used.
- Docs update and test change to adapt to backend node threshold changes.
- Fixing bug in analyze –nodes. The default node steps could not be found.
- Setting dependency of new python bindings version 1.3.1.
- Fixing bug: –shared and –unshared should be considered only when set in the command line by the user. They were always updated, even when absent.
- Fixing bug: –remote predictions were not working when –model was used as training start point.
- Changing the Gazibit report for shared resources to include the model shared url in embedded format.
- Fixing bug: train and tests data could not be read from stdin.
Adding the analyze subcommand. The subcommand presents new features, such as:
--cross-validation that performs k-fold cross-validation, --features that selects the best features to increase accuracy (or any other evaluation metric) using a smart search algorithm and --nodes that selects the node threshold that ensures best accuracy (or any other evaluation metric) in user defined range of nodes.
- Fixing bug: –no-upload flag was not really used.
- Adding the –reports option to generate Gazibit reports.
- Adding the –shared flag to share the created dataset, model and evaluation.
- Fixing bug for model building, when objective field was specified and no –max-category was present the user given objective was not used.
- Fixing bug: max-category data stored even when –max-category was not used.
- Adding –missing-strategy option to allow different prediction strategies when a missing value is found in a split field. Available for local predictions, batch predictions and evaluations.
- Adding new –delete options: –newer-than and –older-than to delete lists of resources according to their creation date.
- Adding –multi-dataset flag to generate a new dataset from a list of equally structured datasets.
- Bug fixing: resume from multi-label processing from dataset was not working.
- Bug fixing: max parallel resource creation check did not check that all the older tasks ended, only the last of the slot. This caused more tasks than permitted to be sent in parallel.
- Improving multi-label training data uploads by zipping the extended file and transforming booleans from True/False to 1/0.
- Bug fixing: dataset objective field is not updated each time –objective is used, but only if it differs from the existing objective.
- Storing the –max-categories info (its number and the chosen other label) in user_metadata.
- Fix when using the combined method in –max-categories models. The combination function now uses confidence to choose the predicted category.
- Allowing full content text fields to be also used as –max-categories objective fields.
- Fix solving objective issues when its column number is zero.
- Adding the –objective-weights option to point to a CSV file containing the weights assigned to each class.
- Adding the –label-aggregates option to create new aggregate fields on the multi label fields such as count, first or last.
- Fix in local random forests’ predictions. Sometimes the fields used in all the models were not correctly retrieved and some predictions could be erroneus.
- Fix to allow the input data for multi-label predictions to be expanded.
- Fix to retrieve from the models definition info the labels that were given by the user in its creation in multi-label models.
- Adding new –balance option to automatically balance all the classes evenly.
- Adding new –weight-field option to use the field contents as weights for the instances.
- Adding new –source-attributes, –ensemble-attributes, –evaluation-attributes and –batch-prediction-attributes options.
- Refactoring –multi-label resources to include its related info in the user_metadata attribute.
- Refactoring the main routine.
- Adding –batch-prediction-tag for delete operations.
- Fix to transmit –training-separator when creating remote sources.
- Fix for multiple multi-label fields: headers did not match rows contents in some cases.
- Fix for datasets generated using the –new-fields option. The new dataset was not used in model generation.
- Adding –multi-label-fields to provide a comma-separated list of multi-label fields in a file.
- Fix for ensembles’ local predictions when order is used in tie break.
- Fix for duplicated model ids in models file.
- Adding new –node-threshold option to allow node limit in models.
- Adding new –model-attributes option pointing to a JSON file containing model attributes for model creation.
- Fix for missing modules during installation.
- Adding the –max-categories option to handle datasets with a high number of categories.
- Adding the –method combine option to produce predictions with the sets of datasets generated using –max-categories option.
- Fixing problem with –max-categories when the categorical field is not a preferred field of the dataset.
- Changing the –datasets option behaviour: it points to a file where dataset ids are stored, one per line, and now it reads all of them to be used in model and ensemble creation.
- Adding confidence to predictions output in full format
- Bug fixing: multi-label predictions failed when the –ensembles option is used to provide the ensemble information
- Bug fixing: –dataset-price could not be set.
- Adding the threshold combination method to the local ensemble.
- Bug fixing: –model-fields option with absolute field names was not compatible with multi-label classification models.
- Changing resource type checking function.
- Bug fixing: evaluations did not use the given combination method.
- Bug fixing: evaluation of an ensemble had turned into evaluations of its
- Adding pruning to the ensemble creation configuration options
- Changing fields_map column order: previously mapped dataset column number to model column number, now maps model column number to dataset column number.
- Adding evaluations to multi-label models.
- Bug fixing: unicode characters greater than ascii-127 caused crash in multi-label classification
- Adapting to predictions issued by the high performance prediction server and the 0.9.0 version of the python bindings.
- Support for shared models using the same version on python bindings.
- Support for different server names using environment variables.
- Adding ensembles’ predictions for multi-label objective fields
- Bug fixing: in evaluation mode, evaluation for –dataset and –number-of-models > 1 did not select the 20% hold out instances to test the generated ensemble.
- Adding text analysis through the corresponding bindings
- Adding support for multi-label objective fields
- Adding –prediction-headers and –prediction-fields to improve –prediction-info formatting options for the predictions file
- Adding the ability to read –test input data from stdin
- Adding –seed option to generate different splits from a dataset
- Adding –test-separator flag
- Bug fixing: resume crash when remote predictions were not completed
- Bug fixing: Fields object for input data dict building lacked fields
- Bug fixing: test data was repeated in remote prediction function
- Bug fixing: Adding replacement=True as default for ensembles’ creation
- Adding –max-parallel-evaluations flag
- Bug fixing: matching seeds in models and evaluations for cross validation
- Changing –model-fields and –dataset-fields flag to allow adding/removing fields with +/- prefix
- Refactoring local and remote prediction functions
- Adding ‘full data’ option to the –prediction-info flag to join test input data with prediction results in predictions file
- Fixing errors in documentation and adding install for windows info
- Adding new flag to control predictions file information
- Bug fixing: using default sample-rate in ensemble evaluations
- Adding standard deviation to evaluation measures in cross-validation
- Bug fixing: using only-model argument to download fields in models
- Adding delete for ensembles
- Creating ensembles when the number of models is greater than one
- Remote predictions using ensembles
- Adding cross-validation feature
- Using user locale to create new resources in BigML
- Adding –ensemble flag to use ensembles in predictions and evaluations
- Deep refactoring of main resources management
- Fixing bug in batch_predict for no headers test sets
- Fixing bug for wide dataset’s models than need query-string to retrieve all fields
- Fixing bug in test asserts to catch subprocess raise
- Adding default missing tokens to models
- Adding stdin input for –train flag
- Fixing bug when reading descriptions in –field-attributes
- Refactoring to get status from api function
- Adding confidence to combined predictions
- Evaluations management
- console monitoring of process advance
- resume option
- user defaults
- Refactoring to improve readability
- Improved locale management.
- Adds progressive handling for large numbers of models.
- More options in field attributes update feature.
- New flag to combine local existing predictions.
- More methods in local predictions: plurality, confidence weighted.
- New flag for locale settings configuration.
- Filtering only finished resources.
- Fix to ensure windows compatibility.
- Initial release.