Quick Start
Let’s see some basic usage examples. Check the
installation and
authentication
sections below if you are not familiar with BigML.
Basics
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 .bigmler_outputs
directory.
The new directory will be 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). You can also set it to
full
if you prefer the result to be presented as a row with your test
input data followed by the corresponding prediction. To include a headers row
in the prediction file you can set --prediction-header
. For both the
--prediction-info full
and --prediction-info brief
options, if you
want to include a subset of the fields in your test file you can select them by
setting --prediction-fields
to a comma-separated list of them. Then
bigmler --train data/iris.csv --test data/test_iris.csv \
--prediction-info full --prediction-header \
--prediction-fields 'petal length','petal width'
will include in the generated predictions file a headers row
petal length,petal width,species,confidence
and only the values of petal length
and petal width
will be shown
before the objective field prediction species
.
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. You can also use as selector the field column number
instead of the name (when –no-train-header is used, for instance).
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'
The model’s predictions in BigMLer are based on the mean of the distribution
of training values in the predicted node. In case you would like to use the
median instead, you could just add the --median
flag to your command
bigmler --train data/grades.csv --test data/test_grades.csv \
--median
Note that this flag can only be applied to regression models.
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
or you can also read the test info from there
cat data/test_iris.csv | bigmler --train data/iris.csv --test
BigMLer will try to use the locale of the model both to create a new source
(if the --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 the .bigmler_outputs
folder in 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., .bigmler_outputs/TueNov1312_003451/model_50c23e5e035d07305a00004f)
by setting the flag --store
.
Remote Predictions
All the predictions we saw in the previous section are computed locally in
your computer. BigMLer allows you to ask for a remote computation by adding
the --remote
flag. Remote computations are treated as batch computations.
This means that your test data will be loaded in BigML as a regular source and
the corresponding dataset will be created and fed as input data to your
model to generate a remote batch prediction
object. BigMLer will download
the predictions file created as a result of this batch prediction
and
save it to local storage just as it did for local predictions
bigmler --train data/iris.csv --test data/test_iris.csv \
--remote --output my_dir/remote_predictions.csv
This command will create a source, dataset and model for your training data,
a source and dataset for your test data and a batch prediction using the model
and the test dataset. The results will be stored in the
my_dir/remote_predictions.csv
file. If you prefer the result not to be
dowloaded but to be stored as a new dataset remotely, add --no-csv
and
to-dataset
to the command line. This can be specially helpful when
dealing with a high number of scores or when adding to the final result
the original dataset fields with --prediction-info full
, that may result
in a large CSV to be created as output. Other output configurations can be
set by using the --batch-prediction-attributes
option pointing to a JSON
file that contains the desired attributes, like:
{"probabilities": true,
"all_fields": true}
In case you prefer BigMLer to issue
one-by-one remote prediction calls, you can use the --no-batch
flag
bigmler --train data/iris.csv --test data/test_iris.csv \
--remote --no-batch
External Connectors
Data can be uploaded from local and remote public files in BigML as you will
see in the sources section. It can also be extracted
from an external database manager like PostgreSQL, MySQL, Elasticsearch or
SQL Server. An externalconnector
resource can be created in BigML to use it
as data feed.
bigmler connector --host my_data.hostname.com \
--port 1234 \
--engine postgresql \
--user my_username \
--password my_password \
--database my_database \
--output-dir out
This command will generate the externalconnector
and the corresponding
external connector ID will be stored in the external_connector
file of
your out
directory. Using this ID as reference and the query of choice
when creating a source
in BigML, you will be able to connect and upload
data to the platform.
Remote Sources
You can create models using remote sources as well. You just need a valid URL
that points to your data.
BigML recognizes a growing list of schemas (http, https, s3,
azure, odata, etc). For example
bigmler --train https://test:test@static.bigml.com/csv/iris.csv
bigmler --train "s3://bigml-public/csv/iris.csv?access-key=[your-access-key]&secret-key=[your-secret-key]"
bigmler --train azure://csv/diabetes.csv?AccountName=bigmlpublic
bigmler --train odata://api.datamarket.azure.com/www.bcn.cat/BCNOFFERING0005/v1/CARRegistration?$top=100
Also, you can use an existing connector to an external source (see the
external connectors section). The connector
ID and the particular query must be placed in a JSON file:
bigmler --train my_connector.json
where the JSON file should contain the following structure:
{"source": "postgresql",
"externalconnector_id": "51901f4337203f3a9a000215",
"query": "select * from my_table"}
Can you imagine how powerful this feature is? You can create predictive
models for huge
amounts of data without using you local CPU, memory, disk or bandwidth.
Welcome to the cloud!!!
Composite Sources
A Composite Source is an arbitrary collection of other BigML Sources.
The Sources in a composite are called components.
When all the components have the same fields,
the composite itself will inherit those fields, and you will be able to
create a dataset from it: the result will just be the concatenation of all
the rows extracted from each component source inside the composite.
You could put together a list of CSV sources, or maybe a couple of CSV files
and an ARFF file with the same exact fields, and the resulting composite
will inherit those fields and behave like a single source for all practical
purposes.
As any other source, a (possibly empty) composite is created open, meaning
that you can modify it. In the case of composites, modifying it means
performing one of the following operations:
bigmler source --source source/4f603fe203ce89bb2d000000 \
--add-sources source/4f603fe203ce89bb2d000001,source/4f603fe203ce89bb2d000002 \
--output-dir final-composite
bigmler source --source source/4f603fe203ce89bb2d000000 \
--remove-sources source/4f603fe203ce89bb2d000001,source/4f603fe203ce89bb2d000002 \
--output-dir final-composite
bigmler source --source source/4f603fe203ce89bb2d000000 \
--replace-sources source/4f603fe203ce89bb2d000001,source/4f603fe203ce89bb2d000002 \
--output-dir final-composite
A source can belong to as many composites as you wish,
and composites can be nested, with the only limitation that a composite
can only be a component if it’s closed (non-editable).
When a source belongs to one or more composites, it cannot be modified,
regardless of whether it’s open or closed. That way all composites see the same
version of the source all the time.
As you add or remove components to a composite, it will check the
compatibility of the fields of all its components, and update its own set
of fields. Thus, adding and removing sources to a composite is in this sense
analogous to changing the parsing specification of, say, a CSV, in the sense
that that is also an operation that can potentially change the collection
of fields (and even the number of rows) extracted to the CSV.
Once you have finished adding components to a composite and want to use it
to create datasets, you must close it. When you close a composite, all its
components will be automatically closed for you.
Unlike all other kinds of source, composites created this way must be
explicitly closed by an API call or UI action in order to create a dataset.
That is mainly to avoid accidentally closing a composite that is being worked
on by several collaborators, or by mistake. Since composites can have a huge
number of components and closing them also closes all of them, it may be
relatively slow.
As an alternative to combining pre-existing sources into a composite,
one can also upload a zip or tar file containing more than one file.
BigML will then automatically create one source for each file inside
the archive, and put them all together in a composite source.
Annotated images as Composite Sources
BigML allows to use images too to build your Machine Learning models.
In order to use images in BigML, each image file needs to be uploaded and
transformed in a Source object, and the collection of images that will become
your training data is handled in BigML as a collection of Sources. However,
this collection of sources is in turn a Source (to be precise, a
Composite Source
). Each row in a Composite Source can contain one or more
images, but it can also contain other fields related to those images,
like labels, used in classification, or regions, used in object detection.
When storing images in a repository, is common practice to keep them
in directories or compressed files. The related fields, like labels or regions,
are usually stored as additional files where some attribute points to the image
they refer to. In BigML Composite Sources, though,
images and annotations can be consolidated as different fields
of the composite source, so that every row of data in the composite source
contains the source created by uploading the related image plus the
annotation fields associated to it.
As there’s not a single standard procedure to create and store these image and
annotation files, BigMLer tries to give options that encompass most of
the usual scenarios. We’ll see some examples using the specific
bigmler source
subcommand.
First scenario: We only need to upload images and they are already stored
in a single compressed file.
bigmler source --train my_images.zip --output-dir output
In this case, the my_images.zip
is uploaded and a new composite source
is created containing the images.
Second scenario: Images are stored in a directory.
bigmler source --train ./my_images_directory --output-dir output
The BigMLer command creates a local compressed file that contains the
images stored in the directory given as a --train
option. The compressed
file is stored in the output
directory and then is uploaded to BigML,
resulting in a composite source
.
Third scenario: The images are stored in a directory and they have associated
annotations which have been stored in an annotations JSON file.
bigmler source --train ./my_images_directory \
--annotations-file annotations.json \
--output-dir output
BigML uses a BigML-COCO syntax to provide labels associated to
images. The annotations file should contain a list of dictionaries and
each dictionary corresponds to one of the images. The reference to the
annotated image is provided in the file
attribute.
[{"file": "my_images/image1.jpg",
"label": "label1"}.
{"file": "my_images/image2.jpg",
"label": "label1"},
{"file": "my_images/image3.jpg",
"label": "label2"}]
In this case, the previous bigmler source
command will zip the images
contained in the my_images_directory
, upload them and create the
corresponding composite source, and finally add a new field named label
to the composite source where the labels provided in the annotations.json
file will be updated.
These are the basic scenarios, but other annotations syntaxes, like VOC
or
YOLO
files are also accepted. As in this case, the annotations are
provided separately, in one file per image, you would need to
provide the directory where these files are stored and
the annotations language as options:
bigmler source --train ./my_images_directory \
--annotations-dir ./annotations_directory \
--annotations-language VOC
--output-dir output
The created composite sources are editable up until you close them
explicitly or you create a dataset from them. While editable, more annotations
can be added to an existing source. For instance, to add annotations
to the source generated in the third scenario,
source/61373ea6520f903f48000001
, we could use:
bigmler source --source source/61373ea6520f903f48000001 \
--images-file my_images.zip \
--annotations-file new_annotations.json \
--output-dir output
Ensembles
You can also easily create ensembles. For example, using
bagging is as easy as
bigmler --train data/iris.csv --test data/test_iris.csv \
--number-of-models 10 --sample-rate 0.75 --replacement \
--tag my_ensemble
To create a
random decision forest
just use the –randomize option
bigmler --train data/iris.csv --test data/test_iris.csv \
--number-of-models 10 --sample-rate 0.75 --replacement \
--tag my_random_forest --randomize
The fields to choose from will be randomized at each split creating a random
decision forest that when used together will increase the prediction
performance of the individual models.
To create a boosted trees’ ensemble use the –boosting option
bigmler --train data/iris.csv --test data/test_iris.csv \
--boosting --tag my_boosted_trees
or add the ``–boosting-iterations` limit
bigmler --train data/iris.csv --test data/test_iris.csv \
--booting-iterations 10 --sample-rate 0.75 --replacement \
--tag my_boosted_trees
Once you have an existing ensemble, you can use it to predict.
You can do so with the command
bigmler --ensemble ensemble/51901f4337203f3a9a000215 \
--test data/test_iris.csv
Or if you want to evaluate it
bigmler --ensemble ensemble/51901f4337203f3a9a000215 \
--test data/iris.csv --evaluate
There are some more advanced options that can help you build local predictions
with your ensembles.
When the number of local models becomes quite large holding all the models in
memory may exhaust your resources. To avoid this problem you can use the
--max_batch_models
flag which controls how many local models are held
in memory at the same time
bigmler --train data/iris.csv --test data/test_iris.csv \
--number-of-models 10 --sample-rate 0.75 --max-batch-models 5
The predictions generated when using this option will be stored in a file per
model and named after the
models’ id (e.g. model_50c23e5e035d07305a00004f__predictions.csv”). Each line
contains the prediction, its confidence, the node’s distribution and the node’s
total number of instances. The default value for ``max-batch-models` is 10.
When using ensembles, model’s predictions are combined to issue a final
prediction. There are several different methods to build the combination.
You can choose plurality
, confidence weighted
, probability weighted
or threshold
using the --method
flag
bigmler --train data/iris.csv --test data/test_iris.csv \
--number-of-models 10 --sample-rate 0.75 \
--method "confidence weighted"
For classification ensembles, the combination is made by majority vote:
plurality
weights each model’s prediction as one vote,
confidence weighted
uses confidences as weight for the prediction,
probability weighted
uses the probability of the class in the distribution
of classes in the node as weight, and threshold
uses an integer number
as threshold and a class name to issue the prediction: if the votes for
the chosen class reach the threshold value, then the class is predicted
and plurality for the rest of predictions is used otherwise
bigmler --train data/iris.csv --test data/test_iris.csv \
--number-of-models 10 --sample-rate 0.75 \
--method threshold --threshold 4 --class 'Iris-setosa'
For regression ensembles, the predicted values are averaged: plurality
again weights each predicted value as one,
confidence weighted
weights each prediction according to the associated
error and probability weighted
gives the same results as plurality
.
As in the model’s case, you can base your prediction on the median of the
predicted node’s distribution by adding --median
to your BigMLer command.
It is also possible to enlarge the number of models that build your prediction
gradually. You can build more than one ensemble for the same test data and
combine the votes of all of them by using the flag combine_votes
followed by the comma separated list of directories where predictions are
stored. For instance
bigmler --train data/iris.csv --test data/test_iris.csv \
--number-of-models 20 --sample-rate 0.75 \
--output ./dir1/predictions.csv
bigmler --dataset dataset/50c23e5e035d07305a000056 \
--test data/test_iris.csv --number-of-models 20 \
--sample-rate 0.75 --output ./dir2/predictions.csv
bigmler --combine-votes ./dir1,./dir2
would generate a set of 20 prediction files, one for each model, in ./dir1
,
a similar set in ./dir2
and combine all of them to generate the final
prediction.
Making your Dataset and Model public or share it privately
Creating a model and making it public in BigML’s gallery is as easy as
bigmler --train data/iris.csv --white-box
If you just want to share it as a black-box model just use
bigmler --train data/iris.csv --black-box
If you also want to make public your dataset
bigmler --train data/iris.csv --public-dataset
You can also share your datasets, models and evaluations privately with
whomever you choose by generating a private link. The --shared
flag will
create such a link
bigmler --dataset dataset/534487ef37203f0d6b000894 --shared --no-model
and the link will be listed in the output of the command
bigmler --dataset dataset/534487ef37203f0d6b000894 --shared --no-model
[2014-04-18 09:29:27] Retrieving dataset. https://bigml.com/dashboard/dataset/534487ef37203f0d6b000894
[2014-04-18 09:29:30] Updating dataset. https://bigml.com/dashboard/dataset/534487ef37203f0d6b000894
[2014-04-18 09:29:30] Shared dataset link. https://bigml.com/shared/dataset/8VPwG7Ny39g1mXBRD1sKQLuHrqE
or can also be found in the information pannel for the resource through the
web interface.
Content
Before making your model public, probably you want to add a name, a category,
a description, and tags to your resources. This is easy too. For example
bigmler --train data/iris.csv --name "My model" --category 6 \
--description data/description.txt --tag iris --tag my_tag
Please note:
You can get a full list of BigML category codes here.
Descriptions are provided in a text file that can also include markdown.
Many tags can be added to the same resource.
Use --no_tag
if you do not want default BigMLer tags to be added.
BigMLer will add the name, category, description, and tags to all the
newly created resources in each request.
Projects
Each resource created in BigML can be associated to a project
. Projects are
intended for organizational purposes, and BigMLer can create projects
each time a source
is created using a --project
option. For instance
bigmler --train data/iris.csv --project "my new project"
will first check for the existence of a project by that name. If it exists,
will associate the source, dataset and model resources to this project.
If it doesn’t, a new project
is created and then associated.
You can also associate resources to any project
in your account
by specifying the option --project-id
followed by its id
bigmler --train data/iris.csv --project-id project/524487ef37203f0d6b000894
Note: Once a source
has been associated to a project
, all the resources
derived from this source
will be automatically associated to the same
project
.
You can also create projects or update their properties by using the bigmler
project subcommand. In particular, when projects need
to be created in an organization
, the --organization
option has to
be added to inform about the ID of the organization where the project should
be created:
bigmler project --organization organization/524487ef37203f0d6b000594 \
--name "my new project"
Only allowed users can create projects in organizations
. If you are not the
owner or an administrator, please check your permissions with them first.
You can learn more about organizations at the
API documentation.
You can also create resources in an organization’s project if your user
has the right privileges. In order to do that, you should add the
--org-project
option followed by the organization’s project ID.
bigmler --train data/iris.csv \
--org-project project/524487ef37203f0d6b000894
Using the existing resources in BigML
You don’t need to create a model from scratch every time that you use BigMLer.
You can generate predictions for a test set using a previously generated
model, cluster, etc. The example shows how you would do that for a tree model:
bigmler --model model/50a1f43deabcb404d3000079 --test data/test_iris.csv
You can also use a number of models providing a file with a model/id per line
bigmler --models TueDec0412_174148/models --test data/test_iris.csv
Or all the models that were tagged with a specific tag
bigmler --model-tag my_tag --test data/test_iris.csv
The same can be extended to any other subcomamnd, like bigmler cluster
using the correct option (--cluster cluster/50a1f43deabcb404d3000da2
,
--clusters TueDec0412_174148/clusters
and cluster-tag my_tag
).
Please, check each subcommand available options for details.
You can also use a previously generated dataset to create a new model
bigmler --dataset dataset/50a1f441035d0706d9000371
You can also input the dataset from a file
bigmler --datasets iris_dataset
A previously generated source can also be used to generate a new
dataset and model
bigmler --source source/50a1e520eabcb404cd0000d1
And test sources and datasets can also be referenced by id in new
BigMLer requests for remote predictions
bigmler --model model/52af53a437203f1cfe0001f0 --remote \
--test-source source/52b0cbe637203f1d3e0015db
bigmler --model model/52af53a437203f1cfe0001f0 --remote \
--test-dataset dataset/52b0fb5637203f5c4f000018
Evaluations
BigMLer can also help you to measure the performance of your supervised
models (decision trees, ensembles, deepnets, linear regressions
and logistic regressions). The
simplest way to build a model and evaluate it all at once is
bigmler --train data/iris.csv --evaluate
which will build the source, dataset and model objects for you using 80% of
the data in your training file chosen at random. After that, the remaining 20%
of the data will be run through the model to obtain
the corresponding evaluation.
The same procedure is available for ensembles:
bigmler --train data/iris.csv --number-of-models 10 --evaluate
for deepnets
bigmler deepnet --train data/iris.csv --evaluate
for linear regressions
bigmler linear-regression --train data/iris.csv --evaluate
and for logistic regressions:
bigmler logistic-regression --train data/iris.csv --evaluate
You can use the same procedure with a previously
existing source or dataset
bigmler --source source/50a1e520eabcb404cd0000d1 --evaluate
bigmler --dataset dataset/50a1f441035d0706d9000371 --evaluate
The results of an evaluation are stored both in txt and json files. Its
contents will follow the description given in the
Developers guide, evaluation section
and vary depending on the model being a classification or regression one.
Finally, you can also evaluate a preexisting model using a separate set of
data stored in a file or a previous dataset
bigmler --model model/50a1f43deabcb404d3000079 --test data/iris.csv \
--evaluate
bigmler --model model/50a1f43deabcb404d3000079 \
--test-dataset dataset/50a1f441035d0706d9000371 --evaluate
As for predictions, you can specify a particular file name to store the
evaluation in
bigmler --train data/iris.csv --evaluate --output my_dir/evaluation
Cross-validation
If you need cross-validation techniques to ponder which parameters (like
the ones related to different kinds of pruning) can improve the quality of your
models, you can use the --cross-validation-rate
flag to settle the
part of your training data that will be separated for cross validation. BigMLer
will use a Monte-Carlo cross-validation variant, building 2*n
different
models, each of which is constructed by a subset of the training data,
holding out randomly n%
of the instances. The held-out data will then be
used to evaluate the corresponding model. For instance, both
bigmler --train data/iris.csv --cross-validation-rate 0.02
bigmler --dataset dataset/519029ae37203f3a9a0002bf \
--cross-validation-rate 0.02
will hold out 2% of the training data to evaluate a model built upon the
remaining 98%. The evaluations will be averaged and the result saved
in json and human-readable formats in cross-validation.json
and
cross-validation.txt
respectively. Of course, in this kind of
cross-validation you can choose the number of evaluations yourself by
setting the --number-of-evaluations
flag. You should just keep in mind
that it must be high enough to ensure low variance, for instance
bigmler --train data/iris.csv --cross-validation-rate 0.1 \
--number-of-evaluations 20
The --max-parallel-evaluations
flag will help you limit the number of
parallel evaluation creation calls.
bigmler --train data/iris.csv --cross-validation-rate 0.1 \
--number-of-evaluations 20 --max-parallel-evaluations 2
Configuring Datasets and Models
What if your raw data isn’t necessarily in the format that BigML expects? So we
have good news: you can use a number of options to configure your sources,
datasets, and models.
Most resources in BigML contain information about the fields used in the
resource construction. Sources contain information about the name, label,
description and type of the fields detected in the data you upload.
In addition to that, datasets contain the information of the values that
each field contains, whether they have missing values or errors and even
if they are preferred
fields or non-preferred (fields that are not expected
to convey real information to the model, like user IDs or constant fields).
This information is available in the “fields” attribute of each resource,
but BigMLer can extract it and build a CSV file with a summary of it.
bigmler --source source/50a1f43deabcb404d3010079 \
--export-fields fields_summary.csv \
--output-dir summary
By using this command, BigMLer will create a fields_summary.csv
file
in a summary
output directory. The file will contain a headers row and
the fields information available in the source, namely the field column,
field ID, field name, field label and field description of each field. If you
execute the same command on a dataset
bigmler --dataset dataset/50a1f43deabcb404d3010079 \
--export-fields fields_summary.csv \
--output-dir summary
you will also see the number of missing values and errors found in each field
and an excerpt of the values and errors.
But then, imagine that you want to alter BigML’s default field names
or the ones provided
by the training set header and capitalize them, even to add a label or a
description to each field. You can use several methods. Write a text file
with a change per line as
follows
bigmler --train data/iris.csv --field-attributes fields.csv
where fields.csv
would be
0,'SEPAL LENGTH','label for SEPAL LENGTH','description for SEPAL LENGTH'
1,'SEPAL WIDTH','label for SEPAL WIDTH','description for SEPAL WIDTH'
2,'PETAL LENGTH','label for PETAL LENGTH','description for PETAL LENGTH'
3,'PETAL WIDTH','label for PETAL WIDTH','description for PETAL WIDTH'
4,'SPECIES','label for SPECIES','description for SPECIES'
The number on the left in each line is the column number of the field in your
source and is followed by the new field’s name, label and description.
Similarly you can also alter the auto-detect type behavior from BigML assigning
specific types to specific fields
bigmler --train data/iris.csv --types types.txt
where types.txt
would be
0, 'numeric'
1, 'numeric'
2, 'numeric'
3, 'numeric'
4, 'categorical'
Finally, the same summary file that could be built with the --export-fields
option can be used to modify the updatable information in sources
and datasets. Just edit the CSV file with your favourite editor setting
the new values for the fields and use:
bigmler --source source/50a1f43deabcb404d3010079 \
--import-fields summary/fields_summary.csv
to update the names, labels, descriptions or types of the fields with the ones
in the summary/fields_summary.csv
file.
You could
also use this option to change the preferred
attributes for each
of the fields. This transformation is made at the dataset level,
so in the prior code it will be applied once a dataset is created from
the referred source. You might as well act
on an existing dataset:
bigmler --dataset dataset/50a1f43deabcb404d3010079 \
--import-fields summary/fields_summary.csv
In order to update more detailed
source options, you can use the --source-attributes
option pointing
to a file path that contains the configuration settings to be modified
in JSON format
bigmler --source source/52b8a12037203f48bc00000a \
--source-attributes my_dir/attributes.json --no-dataset
Let’s say this source has a text field with id 000001
. The
attributes.json
to change its text parsing mode to full field contents
would read
{"fields": {"000001": {"term_analysis": {"token_mode": "full_terms_only"}}}}
you can also reference the fields by its column number in this JSON structures.
If the field to be modified is in the second column (column index starts at 0)
then the contents of the attributes.json
file could be as well
{"fields": {"1": {"term_analysis": {"token_mode": "full_terms_only"}}}}
The source-attributes
JSON can contain any of the updatable attributes
described in the
developers section
You can specify the fields that you want to include in the dataset by naming
them explicitly
bigmler --train data/iris.csv \
--dataset-fields 'sepal length','sepal width','species'
or the fields that you want to include as predictors in the model
bigmler --train data/iris.csv --model-fields 'sepal length','sepal width'
You can also specify the chosen fields by adding or removing the ones you
choose to the list of preferred fields of the previous resource. Just prefix
their names with +
or -
respectively. For example,
you could create a model from an existing dataset using all their fields but
the sepal length
by saying
bigmler --dataset dataset/50a1f441035d0706d9000371 \
--model-fields -'sepal length'
When evaluating, you can map the fields of the evaluated model to those of
the test dataset by writing in a file the field column of the model and
the field column of the dataset separated by a comma and using –fields-map
flag to specify the name of the file
bigmler --dataset dataset/50a1f441035d0706d9000371 \
--model model/50a1f43deabcb404d3000079 --evaluate \
--fields-map fields_map.txt
where fields_map.txt
would contain
if the first two fields had been reversed.
Finally, you can also tell BigML whether your training and test set come with a
header row or not. For example, if both come without header
bigmler --train data/iris_nh.csv --test data/test_iris_nh.csv \
--no-train-header --no-test-header
Splitting Datasets
When following the usual proceedings to evaluate your models you’ll need to
separate the available data in two sets: the training set and the test set. With
BigMLer you won’t need to create two separate physical files. Instead, you
can set a --test-split
flag that will set the percentage of data used to
build the test set and leave the rest for training. For instance
bigmler --train data/iris.csv --test-split 0.2 --name iris --evaluate
will build a source with your entire file contents, create the corresponding
dataset and split it in two: a test dataset with 20% of instances and a
training dataset with the remaining 80%. Then, a model will be created based on
the training set data and evaluated using the test set. By default, split is
deterministic, so that every time you issue the same command will get the
same split datasets. If you want to generate
different splits from a unique dataset you can set the --seed
option to a
different string in every call
bigmler --train data/iris.csv --test-split 0.2 --name iris \
--seed my_random_string_382734627364 --evaluate
Advanced Dataset management
As you can find in the BigML’s API documentation on
datasets besides the basic name,
label and description that we discussed in previous sections, there are many
more configurable options in a dataset resource.
As an example, to publish a dataset in the
gallery and set its price you could use
{"private": false, "price": 120.4}
Similarly, you might want to add fields to your existing dataset by combining
some of its fields or simply tagging their rows. Using BigMLer, you can set the
--new-fields
option to a file path that contains a JSON structure that
describes the fields you want to select or exclude from the original dataset,
or the ones you want to combine and
the Flatline expression to
combine them. This structure
must follow the rules of a specific languange described in the Transformations
item of the developers
section
bigmler --dataset dataset/52b8a12037203f48bc00000a \
--new-fields my_dir/generators.json
To see a simple example, should you want to include all the fields but the
one with id 000001
and add a new one with a label depending on whether
the value of the field sepal length
is smaller than 1,
you would write in generators.json
{"all_but": ["000001"], "new_fields": [{"name": "new_field", "field": "(if (< (f \"sepal length\") 1) \"small\" \"big\")"}]}
Or, as another example, to tag the outliers of the same field one coud use
{"new_fields": [{"name": "outlier?", "field": "(if (within-percentiles? \"sepal length\" 0.5 0.95) \"normal\" \"outlier\")"}]}
You can also export the contents of a generated dataset by using the
--to-csv
option. Thus,
bigmler --dataset dataset/52b8a12037203f48bc00000a \
--to-csv my_dataset.csv --no-model
will create a CSV file named my_dataset.csv
in the default directory
created by BigMLer to place the command output files. If no file name is given,
the file will be named after the dataset id.
A dataset can also be generated as the union of several datasets using the
flag --multi-dataset
. The datasets will be read from a file specified
in the --datasets
option and the file must contain one dataset id per line.
bigmler --datasets my_datasets --multi-dataset --no-model
This syntax is used when all the datasets in the my_datasets
file share
a common field structre, so the correspondence of the fields of all the
datasets is straight forward. In the general case, the multi-dataset will
inherit the field structure of the first component dataset.
If you want to build a multi-dataset with
datasets whose fields share not the same column disposition, you can specify
which fields are correlated to the ones of the first dataset
by mapping the fields of the rest of datasets to them.
The option --multi-dataset-attributes
can point to a JSON
file that contains such a map. The command line syntax would then be
bigmler --datasets my_datasets --multi-dataset \
--multi-dataset-attributes my_fields_map.json \
--no-model
and for a simple case where the second dataset had flipped the first and second
fields with respect to the first one, the file would read
{"fields_maps": {"dataset/53330bce37203f222e00004b": {"000000": "000001",
"000001": "000000"}}
}
where dataset/53330bce37203f222e00004b
would be the id of the
second dataset in the multi-dataset.
Model Weights
To deal with imbalanced datasets, BigMLer offers three options: --balance
,
--weight-field
and --objective-weights
.
For classification models, the --balance
flag will cause all the classes
in the dataset to
contribute evenly. A weight will be assigned automatically to each
instance. This weight is
inversely proportional to the number of instances in the class it belongs to,
in order to ensure even distribution for the classes.
You can also use a field in the dataset that contains the weight you would like
to use for each instance. Using the --weight-field
option followed by
the field name or column number will cause BigMLer to use its data as instance
weight. This is valid for both regression and classification models.
The --objective-weights
option is used in classification models to
transmit to BigMLer what weight is assigned to each class. The option accepts
a path to a CSV file that should contain the class
,``weight`` values one
per row
bigmler --dataset dataset/52b8a12037203f48bc00000a \
--objective-weights my_weights.csv
where the my_weights.csv
file could read
Iris-setosa,5
Iris-versicolor,3
so that BigMLer would associate a weight of 5
to the Iris-setosa
class and 3
to the Iris-versicolor
class. For additional classes
in the model, like Iris-virginica
in the previous example,
weight 1
is used as default. All specified weights must be non-negative
numbers (with either integer or real values) and at least one of them must
be non-zero.
Predictions’ missing strategy
Sometimes the available data lacks some of the features our models use to
predict. In these occasions, BigML offers two different ways of handling
input data with missing values, that is to say, the missing strategy. When the
path to the prediction reaches a split point that checks
the value of a field which is missing in your input data, using the
last prediction
strategy the final prediction will be the prediction for
the last node in the path before that point, and using the proportional
strategy it will be a weighted average of all the predictions for the final
nodes reached considering that both branches of the split are possible.
BigMLer adds the --missing-strategy
option, that can be set either to
last
or proportional
to choose the behavior in such cases. Last
prediction is the one used when this option is not used.
bigmler --model model/52b8a12037203f48bc00001a \
--missing-strategy proportional --test my_test.csv
Models with missing splits
Another configuration argument that can change models when
the training data has instances with missing values in some of its features
is --missing-splits
. By setting this flag, the model building algorithm
will be able to include the instances
that have missing values for the field used to split the data in each node
in one of the stemming branches. This will, obviously, affect also the
predictions given by the model for input data with missing values. Here’s an
example to build
a model using missing-splits and predict with it.
bigmler --dataset dataset/52b8a12037203f48bc00023b \
--missing-splits --test my_test.csv
Fitering Sources
Imagine that you have create a new source and that you want to create a
specific dataset filtering the rows of the source that only meet certain
criteria. You can do that using a JSON expresion as follows
bigmler --source source/50a2bb64035d0706db0006cc --json-filter filter.json
where filter.json
is a file containg a expression like this
["<", 7.00, ["field", "000000"]]
or a LISP expression as follows
bigmler --source source/50a2bb64035d0706db0006cc --lisp-filter filter.lisp
where filter.lisp
is a file containing a expression like this
(< 7.00 (field "sepal length"))
For more details, see the BigML’s API documentation on
filtering rows.
Multi-labeled categories in training data
Sometimes the information you want to predict is not a single category but a
set of complementary categories. In this case, training data is usually
presented as a row of features and an objective field that contains the
associated set of categories joined by some kind of delimiter. BigMLer can
also handle this scenario.
Let’s say you have a simple file
color,year,sex,class
red,2000,male,"Student,Teenager"
green,1990,female,"Student,Adult"
red,1995,female,"Teenager,Adult"
with information about a group of people and we want to predict the class
another person will fall into. As you can see, each record has more
than one class
per person (for example, the first person is labeled as
being both a Student
and a Teenager
) and they are all stored in the
class
field by concatenating all the applicable labels using ,
as
separator. Each of these labels is, ‘per se’, an objective to be predicted, and
that’s what we can rely on BigMLer to do.
The simplest multi-label command in BigMLer is
bigmler --multi-label --train data/tiny_multilabel.csv
First, it will analyze the training file to extract all the labels
stored
in the objective field. Then, a new extended file will be generated
from it by adding a new field per label. Each generated field will contain
a boolean set to
True
if the associated label is in the objective field and False
otherwise
color,year,sex,class - Adult,class - Student,class - Teenager
red,2000,male,False,True,True
green,1990,female,True,True,False
red,1995,female,True,False,True
This new file will be fed to BigML to build a source
, a dataset
and
a set of models
using four input fields: the first three fields as
input features and one of the label fields as objective. Thus, each
of the classes that label the training set can be predicted independently using
one of the models.
But, naturally, when predicting a multi-labeled field you expect to obtain
all the labels that qualify the input features at once, as you provide them in
the training data records. That’s also what BigMLer does. The syntax to
predict using
multi-labeled training data sets is similar to the single labeled case
bigmler --multi-label --train data/tiny_multilabel.csv \
--test data/tiny_test_multilabel.csv
the main difference being that the ouput file predictions.csv
will have
the following structure
"Adult,Student","0.34237,0.20654"
"Adult,Teenager","0.34237,0.34237"
where the first column contains the class
prediction and the second one the
confidences for each label prediction. If the models predict True
for
more than one label, the prediction is presented as a sequence of labels
(and their corresponding confidences) delimited by ,
.
As you may have noted, BigMLer uses ,
both as default training data fields
separator and as label separator. You can change this behaviour by using the
--training-separator
, --label-separator
and --test-separator
flags
to use different one-character separators
bigmler --multi-label --train data/multilabel.tsv \
--test data/test_multilabel.tsv --training-separator '\t' \
--test-separator '\t' --label-separator ':'
This command would use the tab
character as train and test data field
delimiter and :
as label delimiter (the examples in the tests set use
,
as field delimiter and ‘:’ as label separator).
You can also choose to restrict the prediction to a subset of labels using
the --labels
flag. The flag should be set to a comma-separated list of
labels. Setting this flag can also reduce the processing time for the
training file, because BigMLer will rely on them to produce the extended
version of the training file. Be careful, though, to avoid typos in the labels
in this case, or no objective fields will be created. Following the previous
example
bigmler --multi-label --train data/multilabel.csv \
--test data/test_multilabel.csv --label-separator ':' \
--labels Adult,Student
will limit the predictions to the Adult
and Student
classes, leaving
out the Teenager
classification.
Multi-labeled predictions can also be computed using ensembles, one for each
label. To create an ensemble prediction, use the --number-of-models
option
that will set the number of models in each ensemble
bigmler --multi-label --train data/multilabel.csv \
--number-of-models 20 --label-separator ':' \
--test data/test_multilabel.csv
The ids of the ensembles will be stored in an ensembles
file in the output
directory, and can be used in other predictions by setting the --ensembles
option
bigmler --multi-label --ensembles multilabel/ensembles \
--test data/test_multilabel.csv
or you can retrieve all previously tagged ensembles with --ensemble-tag
bigmler --multi-label --ensemble-tag multilabel \
--test data/test_multilabel.csv
Multi-labeled resources
The resources generated from a multi-labeled training data file can also be
recovered and used to generate more multi-labeled predictions. As in the
single-labeled case
bigmler --multi-label --source source/522521bf37203f412f000100 \
--test data/test_multilabel.csv
would generate a dataset and the corresponding set of models needed to create
a predictions.csv
file that contains the multi-labeled predictions.
Similarly, starting from a previously created multi-labeled dataset
bigmler --multi-label --dataset source/522521bf37203f412fac0135 \
--test data/test_multilabel.csv --output multilabel/predictions.csv
creates a bunch of models, one per label, and predicts storing the results
of each operation in the multilabel
directory, and finally
bigmler --multi-label --models multilabel/models \
--test data/test_multilabel.csv
will retrieve the set of models created in the last example and use them in new
predictions. In addition, for these three cases you can restrict the labels
to predict to a subset of the complete list available in the original objective
field. The --labels
option can be set to a comma-separated list of the
selected labels in order to do so.
The --model-tag
can be used as well to retrieve multi-labeled
models and predict with them
bigmler --multi-label --model-tag my_multilabel \
--test data/test_multilabel.csv
Finally, BigMLer is also able to handle training files with more than one
multi-labeled field. Using the --multi-label-fields
option you can
settle the fields that will be expanded as containing multiple labels
in the generated source and dataset.
bigmler --multi-label --multi-label-fields class,type \
--train data/multilabel_multi.csv --objective class
This command creates a source (and its corresponding dataset)
where both the class
and type
fields have been analysed
to create a new field per label. Then the --objective
option sets class
to be the objective field and only the models needed to predict this field
are created. You could also create a new multi-label prediction for another
multi-label field, type
in this case, by issuing a new BigMLer command
that uses the previously generated dataset as starting point
bigmler --multi-label --dataset dataset/52cafddb035d07269000075b \
--objective type
This would generate the models needed to predict type
. It’s important to
remark that the models used to predict class
in the first example will
use the rest of fields (including type
as well as the ones generated
by expanding it) to build the prediction tree. If you don’t want this
fields to be used in the model construction, you can set the --model-fields
option to exclude them. For instance, if type
has two labels, label1
and label2
, then excluding them from the models that predict
class
could be achieved using
bigmler --multi-label --dataset dataset/52cafddb035d07269000075b \
--objective class
--model-fields=' -type,-type - label1,-type - label2'
You can also generate new fields applying aggregation functions such as
count
, first
or last
on the labels of the multi label fields. The
option --label-aggregates
can be set to a comma-separated list of these
functions and a new column per multi label field and aggregation function
will be added to your source
bigmler --multi-label --train data/multilabel.csv \
--label-separator ':' --label-aggregates count,last \
--objective class
will generate class - count
and class - last
in addition to the set
of per label fields.
Multi-label evaluations
Multi-label predictions are computed using a set of binary models
(or ensembles), one for
each label to predict. Each model can be evaluated to check its
performance. In order to do so, you can mimic the commands explained in the
evaluations
section for the single-label models and ensembles. Starting
from a local CSV file
bigmler --multi-label --train data/multilabel.csv \
--label-separator ":" --evaluate
will build the source, dataset and model objects for you using a
random 80% portion of data in your training file. After that, the remaining 20%
of the data will be run through each of the models to obtain an evaluation of
the corresponding model. BigMLer retrieves all evaluations and saves
them locally in json and txt format. They are named using the objective field
name and the value of the label that they refer to. Finally, it averages the
results obtained in all the evaluations to generate a mean evaluation stored
in the evaluation.txt
and evaluation.json
files. As an example,
if your objective field name is class
and the labels it contains are
Adult,Student
, the generated files will be
Generated files:
- MonNov0413_201326
evaluations
extended_multilabel.csv
source
evaluation_class_student.txt
models
evaluation_class_adult.json
dataset
evaluation.json
evaluation.txt
evaluation_class_student.json
bigmler_sessions
evaluation_class_adult.txt
You can use the same procedure with a previously
existing multi-label source or dataset
bigmler --multi-label --source source/50a1e520eabcb404cd0000d1 \
--evaluate
bigmler --multi-label --dataset dataset/50a1f441035d0706d9000371 \
--evaluate
Finally, you can also evaluate a preexisting set of models or ensembles
using a separate set of
data stored in a file or a previous dataset
bigmler --multi-label --models MonNov0413_201326/models \
--test data/test_multilabel.csv --evaluate
bigmler --multi-label --ensembles MonNov0413_201328/ensembles \
--dataset dataset/50a1f441035d0706d9000371 --evaluate
High number of Categories
In BigML there’s a limit in the number of categories of a categorical
objective field. This limit is set to ensure the quality of the resulting
models. This may become a restriction when dealing with
categorical objective fields with a high number of categories. To cope with
these cases, BigMLer offers the –max-categories option. Setting to a number
lower than the mentioned limit, the existing categories will be organized in
subsets of that size. Then the original dataset will be copied many times, one
per subset, and its objective field will only keep the categories belonging to
each subset plus a generic ***** other *****
category that will summarize
the rest of categories. Then a model will be created from each dataset and
the test data will be run through them to generate partial predictions. The
final prediction will be extracted by choosing the class with highest
confidence from the distributions obtained for
each model’s prediction ignoring the ***** other ******
generic category.
For instance, to use the same iris.csv
example, you could do
bigmler --train data/iris.csv --max-categories 1 \
--test data/test_iris.csv --objective species
This command would generate a source and dataset object, as usual, but then,
as the total number of categories is three and –max-categories is set to 1,
three more datasets will be created, one per each category. After generating
the corresponding models, the test data will be run through them and their
predictions combined to obtain the final predictions file. The same procedure
would be applied if starting from a preexisting source or dataset using the
--source
or --dataset
options. Please note that the --objective
flag is mandatory in this case to ensure that the right categorical field
is selected as objective field.
--method
option accepts a new combine
value to use such kind of
combination. You can use it if you need to create a new group of predictions
based on the same models produced in the first example. Filling the path to the
model ids file
bigmler --models my_dir/models --method combine \
--test data/new_test.csv
the new predictions will be created. Also, you could use the set of datasets
created in the first case as starting point. Their ids are stored in a
dataset_parts
file that can be found in the output location
bigmler --dataset my_dir/dataset_parts --method combine \
--test data/test.csv
This command would cause a new set of models, one per dataset, to be generated
and their predictions would be combined in a final predictions file.
Advanced subcommands in BigMLer
Connector subcommand
Connections to external databases can be used to upload data to BigML. The
bigmler connector
subcommand can be used to create such connections in the
platform. The result will be an externalconnector
object, that can be
reused to perform queries on the database and upload the results to create
the corresponding source
in BigML.
bigmler connector --host my_data.hostname.com \
--port 1234 \
--engine postgresql \
--user my_username \
--password my_password \
--database my_database \
--output-dir out
As you can see, the options needed to create an external connector are:
the host that publishes the database manager
the port that listens to the requests
the type of database manager: PostgreSQL, MySQL, Elasticsearch or
SQL Server.
- the user and password needed to grant the access to the database
With this information, the command will create an externalconnector
object
that will be assigned an ID. This ID will be the reference to be used when
querying the database for new data. Please, check the remote sources section to see an example of that.
Dataset subcommand
In addition to the main BigMLer capabilities explained so far, there’s a
subcommand bigmler dataset
that can be used to create datasets either
from data files and sources or by transforming datasets.
bigmler dataset --file iris.csv \
--output-dir my_directory
will create a source and a dataset by uploading the iris.csv
file to
BigML.
You can also create datasets by applying many transformations to one or
several existing datasets.
To merge datasets, you can use the --merge
option
bigmler dataset --datasets my_datasets/dataset \
--merge \
--output-dir my_directory
The file my_datasets/dataset
should contain dataset IDs, one per line.
The datasets to be merged are expected to share the same fields structure and
their rows will be just added in a single resulting dataset, whose ID will
be stored in a my_directory/dataset_multi
file.
Datasets can also be juxtaposed.
bigmler dataset --datasets my_datasets/dataset \
--juxtapose \
--output-dir my_directory
In this case, the generated dataset ID will be stored in the
my_directory/dataset_gen
file. Each row of the new dataset
will contain all the fields of the datasets found in my_datasets/dataset
.
If you need to join datasets, you can do so by using an SQL expression like:
bigmler dataset --datasets-json "[{\"id\": \"dataset/5357eb2637203f1668000004\", \"id\": \"dataset/5357eb2637203f1668000007\"}]" \
--sql-query "select A.*,B.* from A join B on A.\`000000\` = \`B.000000\`" \
--output-dir my_directory
the --datasets-json
option should contain a JSON string that describes the
datasets to be used in the SQL query. Letters from A
to Z
are used
to refer to these datasets in the SQL expression. First dataset in the list is
represented by A
, the second by B
, etc.
Similarly, the SQL expression can be used to generate an aggregation.
bigmler dataset --dataset dataset/5357eb2637203f1668000004 \
--sql-query "select A.\`species\`, avg(\`petal length\`) as apl from A group by A.\`species\`" \
--output-dir my_directory
or to use for pivoting
bigmler dataset --dataset dataset/5357eb2637203f1668000004 \
--sql-query "select cat_avg(\`petal length\`, \`species\`, 'Iris-setosa') from A group by A.\`petal width\`" \
--output-dir my_directory
that will create the average of the petal length
field value for the rows
whose species
field contains the Iris-setosa
category.
Analyze subcommand
In addition to the main BigMLer capabilities explained so far, there’s a
subcommand bigmler analyze
with more options to evaluate the performance
of your models. For instance
bigmler analyze --dataset dataset/5357eb2637203f1668000004 \
--cross-validation --k-folds 5
will create a k-fold cross-validation by dividing the data in your dataset in
the number of parts given in --k-folds
. Then evaluations are created by
selecting one of the parts to be the test set and using the rest of data
to build the model for testing. The generated
evaluations are placed in your output directory and its average is stored in
evaluation.txt
and evaluation.json
.
Similarly, you’ll be able to create an evaluation for ensembles. Using the
same command above and adding the options to define the ensembles’ properties,
such as --number-of-models
, --sample-rate
, --randomize
or
--replacement
bigmler analyze --dataset dataset/5357eb2637203f1668000004 \
--cross-validation --k-folds 5 --number-of-models 20
--sample-rate 0.8 --replacement
More insights can be drawn from the bigmler analyze --features
command. In
this case, the aim of the command is to analyze the complete set of features
in your dataset to single out the ones that produce models with better
evaluation scores. In this case, we focus on accuracy
for categorical
objective fields and r-squared
for regressions.
bigmler analyze --dataset dataset/5357eb2637203f1668000004 \
--features
This command uses an algorithm for smart feature selection as described in this
blog post
that evaluates models built by using subsets of features. It starts by
building one model per feature, chooses the subset of features used in the
model that scores best and, from there on, repeats the procedure
by adding another of the available features in the dataset to the chosen
subset. The iteration stops when no improvement in score is found for a number
of repetitions that can be controlled using the --staleness
option
(default is 5
). There’s
also a --penalty
option (default is 0.1%
) that sets the amount that
is substracted from the score per feature added to the
subset. This penalty is intended
to mitigate overfitting, but it also favors models which are quicker to build
and evaluate. The evaluations for the scores are k-fold cross-validations.
The --k-folds
value is set to 5
by default, but you can change it
to whatever suits your needs using the --k-folds
option.
bigmler analyze --dataset dataset/5357eb2637203f1668000004 \
--features --k-folds 10 --staleness 3 --penalty 0.002
Would select the best subset of features using 10-fold cross-validation
and a 0.2%
penalty per feature, stopping after 3 non-improving iterations.
Depending on the machine learning problem you intend to tackle, you might
want to optimize other evaluation metric, such as precision
or
recall
. The --optimize
option will allow you to set the evaluation
metric you’d like to optimize.
bigmler analyze --dataset dataset/5357eb2637203f1668000004 \
--features --optimize recall
For categorical models, the evaluation values are obtained by counting
the positive and negative matches for all the instances in
the test set, but sometimes it can be more useful to optimize the
performance of the model for a single category. This can be specially
important in highly non-balanced datasets or when the cost function is
mainly associated to one of the existing classes in the objective field.
Using ``–optimize-category” you can set the category whose evaluation
metrics you’d like to optimize
bigmler analyze --dataset dataset/5357eb2637203f1668000004 \
--features --optimize recall \
--optimize-category Iris-setosa
You should be aware that the smart feature selection command still generates
a high number of BigML resources. Using k
as the k-folds
number and
n
as the number of explored feature sets, it will be generating k
datasets (1/k``th of the instances each), and ``k * n
models and
evaluations. Setting the --max-parallel-models
and
--max-parallel-evaluations
to higher values (up to k
) can help you
speed up partially the creation process because resources will be created
in parallel. You must keep in mind, though, that this parallelization is
limited by the task limit associated to your subscription or account type.
As another optimization method, the bigmler analyze --nodes
subcommand
will find for you the best performing model by changing the number of nodes
in its tree. You provide the --min-nodes
and --max-nodes
that define
the range and --nodes-step
controls the increment in each step. The command
runs a k-fold evaluation (see --k-folds
option) on a model built with each
node threshold in you range and tries to optimize the evaluation metric you
chose (again, default is accuracy
). If improvement stops (see
the –staleness option) or the node threshold reaches the --max-nodes
limit, the process ends and shows the node threshold that
lead to the best score.
bigmler analyze --dataset dataset/5357eb2637203f1668000004 \
--nodes --min-nodes 10 \
--max-nodes 200 --nodes-step 50
When working with random forest, you can also change the number of
random_candidates
or number of fields chosen at random when the models
in the forest are built. Using bigmler analyze --random-fields
the number
of random_candidates
will range from 1 to the number of fields in the
origin dataset, and BigMLer will cross-validate the random forests to determine
which random_candidates
number gives the best performance.
bigmler analyze --dataset dataset/5357eb2637203f1668000004 \
--random-fields
Please note that, in general, the exact choice of fields selected as random
candidates might be more
important than their actual number. However, in some marginal cases (e.g.
datasets with a high number noise features) the number of random candidates
can impact tree performance significantly.
For any of these options (--features
, --nodes
and --random-fields
)
you can add the --predictions-csv
flag to the bigmler analyze
command. The results will then include a CSV file that stores the predictions
obtained in the evaluations that gave the best score. The file content includes
the data in your original dataset tagged by k-fold and the prediction and
confidence obtained. This file will be placed in an internal folder of your
chosen output directory.
bigmler analyze --dataset dataset/5357eb2637203f1668000004 \
--features --output-dir my_features --predictions-csv
The output directory for this command is my_features
and it will
contain all the information about the resources generated when testing
the different feature combinations
organized in subfolders. The k-fold datasets’
IDs will be stored in an inner test
directory. The IDs of the resources
created when testing each combination of features will be stored in
kfold1
, kfold2
, etc. folders inside the test
directory.
If the best-scoring prediction
models are the ones in the kfold4
folder, then the predictions CSV file
will be stored in a new folder named kfold4_pred
.
Report subcommand
The results of a bigmler analyze --features
or bigmler analyze --nodes
command are a series of k-fold cross-validations made on the training data that
leads to the configuration value that will create the best performant model.
However, the algorithm maximizes only one evaluation metric. To see the global
picture for the rest of metrics at each validation configuration you can build
a graphical report of the results using the report
subcommand. Let’s say
you previously ran
bigmler analyze --dataset dataset/5357eb2637203f1668000004 \
--nodes --output-dir best_recall
and you want to have a look at the results for each node_threshold
configuration. Just say:
bigmler report --from-dir best_recall --port 8080
and the command will traverse the directories in best_recall
and summarize
the results found there in a metrics comparison graphic and an ROC curve if
your
model is categorical. Then a simple HTTP server will be started locally and
bound to a port of your choice, 8080
in the example (8085
will be the
default value), and a new web browser
window will be started to show the results.
You can see an example
built on the well known diabetes dataset.
The HTTP server will create an auxiliary bigmler/reports
directory in the
user’s home directory, where symbolic links to the reports in each output
directory will be stored and served from.
Cluster subcommand
Just as the simple bigmler
command can generate all the
resources leading to finding models and predictions for a supervised learning
problem, the bigmler cluster
subcommand will follow the steps to generate
clusters and predict the centroids associated to your test data. To mimic what
we saw in the bigmler
command section, the simplest call is
bigmler cluster --train data/diabetes.csv
This command will upload the data in the data/diabetes.csv
file and generate
the corresponding source
, dataset
and cluster
objects in BigML. You
can use any of the generated objects to produce new clusters. For instance, you
could set a subgroup of the fields of the generated dataset to produce a
different cluster by using
bigmler cluster --dataset dataset/53b1f71437203f5ac30004ed \
--cluster-fields="-blood pressure"
that would exclude the field blood pressure
from the cluster creation input
fields.
Similarly to the models and datasets, the generated clusters can be shared
using the --shared
option, e.g.
bigmler cluster --source source/53b1f71437203f5ac30004e0 \
--shared
will generate a secret link for both the created dataset and cluster that
can be used to share the resource selectively.
As models were used to generate predictions (class names in classification
problems and an estimated number for regressions), clusters can be used to
predict the subgroup of data that our input data is more similar to.
Each subgroup is represented by its centroid, and the centroid is labelled
by a centroid name. Thus, a cluster would classify our
test data by assigning to each input an associated centroid name. The command
bigmler cluster --cluster cluster/53b1f71437203f5ac30004f0 \
--test data/my_test.csv
would produce a file centroids.csv
with the centroid name associated to
each input. When the command is executed, the cluster information is downloaded
to your local computer and the centroid predictions are computed locally, with
no more latencies involved. Just in case you prefer to use BigML to compute
the centroid predictions remotely, you can do so too
bigmler cluster --cluster cluster/53b1f71437203f5ac30004f0 \
--test data/my_test.csv --remote
would create a remote source and dataset from the test file data,
generate a batch centroid
also remotely and finally download the result
to your computer. If you prefer the result not to be
dowloaded but to be stored as a new dataset remotely, add --no-csv
and
to-dataset
to the command line. This can be specially helpful when
dealing with a high number of scores or when adding to the final result
the original dataset fields with --prediction-info full
, that may result
in a large CSV to be created as output.
The k-means algorithm used in clustering can only use training data that has
no missing values in their numeric fields. Any data that does not comply with
that is discarded in cluster construction, so you should ensure that enough
number of rows in your training data file has non-missing values in their
numeric fields for the cluster to be built and relevant. Similarly, the cluster
cannot issue a centroid prediction for input data that has missing values in
its numeric fields, so centroid predictions will give a “-” string as output
in this case.
You can change the number of centroids used to group the data in the
clustering procedure
bigmler cluster --dataset dataset/53b1f71437203f5ac30004ed \
--k 3
And also generate the datasets associated to each centroid of a cluster.
Using the --cluster-datasets
option
- bigmler cluster –cluster cluster/53b1f71437203f5ac30004f0
–cluster-datasets “Cluster 1,Cluster 2”
you can generate the datasets associated to a comma-separated list of
centroid names. If no centroid name is provided, all datasets are generated.
Similarly, you can generate the models to predict if one instance is associated
to each centroid of a cluster.
Using the --cluster-models
option
- bigmler cluster –cluster cluster/53b1f71437203f5ac30004f0
–cluster-models “Cluster 1,Cluster 2”
you can generate the models associated to a comma-separated list of
centroid names. If no centroid name is provided, all models are generated.
Models can be useful to see which features are important to determine whether
a certain instance belongs to a concrete cluster.
Anomaly subcommand
The bigmler anomaly
subcommand generates all the resources needed to buid
an anomaly detection model and/or predict the anomaly scores associated to your
test data. As usual, the simplest call
bigmler anomaly --train data/tiny_kdd.csv
uploads the data in the data/tiny_kdd.csv
file and generates
the corresponding source
, dataset
and anomaly
objects in BigML. You
can use any of the generated objects to produce new anomaly detectors.
For instance, you could set a subgroup of the fields of the generated dataset
to produce a different anomaly detector by using
bigmler anomaly --dataset dataset/53b1f71437203f5ac30004ed \
--anomaly-fields="-urgent"
that would exclude the field urgent
from the anomaly detector
creation input fields. You can also change the number of top anomalies
enclosed in the anomaly detector list and the number of trees that the anomaly
detector iforest uses. The default values are 10 top anomalies and 128 trees
per iforest:
bigmler anomaly --dataset dataset/53b1f71437203f5ac30004ed \
--top-n 15 --forest-size 50
with this code, the anomaly detector is built using an iforest of 50 trees and
will produce a list of the 15 top anomalies.
Similarly to the models and datasets, the generated anomaly detectors
can be shared using the --shared
option, e.g.
bigmler anomaly --source source/53b1f71437203f5ac30004e0 \
--shared
will generate a secret link for both the created dataset and anomaly detector
that can be used to share the resource selectively.
The anomaly detector can be used to assign an anomaly score to each new
input data set. The anomaly score is a number between 0 (not anomalous)
and 1 (highest anomaly). The command
bigmler anomaly --anomaly anomaly/53b1f71437203f5ac30005c0 \
--test data/test_kdd.csv
would produce a file anomaly_scores.csv
with the anomaly score associated
to each input. When the command is executed, the anomaly detector
information is downloaded
to your local computer and the anomaly score predictions are computed locally,
with no more latencies involved. Just in case you prefer to use BigML
to compute the anomaly score predictions remotely, you can do so too
bigmler anomaly --anomaly anomaly/53b1f71437203f5ac30005c0 \
--test data/my_test.csv --remote
would create a remote source and dataset from the test file data,
generate a batch anomaly score
also remotely and finally
download the result to your computer. If you prefer the result not to be
dowloaded but to be stored as a new dataset remotely, add --no-csv
and
to-dataset
to the command line. This can be specially helpful when
dealing with a high number of scores or when adding to the final result
the original dataset fields with --prediction-info full
, that may result
in a large CSV to be created as output.
Similarly, you can split your data in train/test datasets to build the
anomaly detector and create batch anomaly scores with the test portion of
data
bigmler anomaly --train data/tiny_kdd.csv --test-split 0.2 --remote
or if you want to apply the anomaly detector on the same training data set
to create a batch anomaly score, use:
bigmler anomaly --train data/tiny_kdd.csv --score --remote
To extract the top anomalies as a new dataset, or to exclude from the training
dataset the top anomalies in the anomaly detector, set the
--anomalies-dataset
to ìn
or out
respectively:
bigmler anomaly --dataset dataset/53b1f71437203f5ac30004ed \
--anomalies-dataset out
will create a new dataset excluding the top anomalous instances according
to the anomaly detector.
Sample subcommand
You can extract samples from your datasets in BigML using the
bigmler sample
subcommand. When a new sample is requested, a copy
of the dataset is stored in a special format in an in-memory cache.
This sample can then be used, before its expiration time, to
extract data from the related dataset by setting some options like the
number of rows or the fields to be retrieved. You can either begin from
scratch uploading your data to BigML, creating the corresponding source and
dataset and extracting your sample from it
bigmler sample --train data/iris.csv --rows 10 --row-offset 20
This command will create a source, a dataset, a sample object, whose id will
be stored in the samples
file in the output directory,
and extract 10 rows of data
starting from the 21st that will be stored in the sample.csv
file.
You can reuse an existing sample by using its id in the command.
bigmler sample --sample sample/53b1f71437203f5ac303d5c0 \
--sample-header --row-order-by="-petal length" \
--row-fields "petal length,petal width" --mode linear
will create a new sample.csv
file with a headers row where only the
petal length
and petal width
are retrieved. The --mode linear
option will cause the first available rows to be returned and the
--row-order-by="-petal length"
option returns these rows sorted in
descending order according to the contents of petal length
.
You can also add to the sample rows some statistical information by using the
--stat-field
or --stat-fields
options. Adding them to the command
will generate a stat-info.json
file where the Pearson’s and Spearman’s
correlations, and linear regression terms will be stored in a JSON format.
You can also apply a filter to select the sample rows by the values in
their fields using the --fields-filter
option. This must be set to
a string containing the conditions that must be met using field ids
and values.
bigmler sample --sample sample/53b1f71437203f5ac303d5c0 \
--fields-filter "000001=&!000004=Iris-setosa"
With this command, only rows where field id 000001
is missing and
field id 000004
is not Iris-setosa
will be retrieved. You can check
the available operators and syntax in the
samples’ developers doc .
More available
options can be found in the Samples subcommand Options
section.
Reify subcommand
This subcommand extracts the information in the existing resources to determine
the arguments that were used when they were created,
and generates scripts that could be used to reproduce them. Currently, the
language used in the scripts will be Python
. The usual starting
point for BigML resources is a source
created from inline, local or remote
data. Thus, the script keeps analyzing the chain of calls that led to a
certain resource until the root source
is found.
The simplest example would be:
bigmler reify --id source/55d77ba60d052e23430027bb
that will output:
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""Python code to reify source/5bd431db3980b574bb0145bf
Generated by BigMLer
"""
def main():
from bigml.api import BigML
api = BigML()
source_url1 = "https://static.bigml.com/csv/iris.csv"
source1 = api.create_source(source_url1)
api.ok(source1)
args = \
{'fields': {'000000': {'name': 'sepal length', 'optype': 'numeric'},
'000001': {'name': 'sepal width', 'optype': 'numeric'},
'000002': {'name': 'petal length', 'optype': 'numeric'},
'000003': {'name': 'petal width', 'optype': 'numeric'},
'000004': {'name': 'species',
'optype': 'categorical',
'term_analysis': {'enabled': True}}}}
source2 = api.update_source(source1, args)
api.ok(source2)
if __name__ == "__main__":
main()
According to this output, the source was created from a remote file
located at https://static.bigml.com/csv/iris.csv
and the types of each of it’s fields are described and stored to ensure
that they match the ones in the resource.
This script will be stored in the command output
directory and named reify.py` (you can specify a different name and location
using the --output
option).
Other resources will have more complex workflows and more user-given
attributes. Let’s see for instance the
script to generate an evaluation from a train/test split of a source that
was created using the
bigmler --train data/iris.csv --evaluate
command:
bigmler reify --id evaluation/55d919850d052e234b000833
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""Python code to reify evaluation/5be371a02774cb26da00061c
Generated by BigMLer
"""
def main():
from bigml.api import BigML
api = BigML()
source1_file = "iris.csv"
args = \
{'category': 12,
'description': 'Created using BigMLer',
'fields': {'000000': {'name': 'sepal length', 'optype': 'numeric'},
'000001': {'name': 'sepal width', 'optype': 'numeric'},
'000002': {'name': 'petal length', 'optype': 'numeric'},
'000003': {'name': 'petal width', 'optype': 'numeric'},
'000004': {'name': 'species',
'optype': 'categorical',
'term_analysis': {'enabled': True}}},
'tags': ['BigMLer', 'BigMLer_ThuNov0818_001323']}
source2 = api.create_source(source1_file, args)
api.ok(source2)
args = \
{'category': 12,
'description': 'Created using BigMLer',
'objective_field': {'id': '000004'},
'tags': ['BigMLer', 'BigMLer_ThuNov0818_001323']}
dataset1 = api.create_dataset(source2, args)
api.ok(dataset1)
args = \
{'category': 12,
'description': 'Created using BigMLer',
'sample_rate': 0.8,
'seed': 'BigML, Machine Learning made easy',
'split_candidates': 32,
'tags': ['BigMLer', 'BigMLer_ThuNov0818_001323']}
model1 = api.create_model(dataset1, args)
api.ok(model1)
args = \
{'category': 12,
'description': 'Created using BigMLer',
'fields_map': {'000001': '000001',
'000002': '000002',
'000003': '000003',
'000004': '000004'},
'operating_kind': 'probability',
'out_of_bag': True,
'sample_rate': 0.8,
'seed': 'BigML, Machine Learning made easy',
'tags': ['BigMLer', 'BigMLer_ThuNov0818_001323']}
evaluation1 = api.create_evaluation(model1, dataset1, args)
api.ok(evaluation1)
if __name__ == "__main__":
main()
As you can see, BigMLer has added a default category
,
description
and tags
attributes, has built the model on 80% of the data
and used the out_of_bag
attribute for the
evaluation to use the remaining part of the dataset test data.
The bigmler reify
command can generate also other types of
output depending on the
choice of the --language
option. The available options are python
(the one by default), nb
and whizzml
.
The nb
option will generate a jupyter notebook file.
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Reified resource: evaluation/5be371a02774cb26da00061c"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Remember to set your credentials in the BIGML_USERNAME and BIGML_API_KEY environment variables."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from bigml.api import BigML\n",
"api = BigML()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Add the inputs for the workflow"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"source1_file = \"iris.csv\""
]
},
...
]
}
We can also reify any
resource and obtain the WhizzML script that would recreate it using
--language whizzml
:
;;Step 1
;;WhizzML for resource: BigMLer_ThuNov0818_001323
;;(5 fields (1 categorical, 4 numeric))
;;source/5be371949252734ec7000938
;;created by mmartin
(define source2
(update-and-wait source1
{"fields"
{"000000" {"name" "sepal length" "optype" "numeric"}
"000001" {"name" "sepal width" "optype" "numeric"}
"000002" {"name" "petal length" "optype" "numeric"}
"000003" {"name" "petal width" "optype" "numeric"}
"000004"
{"name" "species"
"optype" "categorical"
"term_analysis" {"enabled" true}}}
"category" 12
"description" "Created using BigMLer"
"tags" ["BigMLer" "BigMLer_ThuNov0818_001323"]}))
;;Step 2
;;WhizzML for resource: BigMLer_ThuNov0818_001323
;;(150 instances, 5 fields (1 categorical, 4 numeric))
;;dataset/5be371972774cb26d5000954
;;created by mmartin
(define dataset1
(create-and-wait-dataset {"source" source2
"description" "Created using BigMLer"
"category" 12
"tags" ["BigMLer" "BigMLer_ThuNov0818_001323"]
"objective_field" {"id" "000004"}}))
;;Step 3
;;WhizzML for resource: BigMLer_ThuNov0818_001323
;;(512-node, pruned, deterministic order, sample rate=0.8)
;;model/5be3719a2774cb26d60020fa
;;created by mmartin
(define model1
(create-and-wait-model {"dataset" dataset1
"description" "Created using BigMLer"
"category" 12
"tags" ["BigMLer" "BigMLer_ThuNov0818_001323"]
"sample_rate" 0.8
"seed" "BigML, Machine Learning made easy"
"split_candidates" 32}))
;;Step 4
;;WhizzML for resource: BigMLer_ThuNov0818_001323
;;(512-node, pruned, deterministic order, sample rate=0.8, operating kind=probability, sample rate=0.2, out of bag)
;;evaluation/5be371a02774cb26da00061c
;;created by mmartin
(define evaluation1
(create-and-wait-evaluation {"description" "Created using BigMLer"
"category" 12
"tags"
["BigMLer" "BigMLer_ThuNov0818_001323"]
"fields_map"
{"000001" "000001"
"000002" "000002"
"000003" "000003"
"000004" "000004"}
"sample_rate" 0.8
"seed" "BigML, Machine Learning made easy"
"operating_kind" "probability"
"out_of_bag" true
"dataset" dataset1
"model" model1}))
(define output-evaluation evaluation1)
Execute subcommand
This subcommand creates and executes scripts in WhizzML (BigML’s automation
language). With WhizzML you can program any specific workflow that involves
Machine Learning resources like datasets, models, etc. You just write a
script using the directives in the
reference manual
and upload it to BigML, where it will be available as one more resource in
your dashboard. Scripts can also be shared and published in the gallery,
so you can reuse other users’ scripts and execute them. These operations
can also be done using the bigmler execute subcommand.
The simplest example is executing some basic code, like adding two numbers:
bigmler execute --code "(+ 1 2)" --output-dir simple_exe
With this command, bigmler will generate a script in BigML whose source code
is the one given as a string in the --code
option. The script ID will
be stored in a file called scripts
in the simple_text
directory. After that, the
script will be executed, so a new resource called execution
will be
created in BigML, and the corresponding ID will be stored in the
execution
file of the output directory.
Similarly, the result of the execution will be stored
in whizzml_results.txt
and whizzml_results.json
(in human-readable format and JSON respectively) in the
directory set in the --output-dir
option. You can also use the code
stored in a file with the --code-file
option.
Adding the --no-execute
flag to the command will cause the process to
stop right after the script creation. You can also compile your code as a
library to be used in many scripts by setting the --to-library
flag.
bigmler execute --code-file my_library.whizzml --to-library
Existing scripts can be referenced for execution with the --script
option
bigmler execute --script script/50a2bb64035d0706db000643
or the script ID can be read from a file:
bigmler execute --scripts simple_exe/scripts
The script we used as an example is very simple and needs no additional
parameter. But, in general, scripts
will have input parameters and output variables. The inputs define the script
signature and must be declared in order to create the script. The outputs
are optional and any variable in the script can be declared to be an output.
Both inputs and outputs can be declared using the --declare-inputs
and
--declare-outputs
options. These options must contain the path
to the JSON file where the information about the
inputs and outputs (respectively) is stored.
bigmler execute --code '(define addition (+ a b))' \
--declare-inputs my_inputs_dec.json \
--declare-outputs my_outputs_dec.json \
--no-execute
in this example, the my_inputs_dec.json
file could contain
[{"name": "a",
"default": 0,
"type": "number"},
{"name": "b",
"default": 0,
"type": "number",
"description": "second number to add"}]
and my_outputs_dec.json
[{"name": "addition",
"type": "number"}]
so that the value of the addition
variable would be returned as
output in the execution results.
Additionally, a script can import libraries. The list of libraries to be
used as imports can be added to the command with the option --imports
followed by a comma-separated list of library IDs.
Once the script has been created and its inputs and outputs declared, to
execute it you’ll need to provide a value for each input. This can be
done using --inputs
, that will also point to a JSON file where
each input should have its corresponding value.
bigmler execute --script script/50a2bb64035d0706db000643 \
--inputs my_inputs.json
where the my_inputs.json
file would contain:
For more details about the syntax to declare inputs and outputs, please
refer to the
Developers documentation.
You can also provide default configuration attributes
for the resources generated in an execution. Add the
--creation-defaults
option followed by the path
to a JSON file that contains a dictionary whose keys are the resource types
to which the configuration defaults apply and whose values are the
configuration attributes set by default.
bigmler execute --code-file my_script.whizzml \
--creation-defaults defaults.json
For instance, if my_script.whizzml
creates an ensemble from a remote
file:
(define file "s3://bigml-public/csv/iris.csv")
(define source (create-and-wait-source {"remote" file}))
(define dataset (create-and wait-dataset {"source" source}))
(define ensemble (create-and-wait-ensemble {"dataset" dataset}))
and my_create_defaults.json
contains
{
"source": {
"project": "project/54d9553bf0a5ea5fc0000016"
},
"ensemble": {
"number_of_models": 100, "sample_rate": 0.9
}
}
the source created by the script will be associated to the given project
and the ensemble will have 100 models and a 0.9 sample rate unless the source
code in your script explicitly specifies a different value, in which case
it takes precedence over these defaults.
Whizzml subcommand
This subcommand creates packages of scripts and libraries in WhizzML
(BigML’s automation
language) based on the information provided by a metadata.json
file. These operations
can also be performed individually using the bigmler execute subcommand, but
bigmler whizzml reads the components of the package, and for each
component analyzes the corresponding metadata.json
file to identify
the kind of code (script or library) that it contains and creates the
corresponding
resource in BigML. The metadata.json
is expected to contain the
name, kind, description, inputs and outputs needed to create the script.
As an example,
{
"name": "Example of whizzml script",
"description": "Test example of a whizzml script that adds two numbers",
"kind": "script",
"source_code": "code.whizzml",
"inputs": [
{
"name": "a",
"type": "number",
"description": "First number"
},
{
"name": "b",
"type": "number",
"description": "Second number"
}
],
"outputs": [
{
"name": "addition",
"type": "number",
"description": "Sum of the numbers"
}
]
}
describes a script whose code is to be found in the code.whizzml
file.
The script will have two inputs a
and b
and one output: addition
.
In order to create this script, you can type the following command:
bigmler whizzml --package-dir my_package --output-dir creation_log
and bigmler will:
look for the metadata.json
file located in the my_package
directory.
parse the JSON, identify that it defines a script and look for its code in
the code.whizzml
file
create the corresponding BigML script resource, adding as arguments the ones
provided in inputs
, outputs
, name
and description
.
Packages can contain more than one script. In this case, a nested directory
structure is expected. The metadata.json
file for a package with many
components should include the name of the directories where these components
can be found:
{
"name": "Best k",
"description": "Library and scripts implementing Pham-Dimov-Nguyen k selection algorithm",
"kind": "package",
"components":[
"best-k-means",
"cluster",
"evaluation",
"batchcentroid"
]
}
In this example, each string in the components
attributes list corresponds
to one directory where a new script or library (with its corresponding
metadata.json
descriptor) is stored. Then, using bigmler whizzml
for this composite package will create each of the component scripts or
libraries. It will also handle dependencies, using the IDs of the created
libraries as imports for the scripts when needed. The metadata.json
that corresponds to a library is simpler than the one used for the script,
the difference being that kind
in this case will be set to library
and no inputs or outputs are provided.
{
"name": "Best K-Means",
"description": "Best K-Means Clustering using the Pham, Dimov, and Nguyen Algorithm",
"kind": "library",
"source_code": "library.whizzml"
}
To include a library in the list of imports of a script, the imports
attribute is used in the script’s metadata.json
. The imports
should be the list of folders that contain each library source code and
metadata.
{
"name": "Compute Best K-means Batchcentroid",
"description": "Basic script to use the best-kmeans library",
"kind": "script",
"source_code": "script.whizzml",
"imports": ["../best-k-means"],
"inputs": [
{
"name": "dataset",
"type": "dataset-id",
"description": "Dataset ID"
},
{
"name": "cluster-args",
"type": "map",
"description": "Map of args for clustering (excluding dataset and k) for k search",
"default": {}
},
{
"name": "k-min",
"type": "number",
"description": "Minimum value of k for search"
},
{
"name": "k-max",
"type": "number",
"description": "Maximum value of k for search"
},
{
"name": "bestcluster-args",
"type": "map",
"description": "Map of args for clustering (excluding dataset and k) for optimal k",
"default": {}
},
{
"name": "clean",
"type": "boolean",
"description": "Delete intermediate objects created during computation"
},
{
"name": "logf",
"type": "boolean",
"description": "Generate log entries"
}
],
"outputs": [
{
"name": "best-batchcentroid",
"type": "string",
"description": "Batchcentroid ID"
}
]
}
Retrain subcommand
This subcommand can be used to retrain an existing modeling resource (model,
ensemble, deepnet, etc.) by adding new data to it. In BigML, resources are
immutable to ensure traceability, but at the same time they are reproducible.
Therefore, any model can be rebuilt using the data stored in a new consolidated
dataset or even from a list of existing datasets. That’s retraining the model
and the bigmler retrain
subcommand provides a simple way to do it.
In the basic use case, different parameters and model types are tried and
evaluated till the best performing model is found. Then you can call:
bigmler retrain --id model/5a3ae0f14006833a070003a4 --add data/iris.csv \
--output-dir retrain_directory
so that the data in your local data/iris.csv
file is uploaded to the
platform and all the steps that led to your existing model are reproduced to
create a new merged dataset that will be used to retrain your model. The
command output will contain the URL that you need to call to ensure you
always use the latest version of your model. The URL will look like:
https://bigml.io/andromeda/model?username=my_user;api_key=my_api_key;limit=1;full=yes;tags=retrain:model/5a3ae0f14006833a070003a4
Instead of using the original model ID, you can choose to add a unique tag
to your modeling resource and use that as reference:
bigmler retrain --ensemble-tag my_ensemble --add data/iris.csv \
--output-dir retrain_directory
in this case, the resource to retrain is an ensemble that has been
previously tagged as my_ensemble
. The bigmler retrain
command will
look for the newest ensemble that contains that tag and after uploading and
consolidating your data with the one previously used in the ensemble, it will
rebuild it. The reference used in the URL that will contain the latest version
of the ensemble will use this tag also as reference:
https://bigml.io/andromeda/ensemble?username=my_user;api_key=my_api_key;limit=1;full=yes;tags=my_ensemble
In a different scenario, you might want to retrain your model from a list
of datasets, for instance training an anomaly detector using the data of the
last 6 months. This means that you don’t want your data to be merged. Rather
you would like to use a window over the list of available datasets.
bigmler retrain --ensemble-tag my_ensemble --add data/iris.csv \
--window-size 6 --output-dir retrain_directory
In this case, adding the --window-size
option to your command will cause
the dataset created by uploading your new data to be added to the list of
datasets as a separate resource. Then model will be rebuilt using the number
of datasets set as --window-size
.
The operations run by bigmler retrain
are mainly run in BigML’s servers
using WhizzML scripts. This scripts are previously created in the user’s
account the first time you run the command, but they can also be recreated
by using the --upgrade
flag in any bigmler retrain
command call.
Delete subcommand
You have seen that BigMLer is an agile tool that empowers you to create a
great number of resources easily. This is a tremedous help, but it also can
lead to a garbage-prone environment. To keep a control of each new created
remote resource use the flag –resources-log followed by the name of the log
file you choose.
bigmler --train data/iris.csv --resources-log my_log.log
Each new resource created by that command will cause its id to be appended as
a new line of the log file.
BigMLer can help you as well in deleting these resources. Using the delete
subcommand there are many options available. For instance, deleting a
comma-separated list of ids
bigmler delete \
--ids source/50a2bb64035d0706db0006cc,dataset/50a1f441035d0706d9000371
deleting resources listed in a file
bigmler delete --from-file to_delete.log
where to_delete.log contains a resource id per line.
As we’ve previously seen, each BigMLer command execution generates a
bunch of remote resources whose ids are stored in files located in a directory
that can be set using the --output-dir
option. The
bigmler delete
subcommand can retrieve the ids stored in such files by
using the --from-dir
option.
bigmler --train data/iris.csv --output my_BigMLer_output_dir
bigmler delete --from-dir my_BigMLer_output_dir
The last command will delete all the remote resources previously generated by
the fist command by retrieving their ids from the files in
my_BigMLer_output_dir
directory.
You can also delete resources based on the tags they are associated to
bigmler delete --all-tag my_tag
or restricting the operation to a specific type
bigmler delete --source-tag my_tag
bigmler delete --dataset-tag my_tag
bigmler delete --model-tag my_tag
bigmler delete --prediction-tag my_tag
bigmler delete --evaluation-tag my_tag
bigmler delete --ensemble-tag my_tag
bigmler delete --batch-prediction-tag my_tag
bigmler delete --cluster-tag my_tag
bigmler delete --centroid-tag my_tag
bigmler delete --batch-centroid-tag my_tag
bigmler delete --anomaly-tag my_tag
bigmler delete --anomaly-score-tag my_tag
bigmler delete --batch-anomaly-score-tag my_tag
bigmler delete --project-tag my_tag
bigmler delete --logistic-regression-tag my_tag
bigmler delete --linear-regression-tag my_tag
bigmler delete --time-series-tag my_tag
bigmler delete --deepnet-tag my_tag
bigmler delete --topic-model-tag my_tag
bigmler delete --topic-distribution-tag my_tag
bigmler delete --association-tag my_tag
You can also delete resources by date. The options --newer-than
and
--older-than
let you specify a reference date. Resources created after and
before that date respectively, will be deleted. Both options can be combined to
set a range of dates. The allowed values are:
dates in a YYYY-MM-DD format
integers, that will be interpreted as number of days before now
resource id, the creation datetime of the resource will be used
Thus,
bigmler delete --newer-than 2
will delete all resources created less than two days ago (now being
2014-03-23 14:00:00.00000, its creation time will be greater
than 2014-03-21 14:00:00.00000).
bigmler delete --older-than 2014-03-20 --newer-than 2014-03-19
will delete all resources created during 2014, March the 19th (creation time
between 2014-03-19 00:00:00 and 2014-03-20 00:00:00) and
bigmler delete --newer-than source/532db2b637203f3f1a000104
will delete all resources created after the source/532db2b637203f3f1a000104
was created.
You can also combine both types of options, to delete sources tagged as
my_tag
starting from a certain date on
bigmler delete --newer-than 2 --source-tag my_tag
And finally, you can filter the type of resource to be deleted using the
--resource-types
option to specify a comma-separated list of resource
types to be deleted
bigmler delete --older-than 2 --resource-types source,model
will delete the sources and models created more than two days ago.
Additionally, you can use the --resource-types
option to tell which
type of resources to exclude from deletion if the --exclude-types
flag
is added to the call.
bigmler delete --older-than 2 --resource-types source,model --exclude-types
That command will delete all the resources that are older than two days except
for sources and models.
You can simulate the a delete subcommand using the --dry-run
flag
bigmler delete --newer-than source/532db2b637203f3f1a000104 \
--source-tag my_source --dry-run
The output for the command will be a list of resources that would be deleted
if the --dry-run
flag was removed. In this case, they will be sources
that contain the tag my_source
and were created after the one given as
--newer-than
value. The first 15 resources will be logged
to console, and the complete list can be found in the bigmler_sessions
file.
A similar option that does not delete the resources immediately is --bin
.
bigmler delete --newer-than 3 --resource-types source \
--source-tag my_source --bin
By setting that flag, all the selected resources are moved to a newly
created Trash bin
project in your account. That allows the user to
inspect the selected resources before deletion and delete them in an efficient
way by deleting the Trash bin
project.
By default, only finished resources are selected to be deleted. If you want
to delete other resources, you can select them by choosing their status:
bigmler delete --older-than 2 --status faulty
would remove all failed resources created more than two days ago.
Also, you can apply a filter based on the filters used in the API list
query strings (see the API documentation).
bigmler delete --filter "name__icontains=iris"
Export subcommand
The bigmler export
subcommand is intended to help generating the code
needed for the models in BigML to be integrated in other applications.
To produce a prediction using a BigML model you just need a function that
receives as argument the new test
case data and returns this prediction (and a confidence). The bigmler export
subcommand will retrieve the JSON information of your existing
decision tree model in BigML and will generate from it this function code and
store it in a file that can be imported or copied directly in your application.
Obviously, the function syntax will depend on the model and the language
used in your application, so these will be the options we need to provide:
bigmler export --model model/532db2b637203f3f1a001304 \
--language javascript --output-dir my_exports
This command will create a javascript version of the function that
produces the predictions and store it in a file named
model_532db2b637203f3f1a001304.js (after the model
ID) in the my_exports directory.
Models can currently exported in Python, Javascript and R. For models
whose fields are numeric or categorical, the command
also supports creating MySQL functions and Tableau separate expressions
for both the prediction and the confidence.
You can also generate the code for all the models in an ensemble in a
single bigmler export command using the –ensemble option followed
by the corresponding ensemble ID. The code for
each model will be stored in a separate file, named after the model ID and
transforming the slash into an underscore.
bigmler export --ensemble ensemble/532db2b637203f3f1a001307 \
--language javascript --output-dir my_ensemble
Project subcommand
Projects are organizational resources and they are usually created at
source-creation time in order to keep together in a separate repo all
the resources derived from a source. However, you can also create a project
or update its properties independently using the bigmler project
subcommand.
bigmler project --name my_project
will create a new project and name it. You can also add other attributes
such as --tag
, --description
or --category
in the project
creation call. You can also add or update any other attribute to
the project using a JSON file with the --project-attributes
option.
bigmler project --project-id project/532db2b637203f3f1a000153 \
--project-attributes my_attributes.json
Association subcommand
Association Discovery is a popular method to find out relations among values
in high-dimensional datasets.
A common case where association discovery is often used is
market basket analysis. This analysis seeks for customer shopping
patterns across large transactional
datasets. For instance, do customers who buy hamburgers and ketchup also
consume bread?
Businesses use those insights to make decisions on promotions and product
placements.
Association Discovery can also be used for other purposes such as early
incident detection, web usage analysis, or software intrusion detection.
In BigML, the Association resource object can be built from any dataset, and
its results are a list of association rules between the items in the dataset.
In the example case, the corresponding
association rule would have hamburguers and ketchup as the items at the
left hand side of the association rule and bread would be the item at the
right hand side. Both sides in this association rule are related,
in the sense that observing
the items in the left hand side implies observing the items in the right hand
side. There are some metrics to ponder the quality of these association rules:
For an association rule, it means the number of instances in the dataset which
contain the rule’s antecedent and rule’s consequent together
over the total number of instances (N) in the dataset.
It gives a measure of the importance of the rule. Association rules have
to satisfy a minimum support constraint (i.e., min_support).
It measures how often a rule can be applied.
under the condition that the instances also contain the rule’s antecedent.
Confidence is computed using the support of the association rule over the
coverage. That is, the percentage of instances which contain the consequent
and antecedent together over the number of instances which only contain
the antecedent.
Confidence is directed and gives different values for the association
rules Antecedent → Consequent and Consequent → Antecedent. Association
rules also need to satisfy a minimum confidence constraint
(i.e., min_confidence).
rule (i.e., the antecedent and consequent appearing together) and what would
be expected if antecedent and consequent where statistically independent.
This is a value between -1 and 1. A positive value suggests a positive
relationship and a negative value suggests a negative relationship.
0 indicates independence.
Lift: how many times more often antecedent and consequent occur together
than expected if they where statistically independent.
A value of 1 suggests that there is no relationship between the antecedent
and the consequent. Higher values suggest stronger positive relationships.
Lower values suggest stronger negative relationships (the presence of the
antecedent reduces the likelihood of the consequent)
As to the items used in association rules, each type of field is parsed to
extract items for the rules as follows:
Categorical: each different value (class) will be considered a separate item.
Text: each unique term will be considered a separate item.
Items: each different item in the items summary will be considered.
Numeric: Values will be converted into categorical by making a
segmentation of the values.
For example, a numeric field with values ranging from 0 to 600 split
into 3 segments:
segment 1 → [0, 200), segment 2 → [200, 400), segment 3 → [400, 600].
You can refine the behavior of the transformation using
discretization
and field_discretizations.
The bigmler association
subcommand will discover the association
rules present in your
datasets. Starting from the raw data in your files:
bigmler association --train my_file.csv
will generate the source
, dataset
and association
objects
required to present the association rules hidden in your data. You can also
limit the number of rules extracted using the --max-k
option
bigmler association --dataset dataset/532db2b637203f3f1a000103 \
--max-k 20
With the prior command only 20 association rules will be extracted. Similarly,
you can change the search strategy used to find them
bigmler association --dataset dataset/532db2b637203f3f1a000103 \
--search-strategy confidence
In this case, the confidence
is used (the default value being
leverage
).
Logistic-regression subcommand
The bigmler logistic-regression
subcommand generates all the
resources needed to buid
a logistic regression model and use it to predict.
The logistic regression model is a supervised
learning method for solving classification problems. It predicts the
objective field class as logistic function whose argument is a linear
combination of the rest of features. The simplest call to build a logistic
regression is
bigmler logistic-regression --train data/iris.csv
uploads the data in the data/iris.csv
file and generates
the corresponding source
, dataset
and logistic regression
objects in BigML. You
can use any of the generated objects to produce new logistic regressions.
For instance, you could set a subgroup of the fields of the generated dataset
to produce a different logistic regression model by using
bigmler logistic-regression --dataset dataset/53b1f71437203f5ac30004ed \
--logistic-fields="-sepal length"
that would exclude the field sepal length
from the logistic regression
model creation input fields. You can also change some parameters in the
logistic regression model, like the bias
(scale of the intercept term),
c
(the strength of the regularization map) or eps
(stopping criteria
for solver).
bigmler logistic-regression --dataset dataset/53b1f71437203f5ac30004ed \
--bias --c 5 --eps 0.5
with this code, the logistic regression is built using an independent term,
the step in the regularization is 5 and the difference between the results
from the current and last iterations is 0.5.
Similarly to the models and datasets, the generated logistic regressions
can be shared using the --shared
option, e.g.
bigmler logistic-regression --source source/53b1f71437203f5ac30004e0 \
--shared
will generate a secret link for both the created dataset and logistic
regressions, that can be used to share the resource selectively.
The logistic regression can be used to assign a prediction to each new
input data set. The command
bigmler logistic-regression \
--logistic-regression logisticregression/53b1f71435203f5ac30005c0 \
--test data/test_iris.csv
would produce a file predictions.csv
with the predictions associated
to each input. When the command is executed, the logistic regression
information is downloaded
to your local computer and the logistic regression predictions are
computed locally,
with no more latencies involved. Just in case you prefer to use BigML
to compute the predictions remotely, you can do so too
bigmler logistic-regression \
--logistic-regression logisticregression/53b1f71435203f5ac30005c0 \
--test data/my_test.csv --remote
would create a remote source and dataset from the test file data,
generate a batch prediction
also remotely and finally
download the result to your computer. If you prefer the result not to be
dowloaded but to be stored as a new dataset remotely, add --no-csv
and
to-dataset
to the command line. This can be specially helpful when
dealing with a high number of scores or when adding to the final result
the original dataset fields with --prediction-info full
, that may result
in a large CSV to be created as output. Other output configurations can be
set by using the --batch-prediction-attributes
option pointing to a JSON
file that contains the desired attributes, like:
{"probabilities": true,
"all_fields": true}
Linear-regression subcommand
The bigmler linear-regression
subcommand generates all the
resources needed to buid
a linear regression model and use it to predict.
The linear regression model is a supervised
learning method for solving regression problems. It predicts the
objective field class as a linear function whose argument are
the rest of features. The simplest call to build a linear
regression is
bigmler linear-regression --train data/grades.csv
uploads the data in the data/grades.csv
file and generates
the corresponding source
, dataset
and linear regression
objects in BigML. You
can use any of the generated objects to produce new linear regressions.
For instance, you could set a subgroup of the fields of the generated dataset
to produce a different linear regression model by using
bigmler linear-regression --dataset dataset/53b1f71437203f5ac30004ed \
--linear-fields="-Prefix"
that would exclude the field Prefix
from the linear regression
model creation input fields. You can also change some parameters in the
linear regression model, like the bias
(intercept term)).
bigmler linear-regression --dataset dataset/53b1f71437203f5ac30004ed \
--no-bias
with this code, the linear regression is built without using an
independent term.
Similarly to models and datasets, the generated linear regressions
can be shared using the --shared
option, e.g.
bigmler linear-regression --source source/53b1f71437203f5ac30004e0 \
--shared
will generate a secret link for both the created dataset and linear
regressions, that can be used to share the resource selectively.
Linear regressions can produce a prediction for each new
input data set. The command
bigmler linear-regression \
--linear-regression linearregression/53b1f71435203f5ac30005c0 \
--test data/test_grades.csv
would produce a file predictions.csv
with the predictions associated
to each input. When the command is executed, the linear regression
information is downloaded
to your local computer and the linear regression predictions are
computed locally,
with no more latencies involved. Just in case you prefer to use BigML
to compute the predictions remotely, you can do so too
bigmler linear-regression
--linear-regression linearregression/53b1f71435203f5ac30005c0 \
--test data/my_test.csv --remote
would create a remote source and dataset from the test file data,
generate a batch prediction
also remotely and finally
download the result to your computer. If you prefer the result not to be
dowloaded but to be stored as a new dataset remotely, add --no-csv
and
to-dataset
to the command line. This can be specially helpful when
dealing with a high number of scores or when adding to the final result
the original dataset fields with --prediction-info full
, that may result
in a large CSV to be created as output. Other output configurations can be
set by using the --batch-prediction-attributes
option pointing to a JSON
file that contains the desired attributes, like:
{"probabilities": true,
"all_fields": true}
Topic Model subcommand
Using this subcommand you can generate all the
resources leading to finding a topic model
and its topic distributions
.
These are unsupervised learning models which find out the topics in a
collection of documents and will then be useful to classify new documents
according to the topics. The bigmler topic-model
subcommand
will follow the steps to generate
topic models
and predict the topic distribution
, or distribution of
probabilities for the new document to be associated to a certain topic. As
shown in the bigmler
command section, the simplest call is
bigmler topic-model --train data/spam.csv
This command will upload the data in the data/spam.csv
file and
generate
the corresponding source
, dataset
and topic model
objects in BigML.
You
can use any of the intermediate generated objects to produce new
topic models. For instance, you
could set a subgroup of the fields of the generated dataset to produce a
different topic model by using
bigmler topic-model --dataset dataset/53b1f71437203f5ac30004ed \
--topic-fields="-Message"
that would exclude the field Message
from the topic model creation input
fields.
Similarly to the models and datasets, the generated topic models can be shared
using the --shared
option, e.g.
bigmler topic-model --source source/53b1f71437203f5ac30004e0 \
--shared
will generate a secret link for both the created dataset and topic model that
can be used to share the resource selectively.
As models were used to generate predictions (class names in classification
problems and an estimated number for regressions), topic models can be used
to classify a new document in the discovered list of topics. The classification
is run by computing the probability for the document to belonging to the topic
group. The command
bigmler topic-model --topic-model topicmodel/58437a277e0a8d38ec028a5f \
--test data/my_test.csv
would produce a file topic_distributions.csv
where each row will contain
the probabilities
associated to each topic for the corresponding test input.
When the command is executed, the topic model information is downloaded
to your local computer and the distributions are computed locally, with
no more latencies involved. Just in case you prefer to use BigML to compute
the topic distributions remotely, you can do so too
bigmler topic-model --topic-model topicmodel/58437a277e0a8d38ec028a5f \
--test data/my_test.csv --remote
would create a remote source and dataset from the test file data,
generate a batch topic distribution
also remotely and finally
download the result
to your computer. If you prefer the result not to be
dowloaded but to be stored as a new dataset remotely, add --no-csv
and
to-dataset
to the command line. This can be specially helpful when
dealing with a high number of scores or when adding to the final result
the original dataset fields with --prediction-info full
, that may result
in a large CSV to be created as output.
Note that the the topics created in the Topic Model resource are now named
after the more frequent terms that they contain. To return to the previous
Topic 0
style naming you can use the --minimum-name-terms
option and
set it to 0
.
Time Series subcommand
Using this subcommand you can generate all the
resources leading to a time series
and its forecasts
.
The time series
is a supervised learning model that works on
an ordered sequence of data to extract the patterns needed to make
forecasts
. The bigmler time-series
subcommand
will follow the steps to generate
time series
and predict the forecasts
for every numeric field in
the original dataset that has been set as objective field. As
shown in the bigmler
command section, the simplest call is
bigmler time-series --train data/grades.csv
This command will upload the data in the data/grades.csv
file and
generate
the corresponding source
, dataset
and time series
objects in BigML.
You
can use any of the intermediate generated objects to produce new
time series. For instance, you
could set a subgroup of the numeric fields in the dataset to be used
as objective fields using the --objectives
option.
bigmler time-series --dataset dataset/53b1f71437203f5ac30004ed \
--objectives "Assignment,Final"
its value is expected to be a comma-separated list of fields.
Similarly to the models and datasets, the generated clusters can be shared
using the --shared
option, e.g.
bigmler time-series --source source/53b1f71437203f5ac30004e0 \
--shared
will generate a secret link for both the created dataset and time series that
can be used to share the resource selectively.
As models were used to generate predictions (class names in classification
problems and an estimated number for regressions), time series can be used
to generate forecasts, that is, to predict the value of each objective
field up till the user-given horizon. The command
bigmler time-series --time-series timeseries/58437a277e0a8d38ec028a5f \
--horizon 10
would produce a file forecast_000001.csv
with ten rows, one per point, and
as many columns as ETS models the time series contains.
When the command is executed, the time series information is downloaded
to your local computer and the forecasts are computed locally, with
no more latencies involved. Just in case you prefer to use BigML to compute
the forecasts remotely, you can do so too
bigmler time-series --time-series timeseries/58437a277e0a8d38ec028a5f \
--horizon 10 --remote
would create a remote forecast with the specified horizon. You can also
specify more complex inputs for the forecast. For instance, you can set a
different horizon to each objective field and you can give some criteria
to select the models used in the forecast. All of this can be done using
the --test
option pointing to a JSON file that should contain the
input to be used in the forecast as described in the
API documentation. As an example,
let’s set a horizon of 5 points for the Final
field and select the
first model in the time series array of ETS models, and also forecast 7
points for the Assignment
field using the model with less aic
(the one
used by default). The command call should then be:
bigmler time-series --time-series timeseries/58437a277e0a8d38ec028a5f \
--test test.json
and the test.json
file should contain the following JSON:
{"Final": {"horizon": 5, "ets_models": {"indices": [0]}},
"Assignment": {"horizon": 7}}
Deepnet subcommand
The bigmler deepnet
subcommand generates all the
resources needed to buid
a deepnet model and use it to predict.
The deepnet model is a supervised
learning method for solving both regression and classification problems. It
uses deep neural networks, a composition of layers of different functions
that when applied to the
input data generate the prediction.
The simplest call to build a deepnet is:
bigmler deepnet --train data/iris.csv
uploads the data in the data/iris.csv
file and generates
the corresponding source
, dataset
and deepnet
objects in BigML. You
can use any of the generated objects to produce new deepnets.
For instance, you could set a subgroup of the fields of the generated dataset
to produce a different deepnet model by using
bigmler deepnet --dataset dataset/53b1f71437203f5ac30004ed \
--deepnet-fields="-sepal length"
that would exclude the field sepal length
from the deepnet
model creation input fields. You can also change some parameters in the
deepnet model, like the number_of_hidden_layers
, max_iterations
or default_numeric_value
. Please check the Deepnets section
of the API documentation for a detailed
description of the available arguments.
bigmler deepnet --dataset dataset/53b1f71437203f5ac30004ed \
--number-of-hidden-layers 3
--max-iterations 10 --default-numeric-value mean
with this code, the deepnet is built using 3 hidden layers, approximations
will stop after 10 iterations and the missing numerics will be filled with
the mean of the rest of values in the field.
Similarly to the models and datasets, the generated deepnets
can be shared using the --shared
option, e.g.
bigmler deepnet --source source/53b1f71437203f5ac30004e0 \
--shared
will generate a secret link for both the created dataset and deepnet,
that can be used to share the resource selectively.
The deepnet can be used to assign a prediction to each new
input data set. The command
bigmler deepnet \
--deepnet deepnet/5331f71435203f5ac30005c0 \
--test data/test_iris.csv
would produce a file predictions.csv
with the predictions associated
to each input. When the command is executed, the deepnet
information is downloaded
to your local computer and the deepnet predictions are
computed locally,
with no more latencies involved. Just in case you prefer to use BigML
to compute the predictions remotely, you can do so too
bigmler deepnet
--deepnet deepnet/53b1f71435203f5ac30005c0 \
--test data/my_test.csv --remote
would create a remote source and dataset from the test file data,
generate a batch prediction
also remotely and finally
download the result to your computer. If you prefer the result not to be
dowloaded but to be stored as a new dataset remotely, add --no-csv
and
to-dataset
to the command line. This can be specially helpful when
dealing with a high number of scores or when adding to the final result
the original dataset fields with --prediction-info full
, that may result
in a large CSV to be created as output. Other output configurations can be
set by using the --batch-prediction-attributes
option pointing to a JSON
file that contains the desired attributes, like:
{"probabilities": true,
"all_fields": true}
Fusion subcommand
The bigmler fusion
subcommand generates all the
resources needed to build
a fusion model and use it to predict.
The fusion model is a supervised
learning method for solving both regression and classification problems. It’s
a model composed of different supervised models, ensembles, deepnets,
logistic regressions, linear regressions or fusions. The prediction obtained
from a fusion will be an aggregation of the predictions of its component
models. The aggregation will take into account the weight associated to each
of the models in the fusion object. If no specific weight is given on creation,
each model in the fusion will be assigned the same weight.
The simplest call to build a fusion is:
bigmler fusion \
--fusion-models deepnet/53b1f71437203f5ac30004ed,model/53b1f71437203f5ac32004e2 \
--output-dir my_fusion
that creates the fusion object for the deepnet
and model
described in
--fusiion-models
. The fusion ID is stored in a fusions
file in the
directory specified in --output-dir
.
As explained, different weights can be applied to the predictions of each
model to generate the final prediction. To set these weights, you can use
a --fusion-models-file
option to point to the JSON file describing the
models and their weights as explained in the
API developers docs.
bigmler fusion --fusion-models-file components.json \
--output-dir my_fusion
An existing fusion can also be used to predict.
bigmler fusion --fusion fusion/53b1f71437203f5ac30004cd \
--test my_test_data.csv \
--output my_predictions.csv
with this code, the my_test_data
file contents are run through the
fusion and a new prediction is asociated to each line in the CSV file. The
results are stored in the my_predictions.csv
file.
The fusion
information is downloaded
to your local computer and the fusion predictions are
computed locally,
with no more latencies involved. Just in case you prefer to use BigML
to compute the predictions remotely, you can do so too
bigmler fusion --fusion fusion/53b1f71437203f5ac30004cd \
--test my_test_data.csv \
--output my_predictions.csv --remote
would create a remote source and dataset from the test file data,
generate a batch prediction
also remotely and finally
download the result to your computer. If you prefer the result not to be
dowloaded but to be stored as a new dataset remotely, add --no-csv
and
to-dataset
to the command line. This can be specially helpful when
dealing with a high number of scores or when adding to the final result
the original dataset fields with --prediction-info full
, that may result
in a large CSV to be created as output. Other output configurations can be
set by using the --batch-prediction-attributes
option pointing to a JSON
file that contains the desired attributes, like:
{"probabilities": true,
"all_fields": true}
PCA subcommand
The bigmler pca
subcommand generates all the
resources needed to buid
a PCA model and use it to predict.
The PCA model is an unsupervised
learning method for dimensionality reduction that tries to find new features
that can maximize the description of the data variation. The new features
are built as linear combinations of the original features in the dataset.
The simplest call to build a PCA is:
bigmler pca --train data/iris.csv
uploads the data in the data/iris.csv
file and generates
the corresponding source
, dataset
and pca
objects in BigML. You
can use any of the generated objects to produce new PCAs.
For instance, you could set a subgroup of the fields of the generated dataset
to produce a different PCA model by using
bigmler pca --dataset dataset/53b1f71437203f5ac30004ed \
--pca-fields="-sepal length"
that would exclude the field sepal length
from the PCA
model creation input fields.
Similarly to the models and datasets, the generated PCAs
can be shared using the --shared
option, e.g.
bigmler pca --source source/53b1f71437203f5ac30004e0 \
--shared
will generate a secret link for both the created dataset and PCA,
that can be used to share the resource selectively.
The PCA can be used to assign a projection (a set of new components)
to each input data set. The command
bigmler pca \
--pca pca/5331f71435203f5ac30005c0 \
--test data/test_iris.csv \
--output projections.csv
would produce a file projections.csv
with the projections associated
to each input. It’s important to remark that to build projections for
a supervised learning problem the objective field should never be part of
the PCA input fields. Including the objective in the PCA would cause leakage.
In order to remove the objective field, you can use the --exclude-objective
flag. Also, the train/test split should be done before creating the PCA from
the training dataset to avoid leakage from the test set data
in the new components.
You can also change some parameters in the
PCA model, like the --max-components
or --variance-threshold
to select the number of components to be used in the projection.
Please check the PCA section
of the API documentation for a detailed
description of the available arguments.
bigmler pca --dataset dataset/53b1f71437203f5ac30004ed \
--max-components 4 \
--test data/test_iris.csv \
--output projections.csv
with this code, only the first 4 components of the PCA are used to generate
projections, reducing thus the dimensionality of the dataset to 4.
When previous command is executed, the PCA
information is downloaded
to your local computer and the PCA projections are
computed locally,
with no more latencies involved. Just in case you prefer to use BigML
to compute the projections remotely, you can do so too
bigmler pca
--pca pca/53b1f71435203f5ac30005c0 \
--test data/my_test.csv --remote
would create a remote source and dataset from the test file data,
generate a batch projection
also remotely and finally
download the result to your computer. If you prefer the result not to be
dowloaded but to be stored as a new dataset remotely, add --no-csv
and
to-dataset
to the command line. Some output format configurations can
be controlled using the --projection-header
option, that causes
the headers of the fields to be placed as a first row in the projections file,
or the --projection-fields
option, that can be set to all
or to
a comma-separated list of fields of the original dataset that will be included
in the projections file before the projection components.
Other output configurations can be
set by using the --batch-projection-attributes
option pointing to a JSON
file that contains the desired attributes, like:
{"output_fields": ["petal length", "sepal length"],
"all_fields": true}