Local Resources¶
All the resources in BigML can be downloaded and used afterwards locally, with
no connection whatsoever to BigML’s servers. This is specially important
for all Supervised and Unsupervised models, that can be used to generate
predictions in any programmable device. The next sections describe how to
do that for each type of resource, but as a general rule, resources can be
exported to a JSON file in your file system using the export method.
api.export('model/5143a51a37203f2cf7000956',
'filename': 'my_dir/my_model.json')
The contents of the generated file can be used just as the remote model
to generate predictions. As you’ll see in next section, the local Model
object can be instantiated by giving the path to this file as first argument:
from bigml.model import Model
local_model = Model("my_dir/my_model.json")
local_model.predict({"petal length": 3, "petal width": 1})
Iris-versicolor
These bindings define a particular class for each type of Machine Learning model that is able to interpret the corresponding JSON and create the local predictions. The classes can be instantiated using:
The ID of the resource: In this case, the class looks for the JSON information of the resource first locally (expecting to find a file in the local storage directory –
./storageby default – whose name is the ID of the model after replacing/by_) and also remotely if absent.
from bigml.model import Model
from bigml.api import BigML
local_model = Model('model/502fdbff15526876610002615')
A dictionary containing the resource information. In this case, the class checks that this information belongs to a finished resource and contains the attributes needed to create predictions, like the fields structure. If any of these attributes is absent, retrieves the ID of the model and tries to download the correct JSON from the API to store it locally for further use.
from bigml.anomaly import Anomaly
from bigml.api import BigML
api = BigML()
anomaly = api.get_anomaly('anomaly/502fdbff15526876610002615',
query_string='only_model=true;limit=-1')
local_anomaly = Anomaly(anomaly)
A path to the file that contains the JSON information for the resource. In this case, the file is read and the same checks mentioned above are done. If any of these checks fails, it tries to retrieve the correct JSON from the API to store it locally for further use.
from bigml.logistic import LogisticRegression
local_logistic_regression = LogisticRegression('./my_logistic.json')
Internally, these classes need a connection object
(api = BigML(storage="./storage")) to:
Set the local storage in your file system.
Download the JSON of the resource if the information provided is not the full finished resource content.
Users can provide the connection as a second argument when instantiating the
class, but if they do and want the resource to be available locally, the
connection object must be created with an storage setting:
from bigml.cluster import Cluster
from bigml.api import BigML
local_cluster = Cluster('cluster/502fdbff15526876610002435',
api=BigML(my_username,
my_api_key
storage="my_storage"))
If no connection is provided, a default connection will be
instantiated internally. This default connection will use ./storage
as default storage directory and the credentials used to connect to
the API when needed are retrieved from the BIGML_USERNAME and
BIGML_API_KEY environment variables. If no credentials are found in your
environment, any attempt to download the information will raise a condition
asking the user to set these variables.
If a connection with no storage information is provided, then the models
will never be stored in your local file system, and will be retrieved from
BigML’s API each time the local model is instantiated.
Ensembles and composite objects, like Fusions, need more than one resource
to be downloaded and stored locally for the class to work. In this case,
the class needs all the component models,
so providing only a local file or a dictionary containing the
JSON for the resource is not enough for the Ensemble or Fusion
objects to be fully instantiated. If you only provide that partial information,
the class will use the internal API connection the first time
to download the components.
However, using the api.export method for ensembles or fusions
will download these component models for you
and will store them in the same directory as the file used to store
the ensemble or fusion information. After that, you can
instantiate the object using the path to the file where the ensemble
or fusion information was stored. The class will look internally for the
rest of components in the same directory and find them, so no connection to
the API will be done.
If you use a tag to label the resource, you can also ask for the last resource that has the tag:
api.export_last('my_tag',
resource_type='ensemble',
'filename': 'my_dir/my_ensemble.json')
and even for a resource inside a project:
api.export_last('my_tag',
resource_type='dataset',
project='project/5143a51a37203f2cf7000959',
'filename': 'my_dir/my_dataset.json')
Local Datasets¶
You can instantiate a local version of a dataset so that you can reproduce its transformations to generate new fields using Flatline expressions.
from bigml.dataset import Dataset
local_dataset = Dataset('dataset/502fdbff15526876610003215')
This will retrieve the remote dataset information, using an implicitly built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return a Dataset object
that will be stored in the ./storage directory. If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second parameter:
from bigml.dataset import Dataset
from bigml.api import BigML
local_dataset = Dataset('dataset/502fdbff15526876610003215',
api=BigML(my_username,
my_api_key,
storage="my_storage"))
or even use the remote dataset information previously retrieved to build the local dataset object:
from bigml.dataset import Dataset
from bigml.api import BigML
api = BigML()
dataset = api.get_dataset('dataset/502fdbff15526876610003215',
query_string='limit=-1')
local_dataset = Dataset(dataset)
As you can see, the query_string used to retrieve the dataset is
limit=-1, which avoids the pagination of fields that is used by default and
includes them all at once. These details are already taken care of in the
two previous examples, where the dataset ID is used as argument.
You can also build a local dataset from a dataset previously retrieved and stored in a JSON file:
from bigml.dataset import Dataset
local_dataset = Dataset('./my_dataset.json')
Adding new properties to an existing dataset is achieved by
defining some expressions based on the fields
of a previously existing origin dataset. The expressions are written using
the Flatline language. These transformations are
stored in a new_fields attribute and the
Dataset object will store them, if available.
That information can be used to reproduce the same transformations
using new inputs. Of course, the fields in the input data to be transformed
are expected to match the fields structure of the dataset that was
used as origin to create the present one.
from bigml.dataset import Dataset
local_dataset = Dataset('./my_dataset.json')
# The dataset in my_dataset.json was created from a dataset whose fields
# were ``foo`` and ``baz``. The transformation that generated the new
# dataset added a new field ``qux`` whose value is ``baz`` divided by 2
input_data_list = [{"foo": "bar", "baz": 32}]
output_data_list = local_dataset.transform(input_data_list)
# output_data_list: [{"foo": "bar", "baz": 32, "qux": 16}]
The Dataset object offers a method to download a sample of the rows
that can be found in the dataset.
from bigml.dataset import Dataset
local_dataset = Dataset('dataset/502fdbff15526876610003215')
rows = local_dataset.get_sample(rows_number=50)
The result will be a list of lists, which are the row values sorted as
described in the fields structure of the dataset. Of course,
this operation cannot be performed locally. BigML’s API will be
called behind the scene to create a Sample object and retrieve the
corresponding rows. Similarly, you can use the get_input_sample
method to get a sample of rows of the origin dataset (if available in BigML).
from bigml.dataset import Dataset
local_dataset = Dataset('dataset/502fdbff15526876610003215')
rows = local_dataset.get_input_sample(rows_number=50)
# these rows will represent the values available in the dataset
# that was used as origin to create dataset/502fdbff15526876610003215
Local Models¶
You can instantiate a local version of a remote model.
from bigml.model import Model
local_model = Model('model/502fdbff15526876610002615')
This will retrieve the remote model information, using an implicitly built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return a Model object
that will be stored in the ./storage directory and
you can use to make local predictions. If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second parameter:
from bigml.model import Model
from bigml.api import BigML
local_model = Model('model/502fdbff15526876610002615',
api=BigML(my_username,
my_api_key,
storage="my_storage"))
or even use the remote model information previously retrieved to build the local model object:
from bigml.model import Model
from bigml.api import BigML
api = BigML()
model = api.get_model('model/502fdbff15526876610002615',
query_string='only_model=true;limit=-1')
local_model = Model(model)
As you can see, the query_string used to retrieve the model has two parts.
They both act on the fields
information that is added to the JSON response. First
only_model=true is used to restrict the fields described in the
fields structure of the response to those used as
predictors in the model. Also
limit=-1 avoids the pagination of fields which is used by default and
includes them all at once. These details are already taken care of in the
two previous examples, where the model ID is used as argument.
Any of these methods will return a Model object that you can use to make
local predictions, generate IF-THEN rules, Tableau rules
or a Python function that implements the model.
You can also build a local model from a model previously retrieved and stored in a JSON file:
from bigml.model import Model
local_model = Model('./my_model.json')
Local Predictions¶
Once you have a local model you can use to generate predictions locally.
local_model.predict({"petal length": 3, "petal width": 1})
Iris-versicolor
Local predictions have three clear advantages:
Removing the dependency from BigML to make new predictions.
No cost (i.e., you do not spend BigML credits).
Extremely low latency to generate predictions for huge volumes of data.
The default output for local predictions is the prediction itself, but you can
also add other properties associated to the prediction, like its
confidence or probability, the distribution of values in the predicted node
(for decision tree models), and the number of instances supporting the
prediction. To obtain a
dictionary with the prediction and the available additional
properties use the full=True argument:
local_model.predict({"petal length": 3, "petal width": 1}, full=True)
that will return:
{'count': 47,
'confidence': 0.92444,
'probability': 0.9861111111111112,
'prediction': u'Iris-versicolor',
'distribution_unit': 'categories',
'path': [u'petal length > 2.45',
u'petal width <= 1.75',
u'petal length <= 4.95',
u'petal width <= 1.65'],
'distribution': [[u'Iris-versicolor', 47]]}
Note that the path attribute for the proportional missing strategy
shows the path leading to a final unique node, that gives the prediction, or
to the first split where a missing value is found. Other optional
attributes are
next which contains the field that determines the next split after
the prediction node and distribution that adds the distribution
that leads to the prediction. For regression models, min and
max will add the limit values for the data that supports the
prediction.
When your test data has missing values, you can choose between last
prediction or proportional strategy to compute the
prediction. The last prediction strategy is the one used by
default. To compute a prediction, the algorithm goes down the model’s
decision tree and checks the condition it finds at each node (e.g.:
‘sepal length’ > 2). If the field checked is missing in your input
data you have two options: by default (last prediction strategy)
the algorithm will stop and issue the last prediction it computed in
the previous node. If you chose proportional strategy instead, the
algorithm will continue to go down the tree considering both branches
from that node on. Thus, it will store a list of possible predictions
from then on, one per valid node. In this case, the final prediction
will be the majority (for categorical models) or the average (for
regressions) of values predicted by the list of predicted values.
You can set this strategy by using the missing_strategy
argument with code 0 to use last prediction and 1 for
proportional.
from bigml.model import LAST_PREDICTION, PROPORTIONAL
# LAST_PREDICTION = 0; PROPORTIONAL = 1
local_model.predict({"petal length": 3, "petal width": 1},
missing_strategy=PROPORTIONAL)
For classification models, it is sometimes useful to obtain a
probability or confidence prediction for each possible class of the
objective field. To do this, you can use the predict_probability
and predict_confidence methods respectively. The former gives a
prediction based on the distribution of instances at the appropriate
leaf node, with a Laplace correction based on the root node
distribution. The latter returns a lower confidence bound on the leaf
node probability based on the Wilson score interval.
Each of these methods take the missing_strategy
argument that functions as it does in predict, and one additional
argument, compact. If compact is False (the default), the
output of these functions is a list of maps, each with the keys
prediction and probability (or confidence) mapped to the
class name and its associated probability (or confidence). Note that these
methods substitute the deprecated multiple parameter in the predict
method functionallity.
So, for example, the following:
local_model.predict_probability({"petal length": 3})
would result in
[{'prediction': u'Iris-setosa',
'probability': 0.0033003300330033},
{'prediction': u'Iris-versicolor',
'probability': 0.4983498349834984},
{'prediction': u'Iris-virginica',
'probability': 0.4983498349834984}]
If compact is True, only the probabilities themselves are
returned, as a list in class name order. Note that, for reference,
the attribute Model.class_names contains the class names in the
appropriate ordering.
To illustrate, the following:
local_model.predict_probability({"petal length": 3}, compact=True)
would result in
[0.0033003300330033, 0.4983498349834984, 0.4983498349834984]
The output of predict_confidence is the same, except that the
output maps are keyed with confidence instead of probability.
For classifications, the prediction of a local model will be one of the
available categories in the objective field and an associated confidence
or probability that is used to decide which is the predicted category.
If you prefer the model predictions to be operated using any of them, you can
use the operating_kind argument in the predict method.
Here’s the example
to use predictions based on confidence:
local_model.predict({"petal length": 3, "petal width": 1},
{"operating_kind": "confidence"})
Previous versions of the bindings had additional arguments in the predict
method that were used to format the prediction attributes. The signature of
the method has been changed to accept only arguments that affect the
prediction itself, (like missing_strategy, operating_kind and
opreating_point) and full which is a boolean that controls whether
the output is the prediction itself or a dictionary will all the available
properties associated to the prediction. Formatting can be achieved by using
the cast_prediction function:
def cast_prediction(full_prediction, to=None,
confidence=False, probability=False,
path=False, distribution=False,
count=False, next=False, d_min=False,
d_max=False, median=False,
unused_fields=False):
whose first argument is the prediction obtained with the full=True
argument, the second one defines the type of output (None to obtain
the prediction output only, “list” or “dict”) and the rest of booleans
cause the corresponding property to be included or not.
Operating point’s predictions¶
In classification problems, Models, Ensembles and Logistic Regressions can be used at different operating points, that is, associated to particular thresholds. Each operating point is then defined by the kind of property you use as threshold, its value and a the class that is supposed to be predicted if the threshold is reached.
Let’s assume you decide that you have a binary problem, with classes True
and False as possible outcomes. Imagine you want to be very sure to
predict the True outcome, so you don’t want to predict that unless the
probability associated to it is over 0.8. You can achieve this with any
classification model by creating an operating point:
operating_point = {"kind": "probability",
"positive_class": "True",
"threshold": 0.8};
to predict using this restriction, you can use the operating_point
parameter:
prediction = local_model.predict(input_data,
operating_point=operating_point)
where inputData should contain the values for which you want to predict.
Local models allow two kinds of operating points: probability and
confidence. For both of them, the threshold can be set to any number
in the [0, 1] range.
Local feature generation for predictions¶
All kind of local models (ensembles, clusters, etc.) offer a prediction-like method that receives the input data to be used as test data and produces the prediction output (prediction, centroid, etc.). However, one of BigML’s capabilities is automatic feature extraction from date-time or image fields. Also, the Flatline language allows the user to create new features to from the raw data to be used in modelling. Thus, your model might use features that have been derived from the original raw data and should be replicated at prediction time.
Local pipelines are objects that will store all the
feature extraction and transformations used to produce the dataset that was
used for training (see Local Pipelines).
These objects provide a .transform method that can be
applied to the raw input data to reproduce the same transformations that
were used to define the training data used by the model from the raw training
data. Every local model class offers a .data_transformations method that
generates a BMLPipeline object, storing these transformations.
The user can apply them before calling the corresponding prediction method.
from bigml.model import Model
local_model = Model('model/502fdbff15526876610002435')
local_pipeline = local_model.data_transformations()
# the pipeline transform method is applied to lists of dictionaries
# (one row per dictionary).
# For a single prediction, a list of one input is sent to be
# transformed and the result will be a list, whose
# first element is used as transformed input data
input_data = local_pipeline.transform(
[{"petal length": 4.4, "sepal width": 3.2}])[0]
prediction = local_model.predict(input_data)
Local Clusters¶
You can also instantiate a local version of a remote cluster.
from bigml.cluster import Cluster
local_cluster = Cluster('cluster/502fdbff15526876610002435')
This will retrieve the remote cluster information, using an implicitly built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return a Cluster object
that will be stored in the ./storage directory and
you can use to make local centroid predictions. If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second
parameter:
from bigml.cluster import Cluster
from bigml.api import BigML
local_cluster = Cluster('cluster/502fdbff15526876610002435',
api=BigML(my_username,
my_api_key
storage="my_storage"))
or even use the remote cluster information previously retrieved to build the local cluster object:
from bigml.cluster import Cluster
from bigml.api import BigML
api = BigML()
cluster = api.get_cluster('cluster/502fdbff15526876610002435',
query_string='limit=-1')
local_cluster = Cluster(cluster)
Note that in this example we used a limit=-1 query string for the cluster
retrieval. This ensures that all fields are retrieved by the get method in the
same call (unlike in the standard calls where the number of fields returned is
limited).
Local clusters provide also methods for the significant operations that can be done using clusters: finding the centroid assigned to a certain data point, sorting centroids according to their distance to a data point, summarizing the centroids intra-distances and inter-distances and also finding the closest points to a given one. The Local Centroids and the Summary generation sections will explain these methods.
Local Centroids¶
Using the local cluster object, you can predict the centroid associated to an input data set:
local_cluster.centroid({"pregnancies": 0, "plasma glucose": 118,
"blood pressure": 84, "triceps skin thickness": 47,
"insulin": 230, "bmi": 45.8,
"diabetes pedigree": 0.551, "age": 31,
"diabetes": "true"})
{'distance': 0.454110207355, 'centroid_name': 'Cluster 4',
'centroid_id': '000004'}
You must keep in mind, though, that to obtain a centroid prediction, input data
must have values for all the numeric fields. No missing values for the numeric
fields are allowed unless you provided a default_numeric_value in the
cluster construction configuration. If so, this value will be used to fill
the missing numeric fields.
As in the local model predictions, producing local centroids can be done independently of BigML servers, so no cost or connection latencies are involved.
Another interesting method in the cluster object is
local_cluster.closests_in_cluster, which given a reference data point
will provide the rest of points that fall into the same cluster sorted
in an ascending order according to their distance to this point. You can limit
the maximum number of points returned by setting the number_of_points
argument to any positive integer.
local_cluster.closests_in_cluster( \
{"pregnancies": 0, "plasma glucose": 118,
"blood pressure": 84, "triceps skin thickness": 47,
"insulin": 230, "bmi": 45.8,
"diabetes pedigree": 0.551, "age": 31,
"diabetes": "true"}, number_of_points=2)
The response will be a dictionary with the centroid id of the cluster an the list of closest points and their distances to the reference point.
{'closest': [ \
{'distance': 0.06912270988567025,
'data': {'plasma glucose': '115', 'blood pressure': '70',
'triceps skin thickness': '30', 'pregnancies': '1',
'bmi': '34.6', 'diabetes pedigree': '0.529',
'insulin': '96', 'age': '32', 'diabetes': 'true'}},
{'distance': 0.10396456577958413,
'data': {'plasma glucose': '167', 'blood pressure': '74',
'triceps skin thickness': '17', 'pregnancies': '1', 'bmi': '23.4',
'diabetes pedigree': '0.447', 'insulin': '144', 'age': '33',
'diabetes': 'true'}}],
'reference': {'age': 31, 'bmi': 45.8, 'plasma glucose': 118,
'insulin': 230, 'blood pressure': 84,
'pregnancies': 0, 'triceps skin thickness': 47,
'diabetes pedigree': 0.551, 'diabetes': 'true'},
'centroid_id': u'000000'}
No missing numeric values are allowed either in the reference data point.
If you want the data points to belong to a different cluster, you can
provide the centroid_id for the cluster as an additional argument.
Other utility methods are local_cluster.sorted_centroids which given
a reference data point will provide the list of centroids sorted according
to the distance to it
local_cluster.sorted_centroids( \
{'plasma glucose': '115', 'blood pressure': '70',
'triceps skin thickness': '30', 'pregnancies': '1',
'bmi': '34.6', 'diabetes pedigree': '0.529',
'insulin': '96', 'age': '32', 'diabetes': 'true'})
{'centroids': [{'distance': 0.31656890408929705,
'data': {u'000006': 0.34571, u'000007': 30.7619,
u'000000': 3.79592, u'000008': u'false'},
'centroid_id': u'000000'},
{'distance': 0.4424198506958207,
'data': {u'000006': 0.77087, u'000007': 45.50943,
u'000000': 5.90566, u'000008': u'true'},
'centroid_id': u'000001'}],
'reference': {'age': '32', 'bmi': '34.6', 'plasma glucose': '115',
'insulin': '96', 'blood pressure': '70',
'pregnancies': '1', 'triceps skin thickness': '30',
'diabetes pedigree': '0.529', 'diabetes': 'true'}}
or points_in_cluster that returns the list of
data points assigned to a certain cluster, given its centroid_id.
centroid_id = "000000"
local_cluster.points_in_cluster(centroid_id)
Local Anomaly Detector¶
You can also instantiate a local version of a remote anomaly.
from bigml.anomaly import Anomaly
local_anomaly = Anomaly('anomaly/502fcbff15526876610002435')
This will retrieve the remote anomaly detector information, using an implicitly
built BigML() connection object (see the Authentication
section for
more details on how to set your credentials) and return an Anomaly object
that will be stored in the ./storage directory and
you can use to make local anomaly scores. If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second
parameter:
from bigml.anomaly import Anomaly
from bigml.api import BigML
local_anomaly = Anomaly('anomaly/502fcbff15526876610002435',
api=BigML(my_username,
my_api_key,
storage="my_storage_dir"))
or even use the remote anomaly information retrieved previously to build the local anomaly detector object:
from bigml.anomaly import Anomaly
from bigml.api import BigML
api = BigML()
anomaly = api.get_anomaly('anomaly/502fcbff15526876610002435',
query_string='limit=-1')
local_anomaly = Anomaly(anomaly)
Note that in this example we used a limit=-1 query string for the anomaly
retrieval. This ensures that all fields are retrieved by the get method in the
same call (unlike in the standard calls where the number of fields returned is
limited).
The anomaly detector object has also the method anomalies_filter
that will build the LISP filter you would need to filter the original
dataset and create a new one excluding
the top anomalies. Setting the include parameter to True you can do the
inverse and create a dataset with only the most anomalous data points.
Local Anomaly Scores¶
Using the local anomaly detector object, you can predict the anomaly score associated to an input data set:
local_anomaly.anomaly_score({"src_bytes": 350})
0.9268527808726705
As in the local model predictions, producing local anomaly scores can be done independently of BigML servers, so no cost or connection latencies are involved.
Local Anomaly caching¶
Anomalies can become quite large objects. That’s why their use of memory
resources can be heavy. If your usual scenario is using many of them
constantly in a disordered way, the best strategy is setting up a cache
system to store them. The local anomaly class provides helpers to
interact with that cache. Here’s an example using Redis.
from anomaly import Anomaly
import redis
r = redis.Redis()
# First build as you would any core Anomaly object:
anomaly = Anomaly('anomaly/5126965515526876630001b2')
# Store a serialized version in Redis
anomaly.dump(cache_set=r.set)
# (retrieve the external rep from its convenient place)
# Speedy Build from external rep
anomaly = Anomaly('anomaly/5126965515526876630001b2', cache_get=r.get)
# Get scores same as always:
anomaly.anomaly_score({"src_bytes": 350})
Local Logistic Regression¶
You can also instantiate a local version of a remote logistic regression.
from bigml.logistic import LogisticRegression
local_log_regression = LogisticRegression(
'logisticregression/502fdbff15526876610042435')
This will retrieve the remote logistic regression information,
using an implicitly built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return a LogisticRegression
object that will be stored in the ./storage directory and
you can use to make local predictions. If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second
parameter:
from bigml.logistic import LogisticRegression
from bigml.api import BigML
local_log_regression = LogisticRegression(
'logisticregression/502fdbff15526876610602435',
api=BigML(my_username, my_api_key, storage="my_storage"))
You can also reuse a remote logistic regression JSON structure as previously retrieved to build the local logistic regression object:
from bigml.logistic import LogisticRegression
from bigml.api import BigML
api = BigML()
logistic_regression = api.get_logistic_regression(
'logisticregression/502fdbff15526876610002435',
query_string='limit=-1')
local_log_regression = LogisticRegression(logistic_regression)
Note that in this example we used a limit=-1 query string for the
logistic regression retrieval. This ensures that all fields are
retrieved by the get method in the same call (unlike in the standard
calls where the number of fields returned is limited).
Local Logistic Regression Predictions¶
Using the local logistic regression object, you can predict the prediction for an input data set:
local_log_regression.predict({"petal length": 2, "sepal length": 1.5,
"petal width": 0.5, "sepal width": 0.7},
full=True)
{'distribution': [
{'category': u'Iris-virginica', 'probability': 0.5041444478857267},
{'category': u'Iris-versicolor', 'probability': 0.46926542042788333},
{'category': u'Iris-setosa', 'probability': 0.02659013168639014}],
'prediction': u'Iris-virginica', 'probability': 0.5041444478857267}
As you can see, the prediction contains the predicted category and the
associated probability. It also shows the distribution of probabilities for
all the possible categories in the objective field. If you only need the
predicted value, you can remove the full argument.
You must keep in mind, though, that to obtain a logistic regression prediction, input data must have values for all the numeric fields. No missing values for the numeric fields are allowed.
For consistency of interface with the Model class, logistic
regressions again have a predict_probability method, which takes
the same argument as Model.predict:
compact. As stated above, missing values are not allowed, and so
there is no missing_strategy argument.
As with local Models, if compact is False (the default), the
output is a list of maps, each with the keys prediction and
probability mapped to the class name and its associated
probability.
So, for example
local_log_regression.predict_probability({"petal length": 2, "sepal length": 1.5,
"petal width": 0.5, "sepal width": 0.7})
[{'category': u'Iris-setosa', 'probability': 0.02659013168639014},
{'category': u'Iris-versicolor', 'probability': 0.46926542042788333},
{'category': u'Iris-virginica', 'probability': 0.5041444478857267}]
If compact is True, only the probabilities themselves are
returned, as a list in class name order, again, as is the case with
local Models.
Operating point predictions are also available for local logistic regressions and an example of it would be:
operating_point = {"kind": "probability",
"positive_class": "True",
"threshold": 0.8}
local_logistic.predict(inputData, operating_point=operating_point)
You can check the
Operating point’s predictions section
to learn about
operating points. For logistic regressions, the only available kind is
probability, that sets the threshold of probability to be reached for the
prediction to be the positive class.
Local Logistic Regression¶
You can also instantiate a local version of a remote logistic regression:
from bigml.logistic import LogisticRegression
local_log_regression = LogisticRegression(
'logisticregression/502fdbff15526876610042435')
This will retrieve the remote logistic regression information,
using an implicitly built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return a LogisticRegression
object that will be stored in the ./storage directory and
you can use to make local predictions. If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second
parameter:
from bigml.logistic import LogisticRegression
from bigml.api import BigML
local_log_regression = LogisticRegression(
'logisticregression/502fdbff15526876610602435',
api=BigML(my_username, my_api_key, storage="my_storage"))
You can also reuse a remote logistic regression JSON structure as previously retrieved to build the local logistic regression object:
from bigml.logistic import LogisticRegression
from bigml.api import BigML
api = BigML()
logistic_regression = api.get_logistic_regression(
'logisticregression/502fdbff15526876610002435',
query_string='limit=-1')
local_log_regression = LogisticRegression(logistic_regression)
Note that in this example we used a limit=-1 query string for the
logistic regression retrieval. This ensures that all fields are
retrieved by the get method in the same call (unlike in the standard
calls where the number of fields returned is limited).
Local Linear Regression Predictions¶
Using the local LinearRegression class, you can predict the prediction for
an input data set:
local_linear_regression.predict({"petal length": 2, "sepal length": 1.5,
"species": "Iris-setosa",
"sepal width": 0.7},
full=True)
{'confidence_bounds': {
'prediction_interval': 0.43783924497784293,
'confidence_interval': 0.2561542783257394},
'prediction': -0.6109005499999999, 'unused_fields': ['petal length']}
To obtain a linear regression prediction, input data can only have missing values for fields that had already some missings in training data.
The full=True in the predict method will cause the prediction to include
confidence bounds when available. Some logistic regressions will not
contain such information by construction. Also, in order to compute these
bounds locally, you will need numpy and scipy in place.
As they are quite heavy libraries, they aren’t automatically installed as
dependencies of these bindings.
Local Deepnet¶
You can also instantiate a local version of a remote Deepnet.
from bigml.deepnet import Deepnet
local_deepnet = Deepnet(
'deepnet/502fdbff15526876610022435')
This will retrieve the remote deepnet information,
using an implicitly built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return a Deepnet
object that will be stored in the ./storage directory and
you can use to make local predictions. If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second
parameter:
from bigml.deepnet import Deepnet
from bigml.api import BigML
local_deepnet = Deepnet(
'deepnet/502fdbff15526876610602435',
api=BigML(my_username, my_api_key, storage="my_storage"))
You can also reuse a remote Deepnet JSON structure as previously retrieved to build the local Deepnet object:
from bigml.deepnet import Deepnet
from bigml.api import BigML
api = BigML()
deepnet = api.get_deepnet(
'deepnet/502fdbff15526876610002435',
query_string='limit=-1')
local_deepnet = Deepnet(deepnet)
Note that in this example we used a limit=-1 query string for the
deepnet retrieval. This ensures that all fields are
retrieved by the get method in the same call (unlike in the standard
calls where the number of fields returned is limited).
Local Deepnet Predictions¶
Using the local deepnet object, you can predict the prediction for an input data set:
local_deepnet.predict({"petal length": 2, "sepal length": 1.5,
"petal width": 0.5, "sepal width": 0.7},
full=True)
{'distribution': [
{'category': u'Iris-virginica', 'probability': 0.5041444478857267},
{'category': u'Iris-versicolor', 'probability': 0.46926542042788333},
{'category': u'Iris-setosa', 'probability': 0.02659013168639014}],
'prediction': u'Iris-virginica', 'probability': 0.5041444478857267}
As you can see, the full prediction contains the predicted category and the
associated probability. It also shows the distribution of probabilities for
all the possible categories in the objective field. If you only need the
predicted value, you can remove the full argument.
To be consistent with the Model class interface, deepnets
have also a predict_probability method, which takes
the same argument as Model.predict:
compact.
As with local Models, if compact is False (the default), the
output is a list of maps, each with the keys prediction and
probability mapped to the class name and its associated
probability.
So, for example
local_deepnet.predict_probability({"petal length": 2, "sepal length": 1.5,
"petal width": 0.5, "sepal width": 0.7})
[{'category': u'Iris-setosa', 'probability': 0.02659013168639014},
{'category': u'Iris-versicolor', 'probability': 0.46926542042788333},
{'category': u'Iris-virginica', 'probability': 0.5041444478857267}]
If compact is True, only the probabilities themselves are
returned, as a list in class name order, again, as is the case with
local Models.
Operating point predictions are also available for local deepnets and an example of it would be:
operating_point = {"kind": "probability",
"positive_class": "True",
"threshold": 0.8};
prediction = local_deepnet.predict(input_data,
operating_point=operating_point)
Local Deepnets for images supervised learning and object detection¶
Deepnets include Convolutional Neural Networks, so they can be used to do classification, regression and object detection based on images. For image classification and regression, the local Deepnets will just need some image as input data when doing predictions. The image file should be provided in input data as the contents to the corresponding image field.
input_data = {"000002": "my_image.jpg"}
prediction = local_deepnet.predict(input_data)
For object detection, as predictions are only based on one image, the input to be provided is the plain image file itself.
prediction = local_deepnet.predict("my_image.jpg")
Also, object detection Deepnets allow some parameters to be set
at creation time. They slightly modify the operation of the Deepnet, so
they are provided as operation_settings.
from bigml.deepnet import Deepnet
local_deepnet = Deepnet("deepnet/62a85964128d1c55610003cd",
operation_settings={"region_score_threshold": 0.6})
prediction = local_deepnet.predict("my_image.jpg")
The operation settings allowed are region_score_threshold, that will set
the minimum accepted score in the predictions and max_objects which will
limit the number of regions returned.
The prediction will contain a list of dictionaries that contain the
label, score and box description of the found regions. Each box object is
an array that contains the xmin, ymin, xmax and ymax
coordinates:
{'prediction': [{'box': [0.67742, 0.30469, 0.79472, 0.37109],
'label': 'eye',
'score': 0.83528},
{'box': [0.3783, 0.27734, 0.50147, 0.35938],
'label': 'eye',
'score': 0.79117},
{'box': [0.67742, 0.77344, 0.739, 0.81445],
'label': 'eye',
'score': 0.45094}]}
Note: Local predictions for deepnets built on images datasets can differ slightly from the predictions obtained by using BigML’s API create prediction call. When uploaded to BigML, images are standardized to a particular resolution and compressed using the JPEG algorithm while local predictions maintain the original image information. That can cause minor variations in regression predictions or the probability associated to classification predictions. Also object detection predictions can differ slightly, specially if low region_threshold_scores are used.
If anything, the local value will always be slightly more accurate, but if you
need to find results as close as possible to the ones produced in remote
predictions, you can use the remote_preprocess function in the deepnet
module.
Local Fusion¶
You can also instantiate a local version of a remote Fusion.
from bigml.fusion import Fusion
local_fusion = Fusion(
'fusion/502fdbff15526876610022438')
This will retrieve the remote fusion information,
using an implicitly built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return a Fusion
object that will be stored in the ./storage directory and
you can use to make local predictions. If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second
parameter:
from bigml.fusion import Fusion
from bigml.api import BigML
local_fusion = Fusion(
'fusion/502fdbff15526876610602435',
api=BigML(my_username, my_api_key, storage="my_storage"))
You can also reuse a remote Fusion JSON structure as previously retrieved to build the local Fusion object:
from bigml.fusion import Fusion
from bigml.api import BigML
api = BigML()
fusion = api.get_fusion(
'fusion/502fdbff15526876610002435',
query_string='limit=-1')
local_fusion = Fusion(fusion)
Note that in this example we used a limit=-1 query string for the
fusion retrieval. This ensures that all fields are
retrieved by the get method in the same call (unlike in the standard
calls where the number of fields returned is limited).
Local Fusion Predictions¶
Using the local fusion object, you can predict the prediction for an input data set:
local_fusion.predict({"petal length": 2, "sepal length": 1.5,
"petal width": 0.5, "sepal width": 0.7},
full=True)
{'prediction': u'Iris-setosa', 'probability': 0.45224}
As you can see, the full prediction contains the predicted category and the
associated probability. If you only need the
predicted value, you can remove the full argument.
To be consistent with the Model class interface, fusions
have also a predict_probability method, which takes
the same argument as Model.predict:
compact.
As with local Models, if compact is False (the default), the
output is a list of maps, each with the keys prediction and
probability mapped to the class name and its associated
probability.
So, for example
local_fusion.predict_probability({"petal length": 2, "sepal length": 1.5,
"petal width": 0.5, "sepal width": 0.7})
[{'category': u'Iris-setosa', 'probability': 0.45224},
{'category': u'Iris-versicolor', 'probability': 0.2854},
{'category': u'Iris-virginica', 'probability': 0.26236}]
If compact is True, only the probabilities themselves are
returned, as a list in class name order, again, as is the case with
local Models.
Operating point predictions are also available with probability as threshold for local fusions and an example of it would be:
operating_point = {"kind": "probability",
"positive_class": "True",
"threshold": 0.8};
prediction = local_fusion.predict(inputData,
operating_point=operating_point)
Local Association¶
You can also instantiate a local version of a remote association resource.
from bigml.association import Association
local_association = Association('association/502fdcff15526876610002435')
This will retrieve the remote association information, using an implicitly
built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return an Association object
that will be stored in the ./storage directory and
you can use to extract the rules found in the original dataset.
If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second
parameter:
from bigml.association import Association
from bigml.api import BigML
local_association = Association('association/502fdcff15526876610002435',
api=BigML(my_username,
my_api_key
storage="my_storage"))
or even use the remote association information retrieved previously to build the local association object:
from bigml.association import Association
from bigml.api import BigML
api = BigML()
association = api.get_association('association/502fdcff15526876610002435',
query_string='limit=-1')
local_association = Association(association)
Note that in this example we used a limit=-1 query string for the
association retrieval. This ensures that all fields are retrieved by the get
method in the
same call (unlike in the standard calls where the number of fields returned is
limited).
The created Association object has some methods to help retrieving the
association rules found in the original data. The get_rules method will
return the association rules. Arguments can be set to filter the rules
returned according to its leverage, strength, support, p_value,
a list of items involved in the rule or a user-given filter function.
from bigml.association import Association
local_association = Association('association/502fdcff15526876610002435')
local_association.get_rules(item_list=["Edible"], min_p_value=0.3)
In this example, the only rules that will be returned by the get_rules
method will be the ones that mention Edible and their p_value
is greater or equal to 0.3.
The rules can also be stored in a CSV file using rules_CSV:
from bigml.association import Association
local_association = Association('association/502fdcff15526876610002435')
local_association.rules_CSV(file_name='/tmp/my_rules.csv',
min_strength=0.1)
This example will store the rules whose strength is bigger or equal to 0.1 in
the /tmp/my_rules.csv file.
You can also obtain the list of items parsed in the dataset using the
get_items method. You can also filter the results by field name, by
item names and by a user-given function:
from bigml.association import Association
local_association = Association('association/502fdcff15526876610002435')
local_association.get_items(field="Cap Color",
names=["Brown cap", "White cap", "Yellow cap"])
This will recover the Item objects found in the Cap Color field for
the names in the list, with their properties as described in the
developers section
Local Association Sets¶
Using the local association object, you can predict the association sets related to an input data set:
local_association.association_set( \
{"gender": "Female", "genres": "Adventure$Action", \
"timestamp": 993906291, "occupation": "K-12 student",
"zipcode": 59583, "rating": 3})
[{'item': {'complement': False,
'count': 70,
'field_id': u'000002',
'name': u'Under 18'},
'rules': ['000000'],
'score': 0.0969181441561211},
{'item': {'complement': False,
'count': 216,
'field_id': u'000007',
'name': u'Drama'},
'score': 0.025050115102862636},
{'item': {'complement': False,
'count': 108,
'field_id': u'000007',
'name': u'Sci-Fi'},
'rules': ['000003'],
'score': 0.02384578264599424},
{'item': {'complement': False,
'count': 40,
'field_id': u'000002',
'name': u'56+'},
'rules': ['000008',
'000020'],
'score': 0.021845366022721312},
{'item': {'complement': False,
'count': 66,
'field_id': u'000002',
'name': u'45-49'},
'rules': ['00000e'],
'score': 0.019657155185835006}]
As in the local model predictions, producing local association sets can be done independently of BigML servers, so no cost or connection latencies are involved.
Local Topic Model¶
You can also instantiate a local version of a remote topic model.
from bigml.topicmodel import TopicModel
local_topic_model = TopicModel(
'topicmodel/502fdbcf15526876210042435')
This will retrieve the remote topic model information,
using an implicitly built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return a TopicModel
object that will be stored in the ./storage directory and
you can use to obtain local topic distributions.
If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second
parameter:
from bigml.topicmodel import TopicModel
from bigml.api import BigML
local_topic_model = TopicModel(
'topicmodel/502fdbcf15526876210042435',
api=BigML(my_username, my_api_key, storage="my_storage"))
You can also reuse a remote topic model JSON structure as previously retrieved to build the local topic model object:
from bigml.topicmodel import TopicModel
from bigml.api import BigML
api = BigML()
topic_model = api.get_topic_model(
'topicmodel/502fdbcf15526876210042435',
query_string='limit=-1')
local_topic_model = TopicModel(topic_model)
Note that in this example we used a limit=-1 query string for the topic
model retrieval. This ensures that all fields are retrieved by the get
method in the
same call (unlike in the standard calls where the number of fields returned is
limited).
Local Topic Distributions¶
Using the local topic model object, you can predict the local topic distribution for an input data set:
local_topic_model.distribution({"Message": "Our mobile phone is free"})
[ { 'name': u'Topic 00', 'probability': 0.002627154266498529},
{ 'name': u'Topic 01', 'probability': 0.003257671290458176},
{ 'name': u'Topic 02', 'probability': 0.002627154266498529},
{ 'name': u'Topic 03', 'probability': 0.1968263976460698},
{ 'name': u'Topic 04', 'probability': 0.002627154266498529},
{ 'name': u'Topic 05', 'probability': 0.002627154266498529},
{ 'name': u'Topic 06', 'probability': 0.13692728036990331},
{ 'name': u'Topic 07', 'probability': 0.6419714165615805},
{ 'name': u'Topic 08', 'probability': 0.002627154266498529},
{ 'name': u'Topic 09', 'probability': 0.002627154266498529},
{ 'name': u'Topic 10', 'probability': 0.002627154266498529},
{ 'name': u'Topic 11', 'probability': 0.002627154266498529}]
As you can see, the topic distribution contains the name of the possible topics in the model and the associated probabilities.
Local Time Series¶
You can also instantiate a local version of a remote time series.
from bigml.timeseries import TimeSeries
local_time_series = TimeSeries(
'timeseries/502fdbcf15526876210042435')
This will create a series of models from
the remote time series information,
using an implicitly built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return a TimeSeries
object that will be stored in the ./storage directory and
you can use to obtain local forecasts.
If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second
parameter:
from bigml.timeseries import TimeSeries
from bigml.api import BigML
local_time_series = TimeSeries( \
'timeseries/502fdbcf15526876210042435',
api=BigML(my_username, my_api_key, storage="my_storage"))
You can also reuse a remote time series JSON structure as previously retrieved to build the local time series object:
from bigml.timeseries import TimeSeries
from bigml.api import BigML
api = BigML()
time_series = api.get_time_series( \
'timeseries/502fdbcf15526876210042435',
query_string='limit=-1')
local_time_series = TimeSeries(time_series)
Note that in this example we used a limit=-1 query string for the time
series retrieval. This ensures that all fields are retrieved by the get
method in the
same call (unlike in the standard calls where the number of fields returned is
limited).
Local Forecasts¶
Using the local time series object, you can forecast any of the objective field values:
local_time_series.forecast({"Final": {"horizon": 5}, "Assignment": { \
"horizon": 10, "ets_models": {"criterion": "aic", "limit": 2}}})
{u'000005': [
{'point_forecast': [68.53181, 68.53181, 68.53181, 68.53181, 68.53181],
'model': u'A,N,N'}],
u'000001': [{'point_forecast': [54.776650000000004, 90.00943000000001,
83.59285000000001, 85.72403000000001,
72.87196, 93.85872, 84.80786, 84.65522,
92.52545, 88.78403],
'model': u'A,N,A'},
{'point_forecast': [55.882820120000005, 90.5255466567616,
83.44908577909621, 87.64524353046498,
74.32914583152592, 95.12372848262932,
86.69298716626228, 85.31630744944385,
93.62385478607113, 89.06905451921818],
'model': u'A,Ad,A'}]}
As you can see, the forecast contains the ID of the forecasted field, the computed points and the name of the models meeting the criterion. For more details about the available parameters, please check the API documentation.
Local PCAs¶
The PCA class will create a local version of a remote PCA.
from bigml.pca import PCA
local_pca = PCA(
'pca/502fdbcf15526876210042435')
This will create an object that stores the remote information that defines
the PCA, needed to generate
projections to the new dimensionally reduced components. The remote resource
is automatically downloaded the first time the the PCA is instantiated by
using an implicitly built
BigML() connection object (see the
Authentication section for more
details on how to set your credentials). The JSON that contains this
information is stored in a ./storage directory, which is the default
choice. If you want to use a
specific connection object to define the credentials for the authentication
in BigML or the directory where the JSON information is stored,
you can set it as the second parameter:
from bigml.pca import PCA
from bigml.api import BigML
local_pca = PCA( \
'timeseries/502fdbcf15526876210042435',
api=BigML(my_username, my_api_key, storage="my_storage"))
You can also reuse a remote PCA JSON structure as previously retrieved to build the local PCA object:
from bigml.pca import PCA
from bigml.api import BigML
api = BigML()
time_series = api.get_pca( \
'pca/502fdbcf15526876210042435',
query_string='limit=-1')
local_pca = PCA(pca)
Note that in this example we used a limit=-1 query string for the PCA
retrieval. This ensures that all fields are retrieved by the get
method in the
same call (unlike in the standard calls where the number of fields returned is
limited).
Local Projections¶
Using the local PCA object, you can compute the projection of an input dataset into the new components:
local_pca.projection({"species": "Iris-versicolor"})
[6.03852, 8.35456, 5.04432, 0.75338, 0.06787, 0.03018]
You can use the max_components and variance_threshold arguments
to limit the number of components generated. You can also use the full
argument to produce a dictionary whose keys are the names of the generated
components.
local_pca.projection({"species": "Iris-versicolor"}, full=yes)
{'PCA1': 6.03852, 'PCA2': 8.35456, 'PCA3': 5.04432, 'PCA4': 0.75338,
'PCA5': 0.06787, 'PCA6': 0.03018}
As in the local model predictions, producing local projections can be done independently of BigML servers, so no cost or connection latencies are involved.
Local Forecasts¶
Using the local time series object, you can forecast any of the objective field values:
local_time_series.forecast({"Final": {"horizon": 5}, "Assignment": { \
"horizon": 10, "ets_models": {"criterion": "aic", "limit": 2}}})
{u'000005': [
{'point_forecast': [68.53181, 68.53181, 68.53181, 68.53181, 68.53181],
'model': u'A,N,N'}],
u'000001': [{'point_forecast': [54.776650000000004, 90.00943000000001,
83.59285000000001, 85.72403000000001,
72.87196, 93.85872, 84.80786, 84.65522,
92.52545, 88.78403],
'model': u'A,N,A'},
{'point_forecast': [55.882820120000005, 90.5255466567616,
83.44908577909621, 87.64524353046498,
74.32914583152592, 95.12372848262932,
86.69298716626228, 85.31630744944385,
93.62385478607113, 89.06905451921818],
'model': u'A,Ad,A'}]}
As you can see, the forecast contains the ID of the forecasted field, the computed points and the name of the models meeting the criterion. For more details about the available parameters, please check the API documentation.
Multi Models¶
Multi Models use a numbers of BigML remote models to build a local version that can be used to generate predictions locally. Predictions are generated combining the outputs of each model.
from bigml.api import BigML
from bigml.multimodel import MultiModel
api = BigML()
model = MultiModel([api.get_model(model['resource']) for model in
api.list_models(query_string="tags__in=my_tag")
['objects']])
model.predict({"petal length": 3, "petal width": 1})
This will create a multi model using all the models that have been previously
tagged with my_tag and predict by combining each model’s prediction.
The combination method used by default is plurality for categorical
predictions and mean value for numerical ones. You can also use confidence
weighted:
model.predict({"petal length": 3, "petal width": 1}, method=1)
that will weight each vote using the confidence/error given by the model
to each prediction, or even probability weighted:
model.predict({"petal length": 3, "petal width": 1}, method=2)
that weights each vote by using the probability associated to the training distribution at the prediction node.
There’s also a threshold method that uses an additional set of options:
threshold and category. The category is predicted if and only if
the number of predictions for that category is at least the threshold value.
Otherwise, the prediction is plurality for the rest of predicted values.
An example of threshold combination method would be:
model.predict({'petal length': 0.9, 'petal width': 3.0}, method=3,
options={'threshold': 3, 'category': 'Iris-virginica'})
When making predictions on a test set with a large number of models,
batch_predict can be useful to log each model’s predictions in a
separated file. It expects a list of input data values and the directory path
to save the prediction files in.
model.batch_predict([{"petal length": 3, "petal width": 1},
{"petal length": 1, "petal width": 5.1}],
"data/predictions")
The predictions generated for each model will be stored in an output
file in data/predictions using the syntax
model_[id of the model]__predictions.csv. For instance, when using
model/50c0de043b563519830001c2 to predict, the output file name will be
model_50c0de043b563519830001c2__predictions.csv. An additional feature is
that using reuse=True as argument will force the function to skip the
creation of the file if it already exists. This can be
helpful when using repeatedly a bunch of models on the same test set.
model.batch_predict([{"petal length": 3, "petal width": 1},
{"petal length": 1, "petal width": 5.1}],
"data/predictions", reuse=True)
Prediction files can be subsequently retrieved and converted into a votes list
using batch_votes:
model.batch_votes("data/predictions")
which will return a list of MultiVote objects. Each MultiVote contains a list
of predictions (e.g. [{'prediction': u'Iris-versicolor', 'confidence': 0.34,
'order': 0}, {'prediction': u'Iris-setosa', 'confidence': 0.25,
'order': 1}]).
These votes can be further combined to issue a final
prediction for each input data element using the method combine
for multivote in model.batch_votes("data/predictions"):
prediction = multivote.combine()
Again, the default method of combination is plurality for categorical
predictions and mean value for numerical ones. You can also use confidence
weighted:
prediction = multivote.combine(1)
or probability weighted:
prediction = multivote.combine(2)
You can also get a confidence measure for the combined prediction:
prediction = multivolte.combine(0, with_confidence=True)
For classification, the confidence associated to the combined prediction
is derived by first selecting the model’s predictions that voted for the
resulting prediction and computing the weighted average of their individual
confidence. Nevertheless, when probability weighted is used,
the confidence is obtained by using each model’s distribution at the
prediction node to build a probability distribution and combining them.
The confidence is then computed as the wilson score interval of the
combined distribution (using as total number of instances the sum of all
the model’s distributions original instances at the prediction node)
In regression, all the models predictions’ confidences contribute to the weighted average confidence.
Local Ensembles¶
Remote ensembles can also be used locally through the Ensemble
class. The simplest way to access an existing ensemble and using it to
predict locally is:
from bigml.ensemble import Ensemble
ensemble = Ensemble('ensemble/5143a51a37203f2cf7020351')
ensemble.predict({"petal length": 3, "petal width": 1})
This is the simpler method to create a local Ensemble. The
Ensemble('ensemble/5143a51a37203f2cf7020351') constructor, that fetches
all the related JSON files and stores them in an ./storage directory. Next
calls to Ensemble('ensemble/50c0de043b5635198300033c') will retrieve the
files from this local storage, so that internet connection will only be needed
the first time an Ensemble is built.
However, that method can only be used to work with the ensembles in our
account in BigML. If we intend to use ensembles created under an
Organization, then
we need to provide the information about the project that the ensemble
is included in. You need to provide a connection object for that:
from bigml.ensemble import Ensemble
from bigml.api import BigML
# connection object that informs about the project ID and the
# directory where the ensemble will be stored for local use
api = BigML(project="project/5143a51a37203f2cf7020001",
storage="my_storage_directory")
ensemble = Ensemble('ensemble/5143a51a37203f2cf7020351', api=api)
ensemble.predict({"petal length": 3, "petal width": 1})
The local ensemble object can be used to manage the
three types of ensembles: Decision Forests (bagging or random) and
the ones using Boosted Trees. Also, you can choose
the storage directory or even avoid storing at all. The àpi connection
object controls the storage strategy through the storage argument.
from bigml.api import BigML
from bigml.ensemble import Ensemble
# api connection using a user-selected storage
api = BigML(storage='./my_storage')
# creating ensemble
ensemble = api.create_ensemble('dataset/5143a51a37203f2cf7000972')
# Ensemble object to predict
ensemble = Ensemble(ensemble, api)
ensemble.predict({"petal length": 3, "petal width": 1},
operating_kind="votes")
In this example, we create
a new ensemble and store its information in the ./my_storage
folder. Then this information is used to predict locally using the number of
votes (one per model) backing each category.
The operating_kind argument overrides the legacy method argument, which
was previously used to define the combiner for the models predictions.
Similarly, local ensembles can also be created by giving a list of models to be combined to issue the final prediction (note: only random decision forests and bagging ensembles can be built using this method):
from bigml.ensemble import Ensemble
ensemble = Ensemble(['model/50c0de043b563519830001c2', \
'model/50c0de043b5635198300031b')]
ensemble.predict({"petal length": 3, "petal width": 1})
or even a JSON file that contains the ensemble resource:
from bigml.api import BigML
api = BigML()
api.export("ensemble/50c0de043b5635198300033c",
"my_directory/my_ensemble.json")
from bigml.ensemble import Ensemble
local_ensemble = Ensemble("./my_directory/my_ensemble.json")
Note: the ensemble JSON structure is not self-contained, meaning that it
contains references to the models that the ensemble is build of, but not the
information of the models themselves.
To use an ensemble locally with no connection to
the internet, you must make sure that not only a local copy of the ensemble
JSON file is available in your computer, but also the JSON files corresponding
to the models in it. The export method takes care of storing the
information of every model in the ensemble and storing it in the same directory
as the ensemble JSON file. The Ensemble class will also look up for the
model files in the same directory when using a path to an ensemble file as
argument.
If you have no memory limitations you can create the ensemble from a list of local model objects. Then, local model objects will be always in memory and will only be instantiated once. This will increase performance for large ensembles:
from bigml.model import Model
model_ids = ['model/50c0de043b563519830001c2', \
'model/50c0de043b5635198300031b']
local_models = [Model(model_id) for model_id in model_ids]
local_ensemble = Ensemble(local_models)
Local Ensemble caching¶
Ensembles can become quite large objects and demand large memory resources.
If your usual scenario is using many of them
constantly in a disordered way, the best strategy is setting up a cache
system to store them. The local ensemble class provides helpers to
interact with that cache. Here’s an example using Redis.
from ensemble import Ensemble
import redis
r = redis.Redis()
# First build as you would any core Ensemble object:
local_ensemble = Ensemble('ensemble/5126965515526876630001b2')
# Store a serialized version in Redis
ensemble.dump(cache_set=r.set)
# (retrieve the external rep from its convenient place)
# Speedy Build from external rep
local_ensemble = Ensemble('ensemble/5126965515526876630001b2', \
cache_get=r.get)
# Get scores same as always:
local_ensemble.predict({"src_bytes": 350})
Local Ensemble’s Predictions¶
As in the local model’s case, you can use the local ensemble to create
new predictions for your test data, and set some arguments to configure
the final output of the predict method.
The predictions’ structure will vary depending on the kind of
ensemble used. For Decision Forests local predictions will just contain
the ensemble’s final prediction if no other argument is used.
from bigml.ensemble import Ensemble
ensemble = Ensemble('ensemble/5143a51a37203f2cf7020351')
ensemble.predict({"petal length": 3, "petal width": 1})
u'Iris-versicolor'
The final prediction of an ensemble is determined
by aggregating or selecting the predictions of the individual models therein.
For classifications, the most probable class is returned if no especial
operating method is set. Using full=True you can see both the predicted
output and the associated probability:
from bigml.ensemble import Ensemble
ensemble = Ensemble('ensemble/5143a51a37203f2cf7020351')
ensemble.predict({"petal length": 3, "petal width": 1}, \
full=True)
{'prediction': u'Iris-versicolor',
'probability': 0.98566}
In general, the prediction in a classification
will be one amongst the list of categories in the objective
field. When each model in the ensemble
is used to predict, each category has a confidence, a
probability or a vote associated to this prediction.
Then, through the collection
of models in the
ensemble, each category gets an averaged confidence, probabiity and number of
votes. Thus you can decide whether to operate the ensemble using the
confidence, the probability or the votes so that the predicted
category is the one that scores higher in any of these quantities. The
criteria can be set using the operating_kind option (default is set to
probability):
ensemble.predict({"petal length": 3, "petal width": 1}, \
operating_kind="votes")
Regression will generate a predictiona and an associated error, however
Boosted Trees don’t have an associated confidence measure, so
only the prediction will be obtained in this case.
For consistency of interface with the Model class, as well as
between boosted and non-boosted ensembles, local Ensembles again have
a predict_probability method. This takes the same optional
arguments as Model.predict: missing_strategy and
compact. As with local Models, if compact is False (the default),
the output is a list of maps, each with the keys prediction and
probability mapped to the class name and its associated
probability.
So, for example:
ensemble.predict_probability({"petal length": 3, "petal width": 1})
[{'category': u'Iris-setosa', 'probability': 0.006733220044732548},
{'category': u'Iris-versicolor', 'probability': 0.9824478534614787},
{'category': u'Iris-virginica', 'probability': 0.0108189264937886}]
If compact is True, only the probabilities themselves are
returned, as a list in class name order, again, as is the case with
local Models.
Operating point predictions are also available for local ensembles and an example of it would be:
operating_point = {"kind": "probability",
"positive_class": "True",
"threshold": 0.8};
prediction = local_ensemble.predict(inputData,
operating_point=operating_point)
You can check the
Operating point’s predictions section
to learn about
operating points. For ensembles, three kinds of operating points are available:
votes, probability and confidence. Votes will use as threshold
the number of models in the ensemble that vote for the positive class.
The other two are already explained in the above mentioned section.
Local Ensemble Predictor¶
Predictions can take longer when the ensemble is formed by a large number of
models or when its models have a high number of nodes. In these cases,
predictions’ speed can be increased and memory usage minimized by using the
EnsemblePredictor object. The basic example to build it is:
from bigml.ensemblepredictor import EnsemblePredictor
ensemble = EnsemblePredictor('ensemble/5143a51a37203f2cf7020351',
"./model_fns_directory")
ensemble.predict({"petal length": 3, "petal width": 1}, full=True)
{'prediction': u'Iris-versicolor', 'confidence': 0.91519}
This constructor has two compulsory attributes: then ensemble ID (or the
corresponding API response) and the path to a directory that contains a file
per each of the ensemble models. Each file stores the predict function
needed to obtain the model’s predictions. As in the Ensemble object, you
can also add an api argument with the connection to be used to download
the ensemble’s JSON information.
The functions stored in this directory are generated automatically the first time you instantiate the ensemble. Once they are generated, the functions are retrieved from the directory.
Note that only last prediction missings strategy is available for these
predictions and the combiners available are plurality, confidence and
distribution but no operating_kind or operating_point options
are provided at present.
Local Supervised Model¶
There’s a general class that will allow you to predict using any supervised model resource, regardless of its particular type (model, ensemble, logistic regression, linear regression or deepnet).
The SupervisedModel object will retrieve the resource information and
instantiate the corresponding local object, so that you can use its
predict method to produce local predictions:
from bigml.supervised import SupervisedModel
local_supervised_1 = SupervisedModel( \
"logisticregression/5143a51a37203f2cf7020351")
local_supervised_2 = SupervisedModel( \
"model/5143a51a37203f2cf7020351")
input_data = {"petal length": 3, "petal width": 1}
logistic_regression_prediction = local_supervised_1.predict(input_data)
model_prediction = local_supervised_2.predict(input_data)
Local BigML Model¶
Following the approach of the local SupervisedModel class, the LocalModel
class will allow you to predict using any BigML model resource,
either supervised or unsupervised.
This class provides two methods: predict and batch_predict with
total abstraction as to the result of the predictions
(real predictions, centroids, anomaly scores, etc.), their parameters and the
format of the prediction result.
The predict method can be used on any type of
model and delegates to the specific method of each local model class.
Therefore, it will be the programmers responsibility to provide
only the parameters accepted in the low level
method and the response will be a dictionary whose contents will vary depending
on the type of prediction. Similarly, the batch_predict method
accepts a list of inputs and adds the prediction information to each
element of the list.
The LocalModel object will retrieve the resource information and
instantiate the corresponding local object, so that you can use its
predict method to produce local predictions:
from bigml.local_model import LocalModel
local_model_1 = LocalModel( \
"logisticregression/5143a51a37203f2cf7020351")
local_model_2 = LocalModel( \
"anomaly/5143a51a37203f2cf7020351")
input_data = {"petal length": 3, "petal width": 1}
logistic_regression_prediction = local_model_1.predict(input_data)
# {"prediction": "Iris-setosa", "probability": 0.56}
anomaly_prediction = local_model_2.predict(input_data)
# {"score": 0.84}
Local Pipelines¶
More often than not, the Machine Learning solution to a problem entails
using data transformations and different models that produce some predictions
or scores. They all are useful information that contributes to the final
Machine Learning based decision. Usually, the training workflow becomes
a sequence of functions, each of which adds new fields to our data: engineered
features, scores, predictions, etc. Of course, once the training sequence
is determined, the same steps will need to be reproduced to create
batch predictions for a new list of test input data rows.
The BMLPipeline class offers the tools to extract that sequence from
the existing BigML objects and create the prediction pipeline.
The first obvious goal that we may have is reproducing the same feature
extraction and transformations that were used when training our data to create
our model. That is achieved by using a BMLPipeline object built
on the training dataset.
from bigml.pipeline.pipeline import BMLPipeline
local_pipeline = BMLPipeline("my transformations pipeline",
["dataset/5143a55637203f2cf7020351"])
Starting from dataset/5143a55637203f2cf7020351
and tracing the previous datasets up till the original source built from
our data, the pipeline will store all the steps that were done
to transform it. Maybe some year, month and day new features were
automatically extracted from our date-time fields, or even
the features corresponding to the histogram of gradients were
obtained from an image field (if your dataset had one of those).
Also, if transformations were defined using Flatline to
generate new fields, they will be detected and stored as a transformation
step. They are all retrieved and ready to be applied to a
list of dictionaries representing your rows information using the
.transform method.
local_pipeline.transform([{"plasma glucose": 130, "bmi":3},
{"age":26, "plasma glucose": 70}])
As a more powerful example, let’s think about an entire workflow where models have been built on a dataset adding a new field with a simple feature engineering transformation, like the ratio of two fields. Suppose a model has been created from the new dataset. Also, an anomaly detector has been created from the same dataset to check whether the new input data is too different from the original examples used to train the model. If the score is low, the model is still valid, so we accept its prediction. If the score is too high, the model predictions might be inaccurate, and we should not rely on them. Therefore, in order to take a decision on what to do for new input data, we will need not only the values of the fields of that new test case but also the prediction (plus the associated probability) and anomaly score that the trained model and anomaly detector provide for it.
To solve the problem, the process will be: on receving new data,
the transformation to generate the ratio between the raw input fields
should be applied and a new ratio field should be added.
After that, both the prediction and the anomaly score should be computed
and they also should be added to the initial data as new fields.
The BMLPipeline class will help us do that.
First, we instantiate the BMLPipeline object by providing the models
that we want it to use and a name for it:
from bigml.pipeline.pipeline import BMLPipeline
local_pipeline = BMLPipeline("my new pipeline",
["model/5143a51a37203f2cf7020351",
"anomaly/5143a51a37203f2cf7027551"])
This code will retrieve all the datasets previous to the model and anomaly
detector construction and will store any transformation that they contain.
It creates a sequence starting on the first dataset that was created to
summarize the uploaded data, adding the datasets that store transformations,
and finally the model and anomaly detector. Every transformation that was
done when training those models, will be reflected as a new step in the
BMLPipeline and every model that was added to the list will also be
added as an additional transformation step: the model will transform
our data by adding its prediction and associated probability and the
anomaly detector will transform the input by adding the computed
anomaly score. The result is obtained using the BMLPipeline object, that
offers a .transform method which accepts a list of input data dictionaries
or a DataFrame. For every row, it will execute the stored transformations
and generate the model’s prediction and the anomaly’s score.
All of them will be added to the original input data.
local_pipeline.transform([{"plasma glucose": 130, "bmi":3},
{"age":26, "plasma glucose": 70}])
"""That could produce a result such as
[{"plasma glucose": 130, "bmi":3, "prediction": "True",
"probability": 0.578, "score": 0.753},
{"age": 26, "plasma glucose": 70, "prediction": "False",
"probability": 0.573, "score": 0.54}]
"""
As for the rest of local resources, you can pass additional arguments to define
the API connection info and/or a cache_get function to be used when
resources are stored in memory caches.
from bigml.pipeline.pipeline import BMLPipeline
local_pipeline = BMLPipeline("my new pipeline",
["model/5143a51a37203f2cf7020351",
"anomaly/5143a51a37203f2cf7027551"],
api=BigML("my user", "my api",
storage="my_storage"))
If no API connection is passed, or if the one given has no
api.storage value, we use the default ./storage directory
followed by the name of the pipeline as storage folder for the
JSON of the resources used in the pipeline.
In this case, four resources will be stored: the dataset created from
the uploaded data, the dataset generated when we added the ratio
field, the model and the anomaly detector. The BMLPipeline object
offers an .export method that can compress the entire directory to
a .zip file whose name is the name of the BMLPipeline
(conveniently encoded) and will be placed in the output_directory
given by the user:
from bigml.pipeline.pipeline import BMLPipeline
local_pipeline = BMLPipeline("my new pipeline",
["model/5143a51a37203f2cf7020351",
"anomaly/5143a51a37203f2cf7027551"]
api=BigML("my user", "my api",
storage="my_storage"))
local_pipeline.export(output_directory="my_export_dir")
In this example, we wil find a my_export_dir/my_new_pipeline.zip file
in the current directory. The file contains a my new pipeline folder where
the four JSONs for the two datasets and two models are stored.
The BMLPipeline provides also methods to dump and load the
data transformers it contains, in order to save them in a cache or in the file
system. As an example, we can create a BMLPipeline, dump its contents to
a file system folder and build a second pipeline from them. The name of
the pipeline will be used as reference to know which object to load.
from bigml.pipeline.pipeline import BMLPipeline
local_pipeline = BMLPipeline("pipeline1",
"model/5143a51a37203f2cf7020351")
local_pipeline.dump("./pipeline1_storage")
# the `pipeline1_storage` folder is created and all the objects
# used in the pipeline are stored there, one file each
new_pipeline = BMLPipeline.load("pipeline1", "./pipeline1_storage")
# a new pipeline has been built with the same properties and steps
# that local_pipeline had
If using a cache system, the same methods described in the `local caching<#local-caching>`_ section are available.
from bigml.pipeline.pipeline import BMLPipeline
local_pipeline = BMLPipeline("pipeline1",
"model/631a6a6f8f679a2d31000445")
import redis
r = redis.Redis()
local_pipeline.dump(cache_set=r.set)
new_pipeline = BMLPipeline("pipeline1", cache_get=r.get)
# the new_pipeline has been recovered from Redis
Sometimes, one may want to aggregate pre-existing transformations
on your original data before loading it to BigML. In that case, you can use
the more general Pipeline class to store any sequence of transformations
made outside of BigML. As both Pipeline and BMLPipeline offer the
.transform method, they are also data transformers, meaning that they
can be used as steps of a more general Pipeline as well.
Thus, combining pre-existing transformations
based on scikit-learn or Pandas with the transformations and models generated
in BigML is totally possible. For that, we will use the
SKDataTransformer and DFDataTransformer classes, which provide a
.transform method too.
As an example of use, we’ll create a Pipeline based on a existing
scikit pipeline.
import pandas as pd
from sklearn.tree import DecisionTreeClassifier
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
from sklearn.pipeline import Pipeline as SKPipeline
# Building a prediction pipeline using a scikit learn
# scaler and decision tree and adding the prediction
# to the initial dataframe
from bigml.pipeline.transformer import Pipeline, SKDataTransformer
from bigml.constants import OUT_NEW_HEADERS
# pre-existing code to build the scikit pipeline
df = pd.read_csv("data/diabetes.csv")
X = df.drop('diabetes', axis=1)
y = df['diabetes']
X_train, X_test, y_train, y_test = train_test_split(X, y,
random_state=0)
pipe = SKPipeline([('scaler', StandardScaler()),
('DTC', DecisionTreeClassifier())])
pipe.fit(X_train, y_train)
# end of pre-existing code
pipeline = Pipeline(
"skpipeline", # pipeline name
steps=[SKDataTransformer(pipe,
"skDTC",
output={OUT_NEW_HEADERS: ["sk_prediction"]})])
# the `pipe` scikit pipeline is wrapped as a SKDataTransformer to offer
# a `.transform` method
pipeline.transform(X_test)
This new pipeline can be combined with a BMLPipeline and will accumulate
the insights of both.
from bigml.pipeline import BMLPipeline
bml_pipeline = BMLPipeline("bml_pipeline",
"anomaly/631a6a6f8f679a2d31000445")
extended_pipeline = Pipeline("extended",
steps=[pipeline, bml_pipeline])
extended_pipeline.transform([{"plasma glucose": 80}])
The same can be done for a Pandas’ pipe sequence
# based on https://www.kdnuggets.com/2021/01/cleaner-data-analysis-pandas-pipes.html
import pandas as pd
import numpy as np
from bigml.pipeline.transformer import DFDataTransformer, Pipeline
marketing = pd.read_csv("./data/DirectMarketing.csv")
# code to define the transformations
def drop_missing(df):
thresh = len(df) * 0.6
df.dropna(axis=1, thresh=thresh, inplace=True)
return df
def remove_outliers(df, column_name):
low = np.quantile(df[column_name], 0.05)
high = np.quantile(df[column_name], 0.95)
return df[df[column_name].between(low, high, inclusive=True)]
def copy_df(df):
return df.copy()
pipeline = Pipeline("pandas_pipeline",
steps=[DFDataTransformer([copy_df,
drop_missing,
(remove_outliers,
['Salary'])])])
# the list of functions are wrapped as a DFDataTransformer to offer
# a `.transform` method that generates the output using Pandas' `.pipe`
marketing_clean = pipeline.transform(marketing)
where again, the pipeline could be combined with any BMLPipeline to
produce a more general transformation sequence.
Of course, new classes could be built to support other transformation tools
and libraries. A new data transformer can be created by deriving the
DataTransformer class and customizing its .data_transform method
to cover the particulars of the functions to be used in the generation of
new fields.
Local Evaluations¶
You can instantiate a local version of an evaluation that will contain the main evaluation metrics.
from bigml.evaluation import Evaluation
local_evaluation = Evaluation('evaluation/502fdbff15526876610003215')
This will retrieve the remote evaluation information, using an implicitly built
BigML() connection object (see the Authentication
section for more
details on how to set your credentials) and return a Dataset object
that will be stored in the ./storage directory. If you want to use a
specific connection object for the remote retrieval or a different storage
directory, you can set it as second parameter:
from bigml.evaluation import Evaluation
from bigml.api import BigML
local_evaluation = Evaluation('evaluation/502fdbff15526876610003215',
api=BigML(my_username,
my_api_key,
storage="my_storage"))
or even use the remote evaluation information previously retrieved to build the local evaluation object:
from bigml.evaluation import Evaluation
from bigml.api import BigML
api = BigML()
evaluation = api.get_evaluation('evaluation/502fdbff15526876610003215')
local_evaluation = Evaluation(evaluation)
You can also build a local evaluation from a previously retrieved and stored evaluation JSON file:
from bigml.evaluation import Evaluation
local_evaluation = Evaluation('./my_dataset.json')
The Evaluation attributes depend on whether it belongs to a regression or a
classification. Regression evaluations will contain r_square,
mean_absolute_error, mean_squared_error. Classification evaluations
will contain accuracy, precision, recall, phi and f_measure
besides the confusion_matrix and a -full attribute that will contain
the entire set of metrics as downloaded from the API.
from bigml.evaluation import Evaluation
local_evaluation = Evaluation('evaluation/502fdbff15526876610003215')
local_evaluation.full # entire model evaluation metrics
if local_evaluation.regression:
local_evaluation.r_squared # r-squared metric value
else:
local_evaluation.confusion_matrix # confusion matrix
local_evaluation.accuracy
Local batch predictions¶
As explained in the 101s provided in the
Quick Start section, batch predictions for a
list of inputs can be obtained by iterating the single predictions discussed
in each different local model. However, we’ve also provided a
homogeneous batch_predict method in the following local objects:
SupervisedModel
Anomaly
Cluster
PCA
TopicModel
which can receive the following parameters:
input_data_list: This can be a list of input data, expressed as a dictionary containing
field_name: field_valuepairs or a Pandas’ DataFrameoutputs: That’s a dictionary that can contain
output_fieldsand/oroutput_headersinformation. Each one is defined by default as the list of prediction keys to be added to the inputs and the list of headers to be used as keys in the output. E.g., for a supervised learning model, the default if no information is provided would be equivalent to{"output_fields": ["prediction", "probability"], "output_headers": ["prediction", "probability"]}and both the prediction and the associated probability would be added to the input data.**kwargs: Any other parameters allowed in the
.predictmethod could be added to the batch prediction too. For instance, we could add the operating kind to a supervised model batch prediction usingoperating_kind=probabilityas argument.
Let’s write some examples. If we are reading data from a CSV, we can use the
csv library and pass the list of inputs as an array to an anomaly detector.
import csv
from bigml.anomaly import Anomaly
input_data_list = []
with open("my_input_data.csv") as handler:
reader = csv.DictReader(handler)
for row_dict in reader:
input_data_list.append(row_dict)
local_anomaly = Anomaly("anomaly/5143a51a37203f2cf7027551")
scored_data_list = local_anomaly.batch_predict(input_data_list)
Or if we are using a Pandas’ DataFrame instead to read the data, we could
also use the DataFrame directly as input argument:
import pandas as pd
from bigml.anomaly import Anomaly
dataframe = pd.read_csv("my_input_data.csv")
local_anomaly = Anomaly("anomaly/5143a51a37203f2cf7027551")
scored_dataframe = local_anomaly.batch_predict(dataframe)
Now, let’s add some complexity and do use a supervised model. We’d like to
add both the predicted value and the associated probability but we’d like
to use an operating point when predicting. The operating point needs
specifying a positive class, the kind of metric to compare (probabily or
confidence) and the threshold to use. We also want the prediction to
be added to the input data using the key sm_prediction. In this case, the
code would be similar to
import pandas as pd
from bigml.supervised import SupervisedModel
dataframe = pd.read_csv("my_input_data.csv")
local_supervised = SupervisedModel("ensemble/5143a51a37203f2cf7027551")
operating_point = {"positive_class": "yes",
"kind": "probability",
"threshold": 0.7}
predicted_dataframe = local_supervised.batch_predict(
dataframe,
outputs={"output_headers": ["sm_prediction", "probability"]},
operating_point=operating_point)
and the result would be like the one below:
>>>predicted_dataframe
pregnancies plasma glucose ... sm_prediction probability
0 6 148 ... true 0.95917
1 1 85 ... false 0.99538
2 8 183 ... true 0.93701
3 1 89 ... false 0.99452
4 0 137 ... true 0.90622
.. ... ... ... ... ...
195 1 117 ... false 0.90906
196 5 123 ... false 0.97179
197 2 120 ... false 0.99300
198 1 106 ... false 0.99452
199 2 155 ... false 0.51737
[200 rows x 11 columns]
Local Shap Wrapper¶
The Shap library accepts customized predict functions as long as they provide
a particular input/output interface that uses numpy arrays. The previously
described local models can be used to generate such an predict funcion.
The ShapWrapper class has been created to help users connect the
Shap library to BigML supervised models and provides the .predict and
.predict_proba functions especially built to be used with that libary.
from bigml.shapwrapper import ShapWrapper
shap_wrapper = ShapWrapper("model/5143a51a37203f2cf7027551")
# computing the Explainer on the X_test numpy array
explainer = shap.Explainer(shap_wrapper.predict,
X_test, algorithm='partition',
feature_names=shap_wrapper.x_headers)
shap_values = explainer(X_test)
Local caching¶
All local models can use an external cache system to manage memory storage and
recovery. The get and set functions of the cache manager should be
passed to the constructor or dump function. Here’s an example on how to
cache a linear regression:
from bigml.linear import LinearRegression
lm = LinearRegression("linearregression/5e827ff85299630d22007198")
lm.predict({"petal length": 4, "sepal length":4, "petal width": 4, \
"sepal width": 4, "species": "Iris-setosa"}, full=True)
import redis
r = redis.Redis()
# First build as you would any core LinearRegression object:
# Store a serialized version in Redis
lm.dump(cache_set=r.set)
# (retrieve the external rep from its convenient place)
# Speedy Build from external rep
lm = LinearRegression("linearregression/5e827ff85299630d22007198", \
cache_get=r.get)
# Get predictions same as always:
lm.predict({"petal length": 4, "sepal length":4, "petal width": 4, \
"sepal width": 4, "species": "Iris-setosa"}, full=True)
Rule Generation¶
You can also use a local model to generate a IF-THEN rule set that can be very helpful to understand how the model works internally.
local_model.rules()
IF petal_length > 2.45 AND
IF petal_width > 1.65 AND
IF petal_length > 5.05 THEN
species = Iris-virginica
IF petal_length <= 5.05 AND
IF sepal_width > 2.9 AND
IF sepal_length > 5.95 AND
IF petal_length > 4.95 THEN
species = Iris-versicolor
IF petal_length <= 4.95 THEN
species = Iris-virginica
IF sepal_length <= 5.95 THEN
species = Iris-versicolor
IF sepal_width <= 2.9 THEN
species = Iris-virginica
IF petal_width <= 1.65 AND
IF petal_length > 4.95 AND
IF sepal_length > 6.05 THEN
species = Iris-virginica
IF sepal_length <= 6.05 AND
IF sepal_width > 2.45 THEN
species = Iris-versicolor
IF sepal_width <= 2.45 THEN
species = Iris-virginica
IF petal_length <= 4.95 THEN
species = Iris-versicolor
IF petal_length <= 2.45 THEN
species = Iris-setosa
Python, Tableau and Hadoop-ready Generation¶
If you prefer, you can also generate a Python function that implements the model
and that can be useful to make the model actionable right away with local_model.python().
local_model.python()
def predict_species(sepal_length=None,
sepal_width=None,
petal_length=None,
petal_width=None):
""" Predictor for species from model/50a8e2d9eabcb404d2000293
Predictive model by BigML - Machine Learning Made Easy
"""
if (petal_length is None):
return 'Iris-virginica'
if (petal_length <= 2.45):
return 'Iris-setosa'
if (petal_length > 2.45):
if (petal_width is None):
return 'Iris-virginica'
if (petal_width <= 1.65):
if (petal_length <= 4.95):
return 'Iris-versicolor'
if (petal_length > 4.95):
if (sepal_length is None):
return 'Iris-virginica'
if (sepal_length <= 6.05):
if (petal_width <= 1.55):
return 'Iris-virginica'
if (petal_width > 1.55):
return 'Iris-versicolor'
if (sepal_length > 6.05):
return 'Iris-virginica'
if (petal_width > 1.65):
if (petal_length <= 5.05):
if (sepal_width is None):
return 'Iris-virginica'
if (sepal_width <= 2.9):
return 'Iris-virginica'
if (sepal_width > 2.9):
if (sepal_length is None):
return 'Iris-virginica'
if (sepal_length <= 6.4):
if (sepal_length <= 5.95):
return 'Iris-versicolor'
if (sepal_length > 5.95):
return 'Iris-virginica'
if (sepal_length > 6.4):
return 'Iris-versicolor'
if (petal_length > 5.05):
return 'Iris-virginica'
The local.python(hadoop=True) call will generate the code that you need
for the Hadoop map-reduce engine to produce batch predictions using Hadoop
streaming .
Saving the mapper and reducer generated functions in their corresponding files
(let’s say /home/hduser/hadoop_mapper.py and
/home/hduser/hadoop_reducer.py) you can start a Hadoop job
to generate predictions by issuing
the following Hadoop command in your system console:
bin/hadoop jar contrib/streaming/hadoop-*streaming*.jar \
-file /home/hduser/hadoop_mapper.py -mapper hadoop_mapper.py \
-file /home/hduser/hadoop_reducer.py -reducer hadoop_reducer.py \
-input /home/hduser/hadoop/input.csv \
-output /home/hduser/hadoop/output_dir
assuming you are in the Hadoop home directory, your input file is in the
corresponding dfs directory
(/home/hduser/hadoop/input.csv in this example) and the output will
be placed at /home/hduser/hadoop/output_dir (inside the dfs directory).
Tableau-ready rules are also available through local_model.tableau() for
all the models except those that use text predictors.
local_model.tableau()
IF ISNULL([petal width]) THEN 'Iris-virginica'
ELSEIF [petal width]>0.8 AND [petal width]>1.75 AND ISNULL([petal length]) THEN 'Iris-virginica'
ELSEIF [petal width]>0.8 AND [petal width]>1.75 AND [petal length]>4.85 THEN 'Iris-virginica'
ELSEIF [petal width]>0.8 AND [petal width]>1.75 AND [petal length]<=4.85 AND ISNULL([sepal width]) THEN 'Iris-virginica'
ELSEIF [petal width]>0.8 AND [petal width]>1.75 AND [petal length]<=4.85 AND [sepal width]>3.1 THEN 'Iris-versicolor'
ELSEIF [petal width]>0.8 AND [petal width]>1.75 AND [petal length]<=4.85 AND [sepal width]<=3.1 THEN 'Iris-virginica'
ELSEIF [petal width]>0.8 AND [petal width]<=1.75 AND ISNULL([petal length]) THEN 'Iris-versicolor'
ELSEIF [petal width]>0.8 AND [petal width]<=1.75 AND [petal length]>4.95 AND [petal width]>1.55 AND [petal length]>5.45 THEN 'Iris-virginica'
ELSEIF [petal width]>0.8 AND [petal width]<=1.75 AND [petal length]>4.95 AND [petal width]>1.55 AND [petal length]<=5.45 THEN 'Iris-versicolor'
ELSEIF [petal width]>0.8 AND [petal width]<=1.75 AND [petal length]>4.95 AND [petal width]<=1.55 THEN 'Iris-virginica'
ELSEIF [petal width]>0.8 AND [petal width]<=1.75 AND [petal length]<=4.95 AND [petal width]>1.65 THEN 'Iris-virginica'
ELSEIF [petal width]>0.8 AND [petal width]<=1.75 AND [petal length]<=4.95 AND [petal width]<=1.65 THEN 'Iris-versicolor'
ELSEIF [petal width]<=0.8 THEN 'Iris-setosa'
END
Summary generation¶
You can also print the model from the point of view of the classes it predicts
with local_model.summarize().
It shows a header section with the training data initial distribution per class
(instances and percentage) and the final predicted distribution per class.
Then each class distribution is detailed. First a header section shows the percentage of the total data that belongs to the class (in the training set and in the predicted results) and the rules applicable to all the the instances of that class (if any). Just after that, a detail section shows each of the leaves in which the class members are distributed. They are sorted in descending order by the percentage of predictions of the class that fall into that leaf and also show the full rule chain that leads to it.
Data distribution:
Iris-setosa: 33.33% (50 instances)
Iris-versicolor: 33.33% (50 instances)
Iris-virginica: 33.33% (50 instances)
Predicted distribution:
Iris-setosa: 33.33% (50 instances)
Iris-versicolor: 33.33% (50 instances)
Iris-virginica: 33.33% (50 instances)
Field importance:
1. petal length: 53.16%
2. petal width: 46.33%
3. sepal length: 0.51%
4. sepal width: 0.00%
Iris-setosa : (data 33.33% / prediction 33.33%) petal length <= 2.45
· 100.00%: petal length <= 2.45 [Confidence: 92.86%]
Iris-versicolor : (data 33.33% / prediction 33.33%) petal length > 2.45
· 94.00%: petal length > 2.45 and petal width <= 1.65 and petal length <= 4.95 [Confidence: 92.44%]
· 2.00%: petal length > 2.45 and petal width <= 1.65 and petal length > 4.95 and sepal length <= 6.05 and petal width > 1.55 [Confidence: 20.65%]
· 2.00%: petal length > 2.45 and petal width > 1.65 and petal length <= 5.05 and sepal width > 2.9 and sepal length > 6.4 [Confidence: 20.65%]
· 2.00%: petal length > 2.45 and petal width > 1.65 and petal length <= 5.05 and sepal width > 2.9 and sepal length <= 6.4 and sepal length <= 5.95 [Confidence: 20.65%]
Iris-virginica : (data 33.33% / prediction 33.33%) petal length > 2.45
· 76.00%: petal length > 2.45 and petal width > 1.65 and petal length > 5.05 [Confidence: 90.82%]
· 12.00%: petal length > 2.45 and petal width > 1.65 and petal length <= 5.05 and sepal width <= 2.9 [Confidence: 60.97%]
· 6.00%: petal length > 2.45 and petal width <= 1.65 and petal length > 4.95 and sepal length > 6.05 [Confidence: 43.85%]
· 4.00%: petal length > 2.45 and petal width > 1.65 and petal length <= 5.05 and sepal width > 2.9 and sepal length <= 6.4 and sepal length > 5.95 [Confidence: 34.24%]
· 2.00%: petal length > 2.45 and petal width <= 1.65 and petal length > 4.95 and sepal length <= 6.05 and petal width <= 1.55 [Confidence: 20.65%]
You can also use local_model.get_data_distribution() and
local_model.get_prediction_distribution() to obtain the training and
prediction basic distribution
information as a list (suitable to draw histograms or any further processing).
The tree nodes’ information (prediction, confidence, impurity and distribution)
can also be retrieved in a CSV format using the method
local_model.tree_CSV(). The output can be sent to a file by providing a
file_name argument or used as a list.
Local ensembles have a local_ensemble.summarize() method too, the output
in this case shows only the data distribution (only available in
Decision Forests) and field importance sections.
For local clusters, the local_cluster.summarize() method prints also the
data distribution, the training data statistics per cluster and the basic
intercentroid distance statistics. There’s also a
local_cluster.statistics_CSV(file_name) method that store in a CSV format
the values shown by the summarize() method. If no file name is provided,
the function returns the rows that would have been stored in the file as
a list.