[
https://issues.apache.org/jira/browse/SYSTEMML-1962?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Niketan Pansare updated SYSTEMML-1962:
--------------------------------------
Description:
The end goal of this JIRA is to support model selection facility similar to
[http://scikit-learn.org/stable/modules/classes.html#module-sklearn.model_selection].
Currently, we support model selection using MLPipeline's cross-validator. For
example: please replace `from pyspark.ml.classification import
LogisticRegression` with `from systemml.mllearn import LogisticRegression` in
the example
http://spark.apache.org/docs/2.1.1/ml-tuning.html#example-model-selection-via-cross-validation.
However, this invokes k-seperate and independent mlcontext calls. This PR
proposes to add a new class `GridSearchCV`, `RandomizedSearchCV` and possibly
bayesian optimization which like mllearn has methods `fit` and `predict`. These
methods internally generate a script that wraps the external script with a
`parfor` when the fit method is called. For example:
{code}
from sklearn import datasets
from systemml.mllearn import GridSearchCV, SVM
iris = datasets.load_iris()
parameters = {'C':[1, 10]}
svm = SVM()
clf = GridSearchClassifierCV(svm, parameters)
clf.fit(iris.data, iris.target)
{code}
would execute the script:
{code}
CVals = matrix("1; 10", rows=2, cols=1)
parfor(i in seq(1, nrow(CVals))) {
C = CVals[i, 1]
# SVM script
}
{code}
This will require:
1. Functionization of the script (for example: L2SVM)
{code}
svm = function(matrix[double] X, matrix[double] Y, double icpt, double tol,
double reg, double maxiter) returns (matrix[double] w) {
if(nrow(X) < 2)
stop("Stopping due to invalid inputs: Not possible to learn a binary
class classifier without at least 2 rows")
check_min = min(Y)
....
w = t(cbind(t(w), t(extra_model_params)))
}
{code}
2. Adding two new java classes in the package `org.apache.sysml.api.ml` called
`GridSearchClassifierCV` which extends `Estimator[GridSearchClassifierCVModel]`
and `GridSearchClassifierCVModel` which `extends
Model[GridSearchClassifierCVModel] with BaseSystemMLClassifierModel`. Then you
will have to implement the abstract methods: fit and transform respectively.
3. Add a python class GridSearchClassifierCV that invokes the above java
classes.
For more details on step 2 and step 3, please read the design documentation of
mllearn API:
https://github.com/apache/systemml/blob/master/src/main/scala/org/apache/sysml/api/ml/BaseSystemMLClassifier.scala#L42
[~dusenberrymw] may be, this can be part of
https://issues.apache.org/jira/browse/SYSTEMML-1159
was:
The end goal of this JIRA is to support model selection facility similar to
[http://scikit-learn.org/stable/modules/classes.html#module-sklearn.model_selection].
Currently, we support model selection using MLPipeline's cross-validator. For
example: please replace `from pyspark.ml.classification import
LogisticRegression` with `from systemml.mllearn import LogisticRegression` in
the example
http://spark.apache.org/docs/2.1.1/ml-tuning.html#example-model-selection-via-cross-validation.
However, this invokes k-seperate and independent mlcontext calls. This PR
proposes to add a new class `GridSearchCV`, `RandomizedSearchCV` and possibly
bayesian optimization which like mllearn has methods `fit` and `predict`. These
methods internally generate a script that wraps the external script with a
`parfor` when the fit method is called. For example:
{code}
from sklearn import datasets
from systemml.mllearn import GridSearchCV, SVM
iris = datasets.load_iris()
parameters = {'C':[1, 10]}
svm = SVM()
clf = GridSearchClassifierCV(svm, parameters)
clf.fit(iris.data, iris.target)
{code}
would execute the script:
{code}
CVals = matrix("1; 10", rows=2, cols=1)
parfor(i in seq(1, nrow(CVals))) {
C = CVals[i, 1]
# SVM script
}
{code}
This will require:
1. Functionization of the script (for example: L2SVM)
{code}
svm = function(matrix[double] X, matrix[double] Y, double icpt, double tol,
double reg, double maxiter) returns (matrix[double] w) {
if(nrow(X) < 2)
stop("Stopping due to invalid inputs: Not possible to learn a binary
class classifier without at least 2 rows")
check_min = min(Y)
....
w = t(cbind(t(w), t(extra_model_params)))
}
{code}
2. Adding two new java classes in the package `org.apache.sysml.api.ml` called
`GridSearchClassifierCV` which extends `Estimator[GridSearchClassifierCVModel]`
and `GridSearchClassifierCVModel` which `extends
Model[GridSearchClassifierCVModel] with BaseSystemMLClassifierModel`. Then you
will have to implement the abstract methods: fit and transform respectively.
3. Add a python class GridSearchClassifierCV that invokes the above java
classes.
For more details on step 2 and step 3, please read the design documentation of
mllearn API:
https://github.com/apache/systemml/blob/master/src/main/scala/org/apache/sysml/api/ml/BaseSystemMLClassifier.scala#L42
[~dusenberrymw]
> Support model-selection via mllearn APIs
> ----------------------------------------
>
> Key: SYSTEMML-1962
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1962
> Project: SystemML
> Issue Type: New Feature
> Reporter: Niketan Pansare
>
> The end goal of this JIRA is to support model selection facility similar to
> [http://scikit-learn.org/stable/modules/classes.html#module-sklearn.model_selection].
> Currently, we support model selection using MLPipeline's cross-validator. For
> example: please replace `from pyspark.ml.classification import
> LogisticRegression` with `from systemml.mllearn import LogisticRegression` in
> the example
> http://spark.apache.org/docs/2.1.1/ml-tuning.html#example-model-selection-via-cross-validation.
>
> However, this invokes k-seperate and independent mlcontext calls. This PR
> proposes to add a new class `GridSearchCV`, `RandomizedSearchCV` and possibly
> bayesian optimization which like mllearn has methods `fit` and `predict`.
> These methods internally generate a script that wraps the external script
> with a `parfor` when the fit method is called. For example:
> {code}
> from sklearn import datasets
> from systemml.mllearn import GridSearchCV, SVM
> iris = datasets.load_iris()
> parameters = {'C':[1, 10]}
> svm = SVM()
> clf = GridSearchClassifierCV(svm, parameters)
> clf.fit(iris.data, iris.target)
> {code}
> would execute the script:
> {code}
> CVals = matrix("1; 10", rows=2, cols=1)
> parfor(i in seq(1, nrow(CVals))) {
> C = CVals[i, 1]
> # SVM script
> }
> {code}
> This will require:
> 1. Functionization of the script (for example: L2SVM)
> {code}
> svm = function(matrix[double] X, matrix[double] Y, double icpt, double tol,
> double reg, double maxiter) returns (matrix[double] w) {
> if(nrow(X) < 2)
> stop("Stopping due to invalid inputs: Not possible to learn a binary
> class classifier without at least 2 rows")
> check_min = min(Y)
> ....
> w = t(cbind(t(w), t(extra_model_params)))
> }
> {code}
> 2. Adding two new java classes in the package `org.apache.sysml.api.ml`
> called `GridSearchClassifierCV` which extends
> `Estimator[GridSearchClassifierCVModel]` and `GridSearchClassifierCVModel`
> which `extends Model[GridSearchClassifierCVModel] with
> BaseSystemMLClassifierModel`. Then you will have to implement the abstract
> methods: fit and transform respectively.
> 3. Add a python class GridSearchClassifierCV that invokes the above java
> classes.
> For more details on step 2 and step 3, please read the design documentation
> of mllearn API:
> https://github.com/apache/systemml/blob/master/src/main/scala/org/apache/sysml/api/ml/BaseSystemMLClassifier.scala#L42
> [~dusenberrymw] may be, this can be part of
> https://issues.apache.org/jira/browse/SYSTEMML-1159
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
