Machine Learning Model Server on Redis
RedisML is a Redis module that implements several machine learning models as Redis data types.
The stored models are fully operational and support performing the prediction/evaluation.
RedisML is a turn key solution for using trained models in a production environment. Allowing loading ML models from any platform immediately ready to serve.
The module includes these primary features:
 Decision Tree ensembles (random forests) classification and regression
 Linear regression
 Logistic regression
 Matrix operations
Building and running

Build a Redis server with support for modules (currently available from the unstable branch).

You'll also need a BLAS library, for example ATLAS. To install ATLAS:
 Ubuntu:
sudo aptget install libatlasbasedev
 CentOS/RHEL/Fedora:
sudo yum install atlasdevel
 Ubuntu:

Build the RedisML module:
git clone https://github.com/RedisLabsModules/redisml.git cd redisml/src make

To load the module, start Redis with the
loadmodule /path/to/redisml/src/redisml.so
option, add it as a directive to the configuration file or send aMODULE LOAD
command.
Decision tree ensembles
Example of use
The following creates a random forest under the key myforest
that consists of three trees with ids ranging from 0 to 2, where each consists of a single numeric splitter and its predicate values. Afterwards, the forest is used to classify two inputs and yield their predictions.
redis> ML.FOREST.ADD myforst 0 . NUMERIC 1 0.1 .l LEAF 1 .r LEAF 0
OK
redis> ML.FOREST.ADD myforst 1 . NUMERIC 1 0.1 .l LEAF 1 .r LEAF 0
OK
redis> ML.FOREST.ADD myforst 2 . NUMERIC 1 0.1 .l LEAF 0 .r LEAF 1
OK
redis> ML.FOREST.RUN myforst 1:0.01 CLASSIFICATION
"1"
redis> ML.FOREST.RUN myforst 1:0.2 CLASSIFICATION
"0"
ML.FOREST.ADD key tree path ((NUMERICCATEGORIC) attr val  LEAF val) [...]
Time complexity: O(M*log(N)) where N is the tree's depth and M is the number of nodes added
Add nodes to a tree in the forest.
This commands adds one or more nodes to the tree in the forest that's stored under key
. Trees are identified by numeric ids, treeid
, that must begin at 0 and incremented by exactly 1 for each new tree.
Each of the nodes is described by its path and definition. The path
argument is the path from the tree's root to the node. A valid path always starts with the period character (.
), which denotes the root. Optionally, the root may be followed by left or right branches , denoted by the characters l
and r
, respectively. For example, the path ".lr" refers to the right child of the root's left child.
A node in the decision tree can either be a splitter or a terminal leaf. Splitter nodes are either numerical or categorical, and are added using the NUMERIC
or CATEGORIC
keywords. Splitter nodes also require specifying the examined attribute (attr
) as well as the value (val
) used in the comparison made during the branching decision. val
is expected to be a doubleprecision floating point value for numerical splitters, and a string for categorical splitter nodes.
The leaves are created with the LEAF
keyword and only require specifying their doubleprecision floating point value (val
).
Return value:
Simple string reply
ML.FOREST.RUN key sample (CLASSIFICATIONREGRESSION)
Time complexity: O(M*log(N)) where N is the depth of the trees and M is the number of trees in the forest
Predicts the classified (discrete) or regressed (continuous) value of a sample using the forest.
The forest that's stored in key
is used for generating the predicted value for the sample
. The sample is given as a string that is a vector of attributevalue pairs in the format of attr:val
. For example, the sample
"gender:male" has a single attribute, gender, whose value is male. A sample may have multiple such attributevalue pairs, and these must be commaseparated (,
) in the string vector. For example, a sample of a 25 years old male is expressed as "gender:male,age:25".
Return value:
Bulk string reply: the predicted value of the sample
Linear regression
Example of use
The first line of the example shows how a linear regression predictor is set to the key named linear
. The predictor has an intercept of 2 and its coefficients are 3, 4 and 5. Once the predicator is ready, it is used to predict the result given the independent variables' values (features) of 1, 1 and 1.
redis> ML.LINREG.SET linear 2 3 4 5
OK
redis> ML.LINREG.PREDICT linear 1 1 1
"14"
ML.LINREG.SET key intercept coefficient [...]
Time complexity: O(N) where N is the number of coefficients
Sets a linear regression predictor.
This command sets or updates the linear regression predictor that's stored in key
. The predictor's intercept is specified by intercept
, followed by one or more coefficient
arguments of the independent variables.
Return value:
Simple string reply
ML.LINREG.PREDICT key feature [...]
Time complexity: O(N) where N is the number features
Predicts the result for a set of features.
The linear regression predictor stored in key
is used for predicting the result based on one or more features that are given by the feature
argument(s).
Return value:
Bulk string reply: the predicted result for the feature set
Logistic regression
Example of use
In this example, the first line shows how a logistic regression predictor is set to the key named logistic
. The predictor has an intercept of 0 and its coefficients are 2 and 2. Once the predicator is ready, it is used to predict the result given the independent variables' values (features) of 3 and 1.
redis> ML.LOGREG.SET logistic 0 2 2
OK
redis> ML.LOGREG.PREDICT logistic 3 1
"0.017986209962091559"
ML.LOGREG.SET key intercept coefficient [...]
Time complexity: O(N) where N is the number of coefficients
Sets a linear regression predictor.
This command sets or updates the logistic regression predictor that's stored in key
. The predictor's intercept is specified by intercept
, followed by one or more coefficient
arguments of the independent variables.
Return value:
Simple string reply
ML.LOGREG.PREDICT key feature [...]
Time complexity: O(N) where N is the number features
Predicts the result for a set of features.
The logistic regression predictor stored in key
is used for predicting the result based on one or more features that are given by the feature
argument(s).
Return value:
Bulk string reply: the predicted result for the feature set
Matrix operations
Example of use
The following example shows how to set two matrices, a
and b
, and then multiply them storing the result in the matrix ab
. Lastly, the contents of ab
are fetched.
redis> ML.MATRIX.SET a 2 3 1 2 5 3 4 6
OK
redis> ML.MATRIX.SET b 3 2 1 2 3 4 7 1
OK
redis> ML.MATRIX.MULTIPLY a b ab
OK
redis> ML.MATRIX.GET ab
1) (integer) 2
2) (integer) 2
3) "42"
4) "15"
5) "57"
6) "28"
ML.MATRIX.SET key n m entry11 .. entrynm
Time complexity: O(N*M) where N is the number of rows and M is the number of columns
Sets a matrix.
Sets key
to store a matrix of n
rows,m
columns and doubleprecision float entries ranging from entry11
to entrynm
.
Return value:
Simple string reply
ML.MATRIX.GET key
Time complexity: O(N*M) where N is the number of rows and M is the number of columns
Get a matrix.
Returns the matrix's dimensions and entries.
Return value:
Array reply. The first two elements are the matrix's rows and columns, respectively, followed by the entries.
ML.MATRIX.ADD matrix1 matrix2 sum
Time complexity: O(N*M) where N is the number of rows and M is the number of columns
Adds matrices.
The result of adding the two matrices stored in matrix1
and matrix2
is set in sum
.
Return value:
Simple string reply
ML.MATRIX.MULTIPLY matrix1 matrix2 product
Time complexity: O(N*M*P) where N and M are numbers of rows and columns in
matrix1
, and P is the number of columns inmatrix2
Multiplies matrices.
The result of multiplying the two matrices stored in matrix1
and matrix2
is set in product
.
Return value:
Simple string reply
ML.MATRIX.SCALE key scalar
Time complexity: O(N*M) where N is the number of rows and M is the number of columns
Scales a matrix.
Updates the entries of the matrix stored in key
by multiplying them with scalar
.
Return value:
Simple string reply
Contributing
Issue reports, pull and feature requests are welcome.
License
AGPLv3  see LICENSE