Create trained models APIedit

Creates a trained model.

Models created in version 7.8.0 are not backwards compatible with older node versions. If in a mixed cluster environment, all nodes must be at least 7.8.0 to use a model stored by a 7.8.0 node.

Requestedit

PUT _ml/trained_models/<model_id>

Prerequisitesedit

Requires the manage_ml cluster privilege. This privilege is included in the machine_learning_admin built-in role.

Descriptionedit

The create trained model API enables you to supply a trained model that is not created by data frame analytics.

Path parametersedit

<model_id>
(Required, string) The unique identifier of the trained model.

Query parametersedit

defer_definition_decompression
(Optional, boolean) If set to true and a compressed_definition is provided, the request defers definition decompression and skips relevant validations. This deferral is useful for systems or users that know a good byte size estimate for their model and know that their model is valid and likely won’t fail during inference.

Request bodyedit

compressed_definition
(Required, string) The compressed (GZipped and Base64 encoded) inference definition of the model. If compressed_definition is specified, then definition cannot be specified.
definition

(Required, object) The inference definition for the model. If definition is specified, then compressed_definition cannot be specified.

Properties of definition
preprocessors

(Optional, object) Collection of preprocessors. See Preprocessor examples.

Properties of preprocessors
frequency_encoding

(Required, object) Defines a frequency encoding for a field.

Properties of frequency_encoding
feature_name
(Required, string) The name of the resulting feature.
field
(Required, string) The field name to encode.
frequency_map
(Required, object map of string:double) Object that maps the field value to the frequency encoded value.
custom
(Optional, Boolean) Boolean value indicating if the analytics job created the preprocessor or if a user provided it. This adjusts the feature importance calculation. When true, the feature importance calculation returns importance for the processed feature. When false, the total importance of the original field is returned. Default is false.
one_hot_encoding

(Required, object) Defines a one hot encoding map for a field.

Properties of one_hot_encoding
field
(Required, string) The field name to encode.
hot_map
(Required, object map of strings) String map of "field_value: one_hot_column_name".
custom
(Optional, Boolean) Boolean value indicating if the analytics job created the preprocessor or if a user provided it. This adjusts the feature importance calculation. When true, the feature importance calculation returns importance for the processed feature. When false, the total importance of the original field is returned. Default is false.
target_mean_encoding

(Required, object) Defines a target mean encoding for a field.

Properties of target_mean_encoding
default_value
(Required, double) The feature value if the field value is not in the target_map.
feature_name
(Required, string) The name of the resulting feature.
field
(Required, string) The field name to encode.
target_map

(Required, object map of string:double) Object that maps the field value to the target mean value.

custom
(Optional, Boolean) Boolean value indicating if the analytics job created the preprocessor or if a user provided it. This adjusts the feature importance calculation. When true, the feature importance calculation returns importance for the processed feature. When false, the total importance of the original field is returned. Default is false.
trained_model

(Required, object) The definition of the trained model.

Properties of trained_model
tree

(Required, object) The definition for a binary decision tree.

Properties of tree
classification_labels
(Optional, string) An array of classification labels (used for classification).
feature_names
(Required, string) Features expected by the tree, in their expected order.
target_type
(Required, string) String indicating the model target type; regression or classification.
tree_structure
(Required, object) An array of tree_node objects. The nodes must be in ordinal order by their tree_node.node_index value.
tree_node

(Required, object) The definition of a node in a tree.

There are two major types of nodes: leaf nodes and not-leaf nodes.

  • Leaf nodes only need node_index and leaf_value defined.
  • All other nodes need split_feature, left_child, right_child, threshold, decision_type, and default_left defined.
Properties of tree_node
decision_type
(Optional, string) Indicates the positive value (in other words, when to choose the left node) decision type. Supported lt, lte, gt, gte. Defaults to lte.
default_left
(Optional, Boolean) Indicates whether to default to the left when the feature is missing. Defaults to true.
leaf_value
(Optional, double) The leaf value of the of the node, if the value is a leaf (in other words, no children).
left_child
(Optional, integer) The index of the left child.
node_index
(Integer) The index of the current node.
right_child
(Optional, integer) The index of the right child.
split_feature
(Optional, integer) The index of the feature value in the feature array.
split_gain
(Optional, double) The information gain from the split.
threshold
(Optional, double) The decision threshold with which to compare the feature value.
ensemble

(Optional, object) The definition for an ensemble model. See Model examples.

Properties of ensemble
aggregate_output

(Required, object) An aggregated output object that defines how to aggregate the outputs of the trained_models. Supported objects are weighted_mode, weighted_sum, and logistic_regression. See Aggregated output example.

Properties of aggregate_output
logistic_regression

(Optional, object) This aggregated_output type works with binary classification (classification for values [0, 1]). It multiplies the outputs (in the case of the ensemble model, the inference model values) by the supplied weights. The resulting vector is summed and passed to a sigmoid function. The result of the sigmoid function is considered the probability of class 1 (P_1), consequently, the probability of class 0 is 1 - P_1. The class with the highest probability (either 0 or 1) is then returned. For more information about logistic regression, see this wiki article.

Properties of logistic_regression
weights
(Required, double) The weights to multiply by the input values (the inference values of the trained models).
weighted_sum

(Optional, object) This aggregated_output type works with regression. The weighted sum of the input values.

Properties of weighted_sum
weights
(Required, double) The weights to multiply by the input values (the inference values of the trained models).
weighted_mode

(Optional, object) This aggregated_output type works with regression or classification. It takes a weighted vote of the input values. The most common input value (taking the weights into account) is returned.

Properties of weighted_mode
weights
(Required, double) The weights to multiply by the input values (the inference values of the trained models).
exponent

(Optional, object) This aggregated_output type works with regression. It takes a weighted sum of the input values and passes the result to an exponent function (e^x where x is the sum of the weighted values).

Properties of exponent
weights
(Required, double) The weights to multiply by the input values (the inference values of the trained models).
classification_labels
(Optional, string) An array of classification labels.
feature_names
(Optional, string) Features expected by the ensemble, in their expected order.
target_type
(Required, string) String indicating the model target type; regression or classification.
trained_models
(Required, object) An array of trained_model objects. Supported trained models are tree and ensemble.
description
(Optional, string) A human-readable description of the inference trained model.
estimated_heap_memory_usage_bytes
(Optional, integer) [7.16.0] Deprecated in 7.16.0. Replaced by model_size_bytes
estimated_operations
(Optional, integer) The estimated number of operations to use the trained model during inference. This property is supported only if defer_definition_decompression is true or the model definition is not supplied.
inference_config

(Required, object) The default configuration for inference. This can be: regression, classification, fill_mask, ner, text_classification, text_embedding or zero_shot_classification. If regression or classification, it must match the target_type of the underlying definition.trained_model. If fill_mask, ner, text_classification, or text_embedding; the model_type must be pytorch.

Properties of inference_config
classification

(Optional, object) Classification configuration for inference.

Properties of classification inference
num_top_classes
(Optional, integer) Specifies the number of top class predictions to return. Defaults to 0.
num_top_feature_importance_values
(Optional, integer) Specifies the maximum number of feature importance values per document. Defaults to 0 which means no feature importance calculation occurs.
prediction_field_type
(Optional, string) Specifies the type of the predicted field to write. Valid values are: string, number, boolean. When boolean is provided 1.0 is transformed to true and 0.0 to false.
results_field
(Optional, string) The field that is added to incoming documents to contain the inference prediction. Defaults to predicted_value.
top_classes_results_field
(Optional, string) Specifies the field to which the top classes are written. Defaults to top_classes.
fill_mask

(Optional, object) Configuration for a fill_mask natural language processing (NLP) task. The fill_mask task works with models optimized for a fill mask action. For example, for BERT models, the following text may be provided: "The capital of France is [MASK].". The response indicates the value most likely to replace [MASK]. In this instance, the most probable token is paris.

Properties of fill_mask inference
tokenization

(Optional, object) Indicates the tokenization to perform and the desired settings.

Properties of tokenization
bert

(Optional, object) BERT-style tokenization is to be performed with the enclosed settings.

Properties of bert
do_lower_case
(Optional, boolean) Specifies if the tokenization lower case the text sequence when building the tokens.
max_sequence_length
(Optional, integer) Specifies the maximum number of tokens allowed to be output by the tokenizer. The default for BERT-style tokenization is 512.
truncate

(Optional, string) Indicates how tokens are truncated when they exceed max_sequence_length. The default value is first.

  • none: No truncation occurs; the inference request receives an error.
  • first: Only the first sequence is truncated.
  • second: Only the second sequence is truncated. If there is just one sequence, that sequence is truncated.

For zero_shot_classification, the hypothesis sequence is always the second sequence. Therefore, do not use second in this case.

with_special_tokens

(Optional, boolean) Tokenize with special tokens. The tokens typically included in BERT-style tokenization are:

  • [CLS]: The first token of the sequence being classified.
  • [SEP]: Indicates sequence separation.
ner

(Optional, object) Configures a named entity recognition (NER) task. NER is a special case of token classification. Each token in the sequence is classified according to the provided classification labels. Currently, the NER task requires the classification_labels Inside-Outside-Beginning (IOB) formatted labels. Only person, organization, location, and miscellaneous are supported.

Properties of ner inference
classification_labels
(Optional, string) An array of classification labels. NER only supports Inside-Outside-Beginning labels (IOB) and only persons, organizations, locations, and miscellaneous. Example: ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC", "B-MISC", "I-MISC"]
tokenization

(Optional, object) Indicates the tokenization to perform and the desired settings.

Properties of tokenization
bert

(Optional, object) BERT-style tokenization is to be performed with the enclosed settings.

Properties of bert
do_lower_case
(Optional, boolean) Specifies if the tokenization lower case the text sequence when building the tokens.
max_sequence_length
(Optional, integer) Specifies the maximum number of tokens allowed to be output by the tokenizer. The default for BERT-style tokenization is 512.
truncate

(Optional, string) Indicates how tokens are truncated when they exceed max_sequence_length. The default value is first.

  • none: No truncation occurs; the inference request receives an error.
  • first: Only the first sequence is truncated.
  • second: Only the second sequence is truncated. If there is just one sequence, that sequence is truncated.

For zero_shot_classification, the hypothesis sequence is always the second sequence. Therefore, do not use second in this case.

with_special_tokens

(Optional, boolean) Tokenize with special tokens. The tokens typically included in BERT-style tokenization are:

  • [CLS]: The first token of the sequence being classified.
  • [SEP]: Indicates sequence separation.
pass_through

(Optional, object) Configures a pass_through task. This task is useful for debugging as no post-processing is done to the inference output and the raw pooling layer results are returned to the caller.

Properties of pass_through inference
tokenization

(Optional, object) Indicates the tokenization to perform and the desired settings.

Properties of tokenization
bert

(Optional, object) BERT-style tokenization is to be performed with the enclosed settings.

Properties of bert
do_lower_case
(Optional, boolean) Specifies if the tokenization lower case the text sequence when building the tokens.
max_sequence_length
(Optional, integer) Specifies the maximum number of tokens allowed to be output by the tokenizer. The default for BERT-style tokenization is 512.
truncate

(Optional, string) Indicates how tokens are truncated when they exceed max_sequence_length. The default value is first.

  • none: No truncation occurs; the inference request receives an error.
  • first: Only the first sequence is truncated.
  • second: Only the second sequence is truncated. If there is just one sequence, that sequence is truncated.

For zero_shot_classification, the hypothesis sequence is always the second sequence. Therefore, do not use second in this case.

with_special_tokens

(Optional, boolean) Tokenize with special tokens. The tokens typically included in BERT-style tokenization are:

  • [CLS]: The first token of the sequence being classified.
  • [SEP]: Indicates sequence separation.
regression

(Optional, object) Regression configuration for inference.

Properties of regression inference
num_top_feature_importance_values
(Optional, integer) Specifies the maximum number of feature importance values per document. By default, it is zero and no feature importance calculation occurs.
results_field
(Optional, string) The field that is added to incoming documents to contain the inference prediction. Defaults to predicted_value.
text_classification

(Optional, object) A text classification task. Text classification classifies a provided text sequence into previously known target classes. A specific example of this is sentiment analysis, which returns the likely target classes indicating text sentiment, such as "sad", "happy", or "angry".

Properties of text_classification inference
classification_labels
(Optional, string) An array of classification labels.
num_top_classes
(Optional, integer) Specifies the number of top class predictions to return. Defaults to all classes (-1).
tokenization

(Optional, object) Indicates the tokenization to perform and the desired settings.

Properties of tokenization
bert

(Optional, object) BERT-style tokenization is to be performed with the enclosed settings.

Properties of bert
do_lower_case
(Optional, boolean) Specifies if the tokenization lower case the text sequence when building the tokens.
max_sequence_length
(Optional, integer) Specifies the maximum number of tokens allowed to be output by the tokenizer. The default for BERT-style tokenization is 512.
truncate

(Optional, string) Indicates how tokens are truncated when they exceed max_sequence_length. The default value is first.

  • none: No truncation occurs; the inference request receives an error.
  • first: Only the first sequence is truncated.
  • second: Only the second sequence is truncated. If there is just one sequence, that sequence is truncated.

For zero_shot_classification, the hypothesis sequence is always the second sequence. Therefore, do not use second in this case.

with_special_tokens

(Optional, boolean) Tokenize with special tokens. The tokens typically included in BERT-style tokenization are:

  • [CLS]: The first token of the sequence being classified.
  • [SEP]: Indicates sequence separation.
text_embedding

(Object, optional) Text embedding takes an input sequence and transforms it into a vector of numbers. These embeddings capture not simply tokens, but semantic meanings and context. These embeddings can be used in a dense vector field for powerful insights.

Properties of text_embedding inference
tokenization

(Optional, object) Indicates the tokenization to perform and the desired settings.

Properties of tokenization
bert

(Optional, object) BERT-style tokenization is to be performed with the enclosed settings.

Properties of bert
do_lower_case
(Optional, boolean) Specifies if the tokenization lower case the text sequence when building the tokens.
max_sequence_length
(Optional, integer) Specifies the maximum number of tokens allowed to be output by the tokenizer. The default for BERT-style tokenization is 512.
truncate

(Optional, string) Indicates how tokens are truncated when they exceed max_sequence_length. The default value is first.

  • none: No truncation occurs; the inference request receives an error.
  • first: Only the first sequence is truncated.
  • second: Only the second sequence is truncated. If there is just one sequence, that sequence is truncated.

For zero_shot_classification, the hypothesis sequence is always the second sequence. Therefore, do not use second in this case.

with_special_tokens

(Optional, boolean) Tokenize with special tokens. The tokens typically included in BERT-style tokenization are:

  • [CLS]: The first token of the sequence being classified.
  • [SEP]: Indicates sequence separation.
zero_shot_classification

(Object, optional) Configures a zero-shot classification task. Zero-shot classification allows for text classification to occur without pre-determined labels. At inference time, it is possible to adjust the labels to classify. This makes this type of model and task exceptionally flexible.

If consistently classifying the same labels, it may be better to use a fine-tuned text classification model.

Properties of zero_shot_classification inference
classification_labels
(Required, array) The classification labels used during the zero-shot classification. Classification labels must not be empty or null and only set at model creation. They must be all three of ["entailment", "neutral", "contradiction"].

This is NOT the same as labels which are the values that zero-shot is attempting to classify.

hypothesis_template

(Optional, string) This is the template used when tokenizing the sequences for classification.

The labels replace the {} value in the text. The default value is: This example is {}.

labels
(Optional, array) The labels to classify. Can be set at creation for default labels, and then updated during inference.
multi_label
(Optional, boolean) Indicates if more than one true label is possible given the input. This is useful when labeling text that could pertain to more than one of the input labels. Defaults to false.
tokenization

(Optional, object) Indicates the tokenization to perform and the desired settings.

Properties of tokenization
bert

(Optional, object) BERT-style tokenization is to be performed with the enclosed settings.

Properties of bert
do_lower_case
(Optional, boolean) Specifies if the tokenization lower case the text sequence when building the tokens.
max_sequence_length
(Optional, integer) Specifies the maximum number of tokens allowed to be output by the tokenizer. The default for BERT-style tokenization is 512.
truncate

(Optional, string) Indicates how tokens are truncated when they exceed max_sequence_length. The default value is first.

  • none: No truncation occurs; the inference request receives an error.
  • first: Only the first sequence is truncated.
  • second: Only the second sequence is truncated. If there is just one sequence, that sequence is truncated.

For zero_shot_classification, the hypothesis sequence is always the second sequence. Therefore, do not use second in this case.

with_special_tokens

(Optional, boolean) Tokenize with special tokens. The tokens typically included in BERT-style tokenization are:

  • [CLS]: The first token of the sequence being classified.
  • [SEP]: Indicates sequence separation.
input

(Required, object) The input field names for the model definition.

Properties of input
field_names
(Required, string) An array of input field names for the model.
location

(Optional, object) The model definition location. If the definition or compressed_definition are not specified, the location is required.

Properties of location
index
(Required, object) Indicates that the model definition is stored in an index. This object must be empty as the index for storing model definitions is configured automatically.
metadata
(Optional, object) An object map that contains metadata about the model.
model_size_bytes
(Optional, integer) The estimated memory usage in bytes to keep the trained model in memory. This property is supported only if defer_definition_decompression is true or the model definition is not supplied.
model_type

(Optional, string) The created model type. By default the model type is tree_ensemble. Appropriate types are:

  • tree_ensemble: The model definition is an ensemble model of decision trees.
  • lang_ident: A special type reserved for language identification models.
  • pytorch: The stored definition is a PyTorch (specifically a TorchScript) model. Currently only NLP models are supported.
tags
(Optional, string) An array of tags to organize the model.

Examplesedit

Preprocessor examplesedit

The example below shows a frequency_encoding preprocessor object:

{
   "frequency_encoding":{
      "field":"FlightDelayType",
      "feature_name":"FlightDelayType_frequency",
      "frequency_map":{
         "Carrier Delay":0.6007414737092798,
         "NAS Delay":0.6007414737092798,
         "Weather Delay":0.024573576178086153,
         "Security Delay":0.02476631010889467,
         "No Delay":0.6007414737092798,
         "Late Aircraft Delay":0.6007414737092798
      }
   }
}

The next example shows a one_hot_encoding preprocessor object:

{
   "one_hot_encoding":{
      "field":"FlightDelayType",
      "hot_map":{
         "Carrier Delay":"FlightDelayType_Carrier Delay",
         "NAS Delay":"FlightDelayType_NAS Delay",
         "No Delay":"FlightDelayType_No Delay",
         "Late Aircraft Delay":"FlightDelayType_Late Aircraft Delay"
      }
   }
}

This example shows a target_mean_encoding preprocessor object:

{
   "target_mean_encoding":{
      "field":"FlightDelayType",
      "feature_name":"FlightDelayType_targetmean",
      "target_map":{
         "Carrier Delay":39.97465788139886,
         "NAS Delay":39.97465788139886,
         "Security Delay":203.171206225681,
         "Weather Delay":187.64705882352948,
         "No Delay":39.97465788139886,
         "Late Aircraft Delay":39.97465788139886
      },
      "default_value":158.17995752420433
   }
}

Model examplesedit

The first example shows a trained_model object:

{
   "tree":{
      "feature_names":[
         "DistanceKilometers",
         "FlightTimeMin",
         "FlightDelayType_NAS Delay",
         "Origin_targetmean",
         "DestRegion_targetmean",
         "DestCityName_targetmean",
         "OriginAirportID_targetmean",
         "OriginCityName_frequency",
         "DistanceMiles",
         "FlightDelayType_Late Aircraft Delay"
      ],
      "tree_structure":[
         {
            "decision_type":"lt",
            "threshold":9069.33437193022,
            "split_feature":0,
            "split_gain":4112.094574306927,
            "node_index":0,
            "default_left":true,
            "left_child":1,
            "right_child":2
         },
         ...
         {
            "node_index":9,
            "leaf_value":-27.68987349695448
         },
         ...
      ],
      "target_type":"regression"
   }
}

The following example shows an ensemble model object:

"ensemble":{
   "feature_names":[
      ...
   ],
   "trained_models":[
      {
         "tree":{
            "feature_names":[],
            "tree_structure":[
               {
                  "decision_type":"lte",
                  "node_index":0,
                  "leaf_value":47.64069875778043,
                  "default_left":false
               }
            ],
            "target_type":"regression"
         }
      },
      ...
   ],
   "aggregate_output":{
      "weighted_sum":{
         "weights":[
            ...
         ]
      }
   },
   "target_type":"regression"
}

Aggregated output exampleedit

Example of a logistic_regression object:

"aggregate_output" : {
  "logistic_regression" : {
    "weights" : [2.0, 1.0, .5, -1.0, 5.0, 1.0, 1.0]
  }
}

Example of a weighted_sum object:

"aggregate_output" : {
  "weighted_sum" : {
    "weights" : [1.0, -1.0, .5, 1.0, 5.0]
  }
}

Example of a weighted_mode object:

"aggregate_output" : {
  "weighted_mode" : {
    "weights" : [1.0, 1.0, 1.0, 1.0, 1.0]
  }
}

Example of an exponent object:

"aggregate_output" : {
  "exponent" : {
    "weights" : [1.0, 1.0, 1.0, 1.0, 1.0]
  }
}

Trained models JSON schemaedit

For the full JSON schema of trained models, click here.