train_crnn

This document has been machine translated.

Runs CRNN model training.

The execution method is JSON-RPC v2.0.

Example request

To execute CRNN model training, specify train_crnn as method in the parameter of prov.process of the Provenance API.

An example request for JSON-RPC is as follows.

{
  "jsonrpc": "2.0",
  "method": "prov.process",
  "params": {
    "method": "train_crnn",
    "params": {
      "output_ddc": "ddc:model",
      "input_ddc": "ddc:train_data",
      "target_column_name": "ox",
      "prediction_time": 3,
      "crnn_param_json": "{ (see parameter entry) }",
      "no_exec": true
    },
  },
  "id": "occurrence_jsonrpc_id"
}

Parameters

The following parameters can be specified for train_crnn.

Parameter name	Data type	Contents	Default value
output_ddc	string	output ddc name	mandatory
output_mode	string	output mode (overwrite or error)	error
input_ddc	string	input ddc name	mandatory
target_column_name	string	target column name for prediction	mandatory
prediction_time	number (integer)	time to predict	mandatory
crnn_param_json	string	CRNN parameters	mandatory
no_exec	boolean	execute asynchronously	false

In crnn_param_json you can specify detailed parameters that control how the model is trained. The items that can be specified are as follows.

Attreibute (key)	Data type	Default value
source.date_parts	array of string	mandatory
source.id_column	string	mandatory
source.columns	array of string	mandatory
models	array of string	mandatory
data.scales	array of object	mandatory
data.air_rank_boundaries	array of number	mandatory
data.per	number	mandatory
data.use_idw	boolean	false
data.spatial_data_block_size	number (integer)	16
data.spatial_data_size	number (integer)	32
training.max_cnn_epoch	number (integer)	2000
training.max_lstm_epoch	number (integer)	51
training.need_continue_training	boolean	true
training.is_softmax	boolean	false
training.imbalance	boolean	true
training.cnn_batch_size	number (integer)	512
training.rnn_batch_size	number (integer)	32
training.time_step	number (integer)	24
training.k_fold	number (integer)	6

source.date_parts.

List column names to be used as time information in the table specified by input_ddc.

source.id_column.

List column names to be used as time information in the table specified by input_ddc.

source.columns.

List column names in the table specified by input_ddc that are used as features for model learning.

models.

Specifies the training target. You can specify one or more of the following values.

value	meaning
pointwise	Train for each measurement point distinguished by source.id_column.
average	learn the average value of all measurement points
maximum	learn the maximum value of all measurement points

data.scales.

Specifies normalization coefficients for each column to be used as features.

data.air_rank_boundaries.

Specify the bounds for each class when training the classification model.

data.per.

Specifies the normalization factor of the prediction target.

training.max_cnn_epoch.

Specifies the maximum number of epochs for training of CNN among CRNN model training.

training.max_lstm_epoch.

The maximum number of epochs for LSTM training among CRNN model training.

training.need_continue_training.

If true is specified, two-step training is performed.

training.is_softmax.

If true, train a classification model. false to train regression model.

training.imbalance.

If true is specified in the classification model, it assumes that the class distribution is biased.

training.cnn_batch_size.

Specifies the batch size for CNN training.

training.rnn_batch_size.

Batch size for RNN training.

training.time_step.

Specify how many steps in the past to use for prediction.

training.k_fold.

Specify the number of divisions for cross-validation during training.

Input data.

The ddc specified for input_ddc should have the following schema.

column name	contents	remarks
start_datetime	start time	required column for event table
end_datetime	end time	required column in event table
(measurement point id)	id representing the measurement point
(attribute 1)	any attribute value (numeric type)
(attribute 2)	any attribute value (numeric type)
(...)	any attribute value (numeric type)

Output data

The schema of ddc output to output_ddc is as follows. This ddc is required when performing predictions with CRNN.

column name	contents	remarks
data_table	name of training data table	real table name will be recorded instead of ddc
meshcode	measurement point id
target_column	name of column for prediction
prediction_time	time to predict
crnn_param_json	CRNN parameters
cnn_model_path	CNN model file name
lstm_model_path	LSTM model file name

Return values

Output ddc information