Skip to content

Latest commit

 

History

History
45 lines (38 loc) · 4.16 KB

gdmix_config.md

File metadata and controls

45 lines (38 loc) · 4.16 KB
Raw

GDMix Configs

To train fixed effect and random effect models using GDMix, users need to provide a GDMix config, which consists of configs for fixed-effect and random-effect models. For distributed training, computing resource configs for Tensorflow and Spark jobs are needed.

GDMix config examples for movieLens with a fixed-effect global model and two random effect per-user and per-movie models are available in directory examples/movielens-100k:

  • lr-movieLens.yaml: train logistic regression models for the global, per-user and per-movie models; spark_config and tfjob_config sections are resources used for distributed training only.
  • detext-movieLens.yaml: train a deep and wide neural network model for the global and logistic regression models for the per-user and per-movie; spark_config and tfjob_config sections are resources used for distributed training only.

Logistic regression models

Fixed-effect config

Required fields:

  • name: name of the model. String.
  • training_data_dir: path to training data directory. String.
  • uid_column_name: unique id column name in the train/validation data.
  • metadata_file: path to an input data tensor metadata file. String.
  • feature_file: path to a feature list file for outputing model in name-term-value format. String.
  • model_type: the model type to train, e.g, logistic regression, linear regression, detext, etc.
  • output_model_dir: model output directory.

Optional fields:

  • validation_data_dir: path to validation data directory. String, default is "".
  • regularize_bias: whether to regularize the intercept. Ususally we do not put regularization on intercept since it is an important feature. Boolean, default is false.
  • l2_reg_weight: weight of L2 regularization for each feature bag. Float, default is 0.001.
  • optimizer: optimizer used in the training, currently support LBFGS only.
  • metric: metric of the model. String, support "auc" and "mse" Default is "auc".
  • copy_to_local: whether copy training data to local disk. Boolean, default is true.
  • label_column_name: label column name in the train/validation data.
  • weight_column_name: weight column name in the train/validation data.
  • prediction_score_column_name: prediction score column name in the generated result file.
  • feature_bag: feature bag name that is used for training and scoring.

Random-effect config

Required fields include all fields from fixed-effect config plus:

  • partition_entity: the column name used to partition data in order to improve random effect model training parallelism. String.
  • num_partitions: number of partitions. Integer.

Optional fields include all fields from fixed-effect config plus:

  • max_training_queue_size: maximum number of training queue size in the producer/consumer model. The trainer is implemented in a producer/consumer model. The producer reads data from hard drive, then the consumers solve the optimization problem for each entity. The blocking queue synchcronizes both sides. Integer, default is 10.
  • num_of_consumers: the number of consumers (processes that optimizes the models). This specifies the parallelism inside a trainer. Integer, default is 2.
  • enable_local_indexing: whether to enable local indexing. Some dataset has large global feature space, but small per entity feature space. For example the total features in a dataset could be on the order of millions, but each member has only hundreds of features. We should re-index the features to save memory footprint and increase the training efficiency. Boolean, default is true.

Neural network model supported by DeText for fixed-effect

Fixed-effect config

Please refer to DeText training manual TRAINING.md for available parameters for the config, the config is a collection of key/value pair, a DeText config example for movieLens data can be found at detext-single-node-movieLens.config