schema.json

{
  "definitions": {},
  "$schema": "http://json-schema.org/draft-07/schema#",
  "$id": "http://example.com/root.json",
  "type": "object",
  "title": "Polyglot",
  "required": [
    "experiments"
  ],
  "properties": {
    "experiments": {
      "$id": "#/properties/experiments",
      "type": "array",
      "title": "Experiments to run by order",
      "description": "Each experiment defines it's models and their hyperparameters",
      "default": null,
      "minItems": 1,
      "items": {
        "$id": "#/properties/experiments/items",
        "type": "object",
        "title": "Required params for defining an experiment",
        "required": [
          "name",
          "models",
          "learning_technique",
          "hyperparameters"
        ],
        "properties": {
          "name": {
            "$id": "#/properties/experiments/items/properties/name",
            "type": "string",
            "title": "UNIQUE name of experiment",
            "description": "MUST BE UNIQUE",
            "default": "",
            "examples": [
              "exp_1"
            ],
            "minLength": 1,
            "pattern": "^(.*)$"
          },
          "models": {
            "$id": "#/properties/experiments/items/properties/models",
            "type": "array",
            "title": "Models to run",
            "description": "each name must match the name of folder under rnnlm/models and the key in the hyperparameters hash",
            "default": null,
            "minItems": 1,
            "uniqueItems": true,
            "additionalItems": false,
            "items": {
              "$id": "#/properties/experiments/items/properties/models/items",
              "type": "string",
              "title": "name of the model to run from rnnlm/models",
              "default": "",
              "minLength": 1,
              "examples": [
                "lstm_fast"
              ],
              "pattern": "^(.*)$"
            }
          },
          "predict_mode": {
            "$id": "#/properties/experiments/items/properties/predict_mode",
            "type": "boolean",
            "title": "Prediction mode",
            "description": "whether to use the model to predict (true) or train (false)",
            "default": false,
            "examples": [
              false
            ]
          },
          "learning_technique": {
            "$id": "#/properties/experiments/items/properties/learning_technique",
            "type": "string",
            "title": "Learning technique",
            "description": "what learning technique to use to train the model",
            "default": "normal",
            "enum": [
              "normal",
              "transfer",
              "multitask"
            ],
            "examples": [
              "normal",
              "transfer",
              "multitask"
            ],
            "pattern": "^(.*)$"
          },
          "hyperparameters": {
            "$id": "#/properties/experiments/items/properties/hyperparameters",
            "type": "object",
            "title": "Hyperparameters Schema",
            "description": "Hyperparameters per model. Each key must match the model name in 'models' array and the folder under rnnlm/models",
            "default": null,
            "required": [
              "shared_params"
            ],
            "properties": {
              "shared_params": {
                "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params",
                "type": "object",
                "title": "Hyperparameters for the shared layer. Shared layer is usually the hidden layers that are share between all tasks",
                "description": "hyperparameters for the shared layer between tasks, if training one task, here you define the parameters for the hidden layer etc.",
                "default": null,
                "required": [
                  "create_model",
                  "data",
                  "train",
                  "arch"
                ],
                "properties": {
                  "create_model": {
                    "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/create_model",
                    "type": "string",
                    "title": "Name of the model that creates the shared layers",
                    "description": "The name of the model that will be used to create the hidden layers that are shared between all tasks.",
                    "default": "",
                    "examples": [
                      "lstm_fast"
                    ],
                    "minLength": 1,
                    "pattern": "^(.*)$"
                  },
                  "data": {
                    "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/data",
                    "type": "object",
                    "title": "Paths and configurations for the data files",
                    "description": "Paths and configurations for the data files",
                    "default": null,
                    "required": [
                      "save_path"
                    ],
                    "properties": {
                      "save_path": {
                        "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/data/properties/save_path",
                        "type": "string",
                        "title": "Path where the results will be saved",
                        "description": "Path where the results will be saved",
                        "default": "",
                        "examples": [
                          "results"
                        ],
                        "minLength": 1,
                        "pattern": "^(.*)$"
                      }
                    }
                  },
                  "train": {
                    "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train",
                    "type": "object",
                    "title": "Training hyperparams that are shared between all tasks",
                    "description": "shared training parameters",
                    "default": null,
                    "required": [
                      "summary_steps",
                      "save_checkpoint_steps",
                      "keep_checkpoint_max",
                      "early_stopping"
                    ],
                    "properties": {
                      "summary_steps": {
                        "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/summary_steps",
                        "type": "integer",
                        "title": "Steps interval for saving TensorBoard summarizes",
                        "description": "Steps interval for saving TensorBoard summarizes",
                        "default": 100,
                        "examples": [
                          100
                        ],
                        "minimum": 0
                      },
                      "save_checkpoint_steps": {
                        "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/save_checkpoint_steps",
                        "type": "integer",
                        "title": "Steps interval for saving checkpoints",
                        "description": "Steps interval for saving checkpoints. A step is batch of data.",
                        "default": 200,
                        "examples": [
                          200
                        ],
                        "minimum": 10
                      },
                      "keep_checkpoint_max": {
                        "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/keep_checkpoint_max",
                        "type": "integer",
                        "title": "The Keep_checkpoint_max Schema",
                        "description": "how many files to keep for checkpoint history",
                        "default": 5,
                        "examples": [
                          5
                        ],
                        "minimum": 1,
                        "maximum": 10
                      },
                      "multitask": {
                        "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/multitask",
                        "type": "object",
                        "title": "Parameters for multitask learning",
                        "description": "Parameters for multitask learning - valid if training with learning_technique = multitask",
                        "default": null,
                        "required": [
                          "num_multitask_epochs",
                          "switch_each_epoch",
                          "switch_each_batch"
                        ],
                        "properties": {
                          "num_multitask_epochs": {
                            "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/multitask/properties/num_multitask_epochs",
                            "type": "integer",
                            "title": "Number of multitask epochs",
                            "description": "Multitask epoch are the num of epochs, this can be also referred as num of dataset context switches",
                            "default": 0,
                            "examples": [
                              20
                            ],
                            "minimum": 0
                          },
                          "switch_each_epoch": {
                            "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/multitask/properties/switch_each_epoch",
                            "type": "boolean",
                            "title": "Whether to change task after each epoch",
                            "default": false,
                            "examples": [
                              true
                            ]
                          },
                          "switch_each_batch": {
                            "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/multitask/properties/switch_each_batch",
                            "type": "boolean",
                            "title": "Whether to change task after each batch",
                            "default": false,
                            "examples": [
                              true
                            ]
                          }
                        }
                      },
                      "early_stopping": {
                        "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/early_stopping",
                        "type": "object",
                        "title": "Early stopping configuration",
                        "description": "Early stopping configuration",
                        "required": [
                          "threshold",
                          "max_steps_without_improvement"
                        ],
                        "properties": {
                          "threshold": {
                            "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/early_stopping/properties/threshold",
                            "type": "number",
                            "title": "The Threshold of decision for identifying a training epoch without improvement",
                            "description": "This parameters is checked with the loss after each batch. Each time there is no improvement of at least this threshold, the estimator increments a count of steps without improvement.",
                            "default": 0.0,
                            "examples": [
                              0.001
                            ],
                            "minimum": 0.0
                          },
                          "max_steps_without_improvement": {
                            "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/early_stopping/properties/max_steps_without_improvement",
                            "type": "integer",
                            "title": "How many steps should the system count of no improvement in training with the threshold, before stopping the training.",
                            "default": 0,
                            "examples": [
                              10000
                            ],
                            "minimum": 1
                          }
                        }
                      },
                      "start_from_experiment": {
                        "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/train/properties/start_from_experiment",
                        "type": "string",
                        "title": "Directory of checkpoint to start from",
                        "description": "Directory of checkpoint to start from, tensors from here will be loaded and the experiment will used pre trained tensors",
                        "default": ""
                      }
                    }
                  },
                  "arch": {
                    "$id": "#/properties/experiments/items/properties/hyperparameters/properties/shared_params/properties/arch",
                    "type": "object",
                    "title": "Parameters for building the architecture of the model. Should by used in model.py of each model.",
                    "description": "Architecture parameters for the model",
                    "default": null,
                    "required": [
                    ],
                    "properties": {}
                  }
                }
              }
            },
            "additionalProperties": {
              "type": "object",
              "title": "Hyperparameters for model schema",
              "required": [
                "data",
                "train"
              ],
              "properties": {
                "data": {
                  "type": "object",
                  "title": "params for the data files",
                  "description": "Contains params for the data files",
                  "required": [
                    "pre_train",
                    "data_path",
                    "train_raw_data_file",
                    "valid_raw_data_file",
                    "test_raw_data_file",
                    "vocab_path_features",
                    "vocab_path_labels",
                    "vocab_size_features",
                    "vocab_size_labels",
                    "shape_size_features",
                    "shape_size_labels",
                    "tf_record_train_file",
                    "tf_record_valid_file",
                    "tf_record_test_file"
                  ],
                  "properties": {
                    "pre_train": {
                      "type": "boolean",
                      "title": "Whether to call pre_train.py of not",
                      "description": "Whether to perform actions before training. If true, the models/model_name/pre_training.py will ran. You can do in pre_training some pre-process and the conversion of the data to tf record format",
                      "default": false,
                      "examples": [
                        false
                      ]
                    },
                    "data_path": {
                      "type": "string",
                      "title": "The path to the data folder, must be relative to the project's root folder",
                      "description": "The path to the data folder, must be relative to the project's root folder",
                      "default": "rnnlm/data",
                      "examples": [
                        "rnnlm/data"
                      ],
                      "minLength": 1,
                      "pattern": "^(.*)$"
                    },
                    "train_raw_data_file": {
                      "type": "string",
                      "title": "The file containing training data that has not been pre processed",
                      "description": "The file containing training data that has not been pre processed",
                      "default": "",
                      "examples": [
                        "train"
                      ],
                      "minLength": 1,
                      "pattern": "^(.*)$"
                    },
                    "valid_raw_data_file": {
                      "type": "string",
                      "title": "The file containing validation data that has not been pre processed",
                      "default": "",
                      "description": "The file containing validation data that has not been pre processed",
                      "examples": [
                        "valid"
                      ],
                      "minLength": 1,
                      "pattern": "^(.*)$"
                    },
                    "test_raw_data_file": {
                      "type": "string",
                      "title": "The file containing test data that has not been pre processed",
                      "description": "The file containing test data that has not been pre processed",
                      "default": "",
                      "minLength": 1,
                      "examples": [
                        "test"
                      ],
                      "pattern": "^(.*)$"
                    },
                    "vocab_path_features": {
                      "type": "string",
                      "title": "Vocab_path for the features, relative to data_path.",
                      "description": "File path relative to data_path. Vocabulary of words that appear as features in the data.",
                      "default": "",
                      "examples": [
                        "wordlist.rnn.final"
                      ],
                      "minLength": 1,
                      "pattern": "^(.*)$"
                    },
                    "vocab_path_labels": {
                      "type": "string",
                      "title": "Vocab_path for the labels, relative to data_path.",
                      "description": "File path relative to data_path. Vocabulary of words that appear as labels in the data.",
                      "default": "",
                      "minLength": 1,
                      "examples": [
                        "wordlist.rnn.final"
                      ],
                      "pattern": "^(.*)$"
                    },
                    "vocab_size_features": {
                      "type": "integer",
                      "title": "Size of the features vocabulary",
                      "description": "Size of the features vocabulary",
                      "default": 0,
                      "minimum": 1,
                      "examples": [
                        10000
                      ]
                    },
                    "vocab_size_labels": {
                      "type": "integer",
                      "title": "Size of the labels vocabulary",
                      "description": "Size of the labels vocabulary",
                      "default": 0,
                      "minimum": 1,
                      "examples": [
                        37
                      ]
                    },
                    "shape_size_features" : {
                      "type": "integer",
                      "title": "Shape size of the features",
                      "description": "Shape size of the features",
                      "default": 0,
                      "minimum": 1,
                      "examples": [
                        20
                      ]
                    },
                    "shape_size_labels" : {
                      "type": "integer",
                      "title": "Shape size of the features",
                      "description": "Shape size of the features",
                      "default": 0,
                      "minimum": 1,
                      "examples": [
                        1
                      ]
                    },
                    "tf_record_train_file": {
                      "type": "string",
                      "title": "File name of the training tf record file. If pre_train is true this file will be created. Relative to data_path.",
                      "description": "File name of the training tf record file. If pre_train is true this file will be created. Relative to data_path.",
                      "default": "",
                      "examples": [
                        "train.tfrecord"
                      ],
                      "minLength": 1,
                      "pattern": "^(.*)$"
                    },
                    "tf_record_valid_file": {
                      "type": "string",
                      "title": "File name of the validation tf record file. If pre_train is true this file will be created. Relative to data_path",
                      "default": "",
                      "description": "File name of the validation tf record file. If pre_train is true this file will be created. Relative to data_path.",
                      "examples": [
                        "valid.tfrecord"
                      ],
                      "minLength": 1,
                      "pattern": "^(.*)$"
                    },
                    "tf_record_test_file": {
                      "type": "string",
                      "title": "File name of the test tf record file. If pre_train is true this file will be created. Relative to data_path",
                      "default": "",
                      "description": "File name of the test tf record file. If pre_train is true this file will be created. Relative to data_path.",
                      "examples": [
                        "test.tfrecord"
                      ],
                      "minLength": 1,
                      "pattern": "^(.*)$"
                    },
                    "shuffle": {
                      "type": "boolean",
                      "title": "Whether to shuffle the data while training or not",
                      "description": "Whether to shuffle the data while training or not. Will shuffles batches of data",
                      "default": false,
                      "examples": [
                        false
                      ]
                    },
                    "shuffle_buffer_size": {
                      "type": "integer",
                      "title": "How many batches to shuffle from",
                      "description": "How many batches to shuffle from",
                      "default": 10000,
                      "minimum": 0,
                      "examples": [
                        10000
                      ]
                    }
                  }
                },
                "train": {
                  "type": "object",
                  "title": "Hyperparameters for training the model",
                  "description": "Contains hyperparameters for training the model",
                  "required": [
                    "batch_size",
                    "num_epochs",
                    "epoch_size_train",
                    "epoch_size_valid",
                    "epoch_size_test"
                  ],
                  "properties": {
                    "batch_size": {
                      "type": "integer",
                      "title": "Batch_size",
                      "description": "Batch size",
                      "default": 64,
                      "minimum": 1,
                      "examples": [
                        64
                      ]
                    },
                    "num_epochs": {
                      "type": "integer",
                      "title": "Num_epochs",
                      "description": "Number of training epochs",
                      "default": 20,
                      "minimum": 1,
                      "examples": [
                        20
                      ]
                    },
                    "epoch_size_train": {
                      "type": "integer",
                      "title": "Number of words in a training epoch",
                      "description": "Number of words in a training epoch",
                      "default": 0,
                      "examples": [
                        1292
                      ]
                    },
                    "epoch_size_valid": {
                      "$id": "#/properties/experiments/items/properties/hyperparameters/properties/lstm_fast/properties/train/properties/epoch_size_valid",
                      "type": "integer",
                      "title": "Number of words in a validation epoch",
                      "description": "Number of words in a validation epoch",
                      "default": 0,
                      "examples": [
                        130
                      ]
                    },
                    "epoch_size_test": {
                      "$id": "#/properties/experiments/items/properties/hyperparameters/properties/lstm_fast/properties/train/properties/epoch_size_test",
                      "type": "integer",
                      "title": "Number of words in a test epoch",
                      "description": "Number of words in a test epoch",
                      "default": 0,
                      "examples": [
                        156
                      ]
                    }
                  }
                }
              }
            }
          }
        }
      }
    }
  }
}