nirdizati_light.predictive_model.common

  1from enum import Enum
  2import torch
  3import numpy as np
  4from funcy import flatten
  5from pandas import DataFrame
  6
  7
  8class ClassificationMethods(Enum):
  9    """
 10    Available classification methods
 11    """
 12    RANDOM_FOREST = 'randomForestClassifier'
 13    KNN = 'knn'
 14    XGBOOST = 'xgboost'
 15    SGDCLASSIFIER = 'SGDClassifier'
 16    PERCEPTRON = 'perceptron'
 17    LSTM = 'lstm'
 18    CUSTOM_PYTORCH = 'customPytorch'
 19    MLP = 'mlp'
 20    SVM = 'svc'
 21    DT = 'DecisionTree'
 22
 23
 24class RegressionMethods(Enum):
 25    """
 26    Available regression methods
 27    """
 28    RANDOM_FOREST = 'randomForestRegressor'
 29
 30
 31def get_tensor(df: DataFrame, prefix_length):
 32    trace_attributes = [att for att in df.columns if 'prefix_' not in att]
 33    event_attributes = [att[:-2] for att in df.columns if att[-2:] == '_1']
 34
 35    reshaped_data = {
 36            trace_index: {
 37                prefix_index:
 38                    list(flatten(
 39                        feat_values if isinstance(feat_values, tuple) else [feat_values]
 40                        for feat_name, feat_values in trace.items()
 41                        if feat_name in trace_attributes + [event_attribute + '_' + str(prefix_index) for event_attribute in event_attributes]
 42                    ))
 43                for prefix_index in range(1, prefix_length + 1)
 44            }
 45            for trace_index, trace in df.iterrows()
 46    }
 47
 48    flattened_features = max(
 49        len(reshaped_data[trace][prefix])
 50        for trace in reshaped_data
 51        for prefix in reshaped_data[trace]
 52    )
 53
 54    tensor = np.zeros((
 55        len(df),                # sample
 56        prefix_length,          # time steps
 57        flattened_features      # features x single time step (trace and event attributes)
 58    ))
 59
 60    for i, trace_index in enumerate(reshaped_data):  # prefix
 61        for j, prefix_index in enumerate(reshaped_data[trace_index]):  # steps of the prefix
 62            for single_flattened_value in range(len(reshaped_data[trace_index][prefix_index])):
 63                tensor[i, j, single_flattened_value] = reshaped_data[trace_index][prefix_index][single_flattened_value]
 64
 65    return tensor
 66
 67def shape_label_df(df: DataFrame):
 68    labels_list = df['label'].tolist()
 69    labels = np.zeros((len(labels_list), int(max(df['label'].nunique(), int(max(df['label'].values))) + 1)))
 70    for label_idx, label_val in enumerate(labels_list):
 71        labels[int(label_idx), int(label_val)] = 1
 72
 73    return labels
 74
 75# General purpose class to wrap a lambda function as a torch module
 76class LambdaModule(torch.nn.Module):
 77    def __init__(self, lambd):
 78        super().__init__()
 79        import types
 80        assert type(lambd) is types.LambdaType
 81        self.lambd = lambd
 82
 83    def forward(self, x):
 84        return self.lambd(x)
 85    
 86# Class for early stopping
 87class EarlyStopper:
 88    def __init__(self, patience=1, min_delta=0):
 89        self.patience = patience
 90        self.min_delta = min_delta
 91        self.counter = 0
 92        self.min_validation_loss = float('inf')
 93
 94    def early_stop(self, validation_loss):
 95        if validation_loss < self.min_validation_loss:
 96            self.min_validation_loss = validation_loss
 97            self.counter = 0
 98        elif validation_loss > (self.min_validation_loss + self.min_delta):
 99            self.counter += 1
100            if self.counter >= self.patience:
101                return True
102        return False