Source code for lightwood.encoder.categorical.onehot

from copy import deepcopy
from typing import Dict, List, Iterable, Tuple

import torch
import numpy as np
from scipy.special import softmax

from lightwood.helpers.log import log
from lightwood.encoder.base import BaseEncoder
from lightwood.helpers.constants import _UNCOMMON_WORD


[docs]class OneHotEncoder(BaseEncoder): """ Creates a one-hot encoding (OHE) for categorical data. One-hot encoding represents categorical information as a vector where each individual dimension corresponds to a category. A category has a 1:1 mapping between dimension indicated by a "1" in that position. For example, imagine 3 categories, :math:`A`, :math:`B`, and :math:`C`; these can be represented as follows: .. math:: A &= [1, 0, 0] \\ B &= [0, 1, 0] \\ C &= [0, 0, 1] The OHE encoder operates in 2 modes: (1) "use_unknown=True": Makes an :math:`N+1` length vector for :math:`N` categories, the first index always corresponds to the unknown category. (2) "use_unknown=False": Makes an :math:`N` length vector for :math:`N` categories, where an empty vector of 0s indicates an unknown/missing category. An encoder can represent a feature column or target column; in this case it represents a target, `is_target` is `True`, and `target_weights`. The `target_weights` parameter enables users to specify how heavily each class should be weighted within a mixer - useful in imbalanced classes. By default, the `StatisticalAnalysis` phase will provide `target_weights` as the relative fraction of each class in the data which is important for imbalanced populations; for example, suppose there is a 80/05/15 imbalanced representation across 3 different classes - `target_weights` will be a vector as such:: target_weights = {"class1": 0.8, "class2": 0.05, "class3": 0.15} Users should note that models will be presented with the inverse of the target weights, `inv_target_weights`, which will perform the 1/target_value_per_class operation. **This means large values will result in small weights for the model**. """ # noqa def __init__( self, is_target: bool = False, target_weights: Dict[str, float] = None, use_unknown: bool = True, ): """ :param is_target: True if this encoder featurizes the target column :param target_weights: Percentage of total population represented by each category (between [0, 1]). :param mode: True uses an extra dimension to account for unknown/out-of-distribution categories """ # noqa super().__init__(is_target) self.map = None # category name -> index self.rev_map = None # index -> category name self.use_unknown = use_unknown # Weight-balance info if encoder represents target self.target_weights = None self.index_weights = None # vector-weights, mapped by class id if self.is_target: self.target_weights = deepcopy(target_weights)
[docs] def prepare(self, priming_data: Iterable[str]): """ Prepares the OHE Encoder by creating a dictionary mapping. Unknown categories must be explicitly handled as python `None` types. """ # noqa if self.is_prepared: raise Exception('You can only call "prepare" once for a given encoder.') unq_cats = np.unique([i for i in priming_data if i is not None]).tolist() if self.use_unknown: log.info("Encoding UNKNOWN categories as index 0") self.map = {cat: indx + 1 for indx, cat in enumerate(unq_cats)} self.map.update({_UNCOMMON_WORD: 0}) self.rev_map = {indx: cat for cat, indx in self.map.items()} else: log.info("Encoding UNKNOWN categories as vector of all 0s") self.map = {cat: indx for indx, cat in enumerate(unq_cats)} self.rev_map = {indx: cat for cat, indx in self.map.items()} # Set the length of output self.output_size = len(self.map) # For target-only, report on relative weights of classes # Each dimension of the inv_target_weights respects `map` if self.is_target: # Equally wt. all classes self.index_weights = torch.ones(size=(self.output_size,)) # If imbalanced detected, weight by inverse if self.target_weights is not None: # Check target weights properly specified if sum([np.isclose(i, 0) for i in self.target_weights.values()]) > 0: raise ValueError('Target weights cannot be 0') for cat in self.map.keys(): if cat != _UNCOMMON_WORD: self.index_weights[self.map[cat]] = self.target_weights[cat] # If using an unknown category, set to smallest possible value if self.use_unknown: self.target_weights[_UNCOMMON_WORD] = np.min(list(self.target_weights.values())) self.index_weights[0] = self.target_weights[_UNCOMMON_WORD] self.is_prepared = True
[docs] def encode(self, column_data: Iterable[str]) -> torch.Tensor: """ Encodes pre-processed data into OHE. Unknown/unrecognized classes vector of all 0s. :param column_data: Pre-processed data to encode :returns: Encoded data of form :math:`N_{rows} x N_{categories}` """ # noqa if not self.is_prepared: raise Exception( 'You need to call "prepare" before calling "encode" or "decode".' ) ret = torch.zeros(size=(len(column_data), self.output_size)) for idx, word in enumerate(column_data): index = self.map.get(word, None) if index is not None: ret[idx, index] = 1 if self.use_unknown and index is None: ret[idx, 0] = 1 return torch.Tensor(ret)
[docs] def decode(self, encoded_data: torch.Tensor): """ Decodes OHE mapping into the original categories. Since this approach uses an argmax, decoding flexibly works either on logits or an explicitly OHE vector. :param: encoded_data: :returns Returns the original category names for encoded data. """ # noqa encoded_data_list = encoded_data.tolist() ret = [] for vector in encoded_data_list: all_zeros = not np.any(vector) if not self.use_unknown and all_zeros: ret.append(_UNCOMMON_WORD) else: ret.append(self.rev_map[np.argmax(vector)]) return ret
[docs] def decode_probabilities(self, encoded_data: torch.Tensor) -> Tuple[List[str], List[List[float]], Dict[int, str]]: """ Provides decoded answers, as well as a probability assignment to each data point. :param encoded_data: the output of a mixer model :returns Decoded values for each data point, Probability vector for each category, and the reverse map of dimension to category name """ # noqa encoded_data_list = encoded_data.tolist() ret = [] probs = [] for vector in encoded_data_list: if not np.any(vector): # Vector of all 0s -> unknown category ret.append(_UNCOMMON_WORD) else: ret.append(self.rev_map[np.argmax(vector)]) probs.append(self._norm_vec(vector)) return ret, probs, self.rev_map
@staticmethod def _norm_vec(vec: List[float]): """ Given a vector, normalizes so that the sum of elements is 1, using softmax. :param vec: Assigned weights for each category """ return softmax(vec)