#!/usr/bin/env python # -*- coding: utf-8 -*- # # Author: Jayant Jain # Copyright (C) 2017 Radim Rehurek # Licensed under the GNU LGPL v2.1 - http://www.gnu.org/licenses/lgpl.html """Python implementation of Poincaré Embeddings. These embeddings are better at capturing latent hierarchical information than traditional Euclidean embeddings. The method is described in detail in `Maximilian Nickel, Douwe Kiela - "Poincaré Embeddings for Learning Hierarchical Representations" `_. The main use-case is to automatically learn hierarchical representations of nodes from a tree-like structure, such as a Directed Acyclic Graph (DAG), using a transitive closure of the relations. Representations of nodes in a symmetric graph can also be learned. This module allows training Poincaré Embeddings from a training file containing relations of graph in a csv-like format, or from a Python iterable of relations. Examples -------- Initialize and train a model from a list .. sourcecode:: pycon >>> from gensim.models.poincare import PoincareModel >>> relations = [('kangaroo', 'marsupial'), ('kangaroo', 'mammal'), ('gib', 'cat')] >>> model = PoincareModel(relations, negative=2) >>> model.train(epochs=50) Initialize and train a model from a file containing one relation per line .. sourcecode:: pycon >>> from gensim.models.poincare import PoincareModel, PoincareRelations >>> from gensim.test.utils import datapath >>> file_path = datapath('poincare_hypernyms.tsv') >>> model = PoincareModel(PoincareRelations(file_path), negative=2) >>> model.train(epochs=50) """ import csv import logging from numbers import Integral import sys import time from collections import defaultdict, Counter import numpy as np from numpy import random as np_random, float32 as REAL from scipy.stats import spearmanr try: from autograd import grad # Only required for optionally verifying gradients while training from autograd import numpy as grad_np AUTOGRAD_PRESENT = True except ImportError: AUTOGRAD_PRESENT = False from gensim import utils, matutils from gensim.models.keyedvectors import KeyedVectors logger = logging.getLogger(__name__) class PoincareModel(utils.SaveLoad): """Train, use and evaluate Poincare Embeddings. The model can be stored/loaded via its :meth:`~gensim.models.poincare.PoincareModel.save` and :meth:`~gensim.models.poincare.PoincareModel.load` methods, or stored/loaded in the word2vec format via `model.kv.save_word2vec_format` and :meth:`~gensim.models.poincare.PoincareKeyedVectors.load_word2vec_format`. Notes ----- Training cannot be resumed from a model loaded via `load_word2vec_format`, if you wish to train further, use :meth:`~gensim.models.poincare.PoincareModel.save` and :meth:`~gensim.models.poincare.PoincareModel.load` methods instead. An important attribute (that provides a lot of additional functionality when directly accessed) are the keyed vectors: self.kv : :class:`~gensim.models.poincare.PoincareKeyedVectors` This object essentially contains the mapping between nodes and embeddings, as well the vocabulary of the model (set of unique nodes seen by the model). After training, it can be used to perform operations on the vectors such as vector lookup, distance and similarity calculations etc. See the documentation of its class for usage examples. """ def __init__(self, train_data, size=50, alpha=0.1, negative=10, workers=1, epsilon=1e-5, regularization_coeff=1.0, burn_in=10, burn_in_alpha=0.01, init_range=(-0.001, 0.001), dtype=np.float64, seed=0): """Initialize and train a Poincare embedding model from an iterable of relations. Parameters ---------- train_data : {iterable of (str, str), :class:`gensim.models.poincare.PoincareRelations`} Iterable of relations, e.g. a list of tuples, or a :class:`gensim.models.poincare.PoincareRelations` instance streaming from a file. Note that the relations are treated as ordered pairs, i.e. a relation (a, b) does not imply the opposite relation (b, a). In case the relations are symmetric, the data should contain both relations (a, b) and (b, a). size : int, optional Number of dimensions of the trained model. alpha : float, optional Learning rate for training. negative : int, optional Number of negative samples to use. workers : int, optional Number of threads to use for training the model. epsilon : float, optional Constant used for clipping embeddings below a norm of one. regularization_coeff : float, optional Coefficient used for l2-regularization while training (0 effectively disables regularization). burn_in : int, optional Number of epochs to use for burn-in initialization (0 means no burn-in). burn_in_alpha : float, optional Learning rate for burn-in initialization, ignored if `burn_in` is 0. init_range : 2-tuple (float, float) Range within which the vectors are randomly initialized. dtype : numpy.dtype The numpy dtype to use for the vectors in the model (numpy.float64, numpy.float32 etc). Using lower precision floats may be useful in increasing training speed and reducing memory usage. seed : int, optional Seed for random to ensure reproducibility. Examples -------- Initialize a model from a list: .. sourcecode:: pycon >>> from gensim.models.poincare import PoincareModel >>> relations = [('kangaroo', 'marsupial'), ('kangaroo', 'mammal'), ('gib', 'cat')] >>> model = PoincareModel(relations, negative=2) Initialize a model from a file containing one relation per line: .. sourcecode:: pycon >>> from gensim.models.poincare import PoincareModel, PoincareRelations >>> from gensim.test.utils import datapath >>> file_path = datapath('poincare_hypernyms.tsv') >>> model = PoincareModel(PoincareRelations(file_path), negative=2) See :class:`~gensim.models.poincare.PoincareRelations` for more options. """ self.train_data = train_data self.kv = PoincareKeyedVectors(size, 0) self.all_relations = [] self.node_relations = defaultdict(set) self._negatives_buffer = NegativesBuffer([]) self._negatives_buffer_size = 2000 self.size = size self.train_alpha = alpha # Learning rate for training self.burn_in_alpha = burn_in_alpha # Learning rate for burn-in self.alpha = alpha # Current learning rate self.negative = negative self.workers = workers self.epsilon = epsilon self.regularization_coeff = regularization_coeff self.burn_in = burn_in self._burn_in_done = False self.dtype = dtype self.seed = seed self._np_random = np_random.RandomState(seed) self.init_range = init_range self._loss_grad = None self.build_vocab(train_data) def build_vocab(self, relations, update=False): """Build the model's vocabulary from known relations. Parameters ---------- relations : {iterable of (str, str), :class:`gensim.models.poincare.PoincareRelations`} Iterable of relations, e.g. a list of tuples, or a :class:`gensim.models.poincare.PoincareRelations` instance streaming from a file. Note that the relations are treated as ordered pairs, i.e. a relation (a, b) does not imply the opposite relation (b, a). In case the relations are symmetric, the data should contain both relations (a, b) and (b, a). update : bool, optional If true, only new nodes's embeddings are initialized. Use this when the model already has an existing vocabulary and you want to update it. If false, all node's embeddings are initialized. Use this when you're creating a new vocabulary from scratch. Examples -------- Train a model and update vocab for online training: .. sourcecode:: pycon >>> from gensim.models.poincare import PoincareModel >>> >>> # train a new model from initial data >>> initial_relations = [('kangaroo', 'marsupial'), ('kangaroo', 'mammal')] >>> model = PoincareModel(initial_relations, negative=1) >>> model.train(epochs=50) >>> >>> # online training: update the vocabulary and continue training >>> online_relations = [('striped_skunk', 'mammal')] >>> model.build_vocab(online_relations, update=True) >>> model.train(epochs=50) """ old_index_to_key_len = len(self.kv.index_to_key) logger.info("loading relations from train data..") for relation in relations: if len(relation) != 2: raise ValueError('Relation pair "%s" should have exactly two items' % repr(relation)) for item in relation: if item in self.kv.key_to_index: self.kv.set_vecattr(item, 'count', self.kv.get_vecattr(item, 'count') + 1) else: self.kv.key_to_index[item] = len(self.kv.index_to_key) self.kv.index_to_key.append(item) self.kv.set_vecattr(item, 'count', 1) node_1, node_2 = relation node_1_index, node_2_index = self.kv.key_to_index[node_1], self.kv.key_to_index[node_2] self.node_relations[node_1_index].add(node_2_index) relation = (node_1_index, node_2_index) self.all_relations.append(relation) logger.info("loaded %d relations from train data, %d nodes", len(self.all_relations), len(self.kv)) self.indices_set = set(range(len(self.kv.index_to_key))) # Set of all node indices self.indices_array = np.fromiter(range(len(self.kv.index_to_key)), dtype=int) # Numpy array of all node indices self._init_node_probabilities() if not update: self._init_embeddings() else: self._update_embeddings(old_index_to_key_len) def _init_embeddings(self): """Randomly initialize vectors for the items in the vocab.""" shape = (len(self.kv.index_to_key), self.size) self.kv.vectors = self._np_random.uniform(self.init_range[0], self.init_range[1], shape).astype(self.dtype) def _update_embeddings(self, old_index_to_key_len): """Randomly initialize vectors for the items in the additional vocab.""" shape = (len(self.kv.index_to_key) - old_index_to_key_len, self.size) v = self._np_random.uniform(self.init_range[0], self.init_range[1], shape).astype(self.dtype) self.kv.vectors = np.concatenate([self.kv.vectors, v]) def _init_node_probabilities(self): """Initialize a-priori probabilities.""" counts = self.kv.expandos['count'].astype(np.float64) self._node_counts_cumsum = np.cumsum(counts) self._node_probabilities = counts / counts.sum() def _get_candidate_negatives(self): """Get candidate negatives of size `self.negative` from the negative examples buffer. Returns ------- numpy.array Array of shape (`self.negative`,) containing indices of negative nodes. """ if self._negatives_buffer.num_items() < self.negative: # cumsum table of counts used instead of the standard approach of a probability cumsum table # this is to avoid floating point errors that result when the number of nodes is very high # for reference: https://github.com/RaRe-Technologies/gensim/issues/1917 max_cumsum_value = self._node_counts_cumsum[-1] uniform_numbers = self._np_random.randint(1, max_cumsum_value + 1, self._negatives_buffer_size) cumsum_table_indices = np.searchsorted(self._node_counts_cumsum, uniform_numbers) self._negatives_buffer = NegativesBuffer(cumsum_table_indices) return self._negatives_buffer.get_items(self.negative) def _sample_negatives(self, node_index): """Get a sample of negatives for the given node. Parameters ---------- node_index : int Index of the positive node for which negative samples are to be returned. Returns ------- numpy.array Array of shape (self.negative,) containing indices of negative nodes for the given node index. """ node_relations = self.node_relations[node_index] num_remaining_nodes = len(self.kv) - len(node_relations) if num_remaining_nodes < self.negative: raise ValueError( 'Cannot sample %d negative nodes from a set of %d negative nodes for %s' % (self.negative, num_remaining_nodes, self.kv.index_to_key[node_index]) ) positive_fraction = float(len(node_relations)) / len(self.kv) if positive_fraction < 0.01: # If number of positive relations is a small fraction of total nodes # re-sample till no positively connected nodes are chosen indices = self._get_candidate_negatives() unique_indices = set(indices) times_sampled = 1 while (len(indices) != len(unique_indices)) or (unique_indices & node_relations): times_sampled += 1 indices = self._get_candidate_negatives() unique_indices = set(indices) if times_sampled > 1: logger.debug('sampled %d times, positive fraction %.5f', times_sampled, positive_fraction) else: # If number of positive relations is a significant fraction of total nodes # subtract positively connected nodes from set of choices and sample from the remaining valid_negatives = np.array(list(self.indices_set - node_relations)) probs = self._node_probabilities[valid_negatives] probs /= probs.sum() indices = self._np_random.choice(valid_negatives, size=self.negative, p=probs, replace=False) return list(indices) @staticmethod def _loss_fn(matrix, regularization_coeff=1.0): """Computes loss value. Parameters ---------- matrix : numpy.array Array containing vectors for u, v and negative samples, of shape (2 + negative_size, dim). regularization_coeff : float, optional Coefficient to use for l2-regularization Returns ------- float Computed loss value. Warnings -------- Only used for autograd gradients, since autograd requires a specific function signature. """ vector_u = matrix[0] vectors_v = matrix[1:] euclidean_dists = grad_np.linalg.norm(vector_u - vectors_v, axis=1) norm = grad_np.linalg.norm(vector_u) all_norms = grad_np.linalg.norm(vectors_v, axis=1) poincare_dists = grad_np.arccosh( 1 + 2 * ( (euclidean_dists ** 2) / ((1 - norm ** 2) * (1 - all_norms ** 2)) ) ) exp_negative_distances = grad_np.exp(-poincare_dists) regularization_term = regularization_coeff * grad_np.linalg.norm(vectors_v[0]) ** 2 return -grad_np.log(exp_negative_distances[0] / (exp_negative_distances.sum())) + regularization_term @staticmethod def _clip_vectors(vectors, epsilon): """Clip vectors to have a norm of less than one. Parameters ---------- vectors : numpy.array Can be 1-D, or 2-D (in which case the norm for each row is checked). epsilon : float Parameter for numerical stability, each dimension of the vector is reduced by `epsilon` if the norm of the vector is greater than or equal to 1. Returns ------- numpy.array Array with norms clipped below 1. """ one_d = len(vectors.shape) == 1 threshold = 1 - epsilon if one_d: norm = np.linalg.norm(vectors) if norm < threshold: return vectors else: return vectors / norm - (np.sign(vectors) * epsilon) else: norms = np.linalg.norm(vectors, axis=1) if (norms < threshold).all(): return vectors else: vectors[norms >= threshold] *= (threshold / norms[norms >= threshold])[:, np.newaxis] vectors[norms >= threshold] -= np.sign(vectors[norms >= threshold]) * epsilon return vectors def save(self, *args, **kwargs): """Save complete model to disk, inherited from :class:`~gensim.utils.SaveLoad`. See also -------- :meth:`~gensim.models.poincare.PoincareModel.load` Parameters ---------- *args Positional arguments passed to :meth:`~gensim.utils.SaveLoad.save`. **kwargs Keyword arguments passed to :meth:`~gensim.utils.SaveLoad.save`. """ self._loss_grad = None # Can't pickle autograd fn to disk attrs_to_ignore = ['_node_probabilities', '_node_counts_cumsum'] kwargs['ignore'] = set(list(kwargs.get('ignore', [])) + attrs_to_ignore) super(PoincareModel, self).save(*args, **kwargs) @classmethod def load(cls, *args, **kwargs): """Load model from disk, inherited from :class:`~gensim.utils.SaveLoad`. See also -------- :meth:`~gensim.models.poincare.PoincareModel.save` Parameters ---------- *args Positional arguments passed to :meth:`~gensim.utils.SaveLoad.load`. **kwargs Keyword arguments passed to :meth:`~gensim.utils.SaveLoad.load`. Returns ------- :class:`~gensim.models.poincare.PoincareModel` The loaded model. """ model = super(PoincareModel, cls).load(*args, **kwargs) model._init_node_probabilities() return model def _prepare_training_batch(self, relations, all_negatives, check_gradients=False): """Create a training batch and compute gradients and loss for the batch. Parameters ---------- relations : list of tuples List of tuples of positive examples of the form (node_1_index, node_2_index). all_negatives : list of lists List of lists of negative samples for each node_1 in the positive examples. check_gradients : bool, optional Whether to compare the computed gradients to autograd gradients for this batch. Returns ------- :class:`~gensim.models.poincare.PoincareBatch` Node indices, computed gradients and loss for the batch. """ batch_size = len(relations) indices_u, indices_v = [], [] for relation, negatives in zip(relations, all_negatives): u, v = relation indices_u.append(u) indices_v.append(v) indices_v.extend(negatives) vectors_u = self.kv.vectors[indices_u] vectors_v = self.kv.vectors[indices_v].reshape((batch_size, 1 + self.negative, self.size)) vectors_v = vectors_v.swapaxes(0, 1).swapaxes(1, 2) batch = PoincareBatch(vectors_u, vectors_v, indices_u, indices_v, self.regularization_coeff) batch.compute_all() if check_gradients: self._check_gradients(relations, all_negatives, batch) return batch def _check_gradients(self, relations, all_negatives, batch, tol=1e-8): """Compare computed gradients for batch to autograd gradients. Parameters ---------- relations : list of tuples List of tuples of positive examples of the form (node_1_index, node_2_index). all_negatives : list of lists List of lists of negative samples for each node_1 in the positive examples. batch : :class:`~gensim.models.poincare.PoincareBatch` Batch for which computed gradients are to be checked. tol : float, optional The maximum error between our computed gradients and the reference ones from autograd. """ if not AUTOGRAD_PRESENT: logger.warning('autograd could not be imported, cannot do gradient checking') logger.warning('please install autograd to enable gradient checking') return if self._loss_grad is None: self._loss_grad = grad(PoincareModel._loss_fn) max_diff = 0.0 for i, (relation, negatives) in enumerate(zip(relations, all_negatives)): u, v = relation auto_gradients = self._loss_grad( np.vstack((self.kv.vectors[u], self.kv.vectors[[v] + negatives])), self.regularization_coeff) computed_gradients = np.vstack((batch.gradients_u[:, i], batch.gradients_v[:, :, i])) diff = np.abs(auto_gradients - computed_gradients).max() if diff > max_diff: max_diff = diff logger.info('max difference between computed gradients and autograd gradients: %.10f', max_diff) assert max_diff < tol, ( 'Max difference between computed gradients and autograd gradients %.10f, ' 'greater than tolerance %.10f' % (max_diff, tol)) def _sample_negatives_batch(self, nodes): """Get negative examples for each node. Parameters ---------- nodes : iterable of int Iterable of node indices for which negative samples are to be returned. Returns ------- list of lists Each inner list is a list of negative samples for a single node in the input list. """ all_indices = [self._sample_negatives(node) for node in nodes] return all_indices def _train_on_batch(self, relations, check_gradients=False): """Perform training for a single training batch. Parameters ---------- relations : list of tuples of (int, int) List of tuples of positive examples of the form (node_1_index, node_2_index). check_gradients : bool, optional Whether to compare the computed gradients to autograd gradients for this batch. Returns ------- :class:`~gensim.models.poincare.PoincareBatch` The batch that was just trained on, contains computed loss for the batch. """ all_negatives = self._sample_negatives_batch(relation[0] for relation in relations) batch = self._prepare_training_batch(relations, all_negatives, check_gradients) self._update_vectors_batch(batch) return batch @staticmethod def _handle_duplicates(vector_updates, node_indices): """Handle occurrences of multiple updates to the same node in a batch of vector updates. Parameters ---------- vector_updates : numpy.array Array with each row containing updates to be performed on a certain node. node_indices : list of int Node indices on which the above updates are to be performed on. Notes ----- Mutates the `vector_updates` array. Required because vectors[[2, 1, 2]] += np.array([-0.5, 1.0, 0.5]) performs only the last update on the row at index 2. """ counts = Counter(node_indices) node_dict = defaultdict(list) for i, node_index in enumerate(node_indices): node_dict[node_index].append(i) for node_index, count in counts.items(): if count == 1: continue positions = node_dict[node_index] # Move all updates to the same node to the last such update, zeroing all the others vector_updates[positions[-1]] = vector_updates[positions].sum(axis=0) vector_updates[positions[:-1]] = 0 def _update_vectors_batch(self, batch): """Update vectors for nodes in the given batch. Parameters ---------- batch : :class:`~gensim.models.poincare.PoincareBatch` Batch containing computed gradients and node indices of the batch for which updates are to be done. """ grad_u, grad_v = batch.gradients_u, batch.gradients_v indices_u, indices_v = batch.indices_u, batch.indices_v batch_size = len(indices_u) u_updates = (self.alpha * (batch.alpha ** 2) / 4 * grad_u).T self._handle_duplicates(u_updates, indices_u) self.kv.vectors[indices_u] -= u_updates self.kv.vectors[indices_u] = self._clip_vectors(self.kv.vectors[indices_u], self.epsilon) v_updates = self.alpha * (batch.beta ** 2)[:, np.newaxis] / 4 * grad_v v_updates = v_updates.swapaxes(1, 2).swapaxes(0, 1) v_updates = v_updates.reshape(((1 + self.negative) * batch_size, self.size)) self._handle_duplicates(v_updates, indices_v) self.kv.vectors[indices_v] -= v_updates self.kv.vectors[indices_v] = self._clip_vectors(self.kv.vectors[indices_v], self.epsilon) def train(self, epochs, batch_size=10, print_every=1000, check_gradients_every=None): """Train Poincare embeddings using loaded data and model parameters. Parameters ---------- epochs : int Number of iterations (epochs) over the corpus. batch_size : int, optional Number of examples to train on in a single batch. print_every : int, optional Prints progress and average loss after every `print_every` batches. check_gradients_every : int or None, optional Compares computed gradients and autograd gradients after every `check_gradients_every` batches. Useful for debugging, doesn't compare by default. Examples -------- .. sourcecode:: pycon >>> from gensim.models.poincare import PoincareModel >>> relations = [('kangaroo', 'marsupial'), ('kangaroo', 'mammal'), ('gib', 'cat')] >>> model = PoincareModel(relations, negative=2) >>> model.train(epochs=50) """ if self.workers > 1: raise NotImplementedError("Multi-threaded version not implemented yet") # Some divide-by-zero results are handled explicitly old_settings = np.seterr(divide='ignore', invalid='ignore') logger.info( "training model of size %d with %d workers on %d relations for %d epochs and %d burn-in epochs, " "using lr=%.5f burn-in lr=%.5f negative=%d", self.size, self.workers, len(self.all_relations), epochs, self.burn_in, self.alpha, self.burn_in_alpha, self.negative ) if self.burn_in > 0 and not self._burn_in_done: logger.info("starting burn-in (%d epochs)----------------------------------------", self.burn_in) self.alpha = self.burn_in_alpha self._train_batchwise( epochs=self.burn_in, batch_size=batch_size, print_every=print_every, check_gradients_every=check_gradients_every) self._burn_in_done = True logger.info("burn-in finished") self.alpha = self.train_alpha logger.info("starting training (%d epochs)----------------------------------------", epochs) self._train_batchwise( epochs=epochs, batch_size=batch_size, print_every=print_every, check_gradients_every=check_gradients_every) logger.info("training finished") np.seterr(**old_settings) def _train_batchwise(self, epochs, batch_size=10, print_every=1000, check_gradients_every=None): """Train Poincare embeddings using specified parameters. Parameters ---------- epochs : int Number of iterations (epochs) over the corpus. batch_size : int, optional Number of examples to train on in a single batch. print_every : int, optional Prints progress and average loss after every `print_every` batches. check_gradients_every : int or None, optional Compares computed gradients and autograd gradients after every `check_gradients_every` batches. Useful for debugging, doesn't compare by default. """ if self.workers > 1: raise NotImplementedError("Multi-threaded version not implemented yet") for epoch in range(1, epochs + 1): indices = list(range(len(self.all_relations))) self._np_random.shuffle(indices) avg_loss = 0.0 last_time = time.time() for batch_num, i in enumerate(range(0, len(indices), batch_size), start=1): should_print = not (batch_num % print_every) check_gradients = bool(check_gradients_every) and (batch_num % check_gradients_every) == 0 batch_indices = indices[i:i + batch_size] relations = [self.all_relations[idx] for idx in batch_indices] result = self._train_on_batch(relations, check_gradients=check_gradients) avg_loss += result.loss if should_print: avg_loss /= print_every time_taken = time.time() - last_time speed = print_every * batch_size / time_taken logger.info( 'training on epoch %d, examples #%d-#%d, loss: %.2f' % (epoch, i, i + batch_size, avg_loss)) logger.info( 'time taken for %d examples: %.2f s, %.2f examples / s' % (print_every * batch_size, time_taken, speed)) last_time = time.time() avg_loss = 0.0 class PoincareBatch: """Compute Poincare distances, gradients and loss for a training batch. Store intermediate state to avoid recomputing multiple times. """ def __init__(self, vectors_u, vectors_v, indices_u, indices_v, regularization_coeff=1.0): """ Initialize instance with sets of vectors for which distances are to be computed. Parameters ---------- vectors_u : numpy.array Vectors of all nodes `u` in the batch. Expected shape (batch_size, dim). vectors_v : numpy.array Vectors of all positively related nodes `v` and negatively sampled nodes `v'`, for each node `u` in the batch. Expected shape (1 + neg_size, dim, batch_size). indices_u : list of int List of node indices for each of the vectors in `vectors_u`. indices_v : list of lists of int Nested list of lists, each of which is a list of node indices for each of the vectors in `vectors_v` for a specific node `u`. regularization_coeff : float, optional Coefficient to use for l2-regularization """ self.vectors_u = vectors_u.T[np.newaxis, :, :] # (1, dim, batch_size) self.vectors_v = vectors_v # (1 + neg_size, dim, batch_size) self.indices_u = indices_u self.indices_v = indices_v self.regularization_coeff = regularization_coeff self.poincare_dists = None self.euclidean_dists = None self.norms_u = None self.norms_v = None self.alpha = None self.beta = None self.gamma = None self.gradients_u = None self.distance_gradients_u = None self.gradients_v = None self.distance_gradients_v = None self.loss = None self._distances_computed = False self._gradients_computed = False self._distance_gradients_computed = False self._loss_computed = False def compute_all(self): """Convenience method to perform all computations.""" self.compute_distances() self.compute_distance_gradients() self.compute_gradients() self.compute_loss() def compute_distances(self): """Compute and store norms, euclidean distances and poincare distances between input vectors.""" if self._distances_computed: return euclidean_dists = np.linalg.norm(self.vectors_u - self.vectors_v, axis=1) # (1 + neg_size, batch_size) norms_u = np.linalg.norm(self.vectors_u, axis=1) # (1, batch_size) norms_v = np.linalg.norm(self.vectors_v, axis=1) # (1 + neg_size, batch_size) alpha = 1 - norms_u ** 2 # (1, batch_size) beta = 1 - norms_v ** 2 # (1 + neg_size, batch_size) gamma = 1 + 2 * ( (euclidean_dists ** 2) / (alpha * beta) ) # (1 + neg_size, batch_size) poincare_dists = np.arccosh(gamma) # (1 + neg_size, batch_size) exp_negative_distances = np.exp(-poincare_dists) # (1 + neg_size, batch_size) Z = exp_negative_distances.sum(axis=0) # (batch_size) self.euclidean_dists = euclidean_dists self.poincare_dists = poincare_dists self.exp_negative_distances = exp_negative_distances self.Z = Z self.gamma = gamma self.norms_u = norms_u self.norms_v = norms_v self.alpha = alpha self.beta = beta self.gamma = gamma self._distances_computed = True def compute_gradients(self): """Compute and store gradients of loss function for all input vectors.""" if self._gradients_computed: return self.compute_distances() self.compute_distance_gradients() # (1 + neg_size, dim, batch_size) gradients_v = -self.exp_negative_distances[:, np.newaxis, :] * self.distance_gradients_v gradients_v /= self.Z # (1 + neg_size, dim, batch_size) gradients_v[0] += self.distance_gradients_v[0] gradients_v[0] += self.regularization_coeff * 2 * self.vectors_v[0] # (1 + neg_size, dim, batch_size) gradients_u = -self.exp_negative_distances[:, np.newaxis, :] * self.distance_gradients_u gradients_u /= self.Z # (1 + neg_size, dim, batch_size) gradients_u = gradients_u.sum(axis=0) # (dim, batch_size) gradients_u += self.distance_gradients_u[0] assert not np.isnan(gradients_u).any() assert not np.isnan(gradients_v).any() self.gradients_u = gradients_u self.gradients_v = gradients_v self._gradients_computed = True def compute_distance_gradients(self): """Compute and store partial derivatives of poincare distance d(u, v) w.r.t all u and all v.""" if self._distance_gradients_computed: return self.compute_distances() euclidean_dists_squared = self.euclidean_dists ** 2 # (1 + neg_size, batch_size) # (1 + neg_size, 1, batch_size) c_ = (4 / (self.alpha * self.beta * np.sqrt(self.gamma ** 2 - 1)))[:, np.newaxis, :] # (1 + neg_size, 1, batch_size) u_coeffs = ((euclidean_dists_squared + self.alpha) / self.alpha)[:, np.newaxis, :] distance_gradients_u = u_coeffs * self.vectors_u - self.vectors_v # (1 + neg_size, dim, batch_size) distance_gradients_u *= c_ # (1 + neg_size, dim, batch_size) nan_gradients = self.gamma == 1 # (1 + neg_size, batch_size) if nan_gradients.any(): distance_gradients_u.swapaxes(1, 2)[nan_gradients] = 0 self.distance_gradients_u = distance_gradients_u # (1 + neg_size, 1, batch_size) v_coeffs = ((euclidean_dists_squared + self.beta) / self.beta)[:, np.newaxis, :] distance_gradients_v = v_coeffs * self.vectors_v - self.vectors_u # (1 + neg_size, dim, batch_size) distance_gradients_v *= c_ # (1 + neg_size, dim, batch_size) if nan_gradients.any(): distance_gradients_v.swapaxes(1, 2)[nan_gradients] = 0 self.distance_gradients_v = distance_gradients_v self._distance_gradients_computed = True def compute_loss(self): """Compute and store loss value for the given batch of examples.""" if self._loss_computed: return self.compute_distances() self.loss = -np.log(self.exp_negative_distances[0] / self.Z).sum() # scalar self._loss_computed = True class PoincareKeyedVectors(KeyedVectors): """Vectors and vocab for the :class:`~gensim.models.poincare.PoincareModel` training class. Used to perform operations on the vectors such as vector lookup, distance calculations etc. (May be used to save/load final vectors in the plain word2vec format, via the inherited methods save_word2vec_format() and load_word2vec_format().) Examples -------- .. sourcecode:: pycon >>> from gensim.test.utils import datapath >>> >>> # Read the sample relations file and train the model >>> relations = PoincareRelations(file_path=datapath('poincare_hypernyms_large.tsv')) >>> model = PoincareModel(train_data=relations) >>> model.train(epochs=50) >>> >>> # Query the trained model. >>> wv = model.kv.get_vector('kangaroo.n.01') """ def __init__(self, vector_size, vector_count, dtype=REAL): super(PoincareKeyedVectors, self).__init__(vector_size, vector_count, dtype=dtype) self.max_distance = 0 def _load_specials(self, *args, **kwargs): super(PoincareKeyedVectors, self)._load_specials(*args, **kwargs) # fixup rename of syn0 if not hasattr(self, 'vectors'): self.vectors = self.__dict__.pop('syn0') @staticmethod def vector_distance(vector_1, vector_2): """Compute poincare distance between two input vectors. Convenience method over `vector_distance_batch`. Parameters ---------- vector_1 : numpy.array Input vector. vector_2 : numpy.array Input vector. Returns ------- numpy.float Poincare distance between `vector_1` and `vector_2`. """ return PoincareKeyedVectors.vector_distance_batch(vector_1, vector_2[np.newaxis, :])[0] @staticmethod def vector_distance_batch(vector_1, vectors_all): """Compute poincare distances between one vector and a set of other vectors. Parameters ---------- vector_1 : numpy.array vector from which Poincare distances are to be computed, expected shape (dim,). vectors_all : numpy.array for each row in vectors_all, distance from vector_1 is computed, expected shape (num_vectors, dim). Returns ------- numpy.array Poincare distance between `vector_1` and each row in `vectors_all`, shape (num_vectors,). """ euclidean_dists = np.linalg.norm(vector_1 - vectors_all, axis=1) norm = np.linalg.norm(vector_1) all_norms = np.linalg.norm(vectors_all, axis=1) return np.arccosh( 1 + 2 * ( (euclidean_dists ** 2) / ((1 - norm ** 2) * (1 - all_norms ** 2)) ) ) def closest_child(self, node): """Get the node closest to `node` that is lower in the hierarchy than `node`. Parameters ---------- node : {str, int} Key for node for which closest child is to be found. Returns ------- {str, None} Node closest to `node` that is lower in the hierarchy than `node`. If there are no nodes lower in the hierarchy, None is returned. """ all_distances = self.distances(node) all_norms = np.linalg.norm(self.vectors, axis=1) node_norm = all_norms[self.get_index(node)] mask = node_norm >= all_norms if mask.all(): # No nodes lower in the hierarchy return None all_distances = np.ma.array(all_distances, mask=mask) closest_child_index = np.ma.argmin(all_distances) return self.index_to_key[closest_child_index] def closest_parent(self, node): """Get the node closest to `node` that is higher in the hierarchy than `node`. Parameters ---------- node : {str, int} Key for node for which closest parent is to be found. Returns ------- {str, None} Node closest to `node` that is higher in the hierarchy than `node`. If there are no nodes higher in the hierarchy, None is returned. """ all_distances = self.distances(node) all_norms = np.linalg.norm(self.vectors, axis=1) node_norm = all_norms[self.get_index(node)] mask = node_norm <= all_norms if mask.all(): # No nodes higher in the hierarchy return None all_distances = np.ma.array(all_distances, mask=mask) closest_child_index = np.ma.argmin(all_distances) return self.index_to_key[closest_child_index] def descendants(self, node, max_depth=5): """Get the list of recursively closest children from the given node, up to a max depth of `max_depth`. Parameters ---------- node : {str, int} Key for node for which descendants are to be found. max_depth : int Maximum number of descendants to return. Returns ------- list of str Descendant nodes from the node `node`. """ depth = 0 descendants = [] current_node = node while depth < max_depth: descendants.append(self.closest_child(current_node)) current_node = descendants[-1] depth += 1 return descendants def ancestors(self, node): """Get the list of recursively closest parents from the given node. Parameters ---------- node : {str, int} Key for node for which ancestors are to be found. Returns ------- list of str Ancestor nodes of the node `node`. """ ancestors = [] current_node = node ancestor = self.closest_parent(current_node) while ancestor is not None: ancestors.append(ancestor) ancestor = self.closest_parent(ancestors[-1]) return ancestors def distance(self, w1, w2): """Calculate Poincare distance between vectors for nodes `w1` and `w2`. Parameters ---------- w1 : {str, int} Key for first node. w2 : {str, int} Key for second node. Returns ------- float Poincare distance between the vectors for nodes `w1` and `w2`. Examples -------- .. sourcecode:: pycon >>> from gensim.test.utils import datapath >>> >>> # Read the sample relations file and train the model >>> relations = PoincareRelations(file_path=datapath('poincare_hypernyms_large.tsv')) >>> model = PoincareModel(train_data=relations) >>> model.train(epochs=50) >>> >>> # What is the distance between the words 'mammal' and 'carnivore'? >>> model.kv.distance('mammal.n.01', 'carnivore.n.01') 2.9742298803339304 Raises ------ KeyError If either of `w1` and `w2` is absent from vocab. """ vector_1 = self.get_vector(w1) vector_2 = self.get_vector(w2) return self.vector_distance(vector_1, vector_2) def similarity(self, w1, w2): """Compute similarity based on Poincare distance between vectors for nodes `w1` and `w2`. Parameters ---------- w1 : {str, int} Key for first node. w2 : {str, int} Key for second node. Returns ------- float Similarity between the between the vectors for nodes `w1` and `w2` (between 0 and 1). Examples -------- .. sourcecode:: pycon >>> from gensim.test.utils import datapath >>> >>> # Read the sample relations file and train the model >>> relations = PoincareRelations(file_path=datapath('poincare_hypernyms_large.tsv')) >>> model = PoincareModel(train_data=relations) >>> model.train(epochs=50) >>> >>> # What is the similarity between the words 'mammal' and 'carnivore'? >>> model.kv.similarity('mammal.n.01', 'carnivore.n.01') 0.25162107631176484 Raises ------ KeyError If either of `w1` and `w2` is absent from vocab. """ return 1 / (1 + self.distance(w1, w2)) def most_similar(self, node_or_vector, topn=10, restrict_vocab=None): """Find the top-N most similar nodes to the given node or vector, sorted in increasing order of distance. Parameters ---------- node_or_vector : {str, int, numpy.array} node key or vector for which similar nodes are to be found. topn : int or None, optional Number of top-N similar nodes to return, when `topn` is int. When `topn` is None, then distance for all nodes are returned. restrict_vocab : int or None, optional Optional integer which limits the range of vectors which are searched for most-similar values. For example, restrict_vocab=10000 would only check the first 10000 node vectors in the vocabulary order. This may be meaningful if vocabulary is sorted by descending frequency. Returns -------- list of (str, float) or numpy.array When `topn` is int, a sequence of (node, distance) is returned in increasing order of distance. When `topn` is None, then similarities for all words are returned as a one-dimensional numpy array with the size of the vocabulary. Examples -------- .. sourcecode:: pycon >>> from gensim.test.utils import datapath >>> >>> # Read the sample relations file and train the model >>> relations = PoincareRelations(file_path=datapath('poincare_hypernyms_large.tsv')) >>> model = PoincareModel(train_data=relations) >>> model.train(epochs=50) >>> >>> # Which words are most similar to 'kangaroo'? >>> model.kv.most_similar('kangaroo.n.01', topn=2) [(u'kangaroo.n.01', 0.0), (u'marsupial.n.01', 0.26524229460827725)] """ if isinstance(topn, Integral) and topn < 1: return [] if not restrict_vocab: all_distances = self.distances(node_or_vector) else: nodes_to_use = self.index_to_key[:restrict_vocab] all_distances = self.distances(node_or_vector, nodes_to_use) if isinstance(node_or_vector, (str, int,)): node_index = self.get_index(node_or_vector) else: node_index = None if not topn: closest_indices = matutils.argsort(all_distances) else: closest_indices = matutils.argsort(all_distances, topn=1 + topn) result = [ (self.index_to_key[index], float(all_distances[index])) for index in closest_indices if (not node_index or index != node_index) # ignore the input node ] if topn: result = result[:topn] return result def distances(self, node_or_vector, other_nodes=()): """Compute Poincare distances from given `node_or_vector` to all nodes in `other_nodes`. If `other_nodes` is empty, return distance between `node_or_vector` and all nodes in vocab. Parameters ---------- node_or_vector : {str, int, numpy.array} Node key or vector from which distances are to be computed. other_nodes : {iterable of str, iterable of int, None}, optional For each node in `other_nodes` distance from `node_or_vector` is computed. If None or empty, distance of `node_or_vector` from all nodes in vocab is computed (including itself). Returns ------- numpy.array Array containing distances to all nodes in `other_nodes` from input `node_or_vector`, in the same order as `other_nodes`. Examples -------- .. sourcecode:: pycon >>> from gensim.test.utils import datapath >>> >>> # Read the sample relations file and train the model >>> relations = PoincareRelations(file_path=datapath('poincare_hypernyms_large.tsv')) >>> model = PoincareModel(train_data=relations) >>> model.train(epochs=50) >>> >>> # Check the distances between a word and a list of other words. >>> model.kv.distances('mammal.n.01', ['carnivore.n.01', 'dog.n.01']) array([2.97422988, 2.83007402]) >>> # Check the distances between a word and every other word in the vocab. >>> all_distances = model.kv.distances('mammal.n.01') Raises ------ KeyError If either `node_or_vector` or any node in `other_nodes` is absent from vocab. """ if isinstance(node_or_vector, str): input_vector = self.get_vector(node_or_vector) else: input_vector = node_or_vector if not other_nodes: other_vectors = self.vectors else: other_indices = [self.get_index(node) for node in other_nodes] other_vectors = self.vectors[other_indices] return self.vector_distance_batch(input_vector, other_vectors) def norm(self, node_or_vector): """Compute absolute position in hierarchy of input node or vector. Values range between 0 and 1. A lower value indicates the input node or vector is higher in the hierarchy. Parameters ---------- node_or_vector : {str, int, numpy.array} Input node key or vector for which position in hierarchy is to be returned. Returns ------- float Absolute position in the hierarchy of the input vector or node. Examples -------- .. sourcecode:: pycon >>> from gensim.test.utils import datapath >>> >>> # Read the sample relations file and train the model >>> relations = PoincareRelations(file_path=datapath('poincare_hypernyms_large.tsv')) >>> model = PoincareModel(train_data=relations) >>> model.train(epochs=50) >>> >>> # Get the norm of the embedding of the word `mammal`. >>> model.kv.norm('mammal.n.01') 0.6423008703542398 Notes ----- The position in hierarchy is based on the norm of the vector for the node. """ if isinstance(node_or_vector, str): input_vector = self.get_vector(node_or_vector) else: input_vector = node_or_vector return np.linalg.norm(input_vector) def difference_in_hierarchy(self, node_or_vector_1, node_or_vector_2): """Compute relative position in hierarchy of `node_or_vector_1` relative to `node_or_vector_2`. A positive value indicates `node_or_vector_1` is higher in the hierarchy than `node_or_vector_2`. Parameters ---------- node_or_vector_1 : {str, int, numpy.array} Input node key or vector. node_or_vector_2 : {str, int, numpy.array} Input node key or vector. Returns ------- float Relative position in hierarchy of `node_or_vector_1` relative to `node_or_vector_2`. Examples -------- .. sourcecode:: pycon >>> from gensim.test.utils import datapath >>> >>> # Read the sample relations file and train the model >>> relations = PoincareRelations(file_path=datapath('poincare_hypernyms_large.tsv')) >>> model = PoincareModel(train_data=relations) >>> model.train(epochs=50) >>> >>> model.kv.difference_in_hierarchy('mammal.n.01', 'dog.n.01') 0.05382517902410999 >>> model.kv.difference_in_hierarchy('dog.n.01', 'mammal.n.01') -0.05382517902410999 Notes ----- The returned value can be positive or negative, depending on whether `node_or_vector_1` is higher or lower in the hierarchy than `node_or_vector_2`. """ return self.norm(node_or_vector_2) - self.norm(node_or_vector_1) class PoincareRelations: """Stream relations for `PoincareModel` from a tsv-like file.""" def __init__(self, file_path, encoding='utf8', delimiter='\t'): """Initialize instance from file containing a pair of nodes (a relation) per line. Parameters ---------- file_path : str Path to file containing a pair of nodes (a relation) per line, separated by `delimiter`. Since the relations are asymmetric, the order of `u` and `v` nodes in each pair matters. To express a "u is v" relation, the lines should take the form `u delimeter v`. e.g: `kangaroo mammal` is a tab-delimited line expressing a "`kangaroo is a mammal`" relation. For a full input file example, see `gensim/test/test_data/poincare_hypernyms.tsv `_. encoding : str, optional Character encoding of the input file. delimiter : str, optional Delimiter character for each relation. """ self.file_path = file_path self.encoding = encoding self.delimiter = delimiter def __iter__(self): """Stream relations from self.file_path decoded into unicode strings. Yields ------- (unicode, unicode) Relation from input file. """ with utils.open(self.file_path, 'rb') as file_obj: if sys.version_info[0] < 3: lines = file_obj else: lines = (line.decode(self.encoding) for line in file_obj) # csv.reader requires bytestring input in python2, unicode input in python3 reader = csv.reader(lines, delimiter=self.delimiter) for row in reader: if sys.version_info[0] < 3: row = [value.decode(self.encoding) for value in row] yield tuple(row) class NegativesBuffer: """Buffer and return negative samples.""" def __init__(self, items): """Initialize instance from list or numpy array of samples. Parameters ---------- items : list/numpy.array List or array containing negative samples. """ self._items = items self._current_index = 0 def num_items(self): """Get the number of items remaining in the buffer. Returns ------- int Number of items in the buffer that haven't been consumed yet. """ return len(self._items) - self._current_index def get_items(self, num_items): """Get the next `num_items` from buffer. Parameters ---------- num_items : int Number of items to fetch. Returns ------- numpy.array or list Slice containing `num_items` items from the original data. Notes ----- No error is raised if less than `num_items` items are remaining, simply all the remaining items are returned. """ start_index = self._current_index end_index = start_index + num_items self._current_index += num_items return self._items[start_index:end_index] class ReconstructionEvaluation: """Evaluate reconstruction on given network for given embedding.""" def __init__(self, file_path, embedding): """Initialize evaluation instance with tsv file containing relation pairs and embedding to be evaluated. Parameters ---------- file_path : str Path to tsv file containing relation pairs. embedding : :class:`~gensim.models.poincare.PoincareKeyedVectors` Embedding to be evaluated. """ items = set() relations = defaultdict(set) with utils.open(file_path, 'r') as f: reader = csv.reader(f, delimiter='\t') for row in reader: assert len(row) == 2, 'Hypernym pair has more than two items' item_1_index = embedding.get_index(row[0]) item_2_index = embedding.get_index(row[1]) relations[item_1_index].add(item_2_index) items.update([item_1_index, item_2_index]) self.items = items self.relations = relations self.embedding = embedding @staticmethod def get_positive_relation_ranks_and_avg_prec(all_distances, positive_relations): """Compute ranks and Average Precision of positive relations. Parameters ---------- all_distances : numpy.array of float Array of all distances (floats) for a specific item. positive_relations : list List of indices of positive relations for the item. Returns ------- (list of int, float) The list contains ranks of positive relations in the same order as `positive_relations`. The float is the Average Precision of the ranking, e.g. ([1, 2, 3, 20], 0.610). """ positive_relation_distances = all_distances[positive_relations] negative_relation_distances = np.ma.array(all_distances, mask=False) negative_relation_distances.mask[positive_relations] = True # Compute how many negative relation distances are less than each positive relation distance, plus 1 for rank ranks = (negative_relation_distances < positive_relation_distances[:, np.newaxis]).sum(axis=1) + 1 map_ranks = np.sort(ranks) + np.arange(len(ranks)) avg_precision = ((np.arange(1, len(map_ranks) + 1) / np.sort(map_ranks)).mean()) return list(ranks), avg_precision def evaluate(self, max_n=None): """Evaluate all defined metrics for the reconstruction task. Parameters ---------- max_n : int, optional Maximum number of positive relations to evaluate, all if `max_n` is None. Returns ------- dict of (str, float) (metric_name, metric_value) pairs, e.g. {'mean_rank': 50.3, 'MAP': 0.31}. """ mean_rank, map_ = self.evaluate_mean_rank_and_map(max_n) return {'mean_rank': mean_rank, 'MAP': map_} def evaluate_mean_rank_and_map(self, max_n=None): """Evaluate mean rank and MAP for reconstruction. Parameters ---------- max_n : int, optional Maximum number of positive relations to evaluate, all if `max_n` is None. Returns ------- (float, float) (mean_rank, MAP), e.g (50.3, 0.31). """ ranks = [] avg_precision_scores = [] for i, item in enumerate(self.items, start=1): if item not in self.relations: continue item_relations = list(self.relations[item]) item_term = self.embedding.index_to_key[item] item_distances = self.embedding.distances(item_term) positive_relation_ranks, avg_precision = \ self.get_positive_relation_ranks_and_avg_prec(item_distances, item_relations) ranks += positive_relation_ranks avg_precision_scores.append(avg_precision) if max_n is not None and i > max_n: break return np.mean(ranks), np.mean(avg_precision_scores) class LinkPredictionEvaluation: """Evaluate reconstruction on given network for given embedding.""" def __init__(self, train_path, test_path, embedding): """Initialize evaluation instance with tsv file containing relation pairs and embedding to be evaluated. Parameters ---------- train_path : str Path to tsv file containing relation pairs used for training. test_path : str Path to tsv file containing relation pairs to evaluate. embedding : :class:`~gensim.models.poincare.PoincareKeyedVectors` Embedding to be evaluated. """ items = set() relations = {'known': defaultdict(set), 'unknown': defaultdict(set)} data_files = {'known': train_path, 'unknown': test_path} for relation_type, data_file in data_files.items(): with utils.open(data_file, 'r') as f: reader = csv.reader(f, delimiter='\t') for row in reader: assert len(row) == 2, 'Hypernym pair has more than two items' item_1_index = embedding.get_index(row[0]) item_2_index = embedding.get_index(row[1]) relations[relation_type][item_1_index].add(item_2_index) items.update([item_1_index, item_2_index]) self.items = items self.relations = relations self.embedding = embedding @staticmethod def get_unknown_relation_ranks_and_avg_prec(all_distances, unknown_relations, known_relations): """Compute ranks and Average Precision of unknown positive relations. Parameters ---------- all_distances : numpy.array of float Array of all distances for a specific item. unknown_relations : list of int List of indices of unknown positive relations. known_relations : list of int List of indices of known positive relations. Returns ------- tuple (list of int, float) The list contains ranks of positive relations in the same order as `positive_relations`. The float is the Average Precision of the ranking, e.g. ([1, 2, 3, 20], 0.610). """ unknown_relation_distances = all_distances[unknown_relations] negative_relation_distances = np.ma.array(all_distances, mask=False) negative_relation_distances.mask[unknown_relations] = True negative_relation_distances.mask[known_relations] = True # Compute how many negative relation distances are less than each unknown relation distance, plus 1 for rank ranks = (negative_relation_distances < unknown_relation_distances[:, np.newaxis]).sum(axis=1) + 1 map_ranks = np.sort(ranks) + np.arange(len(ranks)) avg_precision = ((np.arange(1, len(map_ranks) + 1) / np.sort(map_ranks)).mean()) return list(ranks), avg_precision def evaluate(self, max_n=None): """Evaluate all defined metrics for the link prediction task. Parameters ---------- max_n : int, optional Maximum number of positive relations to evaluate, all if `max_n` is None. Returns ------- dict of (str, float) (metric_name, metric_value) pairs, e.g. {'mean_rank': 50.3, 'MAP': 0.31}. """ mean_rank, map_ = self.evaluate_mean_rank_and_map(max_n) return {'mean_rank': mean_rank, 'MAP': map_} def evaluate_mean_rank_and_map(self, max_n=None): """Evaluate mean rank and MAP for link prediction. Parameters ---------- max_n : int, optional Maximum number of positive relations to evaluate, all if `max_n` is None. Returns ------- tuple (float, float) (mean_rank, MAP), e.g (50.3, 0.31). """ ranks = [] avg_precision_scores = [] for i, item in enumerate(self.items, start=1): if item not in self.relations['unknown']: # No positive relations to predict for this node continue unknown_relations = list(self.relations['unknown'][item]) known_relations = list(self.relations['known'][item]) item_term = self.embedding.index_to_key[item] item_distances = self.embedding.distances(item_term) unknown_relation_ranks, avg_precision = \ self.get_unknown_relation_ranks_and_avg_prec(item_distances, unknown_relations, known_relations) ranks += unknown_relation_ranks avg_precision_scores.append(avg_precision) if max_n is not None and i > max_n: break return np.mean(ranks), np.mean(avg_precision_scores) class LexicalEntailmentEvaluation: """Evaluate reconstruction on given network for any embedding.""" def __init__(self, filepath): """Initialize evaluation instance with HyperLex text file containing relation pairs. Parameters ---------- filepath : str Path to HyperLex text file. """ expected_scores = {} with utils.open(filepath, 'r') as f: reader = csv.DictReader(f, delimiter=' ') for row in reader: word_1, word_2 = row['WORD1'], row['WORD2'] expected_scores[(word_1, word_2)] = float(row['AVG_SCORE']) self.scores = expected_scores self.alpha = 1000 def score_function(self, embedding, trie, term_1, term_2): """Compute predicted score - extent to which `term_1` is a type of `term_2`. Parameters ---------- embedding : :class:`~gensim.models.poincare.PoincareKeyedVectors` Embedding to use for computing predicted score. trie : :class:`pygtrie.Trie` Trie to use for finding matching vocab terms for input terms. term_1 : str Input term. term_2 : str Input term. Returns ------- float Predicted score (the extent to which `term_1` is a type of `term_2`). """ try: word_1_terms = self.find_matching_terms(trie, term_1) word_2_terms = self.find_matching_terms(trie, term_2) except KeyError: raise ValueError("No matching terms found for either %s or %s" % (term_1, term_2)) min_distance = np.inf min_term_1, min_term_2 = None, None for term_1 in word_1_terms: for term_2 in word_2_terms: distance = embedding.distance(term_1, term_2) if distance < min_distance: min_term_1, min_term_2 = term_1, term_2 min_distance = distance assert min_term_1 is not None and min_term_2 is not None vector_1, vector_2 = embedding.get_vector(min_term_1), embedding.get_vector(min_term_2) norm_1, norm_2 = np.linalg.norm(vector_1), np.linalg.norm(vector_2) return -1 * (1 + self.alpha * (norm_2 - norm_1)) * min_distance @staticmethod def find_matching_terms(trie, word): """Find terms in the `trie` beginning with the `word`. Parameters ---------- trie : :class:`pygtrie.Trie` Trie to use for finding matching terms. word : str Input word to use for prefix search. Returns ------- list of str List of matching terms. """ matches = trie.items('%s.' % word) matching_terms = [''.join(key_chars) for key_chars, value in matches] return matching_terms @staticmethod def create_vocab_trie(embedding): """Create trie with vocab terms of the given embedding to enable quick prefix searches. Parameters ---------- embedding : :class:`~gensim.models.poincare.PoincareKeyedVectors` Embedding for which trie is to be created. Returns ------- :class:`pygtrie.Trie` Trie containing vocab terms of the input embedding. """ try: from pygtrie import Trie except ImportError: raise ImportError( 'pygtrie could not be imported, please install pygtrie in order to use LexicalEntailmentEvaluation') vocab_trie = Trie() for key in embedding.key_to_index: vocab_trie[key] = True return vocab_trie def evaluate_spearman(self, embedding): """Evaluate spearman scores for lexical entailment for given embedding. Parameters ---------- embedding : :class:`~gensim.models.poincare.PoincareKeyedVectors` Embedding for which evaluation is to be done. Returns ------- float Spearman correlation score for the task for input embedding. """ predicted_scores = [] expected_scores = [] skipped = 0 count = 0 vocab_trie = self.create_vocab_trie(embedding) for (word_1, word_2), expected_score in self.scores.items(): try: predicted_score = self.score_function(embedding, vocab_trie, word_1, word_2) except ValueError: skipped += 1 continue count += 1 predicted_scores.append(predicted_score) expected_scores.append(expected_score) logger.info('skipped pairs: %d out of %d' % (skipped, len(self.scores))) spearman = spearmanr(expected_scores, predicted_scores) return spearman.correlation