Source code for DLL.MachineLearning.UnsupervisedLearning.DimensionalityReduction._TSNE

import torch

from . import PCA
from ....DeepLearning.Optimisers import SGD


[docs] class TSNE: """ T-distributed Stochastic Neighbor Embedding (T-SNE) class for dimensionality reduction. This implementation is based on `this paper <https://jmlr.org/papers/volume9/vandermaaten08a/vandermaaten08a.pdf>`_ and `this article <https://towardsdatascience.com/t-sne-from-scratch-ft-numpy-172ee2a61df7/>`_. The main difference is that this implementation uses vectorized matrix operations making it considerably faster than the loop approach used in the article. Args: n_components (int): Number of principal components to keep. The number must be a positive integer. init (str, optional): The method for initializing the embedding. Must be in ``["pca", "random"]``. Defaults to ``"pca"``. p (int, optional): The order of the chosen metric. Must be a positive integer. Defaults to 2, which corresponds to the Euclidian metric. early_exaggeration (float | int, optional): Determines how far apart the clusters are in the embedding space. Must be a positive real number. Defaults to 12.0. perplexity (float | int, optional): Determines how far can samples be from one another to be considered neighbors. Must be a positive real number. Defaults to 30.0. One should consider using something between 5 and 50 to begin with. learning_rate (float | int, optional): Determines how long steps do we take towards the gradient. Must be a positive real number. It is recommended to use a value between 10.0 and 1000.0. Defaults to ``"auto"``, where we use a value of ``max(n_samples / (4 * early_exaggeration), 50)``. Attributes: history (list[float]): The history of KL-divergence loss function each epoch. Available after fitting the model. """ def __init__(self, n_components=2, init="pca", p=2, early_exaggeration=12.0, perplexity=30.0, learning_rate="auto"): if not isinstance(n_components, int) or n_components < 1: raise ValueError("n_components must be a positive integer.") if not init in ["pca", "random"]: raise ValueError('init must be in ["pca", "random"].') if not isinstance(p, int) or p <= 0: raise ValueError("p must be a positive integer.") if not isinstance(early_exaggeration, float | int) or early_exaggeration <= 0: raise ValueError("early_exaggeration must be a positive real number.") if not isinstance(perplexity, float | int) or perplexity <= 0: raise ValueError("perplexity must be a positive real number.") if (learning_rate != "auto" and (not isinstance(learning_rate, float | int) or learning_rate <= 0)): raise ValueError('learning_rate must be a positive real number or "auto".') self.n_components = n_components self.init = init self.p = p self.early_exaggeration = early_exaggeration self.perplexity = perplexity self.learning_rate = learning_rate def _initialize(self, X): if self.init == "pca": return PCA(n_components=self.n_components).fit_transform(X) else: return torch.normal(0, 1e-4, size=(len(X), self.n_components)) def _symmetric_affinities(self, X): affinities = self._pairwise_affinities(X) p_ij_symmetric = (affinities + affinities.T) / (2 * len(affinities)) eps = torch.tensor(torch.finfo(p_ij_symmetric.dtype).tiny) p_ij_symmetric = torch.maximum(p_ij_symmetric, eps) return p_ij_symmetric def _pairwise_affinities(self, X): diff = X.unsqueeze(1) - X.unsqueeze(0) norm = torch.norm(diff, dim=2, p=self.p) variance = self._grid_search(norm) affinities = torch.exp(-norm ** self.p / (2 * variance[:, None] ** 2)) affinities.fill_diagonal_(0) row_sums = affinities.sum(dim=1, keepdim=True) zero_rows = row_sums == 0 eps = torch.finfo(affinities.dtype).tiny row_sums[zero_rows] = eps affinities /= row_sums affinities[affinities == 0] = eps return affinities def _grid_search(self, norm): n_samples = len(norm) std_norm = torch.std(norm, dim=1, keepdim=True) resolution = 200 sigma_search_values = torch.linspace(0.01, 5.0, resolution) sigma_search_values = std_norm * sigma_search_values p_matrix = torch.exp(-norm.unsqueeze(2) ** self.p / (2 * sigma_search_values.pow(2).unsqueeze(1))) eps = torch.tensor(torch.finfo(p_matrix.dtype).tiny) p_matrix[torch.eye(n_samples, dtype=torch.bool)] = eps # do not set to 0 to avoid division by zero errors p_new_matrix = torch.maximum(p_matrix / p_matrix.sum(dim=1, keepdim=True), eps) H_values = -torch.sum(p_new_matrix * torch.log2(p_new_matrix), dim=1) result = torch.abs(torch.log(torch.tensor(self.perplexity)) - H_values * torch.log(torch.tensor(2.0))) sigma_indices = torch.argmin(result, dim=1) return sigma_search_values[torch.arange(n_samples), sigma_indices] def _low_dimensional_affinities(self, low_dimensional_representation): diff = low_dimensional_representation.unsqueeze(1) - low_dimensional_representation.unsqueeze(0) norm = torch.norm(diff, dim=2, p=self.p) result = (1 + norm ** self.p).pow(-1) result.fill_diagonal_(0) result_sum = result.sum() result = result / result_sum if result_sum > 0 else result eps = torch.tensor(torch.finfo(result.dtype).tiny) return torch.maximum(result, eps) def _gradient(self, affinities, low_dimensional_affinities, low_dimensional_representation): diff = low_dimensional_representation.unsqueeze(1) - low_dimensional_representation.unsqueeze(0) A = affinities - low_dimensional_affinities B = (1 + torch.norm(diff, dim=2, p=self.p)).pow(-1) return 4 * torch.sum(((A * B).unsqueeze(2) * diff), dim=1)
[docs] def fit_transform(self, X, epochs=100, verbose=False): """ Fits the T-SNE model to the input data. Args: X (torch.Tensor of shape (n_samples, n_features)): The input data, where each row is a sample and each column is a feature. epochs (int, optional): The number of training epochs after early exaggeration. Must be a positive integer. Defaults to 100. Due to early exaggeration, the embedding is updated epochs + 250 times. verbose (bool, optional): Determines if the loss is printed each epoch. Must be a boolean. Defaults to ``False``. Returns: embedding (torch.tensor): The embedded samples of shape (n_samples, n_components). """ if not isinstance(X, torch.Tensor): raise TypeError("The input matrix must be a PyTorch tensor.") if X.ndim != 2 or X.shape[0] == 1: raise ValueError("The input matrix must be a 2 dimensional tensor with atleast 2 samples.") if not isinstance(epochs, int) or epochs <= 0: raise ValueError("epochs must be a positive integer.") if not isinstance(verbose, bool): raise ValueError("verbose must be a boolean.") result = self._initialize(X) symmetric_affinities = self._symmetric_affinities(X) learning_rate = max(len(X) / (4 * self.early_exaggeration), 50) if self.learning_rate == "auto" else self.learning_rate optimiser = SGD(learning_rate=learning_rate, momentum=0.5) optimiser.initialise_parameters([result]) self.history = [] early_exaggeration = self.early_exaggeration for epoch in range(epochs + 250): if epoch > 250: optimiser.momentum = 0.8 early_exaggeration = 1 low_dimensional_affinities = self._low_dimensional_affinities(result) result.grad = self._gradient(early_exaggeration * symmetric_affinities, low_dimensional_affinities, result) loss = torch.sum(symmetric_affinities * torch.log(symmetric_affinities / low_dimensional_affinities)) self.history.append(loss) if verbose: print(f"Epoch: {epoch + 1} - KL-divergence loss: {loss:.3f}") optimiser.update_parameters() return result
[docs] def fit(self, X, epochs=100, verbose=False): """ Wrapper for the TSNE.fit_transform(X) method. Args: X (torch.Tensor of shape (n_samples, n_features)): The input data, where each row is a sample and each column is a feature. epochs (int, optional): The number of training epochs after early exaggeration. Must be a positive integer. Defaults to 100. verbose (bool, optional): Determines if the loss is printed each epoch. Must be a boolean. Defaults to ``False``. """ return self.fit_transform(X, epochs, verbose)