| 17 | |
| 18 | |
| 19 | class TSNE(BaseEstimator): |
| 20 | y_required = False |
| 21 | |
| 22 | def __init__( |
| 23 | self, n_components=2, perplexity=30.0, max_iter=200, learning_rate=500 |
| 24 | ): |
| 25 | """A t-Distributed Stochastic Neighbor Embedding implementation. |
| 26 | |
| 27 | Parameters |
| 28 | ---------- |
| 29 | max_iter : int, default 200 |
| 30 | perplexity : float, default 30.0 |
| 31 | n_components : int, default 2 |
| 32 | """ |
| 33 | self.max_iter = max_iter |
| 34 | self.perplexity = perplexity |
| 35 | self.n_components = n_components |
| 36 | self.initial_momentum = 0.5 |
| 37 | self.final_momentum = 0.8 |
| 38 | self.min_gain = 0.01 |
| 39 | self.lr = learning_rate |
| 40 | self.tol = 1e-5 |
| 41 | self.perplexity_tries = 50 |
| 42 | |
| 43 | def fit_transform(self, X, y=None): |
| 44 | self._setup_input(X, y) |
| 45 | |
| 46 | Y = np.random.randn(self.n_samples, self.n_components) |
| 47 | velocity = np.zeros_like(Y) |
| 48 | gains = np.ones_like(Y) |
| 49 | |
| 50 | P = self._get_pairwise_affinities(X) |
| 51 | |
| 52 | iter_num = 0 |
| 53 | while iter_num < self.max_iter: |
| 54 | iter_num += 1 |
| 55 | |
| 56 | D = l2_distance(Y) |
| 57 | Q = self._q_distribution(D) |
| 58 | |
| 59 | # Normalizer q distribution |
| 60 | Q_n = Q / np.sum(Q) |
| 61 | |
| 62 | # Early exaggeration & momentum |
| 63 | pmul = 4.0 if iter_num < 100 else 1.0 |
| 64 | momentum = 0.5 if iter_num < 20 else 0.8 |
| 65 | |
| 66 | # Perform gradient step |
| 67 | grads = np.zeros(Y.shape) |
| 68 | for i in range(self.n_samples): |
| 69 | grad = 4 * np.dot((pmul * P[i] - Q_n[i]) * Q[i], Y[i] - Y) |
| 70 | grads[i] = grad |
| 71 | |
| 72 | gains = (gains + 0.2) * ((grads > 0) != (velocity > 0)) + (gains * 0.8) * ( |
| 73 | (grads > 0) == (velocity > 0) |
| 74 | ) |
| 75 | gains = gains.clip(min=self.min_gain) |
| 76 | |