gridsearch.py

import numpy as np
import plotly
from sklearn import svm
from sklearn.linear_model import RidgeClassifier, LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from tqdm import tqdm

from classifier import train, test
from dataset import FaceDataset


def find_svm_hyperparams():
    """
    Finds best SVM hyperparams based on evaluating on test split
    """
    NUM_ITERS = 10
    # coefs = np.arange(-5, 5).astype(np.float)
    coefs = np.linspace(0.25, 1, 10)
    Cs = np.power(2, coefs)
    results = []

    for _ in range(NUM_ITERS):
        data = FaceDataset("embeddings/dev", n=50)
        train_data, train_labels = data.train()
        test_data, test_labels = data.test()
        accs = []
        for c in tqdm(Cs):
            clf = svm.SVC(kernel="linear", C=c)
            clf, _ = train(clf, train_data, train_labels)
            acc, _ = test(clf, test_data, test_labels)
            accs.append(acc)
        results.append(accs)

    results = np.mean(results, axis=0)
    s = plotly.graph_objs.Scatter(x=Cs, y=results)
    plotly.offline.plot([s], filename="svm_linear.html")
    print("C={}".format(Cs[np.argmax(results)]))


def find_logistic_regression_hp():
    NUM_ITERS = 10
    # coefs = np.arange(-5, 5).astype(np.float)
    coefs = np.linspace(1, 5, 10)
    Cs = np.power(2, coefs)
    results = []

    for _ in range(NUM_ITERS):
        data = FaceDataset("embeddings/dev", n=50)
        train_data, train_labels = data.train()
        test_data, test_labels = data.test()
        accs = []
        for c in tqdm(Cs):
            clf = LogisticRegression(solver="lbfgs", multi_class="auto", C=c, max_iter=1000)
            clf, _ = train(clf, train_data, train_labels)
            acc, _ = test(clf, test_data, test_labels)
            accs.append(acc)
        results.append(accs)

    results = np.mean(results, axis=0)
    s = plotly.graph_objs.Scatter(x=Cs, y=results)
    plotly.offline.plot([s], filename="svm_linear.html")
    print("C={}".format(Cs[np.argmax(results)]))


def ridge_regression():
    NUM_ITERS = 10
    # coefs = np.arange(-5, 0).astype(np.float)
    coefs = np.linspace(-10, -5, 10)
    As = np.power(2, coefs)
    results = []

    for _ in range(NUM_ITERS):
        data = FaceDataset("embeddings/dev", n=50)
        train_data, train_labels = data.train()
        test_data, test_labels = data.test()
        accs = []
        for a in tqdm(As):
            clf = RidgeClassifier(alpha=a, solver="lsqr")
            clf, _ = train(clf, train_data, train_labels)
            acc, _ = test(clf, test_data, test_labels)
            accs.append(acc)
        results.append(accs)

    results = np.mean(results, axis=0)
    s = plotly.graph_objs.Scatter(x=As, y=results)
    plotly.offline.plot([s], filename="svm_linear.html")
    print("A={}".format(As[np.argmax(results)]))


def find_linear_svm_hyperparams():
    """
    Fits a linear SVM model
    """
    NUM_ITERS = 10
    # coefs = np.arange(-5, 5).astype(np.float)
    coefs = np.linspace(0.25, 1, 10)
    Cs = np.power(2, coefs)
    results = []

    for _ in range(NUM_ITERS):
        data = FaceDataset("embeddings/dev", n=50)
        train_data, train_labels = data.train()
        test_data, test_labels = data.test()
        accs = []
        for c in tqdm(Cs):
            clf = svm.SVC(kernel="linear", C=c)
            clf, _ = train(clf, train_data, train_labels)
            acc, _ = test(clf, test_data, test_labels)
            accs.append(acc)
        results.append(accs)

    results = np.mean(results, axis=0)
    s = plotly.graph_objs.Scatter(x=Cs, y=results)
    plotly.offline.plot([s], filename="svm_linear.html")
    print("C={}".format(Cs[np.argmax(results)]))


def find_knn_hyperparams():
    """
    Finds best KNN hyperparams
    """
    n_neighbors = np.arange(5, 10)
    ps = np.arange(1, 10)
    results = []

    for p in ps:
        result = []
        for _ in range(10):
            data = FaceDataset("embeddings/known", n=50)
            train_data, train_labels = data.train()
            test_data, test_labels = data.test()
            accs = []
            for n in n_neighbors:
                clf = KNeighborsClassifier(n_neighbors=n, weights="distance", p=p)
                clf, _ = train(clf, train_data, train_labels)
                acc, _ = test(clf, test_data, test_labels)
                accs.append(acc)
            result.append(accs)
        result = np.mean(result, axis=0)
        results.append(result)

    plots = []
    for i in range(len(ps)):
        p = plotly.graph_objs.Scatter(x=n_neighbors, y=results[i], name="p={}".format(ps[i]))
        plots.append(p)

    plotly.offline.plot(plots, filename="knn.html")
    print("C={}".format(n_neighbors[np.argmax(results)]))


if __name__ == '__main__':
    gmm()