Skip to content
Snippets Groups Projects
make_data_simple.py 3.01 KiB
Newer Older
  • Learn to ignore specific revisions
  • abda's avatar
    abda committed
    #%%
    
    import numpy as np
    
    
    def make_data(example_nr, n = 200, noise = 1):
        '''
        Generate data for training a simple neural network.
        
            Arguments:
                example_nr: a number 1 to 3 for each example.
                n: number of points in each class set.
                noise: noise level, best between 0.5 and 2.
            Returns:
                X: 2 x 2n array of points (there are n points in each class).
                T: 2 x 2n target values.
                x: grid points for testing the neural network.
                dim: size of the area covered by the grid points.
    
            Authors: Vedrana Andersen Dahl and Anders Bjorholm Dahl - 25/3-2020
            vand@dtu.dk, abda@dtu.dk
        '''
    
        rg = np.random.default_rng()
        
        dim = (100, 100)
        
        QX, QY = np.meshgrid(range(0, dim[0]), range(0, dim[1]))
        x_grid = np.c_[np.ravel(QX), np.ravel(QY)]
        
        #  Targets: first half class 0, second half class 1
        T = np.vstack((np.tile([True, False], (n, 1)), 
                       np.tile([False, True], (n, 1))))
        
        if example_nr == 1 :  # two separated clusters
    
            X = np.vstack((np.tile([30., 30.], (n, 1)), 
                           np.tile([70., 70.], (n, 1))))
            X += rg.normal(size=X.shape, scale=10*noise)  # add noise
    
        elif example_nr == 2 :  # concentric clusters
    
            rand_ang = 2 * np.pi * rg.uniform(size=n)
            X = np.vstack((30 * np.array([np.cos(rand_ang), np.sin(rand_ang)]).T, 
                           np.tile([0., 0.], (n, 1))))
            X += [50, 50]  # center
            X += rg.normal(size=X.shape, scale=5*noise)# add noise
        
        elif example_nr == 3 :  # 2x2 checkerboard 
            n1 = n//2
            n2 = n//2 + n%2  # if n is odd n2 will have 1 element more
    
            X = np.vstack((np.tile([30., 30.], (n1, 1)), 
                           np.tile([70., 70.], (n2, 1)),
                           np.tile([30. ,70.], (n1, 1)),
                           np.tile([70., 30.], (n2, 1))))
            X += rg.normal(size=X.shape, scale=10*noise)  # add noise
    
        else:
            print('No data returned - example_nr must be 1, 2, or 3')
        
        o = rg.permutation(range(2*n))
        
        return X[o].T, T[o].T, x_grid.T, dim
    
    
    #%% Test of the data generation
    if __name__ == "__main__":
        #%%
    
        import matplotlib.pyplot as plt
        
        n = 1000
        noise = 1
        
        fig, ax = plt.subplots(1, 3)
        for i, a in enumerate(ax):
            example_nr = i + 1
            X, T, x_grid, dim = make_data(example_nr, n, noise)
            a.scatter(X[0][T[0]], X[1][T[0]], c='r', alpha=0.3, s=15)
            a.scatter(X[0][T[1]], X[1][T[1]], c='g', alpha=0.3, s=15)
            a.set_aspect('equal', 'box')
            a.set_title(f'Example {i} data')
        
        plt.show()
        
        
        #%% Before training, you should make data zero mean
        
        c = np.mean(X, axis=1, keepdims=True)
        X_c = X - c
        
        fig, ax = plt.subplots(1,1)
        ax.scatter(X_c[0][T[0]], X_c[1][T[0]], c='r', alpha=0.3, s=15)
        ax.scatter(X_c[0][T[1]], X_c[1][T[1]], c='g', alpha=0.3, s=15)
        ax.set_aspect('equal', 'box')
        plt.title('Zero-mean data')
        plt.show()
    
    
    
    
    
    
    
    
    # %%