

In [1]:

    
%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt

Softmax



In [5]:

    
def softmax(scores):
    exps = np.exp(scores)
    return exps / exps.sum()



In [53]:

    
a = np.array([2., -1, 1])
plt.stem(a)
plt.xlim([-1, len(a)]);



In [54]:

    
a2 = softmax(a)
plt.stem(a2)
plt.xlim([-1, len(a)])
print("Softmax probabilities sum to", np.sum(a2))









    



Softmax probabilities sum to 1.0

Dense (similarities)



In [24]:

    
prototypes_raw = np.array([
    [3, 1, 1],
    [1, 3, 1],
    [1, 1, 3]], dtype=float)
prototype_magnitudes = np.linalg.norm(prototypes_raw, axis=1, keepdims=True)
prototypes = prototypes_raw / prototype_magnitudes
prototypes









    Out[24]:





array([[ 0.90453403,  0.30151134,  0.30151134],
       [ 0.30151134,  0.90453403,  0.30151134],
       [ 0.30151134,  0.30151134,  0.90453403]])



In [25]:

    
def similarities(vec, prototypes):
    return np.dot(prototypes, vec)
similarities([1, 2, 3], prototypes)
similarities([3, 2, 1], prototypes)









    Out[25]:





array([ 3.61813613,  3.01511345,  2.41209076])



In [26]:

    
softmax(similarities([1, 2, 3], prototypes))









    Out[26]:





array([ 0.16213016,  0.29631473,  0.54155511])

Nonlinearity



In [27]:

    
def relu(vector):
    return np.maximum(vector, 0)
x = np.linspace(-2, 2, 100)
plt.plot(x, relu(x))









    Out[27]:





[<matplotlib.lines.Line2D at 0x1083080b8>]



In [28]:

    
def tanh(vector):
    return np.tanh(vector)
plt.plot(x, tanh(x));









    Out[28]:





[<matplotlib.lines.Line2D at 0x10825def0>]

Convolution



In [ ]:

    
from scipy.misc import imsave, fromimage, toimage
from PIL import Image, ImageOps



In [31]:

    
DATA = '../data/'
WIDTH = HEIGHT = 224
def load_and_crop_image(filename, target_size):
    return ImageOps.fit(Image.open(filename), target_size)



In [43]:

    
gatos = np.array([fromimage(load_and_crop_image(DATA+'/gatos/cat.{:04}.jpg'.format(i), (WIDTH, HEIGHT))) for i in range(1,100)])
gatos_nocolor = np.mean(gatos, axis=3)



In [44]:

    
img = gatos_nocolor[0]
toimage(img)









    Out[44]:



In [45]:

    
import scipy.signal

Cambia filt hasta que la imagen parece como lo siguiente



In [49]:

    
filt = np.array([
        [1, 0, -1],
        [1, 0, -1],
        [1, 0, -1]
    ])
convolved = scipy.signal.convolve2d(img, filt)
toimage(convolved)









    Out[49]:



In [49]:

    
filt = np.array([
        [1, 0, -1],
        [1, 0, -1],
        [1, 0, -1]
    ])
convolved = scipy.signal.convolve2d(img, filt)
toimage(convolved)









    Out[49]:



In [ ]: