3.7. Classification Test #2ΒΆ

In [1]:

import sdm as sdmlib
import matplotlib.pyplot as plt
from PIL import Image, ImageDraw, ImageFont
import urllib, cStringIO
import random
from IPython.core.display import display, Image as IPythonImage
%matplotlib inline

In [2]:

width = 30
height = 30
noise_flip = True

In [3]:

def gen_img(letter='A'):
    img = Image.new('RGBA', (30, 30), (255, 255, 255))
    font = ImageFont.truetype('Arial.ttf', 30)
    draw = ImageDraw.Draw(img)
    draw.text((5, 0), letter, (0, 0, 0), font=font)
    return img

In [4]:

def gen_noise_add(img, p=0.15, flip=False):
    img2 = img.copy()
    draw = ImageDraw.Draw(img2)
    for py in xrange(height):
        for px in xrange(width):
            if random.random() < p:
                if flip:
                    pixel = img.getpixel((px, py))
                    value = sum([int(x/255+0.5) for x in pixel[:3]])//3
                    assert value == 0 or value == 1
                    value = (1 - value)*255
                    draw.point((px, py), fill=(value, value, value))
                else:
                    draw.point((px, py), fill=(0, 0, 0))
    return img2

In [5]:

img = gen_img();
img2 = gen_noise_add(img, p=0.05, flip=noise_flip)
plt.subplot(1, 2, 1)
plt.imshow(img)
plt.subplot(1, 2, 2)
plt.imshow(img2);
../_images/notebooks_Classification_Test_2_5_0.png 
In [6]:

def to_bitstring(img):
    v = []
    bs = sdmlib.Bitstring.init_ones(1000)
    for py in xrange(height):
        for px in xrange(width):
            pixel = img.getpixel((px, py))
            value = sum([int(x/255+0.5) for x in pixel[:3]])//3
            assert value == 0 or value == 1
            idx = px+width*py
            assert idx >= 0 and idx < 1000, 'Ops {} {} {}'.format(x, y, idx)
            bs.set_bit(idx, value)
            v.append(value)
    v2 = [bs.get_bit(i) for i in xrange(height*width)]
    assert v == v2
    return bs

In [7]:

def to_img(bs):
    img = Image.new('RGBA', (30, 30), (255, 255, 255))
    draw = ImageDraw.Draw(img)
    for py in xrange(height):
        for px in xrange(width):
            idx = px+width*py
            assert idx >= 0 and idx < 1000, 'Ops {} {} {}'.format(x, y, idx)
            x = 255*bs.get_bit(idx)
            draw.point((px, py), fill=(x, x, x))
    return img

In [8]:

bits = 1000
sample = 1000000
scanner_type = sdmlib.SDM_SCANNER_OPENCL

In [9]:

address_space = sdmlib.AddressSpace.init_from_b64_file('sdm-letters.as')
counter = sdmlib.Counter.create_file('sdm-classification-2', bits, sample)
sdm = sdmlib.SDM(address_space, counter, 451, scanner_type)

In [10]:

def fill_memory(letter, label_bs, p=0.1, n=100):
    cols = 15
    rows = n//cols + 1
    plt.figure(figsize=(20,10))
    for i in xrange(n):
        img = gen_img(letter=letter);
        img2 = gen_noise_add(img, p=p, flip=noise_flip)
        #display(img2)
        plt.subplot(rows, cols, i+1)
        plt.imshow(img2)
        bs = to_bitstring(img2)
        sdm.write(bs, label_bs)
    plt.show()

In [11]:

def read(letter, n=6, p=0.2, radius=None):
    n = 7
    cols = 15
    rows = n//cols + 1
    plt.figure(figsize=(20,10))

    img = gen_img(letter=letter);
    img2 = gen_noise_add(img, p=p, flip=noise_flip)
    plt.subplot(rows, cols, 1)
    plt.imshow(img2)

    bs2 = to_bitstring(img2)
    bs3 = sdm.read(bs2, radius=radius)

    label = min(label_to_bs.items(), key=lambda v: bs3.distance_to(v[1]))
    return label[0]

In [12]:

labels = list('ABCD8OQ0')
label_to_bs = {}
for x in labels:
    label_to_bs[x] = sdmlib.Bitstring.init_random(1000)

In [13]:

for x in labels:
    print 'Training for label {}...'.format(x)
    fill_memory(x, label_to_bs[x])

Training for label A...
../_images/notebooks_Classification_Test_2_13_1.png 
Training for label B...
../_images/notebooks_Classification_Test_2_13_3.png 
Training for label C...
../_images/notebooks_Classification_Test_2_13_5.png 
Training for label D...
../_images/notebooks_Classification_Test_2_13_7.png 
Training for label 8...
../_images/notebooks_Classification_Test_2_13_9.png 
Training for label O...
../_images/notebooks_Classification_Test_2_13_11.png 
Training for label Q...
../_images/notebooks_Classification_Test_2_13_13.png 
Training for label 0...
../_images/notebooks_Classification_Test_2_13_15.png 
In [14]:

read('C')

Out[14]:

'C'
../_images/notebooks_Classification_Test_2_14_1.png 
In [15]:

read('A')

Out[15]:

'A'
../_images/notebooks_Classification_Test_2_15_1.png 
In [16]:

read('D')

Out[16]:

'D'
../_images/notebooks_Classification_Test_2_16_1.png 
In [17]:

read('8')

Out[17]:

'8'
../_images/notebooks_Classification_Test_2_17_1.png 
In [18]:

read('B')

Out[18]:

'B'
../_images/notebooks_Classification_Test_2_18_1.png 
In [19]:

read('O')

Out[19]:

'O'
../_images/notebooks_Classification_Test_2_19_1.png 
In [35]:

read('Q')

Out[35]:

'Q'
../_images/notebooks_Classification_Test_2_20_1.png 
In [25]:

read('0')

Out[25]:

'0'
../_images/notebooks_Classification_Test_2_21_1.png 
In [45]:

read('0', p=0.3)

Out[45]:

'0'
../_images/notebooks_Classification_Test_2_22_1.png