notebook.community

Edit and run



In [1]:

    
!pip freeze









    



Jinja2==2.7.3
MarkupSafe==0.23
Pillow==2.7.0
PyYAML==3.11
Pygments==2.0.2
Theano==0.6.0
argparse==1.3.0
backports.ssl-match-hostname==3.4.0.2
certifi==14.05.14
holoviews==0.7
ipython==3.0.0
jsonschema==2.4.0
mahotas==1.2.4
matplotlib==1.4.3
mistune==0.5
mock==1.0.1
-e git+git@github.com:Neuroglycerin/neukrill-net-tools.git@8769bc55415360058cd2e076c6566f59443b3016#egg=neukrill_net-dev
nose==1.3.4
numpy==1.9.1
param==1.2.1
ptyprocess==0.4
py==1.4.26
-e git+https://github.com/lisa-lab/pylearn2.git@cf3999e7183f8dcaccccf4dfd2a31bbe3a948a97#egg=pylearn2-dev
pyparsing==2.0.3
pytest==2.6.4
python-dateutil==2.4.0
pytz==2014.10
pyzmq==14.5.0
scikit-image==0.10.1
scikit-learn==0.15.2
scipy==0.14.0
six==1.8.0
terminado==0.5
tornado==4.1
wsgiref==0.1.2



In [2]:

    
%pylab









    



Using matplotlib backend: TkAgg
Populating the interactive namespace from numpy and matplotlib



In [3]:

    
%matplotlib inline



In [4]:

    
cd ..









    



/afs/inf.ed.ac.uk/user/s11/s1145806/Documents/git/neukrill-net-work



In [5]:

    
cd ../neukrill-net-work









    



/afs/inf.ed.ac.uk/user/s11/s1145806/Documents/git/neukrill-net-work



In [6]:

    
import sys
import numpy as np
import skimage.io



In [20]:

    
import skimage.transform



In [7]:

    
import imp



In [8]:

    
import neukrill_net.utils as utils
import neukrill_net.image_processing as image_processing
import neukrill_net.augment as augment



In [9]:

    
from IPython.display import display
from IPython.display import Image
from IPython.display import HTML



In [10]:

    
import matplotlib.pyplot as plt
import matplotlib.image as mpimg



In [18]:

    
reload(image_processing)
reload(augment)









    Out[18]:





<module 'neukrill_net.augment' from '/afs/inf.ed.ac.uk/user/s11/s1145806/Documents/git/neukrill-net-tools/neukrill_net/augment.py'>



In [12]:

    
img = skimage.io.imread('data/train/acantharia_protist/100224.jpg')



In [13]:

    
type(img)









    Out[13]:





numpy.ndarray



In [14]:

    
img.shape









    Out[14]:





(66, 59)



In [15]:

    
imgplot = plt.imshow(img)
imgplot.set_cmap('gray')



In [19]:

    
img2 = image_processing.resize_image(img,(48,48))
imgplot = plt.imshow(img2)
imgplot.set_cmap('gray')



In [31]:

    
img3 = np.concatenate((255*np.ones((66,4),dtype='uint8'),img,255*np.ones((66,3),dtype='uint8')),1)
img3 = skimage.transform.resize(img3,(48,48),order=0)
imgplot = plt.imshow(img3)
imgplot.set_cmap('gray')



In [32]:

    
img2 = image_processing.resize_image(img,(96,96))
imgplot = plt.imshow(img2)
imgplot.set_cmap('gray')



In [33]:

    
img2 = image_processing.resize_image(img,(48,48))
imgplot = plt.imshow(img2)
imgplot.set_cmap('gray')



In [35]:

    
augFnc = augment.RandomAugment(rotate=-1,skew=5,flip=1,shunt=0.075,resize=(48,48))



In [36]:

    
for i in range(10):
    imgplot = plt.imshow(augFnc(img2))
    imgplot.set_cmap('gray')
    show()

Let's see if we can fix this



In [15]:

    
imgplot = plt.imshow(img)
imgplot.set_cmap('gray')



In [38]:

    
M = skimage.transform.AffineTransform()



In [42]:

    
M.rotation









    Out[42]:





0.0



In [43]:

    
image = skimage.transform.warp(img, M, cval=1.0)
imgplot = plt.imshow(image)
imgplot.set_cmap('gray')



In [44]:

    
image = skimage.transform.warp(img, M, cval=1.0, order=0)
imgplot = plt.imshow(image)
imgplot.set_cmap('gray')



In [45]:

    
M.rotation = np.pi









    



---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-45-2c3e3f6d74ff> in <module>()
----> 1 M.rotation = np.pi

AttributeError: can't set attribute



In [102]:

    
M = skimage.transform.AffineTransform(rotation=np.pi)



In [51]:

    
imgplot = plt.imshow(img)
imgplot.set_cmap('gray')



In [52]:

    
imgplot = plt.imshow(skimage.img_as_float(img))
imgplot.set_cmap('gray')



In [103]:

    
image = skimage.transform.warp(skimage.img_as_float(img), M, cval=1.0)
imgplot = plt.imshow(image)
imgplot.set_cmap('gray')



In [66]:

    
print skimage.img_as_float(img).dtype
print img.dtype









    



float64
uint8



In [50]:

    
image = skimage.transform.warp(img, M, cval=1.0, order=0)
imgplot = plt.imshow(image)
imgplot.set_cmap('gray')



In [98]:

    
# Create Affine transform with shear
aft = skimage.transform.AffineTransform(shear=5.0/180*np.pi)

# Apply transform to image data
image = skimage.transform.warp(img, aft, cval=1.0)

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')



In [97]:

    
# Create Affine transform with shear
aft = skimage.transform.AffineTransform(shear=np.pi)

# Apply transform to image data
image = skimage.transform.warp(img, aft, cval=1.0, order=0)

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')

print np.allclose(image,img)









    



False



In [105]:

    
# Create Affine transform with shear
aft = skimage.transform.AffineTransform(rotation=0.5*np.pi)

# Apply transform to image data
image = skimage.transform.warp(img, aft, cval=1.0, order=0)

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')

print np.allclose(image,img)









    



False



In [127]:

    
# Create Affine transform with shear
shift_y, shift_x = np.array(image.shape[:2]) / 2.
tf_shift = skimage.transform.SimilarityTransform(translation=[-shift_x, -shift_y])
tf_shift_inv = skimage.transform.SimilarityTransform(translation=[shift_x, shift_y])

aft = skimage.transform.AffineTransform(rotation=0.5*np.pi)

# Apply transform to image data
image = skimage.transform.warp(img, (tf_shift + (aft + tf_shift_inv)), cval=1.0)

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')

print np.allclose(image,img)









    



False



In [51]:

    
imgplot = plt.imshow(img)
imgplot.set_cmap('gray')



In [51]:

    
imgplot = plt.imshow(img)
imgplot.set_cmap('gray')



In [128]:

    
# Create Affine transform with shear
shift_y, shift_x = np.array(image.shape[:2]) / 2.
tf_shift = skimage.transform.SimilarityTransform(translation=[-shift_x, -shift_y])
tf_shift_inv = skimage.transform.SimilarityTransform(translation=[shift_x, shift_y])

aft = skimage.transform.AffineTransform(rotation=0.5*np.pi)

# Apply transform to image data
image = skimage.transform.warp(img, (tf_shift + (aft + tf_shift_inv)), cval=1.0, order=0)

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')

print np.allclose(image,img)









    



False



In [144]:

    
# Create Affine transform with shear
shift_y, shift_x = np.array(image.shape[:2]) / 2.
tf_shift = skimage.transform.SimilarityTransform(translation=[-shift_x, -shift_y])
tf_shift_inv = skimage.transform.SimilarityTransform(translation=[shift_x, shift_y])

aft = skimage.transform.AffineTransform(rotation=0.02*np.pi)

# Apply transform to image data
image = ( 0.5 * skimage.transform.warp(img, (tf_shift + (aft + tf_shift_inv)), cval=1.0, order=1) +
            0.5 * skimage.transform.warp(img, (tf_shift + (aft + tf_shift_inv)), cval=1.0, order=0)
            )

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')

print np.allclose(image,img)









    



False



In [147]:

    
reload(image_processing)









    Out[147]:





<module 'neukrill_net.image_processing' from '/afs/inf.ed.ac.uk/user/s11/s1145806/Documents/git/neukrill-net-tools/neukrill_net/image_processing.py'>



In [148]:

    
# Create Affine transform with shear
image = image_processing.custom_transform_nice_units(img)

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')

print np.allclose(image,img)









    



False



In [151]:

    
# Create Affine transform
image = image_processing.custom_transform_nice_units(img2, rotation=30)

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')



In [240]:

    
reload(image_processing)









    Out[240]:





<module 'neukrill_net.image_processing' from '/afs/inf.ed.ac.uk/user/s11/s1145806/Documents/git/neukrill-net-tools/neukrill_net/image_processing.py'>



In [242]:

    
img2b = image_processing.resize_image(img,(48,48),order=0.75)
imgplot = plt.imshow(img2b)
imgplot.set_cmap('gray')
show()

# Create Affine transform
#image = image_processing.custom_transform_nice_units(img2b, rotation=30, shear=5, order=0.5, translation=(-0.1,0.1))
image = image_processing.custom_transform_nice_units(img2b, scale=1.1)

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')
show()

imgplot = plt.imshow(img)
imgplot.set_cmap('gray')



In [243]:

    
img_ = skimage.io.imread('data/train/hydromedusae_bell_and_tentacles/101932.jpg')



In [245]:

    
img2b = image_processing.resize_image(img_,(48,48),order=0.75)
imgplot = plt.imshow(img2b)
imgplot.set_cmap('gray')
show()

# Create Affine transform
#image = image_processing.custom_transform_nice_units(img2b, rotation=30, shear=5, order=0.5, translation=(-0.1,0.1))
image = image_processing.custom_transform_nice_units(img2b, scale=1.1)

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')
show()

imgplot = plt.imshow(img_)
imgplot.set_cmap('gray')



In [250]:

    
img2b = image_processing.resize_image(img_,(48,48),order=0.75)
imgplot = plt.imshow(img2b)
imgplot.set_cmap('gray')
show()

# Create Affine transform
#image = image_processing.custom_transform_nice_units(img2b, rotation=30, shear=5, order=0.5, translation=(-0.1,0.1))
image = image_processing.custom_transform_nice_units(img2b, scale=1.1)

imgplot = plt.imshow(image)
imgplot.set_cmap('gray')
show()

imgplot = plt.imshow(img_)
imgplot.set_cmap('gray')

Let's try random augmentations again



In [285]:

    
reload(image_processing)
reload(augment)









    Out[285]:





<module 'neukrill_net.augment' from '/afs/inf.ed.ac.uk/user/s11/s1145806/Documents/git/neukrill-net-tools/neukrill_net/augment.py'>



In [255]:

    
imgplot = plt.imshow(img)
imgplot.set_cmap('gray')



In [256]:

    
img2b = image_processing.resize_image(img,(48,48),order=0.75)

imgplot = plt.imshow(img2b)
imgplot.set_cmap('gray')



In [254]:

    
augFnc = augment.RandomAugment(rotate=-1,skew=5,flip=1,shunt=0.075,resize=(48,48))



In [257]:

    
for i in range(10):
    imgplot = plt.imshow(augFnc(img2))
    imgplot.set_cmap('gray')
    show()



In [286]:

    
img2b = image_processing.resize_image(img,(48,48),order=0.75)
augFnc = augment.RandomAugment(flip=1, rotate=-1, scale=0.1, scale_asym=0.04, shear=6, shunt=0.05, noise=0.0001)
for i in range(20):
    imgplot = plt.imshow(augFnc(img2b))
    imgplot.set_cmap('gray')
    show()



In [300]:

    
imgplot = plt.imshow(img)
imgplot.set_cmap('gray')
show()

img_new = image_processing.custom_transform_nice_units(img, rotation=30)
imgplot = plt.imshow(img_new)
imgplot.set_cmap('gray')
show()

dwnimg = image_processing.custom_transform_nice_units(
            image_processing.resize_image(img,(96,96),order=0.75), rotation=30)
imgplot = plt.imshow(dwnimg)
imgplot.set_cmap('gray')
show()

img2b = image_processing.resize_image(img,(192,192),order=0.75)

dwnimg = skimage.transform.downscale_local_mean(
                image_processing.custom_transform_nice_units(img2b, rotation=30), (2,2), cval=255)
imgplot = plt.imshow(dwnimg)
imgplot.set_cmap('gray')
show()

Ok, so supersampling and then interpolating down is no good.

What if we use the magic kernel? http://johncostella.webs.com/magic/



In [301]:

    
import scipy



In [306]:

    
h = 0.5 * np.array([1.0/4, 3.0/4, 3.0/4, 1.0/4])
print h
k = np.outer(h,h)
print k









    



[ 0.125  0.375  0.375  0.125]
[[ 0.015625  0.046875  0.046875  0.015625]
 [ 0.046875  0.140625  0.140625  0.046875]
 [ 0.046875  0.140625  0.140625  0.046875]
 [ 0.015625  0.046875  0.046875  0.015625]]



In [307]:

    
img_dwn = scipy.ndimage.filters.convolve(img, k, mode='constant', cval=255.0)



In [319]:

    
img_dwn[::2,::2].shape









    Out[319]:





(33, 30)



In [320]:

    
imgplot = plt.imshow(img_dwn[::2,::2])
imgplot.set_cmap('gray')
show()



In [324]:

    
imgplot = plt.imshow(img)
imgplot.set_cmap('gray')
show()

# Rotate the original
img_new = image_processing.custom_transform_nice_units(img, rotation=30)
imgplot = plt.imshow(img_new)
imgplot.set_cmap('gray')
show()

# Resize to 96x96, then rotate
img_new = image_processing.custom_transform_nice_units(
            image_processing.resize_image(img,(96,96),order=0.75), rotation=30)
imgplot = plt.imshow(img_new)
imgplot.set_cmap('gray')
show()

# Resize to 192x192, then rotate, then downscale with magic kernal
img_new = image_processing.resize_image(img,(192,192),order=0.75)
img_new = image_processing.custom_transform_nice_units(img_new, rotation=30)
img_new = scipy.ndimage.filters.convolve(img_new, k, mode='constant', cval=1.0)
img_new = img_new[::2,::2]
imgplot = plt.imshow(img_new)
imgplot.set_cmap('gray')
show()



In [327]:

    
imgplot = plt.imshow(img)
imgplot.set_cmap('gray')
show()

# Rotate the original
img_new = image_processing.custom_transform_nice_units(img, rotation=32)
imgplot = plt.imshow(img_new)
imgplot.set_cmap('gray')
show()

# Resize to 96x96, then rotate
img_new = image_processing.custom_transform_nice_units(
            image_processing.resize_image(img,(96,96),order=0.75), rotation=32)
imgplot = plt.imshow(img_new)
imgplot.set_cmap('gray')
show()

# Resize to 192x192, then rotate, then downscale without filtering
img_new = image_processing.resize_image(img,(192,192),order=0.75)
img_new = image_processing.custom_transform_nice_units(img_new, rotation=32)
img_new = img_new[::2,::2]
imgplot = plt.imshow(img_new)
imgplot.set_cmap('gray')
show()