In [1]:

    

%matplotlib inline
from matplotlib import rcParams
rcParams["savefig.dpi"] = 150


    

In [2]:

    

import numpy as np
from itertools import product
import matplotlib.pyplot as pl
from sklearn.decomposition import FastICA


    

In [3]:

    

data = np.load("data/k2/dun/ktwo200000862_c0_p(0,0)_data.npy")[:, 20:30, 35:45]
pred = np.load("data/k2/dun/ktwo200000862_c0_p(0,0)_pre.npy")[:, 20:30, 35:45]


    

In [4]:

    

m = ~np.all(data == 0, axis=(1, 2))
data = data[m]
pred = pred[m]


    

In [5]:

    

diff = (data - pred) / pred


    

In [6]:

    

pl.imshow(diff[0], cmap="gray_r", interpolation="nearest");


    

In [7]:

    

block = diff.reshape((len(diff), -1))


    

In [8]:

    

model = FastICA(n_components=2)
model.fit(block.T)


    

    
Out[8]:





FastICA(algorithm='parallel', fun='logcosh', fun_args=None, max_iter=200,
    n_components=2, random_state=None, tol=0.0001, w_init=None,
    whiten=True)

In [9]:

    

n = len(model.components_)
print(n)


    

    




2

In [10]:

    

pl.figure(figsize=(5, 10))
pl.plot(model.components_.T + 1*np.arange(n));


    

In [14]:

    

v = model.components_[1]
comp = np.abs((np.dot(v, block)).reshape(diff[0].shape))
pl.imshow(comp, cmap="gray_r", interpolation="nearest")


    

    
Out[14]:





<matplotlib.image.AxesImage at 0x10ebe71d0>






    

In [15]:

    

c = np.dot(v, block)
c /= np.sum(c)
pl.plot(np.sum(c * block, axis=1))


    

    
Out[15]:





[<matplotlib.lines.Line2D at 0x10ef7cf98>]






    

In [16]:

    

pl.figure()
pl.imshow(np.log(data[0]), cmap="gray_r", interpolation="nearest")
pl.figure()
pl.imshow(np.log(data[0]), cmap="gray_r", interpolation="nearest")
pl.imshow(comp, cmap="Blues", interpolation="nearest", alpha=0.5);


    

In [ ]:

    




    

In [ ]: