def graph(self, x, y):
plt.gray()
for i in xrange(x * y):
plt.subplot(y, x, i + 1)
plt.imshow(self._data[i][0].reshape(IMG_WIDTH, IMG_HEIGHT))
plt.plot()
def draw_digit(data):
size = int(len(data) ** 0.5)
Z = data.reshape(size, size) # convert from vector to matrix
plt.imshow(Z, interpolation="None")
plt.gray()
plt.tick_params(labelbottom="off")
plt.tick_params(labelleft="off")
def im_plot(x, y, images):
# First we build an array of the indicaes contianing the first ten images for each digit.
vals = np.where(y == 0)[0][:10]
for i in range(1, 10):
vals = np.vstack((vals, np.where(y == i)[0][:10]))
nrows, ncols = vals.shape
plt.figure(figsize=(8.5,8))
plt.gray()
# Build a list of the indices in the order I want.
# plot them, one by one.
for idx, i in enumerate(vals.T.ravel()):
ax = plt.subplot(nrows, ncols, idx + 1)
# We want square images
ax.set_aspect('equal')
# Now show the images, by default pixels are shown as white on black.
# To show black on white, reverse colormap: cmap=plt.cm.gray_r
# To smooth pixelated images: interpolation='nearest'
ax.imshow(images[i])
# No tick marks for small plots
plt.xticks([]) ; plt.yticks([])
# Only put plot lables over columns
if idx < 10:
plt.title(y[i])
def graph_spectrogram(wav_file, wav_folder):
name_save = wav_file.replace(".wav", ".png")
name_save_cv2 = wav_file.replace(".wav", "_cv2.png")
rate, data = get_wav_info(wav_file)
nfft = 256 # Length of the windowing segments
fs = 256 # Sampling frequency
plt.clf()
pxx, freqs, bins, im = plt.specgram(data, nfft, fs)
plt.axis('off')
plt.gray()
plt.savefig(name_save,
dpi=50, # Dots per inch
frameon='false',
aspect='normal',
bbox_inches='tight',
pad_inches=0)
# Expore plote as image
fig = plt.gcf()
fig.canvas.draw()
# Get the RGBA buffer from the figure
w, h = fig.canvas.get_width_height()
buf = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8)
buf.shape = (w, h, 3)
# canvas.tostring_argb give pixmap in ARGB mode. Roll the ALPHA channel to have it in RGBA mode
# buf = np.roll(buf, 2)
cv2.imwrite(name_save_cv2, buf)
def plot_images(data_list, data_shape="auto", fig_shape="auto"):
"""
plotting data on current plt object.
In default,data_shape and fig_shape are auto.
It means considered the data as a sqare structure.
"""
n_data = len(data_list)
if data_shape == "auto":
sqr = int(n_data ** 0.5)
if sqr * sqr != n_data:
data_shape = (sqr + 1, sqr + 1)
else:
data_shape = (sqr, sqr)
plt.figure(figsize=data_shape)
for i, data in enumerate(data_list):
plt.subplot(data_shape[0], data_shape[1], i + 1)
plt.gray()
if fig_shape == "auto":
fig_size = int(len(data) ** 0.5)
if fig_size ** 2 != len(data):
fig_shape = (fig_size + 1, fig_size + 1)
else:
fig_shape = (fig_size, fig_size)
Z = data.reshape(fig_shape[0], fig_shape[1])
plt.imshow(Z, interpolation="nearest")
plt.tick_params(labelleft="off", labelbottom="off")
plt.tick_params(axis="both", which="both", left="off", bottom="off", right="off", top="off")
plt.subplots_adjust(hspace=0.05)
plt.subplots_adjust(wspace=0.05)
def main():
print '-----Question4-----'
#define constant.
N_max = 50
some_threshold = 50
#construct a grid of x,y in range [-2, 1], [-1.5, 1.5].
num_x = 500
num_y = 500
x = np.linspace(-2,1,num_x)
y = np.linspace(-1.5,1.5,num_y)
c = x[:,newaxis]+1j*y[newaxis,:]
#do the iteration
z = c
for v in range(N_max):
z = z**2 + c
#define the boolean mask.
mask = abs(z) < some_threshold
#draw the image and save the result
plt.imshow(mask.T, extent=[-2, 1, -1.5, 1.5])
plt.gray()
plt.savefig('mandelbrot.png')
print '----------------'
def just_do_it(limit_cont):
fig = plt.figure(facecolor='black')
plt.gray()
print("Rozpoczynam przetwarzanie obrazow...")
for i in range(20):
img = data.imread(images[i])
gray_img = to_gray(images[i]) # samoloty1.pdf
#gray_img = to_gray2(images[i], 1001, 0.2, 5, 9, 12) # samoloty2.pdf
#gray_img = to_gray2(images[i], 641, 0.2, 5, 20, 5) # samoloty3.pdf
conts = find_contours(gray_img, limit_cont)
centrs = [find_centroid(cont) for cont in conts]
ax = fig.add_subplot(4,5,i)
ax.set_yticks([])
ax.set_xticks([])
io.imshow(img)
print("Przetworzono: " + images[i])
for n, cont in enumerate(conts):
ax.plot(cont[:, 1], cont[:, 0], linewidth=2)
for centr in centrs:
ax.add_artist(plt.Circle(centr, 5, color='white'))
fig.tight_layout()
#plt.show()
plt.savefig('samoloty3.pdf')
def Question4():
print 'question 4'
#set the grid of x and y
mandelbrot_set=compute_mandeohrot(50,50, 501,501 )
plt.imshow(mandelbrot_set.T,extent=[-2,1,-1.5,1.5])
plt.gray()
plt.savefig('mandelbrot.png')
def mostra_imagens(imagens, tam_patch=64):
"""Display a list of images"""
n_imgs = len(imagens)
fig = plt.figure()
n = 1
for img in imagens:
imagem = img[0]
titulo = img[1]
#####################################
v, h, _ = imagem.shape
# calcula as bordas horizontais
h_m1 = h % tam_patch
h_m2 = h_m1//2
h_m1 -= h_m2
# calcula das bordas verticais
v_m1 = v % tam_patch
v_m2 = v_m1//2
v_m1 -= v_m2
#####################################
a = fig.add_subplot(1,n_imgs,n)
a.set_xticks(np.arange(0+h_m1, 700-h_m2, tam_patch))
a.set_yticks(np.arange(0+v_m1, 460-v_m2, tam_patch))
a.grid(True)
if imagem.ndim == 2:
plt.gray()
plt.imshow(imagem)
a.set_title(titulo)
n += 1
fig.set_size_inches(np.array(fig.get_size_inches()) * n_imgs)
plt.show()
def run_conv_net_image_filtering():
img = get_3_wolves_image()
img = img.astype(theano.config.floatX) / 256.
prepared_image = img.transpose(2, 0, 1).reshape(1, 3, img.shape[0], img.shape[1])
# now image shape is 1x3x639x516 - (batchsize x number_of_feature_maps x height x width)
input = T.tensor4('input')
rng = np.random.RandomState(23455)
filter_symbolic, _ = conv_layer(
input, feature_maps_count_in=3, feature_maps_count_out=2, filter_shape=(9, 9), rng=rng)
filter = theano.function([input], filter_symbolic)
filtered_image = filter(prepared_image)
output0 = filtered_image[0, 0, :, :]
output1 = filtered_image[0, 1, :, :]
plots.subplot(1, 3, 1)
plots.axis('off')
plots.imshow(img)
plots.gray()
plots.subplot(1, 3, 2)
plots.axis('off')
plots.imshow(output0)
plots.subplot(1, 3, 3)
plots.axis('off')
plots.imshow(output1)
plots.show()
def plot_images(images, labels):
fig = plt.figure()
plt.gray()
for i in range(min(9, images.shape[0])):
fig.add_subplot(3, 3, i+1)
show_image(images[i], labels[i])
plt.show()
def q4():
'''question #4 module'''
#grids
pixel=1000.0
x=np.arange(-2,1,3.0/pixel)
y=np.arange(-1.5,1.5,3.0/pixel)
xs,ys=np.meshgrid(x,y)
#intialize
mask=np.empty((pixel,pixel))
mask[:,:]=1
max_N=50
bound=50
c=xs+(1j*ys)
z=c
for t in np.arange(max_N):
z= z**2 + c
mask=mask*(z<bound)
import matplotlib.pyplot as plt
plt.imshow(mask, extent=[-2, 1, -1.5, 1.5])
plt.gray()
plt.savefig('mandelbrot.png')
print ""
print "Question #4"
print "mandelbrot.png has been saved to local directory"
print ""
def test():
saver.restore(sess, FLAGS.save_dir+'/model.ckpt')
batch_x, _ = mnist.test.next_batch(10)
batch_x_att, batch_p = sess.run([x_att, p], {x:batch_x})
A = np.zeros((0, N*N))
for i in range(10):
for k in range(K):
A = np.concatenate([A, batch_x_att[k][i].reshape((1, N*N))], 0)
fig = plt.figure('attended')
plt.gray()
plt.axis('off')
plt.imshow(batchmat_to_tileimg(A, (N, N), (10, K)))
fig.savefig(FLAGS.save_dir+'/attended.png')
"""
P = np.zeros((0, n_in))
for i in range(10):
P = np.concatenate([P, batch_x[i].reshape((1, n_in))], 0)
for k in range(K):
P = np.concatenate([P, batch_pk[k][i].reshape((1, n_in))], 0)
P = np.concatenate([P, batch_p[i].reshape((1, n_in))])
fig = plt.figure('reconstructed')
plt.gray()
plt.axis('off')
plt.imshow(batchmat_to_tileimg(P, (height, width), (10, K+2)))
fig.savefig(FLAGS.save_dir+'/reconstructed.png')
"""
plt.show()
def visualize_world(self, brain):
"""
Show what's going on in the world.
"""
if self.print_features:
projections, activities = brain.cortex.get_index_projections(to_screen=True)
wtools.print_pixel_array_features(
projections,
self.fov_span ** 2 * 2,
0,#self.num_actions,
self.fov_span, self.fov_span,
world_name=self.name)
# Periodically show the state history and inputs as perceived by BECCA
print ''.join(["world is ", str(self.timestep), " timesteps old"])
fig = plt.figure(11)
plt.clf()
plt.plot( self.column_history, 'k.')
plt.title(''.join(['Column history for ', self.name]))
plt.xlabel('time step')
plt.ylabel('position (pixels)')
fig.show()
fig.canvas.draw()
fig = plt.figure(12)
sensed_image = np.reshape(
0.5 * (self.sensors[:len(self.sensors)/2] -
self.sensors[len(self.sensors)/2:] + 1),
(self.fov_span, self.fov_span))
plt.gray()
plt.imshow(sensed_image, interpolation='nearest')
plt.title("Image sensed")
fig.show()
fig.canvas.draw()
def run_denoising():
noisy = prediction_image
fig, ax = plt.subplots(nrows=2, ncols=3, figsize=(8, 5), sharex=True, sharey=True, subplot_kw={'adjustable':'box-forced'})
plt.gray()
ax[0, 0].imshow(noisy)
ax[0, 0].axis('off')
ax[0, 0].set_title('noisy')
ax[0, 1].imshow(denoise_tv_chambolle(noisy, weight=0.1, multichannel=True))
ax[0, 1].axis('off')
ax[0, 1].set_title('TV')
ax[0, 2].imshow(denoise_bilateral(noisy, sigma_range=0.05, sigma_spatial=15))
ax[0, 2].axis('off')
ax[0, 2].set_title('Bilateral')
ax[1, 0].imshow(denoise_tv_chambolle(noisy, weight=0.2, multichannel=True))
ax[1, 0].axis('off')
ax[1, 0].set_title('(more) TV')
ax[1, 1].imshow(denoise_bilateral(noisy, sigma_range=0.1, sigma_spatial=15))
ax[1, 1].axis('off')
ax[1, 1].set_title('(more) Bilateral')
ax[1, 2].imshow(noisy)
ax[1, 2].axis('off')
ax[1, 2].set_title('original')
fig.subplots_adjust(wspace=0.02, hspace=0.2,
top=0.9, bottom=0.05, left=0, right=1)
plt.show()
def display_depth(dev, data, timestamp):
global image_depth
global filenum
filenum = filenum+1
#print data.shape
'''
print type(data)
print 100.0/(-0.00307 * data[240,320] + 3.33)
print 0.1236 * math.tan(data[240,320] / 2842.5 + 1.1863)
print data[240,320]
print '-------'
'''
depth = 100/(-0.00307*data + 3.33)
#savedepth(depth)
t = time.time()
fname = '%s/D%s.nparray'%(directory,t)
np.save(fname, data)
print "Writing %s"%fname
data = frame_convert.pretty_depth(data)
mp.gray()
mp.figure(1)
if image_depth:
image_depth.set_data(data)
else:
image_depth = mp.imshow(data, interpolation='nearest', animated=True)
mp.draw()
def view(stack, image): # function to view an image
''' View a single image from a chosen stack of images
Keyword arguments:
stack -- (str) this should be either 'train', 'test', or 'recon'
number -- (int) this is the index number of the image the stack,
12,000 images in the training set, 1233 in test set
'''
plt.gray()
if stack == 'train':
plt.imshow(np.hstack((
left[image].reshape(64,32),
right[image].reshape(64,32)
))
)
elif stack == 'test':
plt.imshow(test[image].reshape(64,32))
elif stack == 'recon':
plt.imshow(np.hstack((
test[image].reshape(64,32),
reconstructed[image].reshape(64,32)
))
)
def visualizefacenet(fname, imgs, patches_left, patches_right,
true_label, predicted_label):
"""Builds a plot of facenet with attention per RNN step and
classification result
"""
nsamples = imgs.shape[0]
nsteps = patches_left.shape[1]
is_correct = true_label == predicted_label
w = nsteps + 2 + (nsteps % 2)
h = nsamples * 2
plt.clf()
plt.gray()
for i in range(nsamples):
plt.subplot(nsamples, w//2, i*w//2 + 1)
plt.imshow(imgs[i])
msg = ('Prediction: ' + predicted_label[i] + ' TrueLabel: ' +
true_label[i])
if is_correct[i]:
plt.title(msg,color='green')
else:
plt.title(msg,color='red')
plt.axis('off')
for j in range(nsteps):
plt.subplot(h, w, i*2*w + 2 + 1 + j)
plt.imshow(patches_left[i, j])
plt.axis('off')
plt.subplot(h, w, i*2*w + 2 + 1 + j + w)
plt.imshow(patches_right[i, j])
plt.axis('off')
plt.show()
plt.savefig(fname)
def run_script(args):
matplotlib.interactive(False)
Rprop0 = args.Rpp0
Rprop1 = args.Rpp1
theta = np.arange(args.theta[0], args.theta[1], args.theta[2])
warray_amp = create_theta_spike(args.pad,
Rprop0, Rprop1, theta,
args.f, args.points, args.reflectivity_method)
fig = plt.figure()
ax1 = fig.add_subplot(111)
plt.gray()
aspect = float(warray_amp.shape[1]) / warray_amp.shape[0]
ax1.imshow(warray_amp, aspect=aspect, cmap=args.colour)
plt.title(args.title % locals())
plt.ylabel('time (ms)')
plt.xlabel('trace')
fig_path = tempfile.mktemp('.jpeg')
plt.savefig(fig_path)
with open(fig_path, 'rb') as fd:
data = fd.read()
unlink(fig_path)
return data
def compute_mandelbrot():
N_max = 50
some_threshold = 50
grid_interval = 1000
# construct the grid
x = np.linspace(-2, 1, grid_interval)
y = np.linspace(-1.5, 1.5, grid_interval)
c = x[:, np.newaxis] + 1j*y[np.newaxis, :]
# do the iteration
z = c
for v in range(N_max):
with np.errstate(all = "ignore"): # catches overflow and invalid value errors
z = z**2 + c
with np.errstate(all = "ignore"): # catches overflow and invalid value errors
# form a 2-D boolean mask
mask = (abs(z) < some_threshold)
# save the result to an image
plt.imshow(mask.T, extent = [-2, 1, -1.5, 1.5])
plt.gray()
plt.savefig('mandelbrot.png')
def test_warp_from_file():
"""File to ndarray"""
with rasterio.open('rasterio/tests/data/RGB.byte.tif') as src:
dst_transform = [-8789636.708, 300.0, 0.0, 2943560.235, 0.0, -300.0]
dst_crs = dict(
proj='merc',
a=6378137,
b=6378137,
lat_ts=0.0,
lon_0=0.0,
x_0=0.0,
y_0=0,
k=1.0,
units='m',
nadgrids='@null',
wktext=True,
no_defs=True)
destin = numpy.empty(src.shape, dtype=numpy.uint8)
reproject(
rasterio.band(src, 1),
destin,
dst_transform=dst_transform,
dst_crs=dst_crs)
assert destin.any()
try:
import matplotlib.pyplot as plt
plt.imshow(destin)
plt.gray()
plt.savefig('test_warp_from_filereproject.png')
except:
pass
def save_image(img, fn): if len(img) == 96*96: plt.imshow(to_matrix(img)) else: plt.imshow(img) plt.gray() plt.savefig(fn)
def test():
saver.restore(sess, FLAGS.save_dir+'/model.ckpt')
batch_x, batch_y = mnist.test.next_batch(100)
"""
fig = plt.figure('original')
plt.gray()
plt.axis('off')
plt.imshow(batchmat_to_tileimg(batch_x, (height, width), (10, 10)))
fig.savefig(FLAGS.save_dir+'/original.png')
fig = plt.figure('reconstructed')
plt.gray()
plt.axis('off')
p_recon = sess.run(p_x, {x:batch_x, y:batch_y})
plt.imshow(batchmat_to_tileimg(p_recon, (height, width), (10, 10)))
fig.savefig(FLAGS.save_dir+'/reconstructed.png')
"""
batch_z = np.random.normal(size=(100, 50))
batch_y = np.zeros((100, 10))
for i in range(10):
batch_y[i*10:(i+1)*10, i] = 1.0
fig = plt.figure('generated')
plt.gray()
plt.axis('off')
p_gen = sess.run(p_x, {z:batch_z, y:batch_y, is_train:False})
plt.imshow(batchmat_to_tileimg(p_gen, (height, width), (10, 10)))
fig.savefig(FLAGS.save_dir+'/generated.png')
plt.show()
def plot(image, invert = False, cmap=plt.cm.binary):
if invert:
image = np.ones(len(image)) - image
plt.gray()
plt.imshow(image, cmap=cmap)
def Plot_harris_points(img, filtered_coords):
plt.figure()
plt.gray()
plt.imshow(img)
plt.plot([p[1] for p in filtered_coords], [p[0] for p in filtered_coords], '.')
plt.axis('off')
plt.show()
def Mandelbrot(rows= 1000, cols = 1000):
print " **** Question 4 ****"
print "Please wait patiently while the calculation is underway."
#initialize grid and params
threshold = 50
N_max = 50
x = np.linspace(-2,1,cols)
y = np.linspace(-1.5, 1.5,rows)
mask = np.ones((rows,cols))
x_elements, y_elements = np.meshgrid(x,y)
C = x_elements + 1j*y_elements
#initialize
Z = C
# ignore runtime\overflow error. we can do this since we do not care about about the actual entries of z besides diverging\not diverging.
np.seterr(all='ignore')
# loop and change mask.
for v in range(N_max):
Z = Z**2 + C
mask = mask*(np.abs(Z)<threshold)
plt.imshow(mask, extent=[-2, 1, -1.5, 1.5])
plt.gray()
plt.savefig('mandelbrot.png')
def pltImAndCoords(h5file, frame, coordFiles, printerFriendly=True):
'''read image coordinates, plot them together with the raw data
save output as .png or .tiff or display with matplotlib'''
# read input data
rawData = h5.readHDF5Frame(h5file, frame)
coordDat = [(coord2im.readCoordsFile(f)) for f in coordFiles]
if printerFriendly:
rawData = -rawData
plt.imshow(rawData, interpolation='nearest')
plt.colorbar()
plt.gray()
axx = plt.axis()
markers = ['r+', 'bx', 'go', 'k,', 'co', 'yo']
for i in range(len(coordFiles)):
c, w, h = coordDat[i]
cc = np.array(c)
idxs = (cc[:,2] == frame)
cc = cc[idxs]
ll = coordFiles[i][coordFiles[i].rfind('/')+1:]
plt.plot(cc[:,0], cc[:,1], markers[i], label=ll)
plt.legend()
plt.axis(axx) # dont change axis by plotting coordinates
plt.show()
def display_image(img): if len(img) == 96*96: plt.imshow(to_matrix(img)) else: plt.imshow(img) plt.gray() plt.show()
def denoising(astro):
noisy = astro + 0.6 * astro.std() * np.random.random(astro.shape)
noisy = np.clip(noisy, 0, 1)
fig, ax = plt.subplots(nrows=2, ncols=3, figsize=(8, 5), sharex=True,
sharey=True, subplot_kw={'adjustable': 'box-forced'})
plt.gray()
ax[0, 0].imshow(noisy)
ax[0, 0].axis('off')
ax[0, 0].set_title('noisy')
ax[0, 1].imshow(denoise_tv_chambolle(noisy, weight=0.1, multichannel=True))
ax[0, 1].axis('off')
ax[0, 1].set_title('TV')
ax[0, 2].imshow(denoise_bilateral(noisy, sigma_range=0.05, sigma_spatial=15))
ax[0, 2].axis('off')
ax[0, 2].set_title('Bilateral')
ax[1, 0].imshow(denoise_tv_chambolle(noisy, weight=0.2, multichannel=True))
ax[1, 0].axis('off')
ax[1, 0].set_title('(more) TV')
ax[1, 1].imshow(denoise_bilateral(noisy, sigma_range=0.1, sigma_spatial=15))
ax[1, 1].axis('off')
ax[1, 1].set_title('(more) Bilateral')
ax[1, 2].imshow(astro)
ax[1, 2].axis('off')
ax[1, 2].set_title('original')
fig.tight_layout()
plt.show()
def plot_raw(data, spectrum=None):
"""This function plots image data and spectrum. If the spectrum is not
given, it is computed from the image data."""
if spectrum==None:
spectrum = spectrum_from_raw(data)
plot.figure()
plot.title("Raw data with summation lines")
plot.imshow(data)
plot.gray()
X = np.linspace(0,data.shape[1],1000)
b = CURVE_B
A = np.linspace(CURVE_A_MIN, CURVE_A_MAX, data.shape[0])
for y in xrange(0, data.shape[0], 50):
a = A[y]
Y = a*(X-b)**2 + y
plot.plot(X,Y.round(),color="white")
plot.xlim(0,data.shape[1])
plot.ylim(0,data.shape[0])
plot.axvline(SLIT_DIVISION)
plot.figure()
plot.title("Spectrum normalized to 550nm")
plot.xlabel("Wavelength (nm)")
plot.ylabel("Normalized intensity")
X = np.arange(len(spectrum))
X = calibration(X)
plot.plot(X,spectrum[:,0], "-", linewidth=2, color="black", label="Wide slit")
plot.plot(X,spectrum[:,1], ":", linewidth=2, color="black", label="Narrow slit")
plot.xlim(X.min(), X.max())
plot.legend()
# plot.axvline(435.8)
# plot.axvline(611.6)
# plot.axvline(487)
plot.show()
from sklearn.ensemble import AdaBoostClassifier #AdaBoost from xgboost import XGBClassifier #XGBoost from sklearn.discriminant_analysis import LinearDiscriminantAnalysis import matplotlib.pyplot as plt # 加载数据 digits = load_digits() data = digits.data # 数据探索 print(data.shape) # 查看第一幅图像 print(digits.images[0]) # 第一幅图像代表的数字含义 print(digits.target[0]) # 将第一幅图像显示出来 """ plt.gray() plt.imshow(digits.images[0]) plt.show() """ # 分割数据，将25%的数据作为测试集，其余作为训练集 train_x, test_x, train_y, test_y = train_test_split(data, digits.target, test_size=0.25, random_state=33) # 采用Z-Score规范化 ss = preprocessing.StandardScaler() train_ss_x = ss.fit_transform(train_x) test_ss_x = ss.transform(test_x) # 创建LR分类器 lr = LogisticRegression(solver='liblinear', multi_class='auto') #数据集比较小，使用liblinear，数据集大使用 sag或者saga
validation_data=(x_test, x_test),
callbacks=[TensorBoard(log_dir='/tmp/autoencoder')])
#decode some inputs
#note take from test set
encoder = keras.Model(input_img, encoded)
encoded_imgs = encoder.predict(x_test)
#preview of encoded images
n = 10
plt.figure(figsize=(20, 8))
for i in range(1, n + 1):
ax = plt.subplot(1, n, i)
plt.imshow(encoded_imgs[i].reshape((4, 4 * 8)).T)
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.show()
decoded_imgs = autoencoder.predict(x_test)
# train_loss = tf.keras.losses.mae(decoded_imgs,x_test_temp)
# plt.hist(train_loss, bins=50)
# plt.xlabel("Train loss")
# plt.ylabel("No of examples")
total_value_x_test = np.sum(x_test[0])
total_value = np.sum(decoded_imgs[0])
print("Total value test", total_value_x_test)
print("Total value", total_value)
def gaussian(im):
b = array([[2, 4, 5, 4, 2], [4, 9, 12, 9, 4], [5, 12, 15, 12, 5],
[4, 9, 12, 9, 4], [2, 4, 5, 4, 2]]) / 156
k = zeros(im.shape)
k[:b.shape[0], :b.shape[1]] = b
fim = fft2(im)
fk = fft2(k)
fil_im = ifft2(fim * fk)
return abs(fil_im).astype(
int) #returning numpy array with absolute interger values
if __name__ == "__main__":
from sys import argv
if len(argv) < 2: #counting the length of arguments
print "Usage: python %s <image>" % argv[0]
exit()
im = array(Image.open(argv[1])) #taking input
im = im[:, :, 0]
gray() #changing to grey
subplot(1, 2, 1) #creating a plot to obtain image
imshow(im)
axis('off')
title('Original')
subplot(1, 2, 2)
imshow(gaussian(im)) #gaussian function is called here
axis('off')
title('Filtered')
show()
def plot_results_per_target(options):
"""
From a folder of SPServer results, creates a results file for CASP.
"""
# Get parameters
input_dir = options.input_dir
output_dir = options.output_dir
create_directory(output_dir)
scoring_functions = ['ES3DC', 'PAIR', 'ZES3DC', 'ZPAIR']
# Read results
native_results_file = os.path.join(input_dir, 'Native.csv')
nearnative_results_file = os.path.join(input_dir, 'NearNative.csv')
wrong_results_file = os.path.join(input_dir, 'Decoy.csv')
native_results_df = pd.read_csv(native_results_file, sep='\t', index_col=0)
nearnative_results_df = pd.read_csv(nearnative_results_file,
sep='\t',
index_col=0)
wrong_results_df = pd.read_csv(wrong_results_file, sep='\t', index_col=0)
nearnative_results_df['TypeStructure'] = 'Near-native'
nearnative_results_df['Color'] = 'blue'
wrong_results_df['TypeStructure'] = 'Wrong'
wrong_results_df['Color'] = 'red'
results_df = pd.concat([nearnative_results_df,
wrong_results_df]).reset_index(drop=True)
new_df = results_df['PDB'].str.split('_', n=1, expand=True)
results_df['Target'] = new_df[0]
targets = set(results_df['Target'])
results_df.rename(columns={
'SPS_es3dc': 'ES3DC',
'SPS_Pair': 'PAIR',
'SPS_zes3dc': 'ZES3DC',
'SPS_Zpair': 'ZPAIR',
'GDT': 'GDT_TS',
'TM': 'TM score'
},
inplace=True)
# Plot results per target
results_per_target_dir = os.path.join(output_dir, 'results_per_target')
create_directory(results_per_target_dir)
results_per_target_file = os.path.join(output_dir,
'results_per_target.txt')
columns = [
'Target', 'ZES3DC-GDT', 'ZES3DC-TM', 'ZES3DC-QCS', 'ZPAIR-GDT',
'ZPAIR-TM', 'ZPAIR-QCS'
]
results_per_target_df = pd.DataFrame(columns=columns)
metric_to_top = {'GDT_TS': 100, 'TM score': 1, 'QCS': 100}
metric_to_bottom = {'GDT_TS': 0, 'TM score': 0, 'QCS': 0}
for target in sorted(targets):
target_df = results_df[results_df['Target'] == target].fillna(0)
r_values = []
for spserver_scoring_function in ['ZES3DC', 'ZPAIR']:
for metric in ['GDT_TS', 'TM score', 'QCS']:
#for metric in ['GDT_TS']:
# Calculate pearson correlation
scores_spserver = target_df[spserver_scoring_function].astype(
float).tolist()
scores_metric = target_df[metric].astype(float).tolist()
r_value, p_value = pearsonr(scores_spserver, scores_metric)
r_value = round(r_value, 2)
r_values.append(r_value)
print('Correlation for energies {} and {}: {}'.format(
spserver_scoring_function, metric, r_value))
# Make plot
output_plot = os.path.join(
results_per_target_dir,
'{}_{}_vs_{}.png'.format(target, spserver_scoring_function,
metric))
fig = pylab.figure(dpi=300)
#sns.set_context("paper")
#ax = sns.lmplot(x=spserver_scoring_function, y=metric, data=target_df, hue="TypeStructure", palette=['blue', 'red'], markers="+")
ax = sns.regplot(
target_df[spserver_scoring_function],
target_df[metric],
color="black",
order=1,
marker="+",
label="R: {}".format(r_value),
scatter_kws={'facecolors': target_df['Color']})
ax.set_ylim(metric_to_bottom[metric], metric_to_top[metric])
# Make a legend
# groupby and plot points of one color
target_df[target_df['Color'] == 'blue'].plot(
kind='scatter',
x=spserver_scoring_function,
y=metric,
c='blue',
ax=ax,
marker="+",
label='Near-native',
zorder=3)
target_df[target_df['Color'] == 'red'].plot(
kind='scatter',
x=spserver_scoring_function,
y=metric,
c='red',
ax=ax,
marker="+",
label='Wrong',
zorder=3)
handles, labels = ax.get_legend_handles_labels(
) # How to remove the default title: https://stackoverflow.com/questions/51579215/remove-seaborn-lineplot-legend-title?rq=1
ax.legend(handles=plt.plot([], ls="-", color='black') +
handles[1:],
labels=labels,
loc="best")
#plt.title('{} vs {}'.format(spserver_scoring_function, metric), fontweight='bold')
plt.gray()
plt.grid(True, color='w', linestyle='-', linewidth=1, zorder=0)
plt.gca().patch.set_facecolor('#EAEAF1')
plt.xlabel(spserver_scoring_function, fontweight='bold')
plt.ylabel(metric, fontweight='bold')
pylab.savefig(output_plot, format='png')
plt.clf()
# Save the r values in the table
results = [target] + r_values
df2 = pd.DataFrame([results], columns=columns)
results_per_target_df = results_per_target_df.append(df2)
# Plot pair plot
# output_plot = os.path.join(results_per_target_dir, '{}.png'.format(target))
# fig, axs = pylab.subplots(ncols=2, dpi=300, figsize=(9, 4))
# plt.rcParams['axes.grid'] = True
# sns.regplot(target_df['ZES3DC'], target_df['GDT_TS'], color="black", order=1, marker="+", label="R: {}".format(r_value), scatter_kws={'facecolors':target_df['Color']}, ax=axs[0])
# sns.regplot(target_df['ZPAIR'], target_df['GDT_TS'], color="black", order=1, marker="+", label="R: {}".format(r_value), scatter_kws={'facecolors':target_df['Color']}, ax=axs[1])
# target_df[target_df['Color'] == 'blue'].plot(kind = 'scatter', x="ZES3DC", y="GDT_TS", c = 'blue', ax = axs[0], marker="+", label = 'Near-native', zorder = 3)
# target_df[target_df['Color'] == 'red'].plot(kind = 'scatter', x="ZES3DC", y="GDT_TS", c = 'red', ax = axs[0], marker="+", label = 'Wrong', zorder = 3)
# target_df[target_df['Color'] == 'blue'].plot(kind = 'scatter', x="ZPAIR", y="GDT_TS", c = 'blue', ax = axs[1], marker="+", label = 'Near-native', zorder = 3)
# target_df[target_df['Color'] == 'red'].plot(kind = 'scatter', x="ZPAIR", y="GDT_TS", c = 'red', ax = axs[1], marker="+", label = 'Wrong', zorder = 3)
# handles, labels = axs[0].get_legend_handles_labels() # How to remove the default title: https://stackoverflow.com/questions/51579215/remove-seaborn-lineplot-legend-title?rq=1
# axs[0].legend(handles=plt.plot([],ls="-", color='black') + handles[1:], labels=labels, loc="best")
# handles, labels = axs[1].get_legend_handles_labels() # How to remove the default title: https://stackoverflow.com/questions/51579215/remove-seaborn-lineplot-legend-title?rq=1
# axs[1].legend(handles=plt.plot([],ls="-", color='black') + handles[1:], labels=labels, loc="best")
# #plt.gray()
# #plt.grid(True, color='w', linestyle='-', linewidth=1, zorder=0)
# #plt.gca().patch.set_facecolor('#EAEAF1')
# pylab.savefig(output_plot, format='png')
# plt.clf()
# Write results
results_per_target_df.to_csv(results_per_target_file,
sep='\t',
index=False)
return
def snr_peakstddev(array, source_xy, fwhm, out_coor=False, plot=False,
verbose=False):
"""Calculates the S/N (signal to noise ratio) of a single planet in a
post-processed (e.g. by LOCI or PCA) frame. The signal is taken as the ratio
of pixel value of the planet (test speckle) and the noise computed as the
standard deviation of the pixels in an annulus at the same radial distance
from the center of the frame. The diameter of the signal aperture and the
annulus width is in both cases 1 FWHM ~ 1 lambda/D.
Parameters
----------
array : array_like, 2d
Post-processed frame where we want to measure S/N.
source_xy : tuple of floats
X and Y coordinates of the planet or test speckle.
fwhm : float
Size in pixels of the FWHM.
out_coor: bool, optional
If True returns back the S/N value and the y, x input coordinates.
plot : bool, optional
Plots the frame and the apertures considered for clarity.
verbose: bool, optional
Chooses whether to print some intermediate results or not.
Returns
-------
snr : float
Value of the S/N for the given planet or test speckle.
"""
sourcex, sourcey = source_xy
centery, centerx = frame_center(array)
rad = dist(centery,centerx,sourcey,sourcex)
array = array + np.abs(array.min())
inner_rad = np.round(rad) - fwhm/2
an_coor = get_annulus(array, inner_rad, fwhm, output_indices=True)
ap_coor = draw.circle(sourcey, sourcex, int(np.ceil(fwhm/2)))
array2 = array.copy()
array2[ap_coor] = array[an_coor].mean() # we 'mask' the flux aperture
stddev = array2[an_coor].std()
peak = array[sourcey, sourcex]
snr = peak / stddev
if verbose:
msg = "S/N = {:.3f}, Peak px = {:.3f}, Noise = {:.3f}"
print(msg.format(snr, peak, stddev))
if plot:
_, ax = plt.subplots(figsize=(6, 6))
ax.imshow(array, origin='lower', interpolation='nearest')
circ = plt.Circle((centerx, centery), radius=inner_rad, color='r',
fill=False)
ax.add_patch(circ)
circ2 = plt.Circle((centerx, centery), radius=inner_rad+fwhm, color='r',
fill=False)
ax.add_patch(circ2)
aper = plt.Circle((sourcex, sourcey), radius=fwhm/2., color='b',
fill=False) # Coordinates (X,Y)
ax.add_patch(aper)
plt.show()
plt.gray()
if out_coor:
return sourcey, sourcex, snr
else:
return snr
def draw_images(img, name, data_output_path, vmin=0, vmax=1):
plt.gray()
plt.imshow(img, vmin=vmin, vmax=vmax)
plt.axis('off')
plt.savefig(data_output_path + name, bbox_inches='tight', pad_inches=0)
plt.close()
from keras.layers import Input, Dense from keras.models import Model from keras.datasets import mnist from keras import backend as K import numpy as np import matplotlib.pyplot as plt # this is the size of our encoded representations encoding_dim = 32 # 32 floats -> compression factor 24.5, assuming the input is 784 floats # this is our input placeholder; 784 = 28 x 28 input_img = Input(shape=(784, )) my_epochs = 100 # "encoded" is the encoded representation of the inputs encoded = Dense(encoding_dim * 4, activation='relu')(input_img) encoded = Dense(encoding_dim * 2, activation='relu')(encoded) encoded = Dense(encoding_dim, activation='relu')(encoded) # "decoded" is the lossy reconstruction of the input decoded = Dense(encoding_dim * 2, activation='relu')(encoded) decoded = Dense(encoding_dim * 4, activation='relu')(decoded) decoded = Dense(784, activation='sigmoid')(decoded) # this model maps an input to its reconstruction autoencoder = Model(input_img, decoded) # Separate Encoder model # this model maps an input to its encoded representation encoder = Model(input_img, encoded) # Separate Decoder model # create a placeholder for an encoded (32-dimensional) input encoded_input = Input(shape=(encoding_dim, )) # retrieve the layers of the autoencoder model decoder_layer1 = autoencoder.layers[-3] decoder_layer2 = autoencoder.layers[-2]
def solex_proc(serfile, shift, flag_display, ratio_fixe):
"""
----------------------------------------------------------------------------
Reconstuit l'image du disque a partir de l'image moyenne des trames et
des trames extraite du fichier ser
avec un fit polynomial
Corrige de mauvaises lignes et transversallium
basefich: nom du fichier de base de la video sans extension, sans repertoire
shift: ecart en pixel par rapport au centre de la raie pour explorer
longueur d'onde decalée
----------------------------------------------------------------------------
"""
plt.gray() #palette de gris si utilise matplotlib pour visu debug
global mylog
mylog = []
WorkDir = os.path.dirname(serfile) + "/"
os.chdir(WorkDir)
base = os.path.basename(serfile)
basefich = os.path.splitext(base)[0]
#affiche ou pas l'image du disque qui se contruit en temps reel
#gain de temps si affiche pas, mis en dur dans le script
#print(flag_display)
#flag_display=False
"""
----------------------------------------------------------------------------
Calcul polynome ecart sur une image au centre de la sequence
----------------------------------------------------------------------------
"""
savefich = basefich + '_mean'
ImgFile = savefich + '.fits'
#ouvre image _mean qui la moyenne de toutes les trames spectrales du fichier ser
hdulist = fits.open(ImgFile)
hdu = hdulist[0]
myspectrum = hdu.data
ih = hdu.header['NAXIS2']
iw = hdu.header['NAXIS1']
myimg = np.reshape(myspectrum, (ih, iw))
y1, y2 = detect_bord(myimg, axis=1, offset=5)
toprint = 'Limites verticales y1,y2 : ' + str(y1) + ' ' + str(y2)
print(toprint)
mylog.append(toprint + '\n')
PosRaieHaut = y1
PosRaieBas = y2
"""
-----------------------------------------------------------
Trouve les min intensité de la raie
-----------------------------------------------------------
"""
# construit le tableau des min de la raie a partir du haut jusqu'en bas
"""
MinOfRaie=[]
for i in range(PosRaieHaut,PosRaieBas):
line_h=myimg[i,:]
MinX=line_h.argmin()
MinOfRaie.append([MinX,i])
#print('MinOfRaie x,y', MinX,i)
np_m=np.asarray(MinOfRaie)
xm,ym=np_m.T
#best fit d'un polynome degre 2, les lignes y sont les x et les colonnes x sont les
p=np.polyfit(ym,xm,2)
"""
MinX = np.argmin(myimg, axis=1)
MinX = MinX[PosRaieHaut:PosRaieBas]
IndY = np.arange(PosRaieHaut, PosRaieBas, 1)
LineRecal = 1
#best fit d'un polynome degre 2, les lignes y sont les x et les colonnes x sont les y
p = np.polyfit(IndY, MinX, 2)
#p=np.polyfit(ym,xm,2)
#calcul des x colonnes pour les y lignes du polynome
a = p[0]
b = p[1]
c = p[2]
fit = []
#ecart=[]
for y in range(0, ih):
x = a * y**2 + b * y + c
deci = x - int(x)
fit.append([int(x) - LineRecal, deci, y])
#ecart.append([x-LineRecal,y])
toprint = 'Coef A2,A1,A0 :' + str(a) + ' ' + str(b) + ' ' + str(c)
print(toprint)
mylog.append(toprint + '\n')
np_fit = np.asarray(fit)
xi, xdec, y = np_fit.T
xdec = xi + xdec + LineRecal
xi = xi + LineRecal
#imgplot1 = plt.imshow(myimg)
#plt.scatter(xm,ym,s=0.1, marker='.', edgecolors=('blue'))
#plt.scatter(xi,y,s=0.1, marker='.', edgecolors=('red'))
#plt.scatter(xdec,y,s=0.1, marker='.', edgecolors=('green'))
#plt.show()
#on sauvegarde les params de reconstrution
#reconfile='recon_'+basefich+'.txt'
#np.savetxt(reconfile,ecart,fmt='%f',header='fichier recon',footer=str(LineRecal))
"""
----------------------------------------------------------------------------
----------------------------------------------------------------------------
Applique les ecarts a toute les lignes de chaque trame de la sequence
----------------------------------------------------------------------------
----------------------------------------------------------------------------
"""
#ouverture et lecture de l'entete du fichier ser
f = open(serfile, "rb")
b = np.fromfile(serfile, dtype='int8', count=4)
offset = 14
b = np.fromfile(serfile, dtype=np.uint32, count=1, offset=offset)
#print (LuID[0])
offset = offset + 4
b = np.fromfile(serfile, dtype='uint32', count=1, offset=offset)
#print(ColorID[0])
offset = offset + 4
b = np.fromfile(serfile, dtype='uint32', count=1, offset=offset)
#print(little_Endian[0])
offset = offset + 4
Width = np.fromfile(serfile, dtype='uint32', count=1, offset=offset)
Width = Width[0]
#print('Width :', Width)
offset = offset + 4
Height = np.fromfile(serfile, dtype='uint32', count=1, offset=offset)
Height = Height[0]
#print('Height :',Height)
offset = offset + 4
PixelDepthPerPlane = np.fromfile(serfile,
dtype='uint32',
count=1,
offset=offset)
PixelDepthPerPlane = PixelDepthPerPlane[0]
#print('PixelDepth :',PixelDepthPerPlane)
offset = offset + 4
FrameCount = np.fromfile(serfile, dtype='uint32', count=1, offset=offset)
FrameCount = FrameCount[0]
#print('nb de frame :',FrameCount)
toprint = ('width, height : ' + str(Width) + ' ' + str(Height))
print(toprint)
mylog.append(toprint + '\n')
toprint = ('Nb frame : ' + str(FrameCount))
print(toprint)
mylog.append(toprint + '\n')
count = Width * Height # Nombre d'octet d'une trame
FrameIndex = 1 # Index de trame, on evite les deux premieres
offset = 178 # Offset de l'entete fichier ser
if Width > Height:
flag_rotate = True
ih = Width
iw = Height
else:
flag_rotate = False
iw = Width
ih = Height
#debug
ok_resize = True
if flag_display:
cv2.namedWindow('disk', cv2.WINDOW_NORMAL)
FrameMax = FrameCount
cv2.resizeWindow('disk', FrameMax, ih)
cv2.moveWindow('disk', 100, 0)
#initialize le tableau qui va recevoir la raie spectrale de chaque trame
Disk = np.zeros((ih, FrameMax), dtype='uint16')
cv2.namedWindow('image', cv2.WINDOW_NORMAL)
cv2.moveWindow('image', 0, 0)
cv2.resizeWindow('image', int(iw), int(ih))
else:
#Disk=np.zeros((ih,1), dtype='uint16')
FrameMax = FrameCount
Disk = np.zeros((ih, FrameMax), dtype='uint16')
# init vector to speed up from Andrew Smiths
ind_l = (np.asarray(fit)[:, 0] + np.ones(ih) *
(LineRecal + shift)).astype(int)
ind_r = (ind_l + np.ones(ih)).astype(int)
left_weights = np.ones(ih) - np.asarray(fit)[:, 1]
right_weights = np.ones(ih) - left_weights
# lance la reconstruction du disk a partir des trames
while FrameIndex < FrameCount:
#t0=float(time.time())
img = np.fromfile(serfile, dtype='uint16', count=count, offset=offset)
img = np.reshape(img, (Height, Width))
if flag_rotate:
img = np.rot90(img)
if flag_display:
cv2.imshow('image', img)
if cv2.waitKey(1) == 27:
cv2.destroyAllWindows()
sys.exit()
#new code from Andraw Smiths to speed up reconstruction
# improve speed here
left_col = img[np.arange(ih), ind_l]
right_col = img[np.arange(ih), ind_r]
IntensiteRaie = left_col * left_weights + right_col * right_weights
#ajoute au tableau disk
Disk[:, FrameIndex] = IntensiteRaie
"""
---------------------------------------------------------
original code
---------------------------------------------------------
IntensiteRaie=np.empty(ih,dtype='uint16')
for j in range(0,ih):
dx=fit[j][0]+shift
deci=fit[j][1]
try:
IntensiteRaie[j]=(img[j,LineRecal+dx] *(1-deci)+deci*img[j,LineRecal+dx+1])
if img[j,LineRecal+dx]>=65000:
IntensiteRaie[j]=64000
#print ('intensite : ', img[j,LineRecal+dx])
except:
IntensiteRaie[j]=IntensiteRaie[j-1]
#ajoute au tableau disk
Disk[:,FrameIndex]=IntensiteRaie
#cv2.resizeWindow('disk',i-i1,ih)
if ok_resize==False:
Disk=Disk[1:,FrameIndex:]
#Disp=Disk
"""
if flag_display and FrameIndex % 5 == 0:
cv2.imshow('disk', Disk)
if cv2.waitKey(1) == 27: # exit if Escape is hit
cv2.destroyAllWindows()
sys.exit()
FrameIndex = FrameIndex + 1
offset = 178 + FrameIndex * count * 2
#ferme fichier ser
f.close()
#sauve fichier disque reconstruit
hdu.header['NAXIS1'] = FrameCount - 1
DiskHDU = fits.PrimaryHDU(Disk, header=hdu.header)
DiskHDU.writeto(basefich + '_img.fits', overwrite='True')
if flag_display:
cv2.destroyAllWindows()
"""
--------------------------------------------------------------------
--------------------------------------------------------------------
on passe au calcul des mauvaises lignes et de la correction geometrique
--------------------------------------------------------------------
--------------------------------------------------------------------
"""
iw = Disk.shape[1]
ih = Disk.shape[0]
img = Disk
y1, y2 = detect_bord(img, axis=1, offset=5) # bords verticaux
#detection de mauvaises lignes
# somme de lignes projetées sur axe Y
ysum = np.mean(img, 1)
#plt.plot(ysum)
#plt.show()
# ne considere que les lignes du disque avec marge de 15 lignes
ysum = ysum[y1 + 15:y2 - 15]
# filtrage sur fenetre de 31 pixels, polynome ordre 3 (etait 101 avant)
yc = savgol_filter(ysum, 31, 3)
# divise le profil somme par le profil filtré pour avoir les hautes frequences
hcol = np.divide(ysum, yc)
# met à zero les pixels dont l'intensité est inferieur à 1.03 (3%)
hcol[abs(hcol - 1) <= 0.03] = 0
# tableau de zero en debut et en fin pour completer le tableau du disque
a = [0] * (y1 + 15)
b = [0] * (ih - y2 + 15)
hcol = np.concatenate((a, hcol, b))
#plt.plot(hcol)
#plt.show()
# creation du tableau d'indice des lignes a corriger
l_col = np.where(hcol != 0)
listcol = l_col[0]
# correction de lignes par filtrage median 13 lignes, empririque
for c in listcol:
m = img[c - 7:c + 6, ]
s = np.median(m, 0)
img[c - 1:c, ] = s
#sauvegarde le fits
DiskHDU = fits.PrimaryHDU(img, header=hdu.header)
DiskHDU.writeto(basefich + '_corr.fits', overwrite='True')
"""
------------------------------------------------------------
calcul de la geometrie si on voit les bords du soleil
sinon on applique un facteur x=0.5
------------------------------------------------------------
"""
img2 = np.copy(img)
print()
#methode des limbes
#NewImg, newiw,flag_nobords,cercle =circularise(img2,iw,ih,ratio_fixe)
#methode fit ellipse
zexcl = 0.01 #zone d'exclusion des points contours
EllipseFit, section = detect_fit_ellipse(img, y1, y2, zexcl)
ratio = EllipseFit[2] / EllipseFit[1]
NewImg, newiw = circularise2(img2, iw, ih, ratio)
if section == 0:
flag_nobords = True
else:
flag_nobords = False
# sauve l'image circularisée
frame = np.array(NewImg, dtype='uint16')
hdu.header['NAXIS1'] = newiw
DiskHDU = fits.PrimaryHDU(frame, header=hdu.header)
DiskHDU.writeto(basefich + '_circle.fits', overwrite='True')
#on fit un cercle !!!
#CercleFit=detect_fit_cercle (frame,y1,y2)
#print(CercleFit)
#print()
print()
"""
NewImg, newiw =circularise2(img,iw,ih,ratio)
frame=np.array(NewImg, dtype='uint16')
hdu.header['NAXIS1']=newiw
DiskHDU=fits.PrimaryHDU(frame,header=hdu.header)
DiskHDU.writeto(basefich+'_circle2.fits',overwrite='True')
#on fit un cercle
CercleFit=detect_fit_cercle (frame,y1,y2)
print(CercleFit)
print()
#EllipseFit,section=detect_fit_ellipse(frame,y1,y2,0.01)
#x0=EllipseFit[0][0]
#y0=EllipseFit[0][1]
#diam_cercle=EllipseFit[2]
#cercle=[x0,y0, diam_cercle]
"""
"""
--------------------------------------------------------------
on echaine avec la correction de transversallium
--------------------------------------------------------------
"""
# on cherche la projection de la taille max du soleil en Y
y1, y2 = detect_bord(frame, axis=1, offset=0)
#print ('flat ',y1,y2)
# si mauvaise detection des bords en x alors on doit prendre toute l'image
if flag_nobords:
ydisk = np.median(img, 1)
else:
#plt.hist(frame.ravel(),bins=1000,)
#plt.show()
#plt.hist(frame.ravel(),bins=1000,cumulative=True)
#plt.show()
#seuil_bas=np.percentile(frame,25)
seuil_haut = np.percentile(frame, 97)
#print ('Seuils de flat: ',seuil_bas, seuil_haut)
#print ('Seuils bas x: ',seuil_bas*4)
#print ('Seuils haut x: ',seuil_haut*0.25)
#myseuil=seuil_haut*0.2
myseuil = seuil_haut * 0.5
# filtre le profil moyen en Y en ne prenant que le disque
ydisk = np.empty(ih + 1)
for j in range(0, ih):
temp = np.copy(frame[j, :])
temp = temp[temp > myseuil]
if len(temp) != 0:
ydisk[j] = np.median(temp)
else:
ydisk[j] = 1
y1 = y1
y2 = y2
ToSpline = ydisk[y1:y2]
Smoothed2 = savgol_filter(ToSpline, 301,
3) # window size, polynomial order
#best fit d'un polynome degre 4
np_m = np.asarray(ToSpline)
ym = np_m.T
xm = np.arange(y2 - y1)
p = np.polyfit(xm, ym, 4)
#calcul des x colonnes pour les y lignes du polynome
a = p[0]
b = p[1]
c = p[2]
d = p[3]
e = p[4]
Smoothed = []
for x in range(0, y2 - y1):
y = a * x**4 + b * x**3 + c * x**2 + d * x + e
Smoothed.append(y)
"""
plt.plot(ToSpline)
plt.plot(Smoothed)
plt.plot(Smoothed2)
plt.show()
"""
# divise le profil reel par son filtre ce qui nous donne le flat
hf = np.divide(ToSpline, Smoothed2)
# elimine possible artefact de bord
hf = hf[5:-5]
#reconstruit le tableau du pofil complet
a = [1] * (y1 + 5)
b = [1] * (ih - y2 + 5)
hf = np.concatenate((a, hf, b))
Smoothed = np.concatenate((a, Smoothed, b))
ToSpline = np.concatenate((a, ToSpline, b))
Smoothed2 = np.concatenate((a, Smoothed2, b))
"""
plt.plot(ToSpline)
plt.plot(Smoothed2)
plt.show()
plt.plot(hf)
plt.show()
"""
# genere tableau image de flat
flat = []
for i in range(0, newiw):
flat.append(hf)
np_flat = np.asarray(flat)
flat = np_flat.T
#evite les divisions par zeros...
flat[flat == 0] = 1
"""
plt.imshow(flat)
plt.show()
"""
# divise image par le flat
BelleImage = np.divide(frame, flat)
frame = np.array(BelleImage, dtype='uint16')
# sauvegarde de l'image deflattée
DiskHDU = fits.PrimaryHDU(frame, header=hdu.header)
DiskHDU.writeto(basefich + '_flat.fits', overwrite='True')
"""
-----------------------------------------------------------------------
correction de Tilt du disque
-----------------------------------------------------------------------
"""
img = frame
if not (flag_nobords) and abs(section) < 0.1:
# correction de slant uniquement si on voit les limbes droit/gauche
# trouve les coordonnées y des bords du disque dont on a les x1 et x2
# pour avoir les coordonnées y du grand axe horizontal
# on cherche la projection de la taille max du soleil en Y et en X
BackGround = 1000
x1, x2 = detect_bord(frame, axis=0, offset=0)
y_x1, y_x2 = detect_y_of_x(img, x1, x2)
# test que le grand axe de l'ellipse est horizontal
if abs(y_x1 - y_x2) > 5:
#calcul l'angle et fait une interpolation de slant
dy = (y_x2 - y_x1)
dx = (x2 - x1)
TanAlpha = (-dy / dx)
AlphaRad = math.atan(TanAlpha)
AlphaDeg = math.degrees(AlphaRad)
print('tan ', TanAlpha)
toprint = 'Angle slant limbes: ' + "{:+.2f}".format(AlphaDeg)
print(toprint)
mylog.append(toprint + '\n')
#decale lignes images par rapport a x1
colref = x1
NewImg = np.empty((ih, newiw))
for i in range(0, newiw):
x = img[:, i]
NewImg[:, i] = x
y = np.arange(0, ih)
dy = (i - colref) * TanAlpha
#print (dy)
#ycalc=[]
ycalc = y + np.ones(ih) * dy # improvements TheSmiths
#x et y sont les valeurs de la ligne originale avant decalge
#for j in range(0, len(y)):
#ycalc.append(y[j]+dy)
f = interp1d(ycalc,
x,
kind='linear',
fill_value=(BackGround, BackGround),
bounds_error=False)
xcalc = f(y)
NewLine = xcalc
NewImg[:, i] = NewLine
NewImg[NewImg <= 0] = 0 #modif du 19/05/2021 etait a 1000
img = NewImg
"""
#decale lignes images par rapport a x1
AlphaRad=np.deg2rad(angle_slant)
TanAlpha=math.tan(AlphaRad)
print()
print ('tan ellipse', TanAlpha)
print ('Angle_slant ellipse :',angle_slant)
colref=x1
NewImg=np.empty((ih,newiw))
for i in range(0,newiw):
x=img[:,i]
NewImg[:,i]=x
y=np.arange(0,ih)
dy=(i-colref)*TanAlpha
#print (dy)
ycalc=[]
#x et y sont les valeurs de la ligne originale avant decalge
for j in range(0, len(y)):
ycalc.append(y[j]+dy)
f=interp1d(ycalc,x,kind='linear',fill_value=(BackGround,BackGround),bounds_error=False)
xcalc=f(y)
NewLine=xcalc
NewImg[:,i]=NewLine
NewImg[NewImg<=0]=0 #modif du 19/05/2021 etait a 1000
img=NewImg
"""
# refait un calcul de mise a l'echelle
# le slant peut avoir legerement modifié la forme
#methode fit ellipse
print()
print('recircularise apres slant')
if (abs(section) < 0.1):
zexcl = 0.1
else:
zexcl = 0.01
EllipseFit, section = detect_fit_ellipse(img, y1, y2, zexcl)
ratio = EllipseFit[2] / EllipseFit[1]
NewImg, newiw = circularise2(img, newiw, ih, ratio)
flag_nobords = False
print()
img = np.copy(NewImg)
#on refait un test de fit ellipse
print('ellipse finale')
EllipseFit, section = detect_fit_ellipse(img, y1, y2, zexcl)
xc = int(EllipseFit[0][0])
yc = int(EllipseFit[0][1])
wi = int(EllipseFit[1])
he = int(EllipseFit[2])
cercle = [xc, yc, wi, he]
print()
# sauvegarde en fits de l'image finale
frame = np.array(img, dtype='uint16')
DiskHDU = fits.PrimaryHDU(frame, header=hdu.header)
DiskHDU.writeto(basefich + '_recon.fits', overwrite='True')
with open(basefich + '.txt', "w") as logfile:
logfile.writelines(mylog)
return frame, hdu.header, cercle
def main():
plt.gray()
exp = WaveletDenoise(imgutils.Image(imgutils.galaxy()[::-1]))
exp.gui.start()
#Read in a text file
inputimage = np.loadtxt("inputData.txt");
cpuimage = np.loadtxt("cpuoutput.txt");
gpuimage = np.loadtxt("gpuoutput.txt");
test = np.random.random((100,101))
#Display the arrays
#fig, ax = plt.subplots();
#ax.imshow(npcpuimage)
#plt.show()
f0 = plt.figure()
plt.imshow(inputimage,cmap=plt.gray())
plt.colorbar()
plt.suptitle("Input Image")
f1 = plt.figure()
plt.imshow(cpuimage,cmap=plt.gray())
plt.colorbar()
plt.suptitle("CPU Output")
f2 = plt.figure()
plt.imshow(gpuimage,cmap=plt.gray())
plt.colorbar()
plt.suptitle("GPU Output")
def show(img):
plt.gray()
plt.imshow(img)
plt.show()
early = first
i = 0
while i < 30:
current = updateData(kspace, pattern, current, 1)
current = waveletShrinkage(current, 0.001)
if (i == 0):
early = current
i += 1
#current = updateData(kspace, current, 0.1)
# todo:
# - impleemnt with conjugate transpose
# - check shrinkage with 0
# - create smaller phantom to speed computation time up
fig = pyplot.figure(dpi=90)
pyplot.subplot(221)
pyplot.set_cmap(pyplot.gray())
pyplot.imshow(abs(recon))
pyplot.subplot(222)
pyplot.set_cmap(pyplot.gray())
pyplot.imshow(abs(first))
pyplot.subplot(223)
pyplot.set_cmap(pyplot.gray())
pyplot.imshow(abs(early))
pyplot.subplot(224)
pyplot.set_cmap(pyplot.gray())
pyplot.imshow(abs(current))
pyplot.show()
print("After crop: ", img.shape, img.dtype)
pyplot.figure()
pyplot.imshow(img)
pyplot.axis('on')
pyplot.title('Cropped')
# switch to CHW
img = img.swapaxes(1, 2).swapaxes(0, 1)
pyplot.figure()
for i in range(3):
# For some reason, pyplot subplot follows Matlab's indexing
# convention (starting with 1). Well, we'll just follow it...
pyplot.subplot(1, 3, i + 1)
pyplot.imshow(img[i])
pyplot.axis('off')
pyplot.gray()
pyplot.title('RGB channel %d' % (i + 1))
# switch to BGR
img = img[(2, 1, 0), :, :]
# remove mean for better results
img = img * 255 - mean
# add batch size
img = img[np.newaxis, :, :, :].astype(np.float32)
print("NCHW: ", img.shape, img.dtype)
# initialize the neural net
with open(INIT_NET, 'rb') as f:
init_net = f.read()
kernel = zeros(im.shape)
kernel[:b.shape[0], :b.shape[1]] = b
fim = fft2(im)
fkernel = fft2(kernel)
fil_im = ifft2(fim * fkernel)
return abs(fil_im).astype(int)
if __name__ == "__main__":
from sys import argv
if len(argv) < 2:
print "Usage: python %s <image>" % argv[0]
exit()
im = array(Image.open(argv[1]))
im = im[:, :, 0]
gray()
subplot(1, 2, 1)
imshow(im)
axis('off')
title('Original')
subplot(1, 2, 2)
imshow(gaussian(im))
axis('off')
title('Filtered')
show()
# create subplot object
fig, axes = plt.subplots(
2,
5, # allocate subplot object(2x5) to axes
subplot_kw={
'xticks': (),
'yticks': ()
})
for ax, img in zip(axes.ravel(), digits.images):
ax.imshow(img)
plt.gray() # plot graph
plt.show() # plot graph
from sklearn.manifold import TSNE
# create model and learning
tsne = TSNE(random_state=0)
digits_tsne = tsne.fit_transform(digits.data)
colors = [
"#476A2A", "#7851B8", "#BD3430", "#4A2D4E", "#875525", "#A83683",
"#4E655E", "#853541", "#3A3120", "#535D8E"
]
def decodeQuad(self, quads, gray):
"""
decode the Quad
:param quads: array of quad which have four points
:param gray: gray picture
:return: array of detection
"""
detections = []
points = []
whitepoint = []
for quad in quads:
dd = 2 * self._blackBorder + self._d # tagFamily.d
blackvalue = []
whitevalue = []
tagcode = 0
for iy in range(dd):
for ix in range(dd):
x = (ix + 0.5) / dd
y = (iy + 0.5) / dd
point = np.int32(self._interpolate(quad, (x, y)))
points.append(point)
value = gray[point[0], point[1]]
if ((iy == 0 or iy == dd - 1)
or (ix == 0 or ix == dd - 1)):
blackvalue.append(value)
elif ((iy == 1 or iy == dd - 2)
or (ix == 1 or ix == dd - 2)):
whitevalue.append(value)
else:
continue
threshold = (np.average(blackvalue) + np.average(whitevalue)) / 2
for iy in range(dd):
for ix in range(dd):
if ((iy == 0 or iy == dd - 1)
or (ix == 0 or ix == dd - 1)):
continue
x = (ix + 0.5) / dd
y = (iy + 0.5) / dd
point = np.int32(self._interpolate(quad, (x, y)))
value = gray[point[0], point[1]]
tagcode = tagcode << 1
if value > threshold:
if (self._debug):
whitepoint.append(point)
tagcode |= 1
tagcode = hex(tagcode)
detection = self._decode(tagcode, quad)
if detection.good == True:
detection.addHomography()
detections.append(detection)
if self._debug and len(points) != 0:
plt.figure().set_size_inches(19.2, 10.8)
plt.subplot(121)
showpoint = np.array(points)
plt.plot(showpoint[:, 1], showpoint[:, 0], 'rx')
plt.imshow(gray)
plt.gray()
showpoint = np.array(whitepoint)
plt.subplot(122)
plt.plot(showpoint[:, 1], showpoint[:, 0], 'rx')
plt.imshow(gray)
plt.gray()
plt.show()
return detections
def imshow(image):
"""Show a [-1.0, 1.0] image."""
if image.ndim == 3 and image.shape[2] == 1: # for gray image
image = np.array(image, copy=True)
image.shape = image.shape[0:2]
plt.imshow(to_range(image), cmap=plt.gray())
def plot_generated_batch(generator_model):
import matplotlib.pyplot as plt
index = np.arange(Cloud.shape[0])
np.random.shuffle(index)
index = list(np.sort(index[:10]))
future_frames = generator_model.predict([Cloud[index], Cloud_[index]],
batch_size=10,
verbose=1)
plt.figure(figsize=(20, 20))
t = 0
for i in index:
ax = plt.subplot(10, 9, t * 9 + 1)
plt.imshow(Noncloud[i, 0].reshape((
64, 64,
IMAGE_CHANNEL)) if IMAGE_CHANNEL > 1 else Noncloud[i, 0].reshape((
64, 64)))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
ax = plt.subplot(10, 9, t * 9 + 2)
plt.imshow(Noncloud[i, 1].reshape((
64, 64,
IMAGE_CHANNEL)) if IMAGE_CHANNEL > 1 else Noncloud[i, 1].reshape((
64, 64)))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
ax = plt.subplot(10, 9, t * 9 + 3)
plt.imshow(Noncloud[i, 2].reshape((
64, 64,
IMAGE_CHANNEL)) if IMAGE_CHANNEL > 1 else Noncloud[i, 2].reshape((
64, 64)))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
ax = plt.subplot(10, 9, t * 9 + 4)
plt.imshow(Noncloud[i, 3].reshape((
64, 64,
IMAGE_CHANNEL)) if IMAGE_CHANNEL > 1 else Noncloud[i, 3].reshape((
64, 64)))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
ax = plt.subplot(10, 9, t * 9 + 5)
plt.imshow(future_frames[t, 0].reshape((
64, 64, IMAGE_CHANNEL
)) if IMAGE_CHANNEL > 1 else future_frames[t, 0].reshape((64, 64)))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
ax = plt.subplot(10, 9, t * 9 + 6)
plt.imshow(future_frames[t, 1].reshape((
64, 64, IMAGE_CHANNEL
)) if IMAGE_CHANNEL > 1 else future_frames[t, 1].reshape((64, 64)))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
ax = plt.subplot(10, 9, t * 9 + 7)
plt.imshow(future_frames[t, 2].reshape((
64, 64, IMAGE_CHANNEL
)) if IMAGE_CHANNEL > 1 else future_frames[t, 2].reshape((64, 64)))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
ax = plt.subplot(10, 9, t * 9 + 8)
plt.imshow(future_frames[t, 3].reshape((
64, 64, IMAGE_CHANNEL
)) if IMAGE_CHANNEL > 1 else future_frames[t, 3].reshape((64, 64)))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
ax = plt.subplot(10, 9, t * 9 + 9)
plt.imshow(Cloud_[i, 0].reshape((
64, 64,
IMAGE_CHANNEL)) if IMAGE_CHANNEL > 1 else Cloud_[i,
0].reshape((64,
64)))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
t += 1
plt.show()
def cae_mnist_encoding(train_percentage=0.1, test_percentage=0.1):
mnist = input_data.read_data_sets("MNIST_data/", one_hot=False)
train_m = int(mnist.train.num_examples * train_percentage)
test_m = int(mnist.test.num_examples * test_percentage)
validation_m = mnist.validation.num_examples
auto = ContractiveAutoencoder()
learning_rate = 0.001
optimizer_loss = tf.train.AdamOptimizer(learning_rate).minimize(auto.loss)
# Add ops to save and restore all the variables.
saver = tf.train.Saver()
sess = tf.Session()
sess.run(tf.global_variables_initializer())
batch_size = 100
epochs = 5
minimum_loss = np.inf
for epoch_i in range(epochs):
val_loss_list = []
for batch_i in range(train_m // batch_size):
batch_x, batch_y = mnist.train.next_batch(batch_size)
sess.run(optimizer_loss,
feed_dict={
auto.x: batch_x,
auto.y: batch_y
})
for batch_i in range(validation_m // batch_size):
batch_x, batch_y = mnist.validation.next_batch(batch_size)
val_loss = sess.run([auto.loss],
feed_dict={
auto.x: batch_x,
auto.y: batch_y
})
val_loss_list.append(val_loss)
validation_loss = np.mean(val_loss_list)
print(epoch_i, validation_loss)
if validation_loss < minimum_loss:
minimum_loss = validation_loss
save_path = saver.save(sess, "./models/cae_mnist/model.ckpt")
# Encode training and testing samples
# Save the encoded tensors
saver.restore(sess, "./models/cae_mnist/model.ckpt")
encoding_train_imgs_path = './data/MNIST_encoding/cae_train.encoding'
encoding_test_imgs_path = './data/MNIST_encoding/cae_test.encoding'
train_labels_path = './data/MNIST_encoding/cae_train.labels'
test_labels_path = './data/MNIST_encoding/cae_test.labels'
encoded_imgs_list = []
labels_list = []
for batch_i in range(train_m // batch_size):
batch_x, batch_y = mnist.train.next_batch(batch_size)
encoded_batches = sess.run(auto.encoded_x,
feed_dict={
auto.x: batch_x,
auto.y: batch_y
})
encoded_imgs_list.append(encoded_batches)
labels_list.append(batch_y)
for batch_i in range(validation_m // batch_size):
batch_x, batch_y = mnist.validation.next_batch(batch_size)
encoded_batches = sess.run(auto.encoded_x,
feed_dict={
auto.x: batch_x,
auto.y: batch_y
})
encoded_imgs_list.append(encoded_batches)
labels_list.append(batch_y)
encoded_train_imgs = np.array(encoded_imgs_list)
m, n, d = encoded_train_imgs.shape
encoded_train_imgs = encoded_train_imgs.reshape(m * n, d)
print(encoded_train_imgs.shape)
train_labels = np.array(labels_list).flatten()
print(train_labels.shape)
# Save the encoded imgs
pickle.dump(encoded_train_imgs, open(encoding_train_imgs_path, 'wb'))
pickle.dump(train_labels, open(train_labels_path, 'wb'))
encoded_imgs_list = []
labels_list = []
for batch_i in range(test_m // batch_size):
batch_x, batch_y = mnist.test.next_batch(batch_size)
encoded_batches = sess.run(auto.encoded_x,
feed_dict={
auto.x: batch_x,
auto.y: batch_y
})
encoded_imgs_list.append(encoded_batches)
labels_list.append(batch_y)
encoded_test_imgs = np.array(encoded_imgs_list)
m, n, d = encoded_test_imgs.shape
encoded_test_imgs = encoded_test_imgs.reshape(m * n, d)
print(encoded_test_imgs.shape)
test_labels = np.array(labels_list).flatten()
print(test_labels.shape)
# Save the encoded imgs
pickle.dump(encoded_test_imgs, open(encoding_test_imgs_path, 'wb'))
pickle.dump(test_labels, open(test_labels_path, 'wb'))
# Show 10 reconstructed images
n = 32
x_test, _ = mnist.test.next_batch(n)
reconstructed_imgs = sess.run(auto.reconstructed_x,
feed_dict={auto.x: x_test})
n = 10
plt.figure(figsize=(20, 4))
for i in range(n):
# display original
ax = plt.subplot(2, n, i + 1)
plt.imshow(x_test[i].reshape(28, 28))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
# display reconstruction
ax = plt.subplot(2, n, i + 1 + n)
plt.imshow(reconstructed_imgs[i].reshape(28, 28))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.savefig('./tmp/cae_mnist.png')
def plot_2D_clustering(X, labels):
plt.figure()
plt.scatter(X[:, 0], X[:, 1], c=labels.astype(np.float))
plt.gray()
plt.show()
from scipy import signal, misc
import matplotlib.pyplot as plt
from scipy import ndimage
col = 256
row = 256
cx = 128
cy = 128
D0 = 50 # 원의 크기
N = 2 # 필터링 정도
W = 30 #링의 두께
BW_L = np.zeros(shape=[col, row], dtype=np.float32)
BW_H = np.zeros(shape=[col, row], dtype=np.float32)
BW_BP = np.zeros(shape=[col, row], dtype=np.float32)
BW_BS = np.zeros(shape=[col, row], dtype=np.float32)
for x in range(col):
for y in range(row):
D = np.sqrt((cx - x)**2 + (cy - y)**2)
s = (W * D) / (D**2 - D0**2)
BW_L[x, y] = 1 / (1 + pow(D / D0, 2 * N))
BW_H[x, y] = 1 / (1 + pow(D0 / D, 2 * N))
BW_BS[x, y] = 1 / (1 + pow(s, 2 * N))
BW_BP = 1 - BW_BS
plt.subplot(221), plt.imshow(BW_L), plt.gray(), plt.axis('off'), plt.title(
'Butterworth Lowpass')
plt.subplot(222), plt.imshow(BW_H), plt.gray(), plt.axis('off'), plt.title(
'Butterworth Highpass')
plt.subplot(223), plt.imshow(BW_BP), plt.gray(), plt.axis('off'), plt.title(
'Butterworth Bandpass')
plt.subplot(224), plt.imshow(BW_BS), plt.gray(), plt.axis('off'), plt.title(
'Butterworth Bandstop')
plt.show()
from scipy import ndimage, misc import numpy.fft import matplotlib.pyplot as plt fig, (ax1, ax2) = plt.subplots(1, 2) plt.gray() # show the filtered result in grayscale ascent = misc.ascent() input_ = numpy.fft.fft2(ascent) result = ndimage.fourier_uniform(input_, size=20) result = numpy.fft.ifft2(result) ax1.imshow(ascent) ax2.imshow(result.real) # the imaginary part is an artifact plt.show()
def plot_RFs(name, nade, sample_shape, rows=5, cols=10):
rf_sizes = []
W = nade.W1
for i in range(W.shape[1]):
rf_sizes.append((i, -(W[:, i]**2).sum()))
rf_sizes.sort(key=lambda x: x[1])
plt.figure(figsize=(0.5 * cols, 0.5 * rows), dpi=100)
plt.gray()
for i in range(rows):
for j in range(cols):
n = i * cols + j
plt.subplot(rows, cols, n + 1)
rf = nade.W1[:, rf_sizes[n][0]]
#plot_RF(rf, sample_shape)
plot_RF_of_ps(np.exp(rf) / (1 + np.exp(rf)), sample_shape)
plt.tight_layout(0.2, 0.2, 0.2)
#plt.savefig(os.path.join(DESTINATION_PATH, name + "_W1.pdf"))
plt.figure(figsize=(0.5 * cols, 0.5 * rows), dpi=100)
plt.gray()
for i in range(rows):
for j in range(cols):
n = i * cols + j
plt.subplot(rows, cols, n + 1)
rf = np.resize(nade.Wflags[:, rf_sizes[n][0]],
np.prod(sample_shape)).reshape(sample_shape)
plot_RF(rf, sample_shape)
#plot_RF_of_ps(np.exp(rf)/(1+np.exp(rf)), sample_shape)
plt.tight_layout(0.2, 0.2, 0.2)
#plt.savefig(os.path.join(DESTINATION_PATH, name + "_Wflags.pdf"))
plt.figure(figsize=(0.5 * cols, 0.5 * rows), dpi=100)
plt.gray()
for i in range(rows):
for j in range(cols):
n = i * cols + j
plt.subplot(rows, cols, n + 1)
w = nade.W1[:, rf_sizes[n][0]]
f = nade.Wflags[:, rf_sizes[n][0]]
x = np.log((1 - np.exp(f - 1)) / (1 + np.exp(w + f)))
y = w - x
rf = np.resize(x, np.prod(sample_shape)).reshape(sample_shape)
#plot_RF_of_ps(rf, sample_shape)
plot_RF(rf, sample_shape)
plt.tight_layout(0.2, 0.2, 0.2)
plt.figure(figsize=(0.5 * cols, 0.5 * rows), dpi=100)
plt.gray()
for i in range(rows):
for j in range(cols):
n = i * cols + j
plt.subplot(rows, cols, n + 1)
w = nade.W1[:, rf_sizes[n][0]]
f = nade.Wflags[:, rf_sizes[n][0]]
x = np.log((1 - np.exp(f - 1)) / (1 + np.exp(w + f)))
y = w - x
rf = np.resize(y, np.prod(sample_shape)).reshape(sample_shape)
#plot_RF_of_ps(rf, sample_shape)
plot_RF(rf, sample_shape)
plt.tight_layout(0.2, 0.2, 0.2)
plt.show()
def main():
# this is the size of our encoded representations
input_img = Input(shape=(1375, ))
encoding_model = Model(input_img, get_encoding_network(input_img))
encoding_model.summary()
input_code = Input(shape=(49, ))
decoding_model = Model(input_code, get_deconding_network(input_code))
auto_encoder_model = Model(input_img,
decoding_model(encoding_model(input_img)))
print('***** encoding_model *******')
encoding_model.summary()
print('***** decoding_model *******')
decoding_model.summary()
print('***** auto_encoder_model*******')
auto_encoder_model.summary()
print('********************************')
autoencoder = auto_encoder_model
# this model maps an input to its encoded representation
encoder = encoding_model
decoder = decoding_model
autoencoder.compile(optimizer=Adam(lr=0.0001), loss='mse')
autoencoder.summary()
print('loading file from server')
all_subjects = [
"gcd",
]
add_time_domain_noise = False
number_of_k_fold = 10
downsample_params = 8
current_experiment_setting = "Color116ms"
cross_validation_iter = 1
train_data, train_tags, test_data_with_noise, test_tags, noise_shifts = prepare_data_for_experiment(
all_subjects,
add_time_domain_noise=add_time_domain_noise,
current_experiment_setting=current_experiment_setting,
downsample_params=downsample_params,
number_of_k_fold=number_of_k_fold,
cross_validation_iter=cross_validation_iter)
print('processed file, now running AE')
x_train = train_data.reshape(train_data.shape[0] * train_data.shape[1],
train_data.shape[2] * train_data.shape[3])
x_tst = test_data_with_noise[0]
x_test = x_tst.reshape(x_tst.shape[0] * x_tst.shape[1],
x_tst.shape[2] * x_tst.shape[3])
random_number = str(dt.microsecond)
reduce_lr = ReduceLROnPlateau(monitor='loss',
factor=0.2,
patience=5,
min_lr=0.00001)
callback = keras.callbacks.ModelCheckpoint(
r"c:\temp\weights_" + random_number +
".{epoch:02d}-{val_loss:.2f}.hdf5",
monitor='val_loss',
verbose=0,
save_best_only=True,
save_weights_only=False,
mode='auto',
period=1)
autoencoder.fit(x_train,
x_train,
epochs=30,
batch_size=32,
shuffle=True,
validation_data=(x_test, x_test),
callbacks=[callback, reduce_lr])
encoded_imgs = encoder.predict(x_test)
decoded_imgs = decoder.predict(encoded_imgs)
# use Matplotlib (don't ask)
n = 10 # how many digits we will display
plt.figure(figsize=(20, 4))
for i in range(n):
# display original
ax = plt.subplot(2, n, i + 1)
plt.imshow(x_test[i].reshape(25, 55))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
# display reconstruction
ax = plt.subplot(2, n, i + 1 + n)
plt.imshow(decoded_imgs[i].reshape(25, 55))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.show()
def cnn_nca_mnist_pretrain(trial, train_percentage=0.1, test_percentage=0.1):
mnist = input_data.read_data_sets("MNIST_data/", one_hot=False)
train_m = int(mnist.train.num_examples * train_percentage)
test_m = int(mnist.test.num_examples * test_percentage)
validation_m = mnist.validation.num_examples
auto = Autoencoder()
learning_rate = 0.001
optimizer_rec_error = tf.train.AdamOptimizer(learning_rate).minimize(
auto.reconstruction_error)
# Add ops to save and restore all the variables.
saver = tf.train.Saver()
sess = tf.Session()
sess.run(tf.global_variables_initializer())
# Pre-train step
batch_size = 100
epochs = 100
rec_error = np.inf
for epoch_i in range(epochs):
for batch_i in range(train_m // batch_size):
batch_x, batch_y = mnist.train.next_batch(batch_size)
sess.run(optimizer_rec_error,
feed_dict={
auto.x: batch_x,
auto.y: batch_y
})
validation_loss, reconstruction_error, nca_obj = cal_loss(
auto, sess, mnist.validation, validation_m, batch_size)
print(epoch_i, validation_loss, reconstruction_error, nca_obj)
if reconstruction_error < rec_error:
rec_error = reconstruction_error
save_path = saver.save(sess, "./models/tf_mnist/model.ckpt")
# Report the loss
validation_loss, reconstruction_error, nca_obj = cal_loss(
auto, sess, mnist.test, test_m, batch_size)
print("Report the test loss of the best model: ")
print(validation_loss, reconstruction_error, nca_obj)
# Show 10 reconstructed images
n = 10
x_test, _ = mnist.test.next_batch(n)
reconstructed_imgs = sess.run(auto.reconstructed_x,
feed_dict={auto.x: x_test})
plt.figure(figsize=(20, 4))
for i in range(n):
# display original
ax = plt.subplot(2, n, i + 1)
plt.imshow(x_test[i].reshape(28, 28))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
# display reconstruction
ax = plt.subplot(2, n, i + 1 + n)
plt.imshow(reconstructed_imgs[i].reshape(28, 28))
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.savefig('./tmp/tf_mnist.png')
def main():
global vmin, vmax
mnist = True
bsds = False
if mnist:
dataset_file = os.path.join(os.environ["DATASETSPATH"],
"original_NADE/binarized_mnist.hdf5")
test_dataset = Data.BigDataset(dataset_file, "test", "data")
nade2hl = load_model(
os.path.join(os.environ["RESULTSPATH"],
"orderless/mnist/2hl/NADE.hdf5"))
#plot_MNIST_results()
#exit()
#plot_examples(nade2hl, test_dataset, (28,28), "mnist_examples.pdf", rows = 5, cols=10)
#plot_samples(nade2hl, (28,28), "mnist_samples_2hl.pdf", rows = 5, cols=10)
plot_RFs("MNIST_2hl", nade2hl, sample_shape=(28, 28))
#nade2hl = load_model(os.path.join(os.environ["RESULTSPATH"], "orderless/mnist/2hl/NADE.hdf5"))
#np.random.seed(1) #1,17,43,49
#nade2hl.setup_n_orderings(10)
#inpaint_digits(test_dataset, (28,28), nade2hl, delete_shape=(10,10), n_examples = 6, n_samples = 5)
if bsds:
vmin = -0.5
vmax = 0.5
np.random.seed(1)
dataset_file = os.path.join(
os.environ["DATASETSPATH"],
"natural_images/BSDS300/BSDS300_63_no_DC_val.hdf5")
test_dataset = Data.BigDataset(dataset_file, "test/.", "patches")
nade = load_model(
os.path.join(os.environ["RESULTSPATH"],
"orderless/BSDS300/6hl/NADE.hdf5"))
plot_examples(nade,
test_dataset, (8, 8),
"BSDS_data.pdf",
rows=5,
cols=10)
plot_samples(nade, (8, 8), "BSDS_6hl_samples.pdf", rows=5, cols=10)
nade = load_model(
os.path.join(os.environ["RESULTSPATH"],
"orderless/BSDS300/2hl/NADE.hdf5"))
plot_RFs("BSDS_2hl", nade, sample_shape=(8, 8))
exit()
np.random.seed(8341)
show_RFs = False
print_likelihoods = False
print_mixture_likelihoods = True
show_data = False
show_samples = False
denoise_samples = False
n_orderings = 6
n_samples = 10
#sample_shape = (28, 28)
sample_shape = (8, 8)
parser = OptionParser(
usage="usage: %prog [options] dataset_file nade_path")
parser.add_option("--training_samples",
dest="training_route",
default="train")
parser.add_option("--validation_samples",
dest="validation_route",
default="validation")
parser.add_option("--test_samples", dest="test_route", default="test")
parser.add_option("--samples_name", dest="samples_name", default="data")
parser.add_option("--normalize",
dest="normalize",
default=False,
action="store_true")
parser.add_option("--add_dimension",
dest="add_dimension",
default=False,
action="store_true")
(options, args) = parser.parse_args()
dataset_filename = os.path.join(os.environ["DATASETSPATH"], args[0])
model_filename = os.path.join(os.environ["RESULTSPATH"], args[1])
print(model_filename)
try:
hdf5_route = "/highest_validation_likelihood/parameters"
params = Results.Results(model_filename).get(hdf5_route)
except:
hdf5_route = "/final_model/parameters"
params = Results.Results(model_filename).get(hdf5_route)
model_class = getattr(NADE, params["__class__"])
nade = model_class.create_from_params(params)
#Load datasets
print("Loading datasets")
dataset_file = os.path.join(os.environ["DATASETSPATH"], dataset_filename)
training_dataset = Data.BigDataset(dataset_file, options.training_route,
options.samples_name)
validation_dataset = Data.BigDataset(dataset_file,
options.validation_route,
options.samples_name)
test_dataset = Data.BigDataset(dataset_file, options.test_route,
options.samples_name)
n_visible = training_dataset.get_dimensionality(0)
if options.normalize:
mean, std = Data.utils.get_dataset_statistics(training_dataset)
training_dataset = Data.utils.normalise_dataset(
training_dataset, mean, std)
validation_dataset = Data.utils.normalise_dataset(
validation_dataset, mean, std)
test_dataset = Data.utils.normalise_dataset(test_dataset, mean, std)
#Setup a list with n_orderings orderings
print("Creating random orderings")
orderings = list()
#orderings.append(range(nade.n_visible))
for i in range(n_orderings):
o = range(nade.n_visible)
np.random.shuffle(o)
orderings.append(o)
#Print average loglikelihood and se for several orderings
if print_likelihoods:
nade.setup_n_orderings(orderings=orderings)
ll = nade.get_average_loglikelihood_for_dataset(test_dataset)
print("Mean test-loglikelihood (%d orderings): %.2f" % (orderings, ll))
if print_mixture_likelihoods:
#d = test_dataset.sample_data(1000)[0].T
d = test_dataset.get_data()[0].T
for n in [1, 2, 4, 8, 16, 32, 64, 128]:
#nade.setup_n_orderings(n)
multi_ord = nade.logdensity(d)
print(n, np.mean(multi_ord), scipy.stats.sem(multi_ord))
if show_RFs:
rf_sizes = []
W = nade.W1.get_value()
for i in range(W.shape[1]):
rf_sizes.append((i, -(W[:, i]**2).sum()))
rf_sizes.sort(key=lambda x: x[1])
plt.figure()
plt.gray()
for i in range(10):
for j in range(10):
n = i * 10 + j
plt.subplot(10, 10, n + 1)
rf = nade.Wflags.get_value()[:, rf_sizes[n][0]]
plot_RF(rf, sample_shape)
plt.figure()
plt.gray()
for i in range(10):
for j in range(10):
n = i * 10 + j
plt.subplot(10, 10, n + 1)
rf = np.resize(nade.W1.get_value()[:, rf_sizes[n][0]],
np.prod(sample_shape)).reshape(sample_shape)
plot_RF(rf, sample_shape)
plt.show()
#Show some samples
if show_samples:
images = []
for row, o in enumerate(orderings):
samples = nade.sample(n_samples)
for i in range(samples.shape[1]):
nade.setup_n_orderings(n=1)
sample = samples[:, i]
dens = nade.logdensity(sample[:, np.newaxis])
if options.add_dimension:
sample_extra_dim = np.resize(sample, len(sample) + 1)
sample_extra_dim[-1] = -sample.sum()
images.append((sample_extra_dim, dens))
else:
images.append((sample, dens))
images.sort(key=lambda x: -x[1])
plt.figure()
plt.gray()
for row, o in enumerate(orderings):
for i in range(samples.shape[1]):
plt.subplot(n_orderings, n_samples, row * n_samples + i + 1)
im_ll = images[row * n_samples + i]
plot_sample(im_ll[0], sample_shape)
plt.title("%.1f" % im_ll[1], fontsize=9)
plt.show()
#Show some data
if show_data:
images = []
for row, o in enumerate(orderings):
samples = test_dataset.sample_data(n_samples)[0].T
for i in range(samples.shape[1]):
nade.setup_n_orderings(n=1)
sample = samples[:, i]
dens = nade.logdensity(sample[:, np.newaxis])
if options.add_dimension:
sample_extra_dim = np.resize(sample, len(sample) + 1)
sample_extra_dim[-1] = -sample.sum()
images.append((sample_extra_dim, dens))
else:
images.append((sample, dens))
images.sort(key=lambda x: -x[1])
plt.figure()
plt.gray()
for row, o in enumerate(orderings):
for i in range(samples.shape[1]):
plt.subplot(n_orderings, n_samples, row * n_samples + i + 1)
im_ll = images[row * n_samples + i]
plot_sample(im_ll[0], sample_shape)
plt.title("%.1f" % im_ll[1], fontsize=9)
plt.show()
#Get a sample and clean it by taking pixels randomly and assigning the most probably value given all the others
if denoise_samples:
n = 10
nade.set_ordering(orderings[-1])
sample = nade.sample(n)
plt.figure()
plt.gray()
for it in range(10):
for s in range(n):
logdensities = nade.logdensity(sample)
plt.subplot(n, 10, s * 10 + it + 1)
plot_sample(sample[:, s], sample_shape)
plt.title("%.1f" % logdensities[s])
#i = np.random.randint(sample.shape[0])
for i in range(sample.shape[0]):
j = np.random.randint(sample.shape[0])
#mask = sample
mask = np.ones_like(sample)
mask[j] = 0
sample[j] = 0
ps = nade.conditional(sample, mask)
#sample[j] = ps[j] > np.random.rand()
sample[j] = ps[j] > 0.5
plt.show()
def latent_traversal(self):
limit = 1
pad = 0
random_index = np.random.randint(low=0,
high=self.data.x_test.shape[0],
size=2)
z, z_mean, z_log_var = self.network.vae.encoder(
self.data.x_test[random_index])
z = z.numpy()
fig = plt.figure(figsize=(18, 10))
ax = fig.add_subplot(111)
ax.set_yticks([])
ax.set_xticks([-100, 100])
ax.set_xticklabels(
[r"$\mu_{z_{j}}-$" + str(limit), r"$\mu_{z_{j}}+$" + str(limit)],
fontsize=16)
n_col = 10
n_row = z.shape[1]
spec = gridspec.GridSpec(nrows=n_row,
ncols=n_col,
width_ratios=np.ones((n_col, )),
height_ratios=np.ones((n_row, )),
wspace=0,
hspace=0)
for i in range(n_row):
start = z[0][i] - limit
end = z[0][i] + limit
int1 = np.linspace(start, z[0][i], num=5, endpoint=False)
int2 = np.linspace(z[0][i], end, num=5,
endpoint=False) # detta måste ändras
grid_z = np.concatenate((int1, int2), axis=0)
for j in range(n_col):
f_ax1 = fig.add_subplot(spec[i, j])
if j == 0:
string_i = str(i)
string_list = [int(string_i) for string_i in str(string_i)]
if len(string_list) == 2:
f_ax1.annotate(r"$z_{}$".format(string_list[0]) +
r"$_{}$".format(string_list[1]),
xy=(0, 0.5),
xytext=(-f_ax1.yaxis.labelpad - pad, 0),
xycoords=f_ax1.yaxis.label,
textcoords='offset points',
size='large',
fontsize=20,
ha='right',
va='center')
else:
f_ax1.annotate(r"$z_{}$".format(string_list[0]),
xy=(0, 0.5),
xytext=(-f_ax1.yaxis.labelpad - pad, 0),
xycoords=f_ax1.yaxis.label,
textcoords='offset points',
size='large',
fontsize=20,
ha='right',
va='center')
f_ax1.set_yticks([])
f_ax1.set_xticks([])
f_ax1.set_aspect('equal')
z[0][i] = grid_z[j]
x_rec = self.network.vae.decoder(z)
x_rec = x_rec.numpy()
plt.gray()
f_ax1.imshow(x_rec[0].reshape(64, 64), aspect='auto')
plt.savefig(self.path + "/figures/" + "traversal" + self.path + ".png")
plt.show()
def inpaint_digits(dataset,
shape,
model,
n_examples=5,
delete_shape=(10, 10),
n_samples=5,
name="inpaint_digits"):
#Load a few digits from the test dataset (as rows)
data = dataset.sample_data(1000)[0]
#data = data[[1,12,17,81,88,102],:]
data = data[[1, 12, 17, 81, 88, 37], :]
n_examples = data.shape[0]
#Generate a random region to delete
regions = [(np.random.randint(shape[0] - delete_shape[0] + 1),
np.random.randint(shape[1] - delete_shape[1] + 1))
for i in range(n_examples)]
print(regions)
regions = [(11, 5), (11, 5), (11, 5), (4, 13), (4, 13), (4, 13)]
#Generate masks
def create_mask(x, y):
mask = np.ones(shape)
mask[y:y + delete_shape[1], x:x + delete_shape[0]] = 0
return mask.flatten()
masks = [create_mask(x, y) for (x, y) in regions]
#Hollow
def hollow(example, mask):
hollowed = example.copy()
return hollowed * mask
hollowed = [hollow(data[i, :], mask) for i, mask in enumerate(masks)]
densities = model.logdensity(data.T)
#Calculate the marginal probability under a nade
marginal_densities = [
model.marginal_density(h, mask) for h, mask in zip(hollowed, masks)
]
#Generate some samples
samples = [
model.sample_conditional(h, mask, n_samples=n_samples)
for h, mask in zip(hollowed, masks)
]
#samples = [model.sample_conditional_max(h, mask, n_samples=n_samples) for h, mask in zip(hollowed, masks)]
#Plot it all
matplotlib.rcParams.update({'font.size': 8})
plt.figure(figsize=(5, 5), dpi=100)
plt.gray()
cols = 2 + n_samples
for row in range(n_examples):
# Original
plt.subplot(n_examples, cols, row * cols + 1)
plot_sample(data[row, :], shape, origin="upper")
plt.title("%.2f" % densities[row])
# Marginalization region
plt.subplot(n_examples, cols, row * cols + 2)
plot_sample(hollowed[row], shape, origin="upper")
plt.gca().add_patch(
plt.Rectangle(regions[row],
delete_shape[0],
delete_shape[1],
facecolor="red",
edgecolor="red"))
plt.title("%.2f" % marginal_densities[row])
# Samples
for j in range(n_samples):
plt.subplot(n_examples, cols, row * cols + 3 + j)
plot_sample(samples[row][:, j], shape, origin="upper")
plt.subplots_adjust(left=0.01,
right=0.99,
top=0.95,
bottom=0.01,
hspace=0.40,
wspace=0.04)
plt.savefig(os.path.join(DESTINATION_PATH, name + ".pdf"))
def plot_digit(feature_vector):
plt.gray()
plt.matshow(feature_vector.reshape(28,28))
plt.show()
dataTemp["Micro_precision"] = precision_score(y_true,
y_pred,
average='micro')
dataTemp["Micro_recall"] = recall_score(y_true, y_pred, average='micro')
dataTemp["Micro_f1-score"] = f1_score(y_true, y_pred, average='micro')
resultDimen = resultDimen.append(dataTemp, ignore_index=True)
# print('------Weighted-----')
# print('Weighted_precision',precision_score(y_true,y_pred,average='weighted'))
# print('Weighted_recall', recall_score(y_true, y_pred, average='weighted'))
# print('Weighted_f1-score', f1_score(y_true, y_pred, average='weighted'))
#
# print('------Macro-----')
# print('Macro_precision', precision_score(y_true, y_pred, average='macro'))
# print('Macro_recall', recall_score(y_true, y_pred, average='macro'))
# print('Macro_f1-score', f1_score(y_true, y_pred, average='macro'))
#
# print('------Micro-----')
# print('Micro_precision', precision_score(y_true, y_pred, average='micro'))
# print('Micro_recall', recall_score(y_true, y_pred, average='micro'))
# print('Micro_f1-score', f1_score(y_true, y_pred, average='micro'))
plt.matshow(conf_matrix, cmap=plt.gray())
plt.title(str(list))
plt.savefig("out/" + str(list) + ".png")
plt.show()
print(resultDimen)
resultDimen.to_csv("out/result.csv")
