def glow(src, dst):
graph = Image.open(src)
size = graph.size
width = size[0]
height = size[1]
source = graph.convert('RGB')
Gauss = graph.convert('RGB')
source = ny.double(ny.array(source))
Gauss = Gauss.filter(MyGaussianBlur(radius=15))
Gauss = ny.double(ny.array(source))
Result = ny.zeros([height, width, 3])
for row in range(height):
for col in range(width):
for k in range(3):
if source[row, col, k] <= 128:
value = Gauss[row, col, k] * source[row, col, k] / 128
Result[row, col, k] = min(255, max(0, value))
else:
value = 255 - (255 - Gauss[row, col, k]) * (255 - source[row, col, k]) / 128
Result[row, col, k] = min(255, max(0, value))
result = Image.fromarray(ny.uint8(Result)).convert('RGB')
result.save(dst)
return 0
def sig_lmc(C, A):
'''
This a function that using Lumped Markov chain to calculate
the significance of clusters in a give communinity structure.
refer to "Piccardi 2011 in PloS one".
Here we normalize the original definition of persistence by
the size of the corresponding cluster to get a better
INPUT:
"A" is a N-by-N weighted adjacency matrix
"C" is a N-by-1 partition(cluster) vector
OUTPUT:
normalized persistence probability of all clusters
'''
'''
Transition Matrix
'''
C = np.asarray(C)
A = np.double(A)
P = np.linalg.solve(np.diag(np.sum(A,axis = 1)),A)
[eval, evec] = linalg.eigs(P.T, 1)
if min(evec)<0:
evec = -evec
pi = np.double(evec.T)
num_node = np.double(np.shape(A)[0])
cl_label = np.double(np.unique(C))
num_cl = len(cl_label)
H = np.zeros((num_node, num_cl),dtype = np.double)
for i in range(num_cl):
H[:, i] = np.double((C==cl_label[i]))
# Transition matrix of the lumped Markov chain
Q = np.dot(np.dot(np.dot(np.linalg.solve(np.diag(np.dot(pi,H).flatten()),H.T),np.diag(pi.flatten())),P),H)
persistence = np.multiply(np.divide(np.diag(Q), np.sum(H,axis = 0)),np.sum(H))
return persistence
def apodization( name = 'butter.32', shape= [2048,2048], radius=None ):
""" apodization( name = 'butter.32', size = [2048,2048], radius=None )
Provides a 2-D filter or apodization window for Fourier filtering or image clamping.
Radius = None defaults to shape/2
Valid names are:
'hann' - von Hann cosine window on radius
'hann_square' as above but on X-Y
'hamming' - good for apodization, nonsense as a filter
'butter.X' Butterworth multi-order filter where X is the order of the Lorentzian
'butter_square.X' Butterworth in X-Y
'gauss_trunc' - truncated gaussian, higher performance (smaller PSF) than hann filter
'gauss' - regular gaussian
NOTE: There are windows in scipy.signal for 1D-filtering...
WARNING: doesn't work properly for odd image dimensions
"""
# Make meshes
shape = np.asarray( shape )
if radius is None:
radius = shape/2.0
else:
radius = np.asarray( radius, dtype='float' )
# DEBUG: Doesn't work right for odd numbers
[xmesh,ymesh] = np.meshgrid( np.arange(-shape[1]/2,shape[1]/2), np.arange(-shape[0]/2,shape[0]/2) )
r2mesh = xmesh*xmesh/( np.double(radius[0])**2 ) + ymesh*ymesh/( np.double(radius[1])**2 )
try:
[name, order] = name.lower().split('.')
order = np.double(order)
except ValueError:
order = 1
if name == 'butter':
window = np.sqrt( 1.0 / (1.0 + r2mesh**order ) )
elif name == 'butter_square':
window = np.sqrt( 1.0 / (1.0 + (xmesh/radius[1])**order))*np.sqrt(1.0 / (1.0 + (ymesh/radius[0])**order) )
elif name == 'hann':
cropwin = ((xmesh/radius[1])**2.0 + (ymesh/radius[0])**2.0) <= 1.0
window = cropwin.astype('float') * 0.5 * ( 1.0 + np.cos( 1.0*np.pi*np.sqrt( (xmesh/radius[1])**2.0 + (ymesh/radius[0])**2.0 ) ) )
elif name == 'hann_square':
window = ( (0.5 + 0.5*np.cos( np.pi*( xmesh/radius[1]) ) ) *
(0.5 + 0.5*np.cos( np.pi*( ymesh/radius[0] ) ) ) )
elif name == 'hamming':
cropwin = ((xmesh/radius[1])**2.0 + (ymesh/radius[0])**2.0) <= 1.0
window = cropwin.astype('float') * ( 0.54 + 0.46*np.cos( 1.0*np.pi*np.sqrt( (xmesh/radius[1])**2.0 + (ymesh/radius[0])**2.0 ) ) )
elif name == 'hamming_square':
window = ( (0.54 + 0.46*np.cos( np.pi*( xmesh/radius[1]) ) ) *
(0.54 + 0.46*np.cos( np.pi*( ymesh/radius[0] ) ) ) )
elif name == 'gauss' or name == 'gaussian':
window = np.exp( -(xmesh/radius[1])**2.0 - (ymesh/radius[0])**2.0 )
elif name == 'gauss_trunc':
cropwin = ((0.5*xmesh/radius[1])**2.0 + (0.5*ymesh/radius[0])**2.0) <= 1.0
window = cropwin.astype('float') * np.exp( -(xmesh/radius[1])**2.0 - (ymesh/radius[0])**2.0 )
elif name == 'lanczos':
print( "TODO: Implement Lanczos window" )
return
else:
print( "Error: unknown filter name passed into apodization" )
return
return window
def reconstruct(pitch,fs,coffs,syllable): gain = coffs[0] coffs[0] = 1; x = np.double(fs)/np.double(pitch) num = np.ceil((pitch*3)/10.0) #300ms thus if syllable != "s": ex_input = np.zeros(num*x) for i in range (0,int(num)): ex_input[(i*x)-1] = 1 # Filtering the signal out = signal.filtfilt([gain],coffs,ex_input) d_num = [1] d_den = [1,-0.9] out = signal.filtfilt(d_num,d_den,out) # De-emphasis out = np.int16(out/np.max(np.abs(out)) * 32767) else: ex_input = np.random.normal(0,1,num) out = signal.filtfilt([gain],coffs,ex_input) out = np.int16(out/np.max(np.abs(out)) * 32767) return out
def openDATfile(filename,ftype,srate=25000):
fh = open(filename,'r')
fh.seek(0)
if ftype == 'amp':
data = np.fromfile(fh, dtype=np.int16)
fh.close()
data = np.double(data)
data *= 0.195 # according the Intan, the output should be multiplied by 0.195 to be converted to micro-volts
elif ftype == 'adc':
data = np.fromfile(fh, dtype=np.uint16)
fh.close()
data = np.double(data)
data *= 0.000050354 # according the Intan, the output should be multiplied by 0.195 to be converted to micro-volts
data -= np.mean(data)
elif ftype == 'aux':
data = np.fromfile(fh, dtype=np.uint16)
fh.close()
data = np.double(data)
data *= 0.0000748 # according the Intan, the output should be multiplied by 0.195 to be converted to micro-volts
elif ftype == 'time':
data = np.fromfile(fh, dtype=np.int32)
fh.close()
data = np.double(data)
data /= srate # according the Intan, the output should be multiplied by 0.195 to be converted to micro-volts
return data
def _save_hdf5_object(object, filename):
"""
Save a class object in hdf5 format.
Parameters
----------
object: class instance
A class object whose attributes would be saved in a dictionary format.
filename: str
The file name to save to
"""
items = vars(object)
attrs = [name for name in items]
with h5py.File(filename, 'w') as hf:
for attr in attrs:
data = items[attr]
# If data is a single number, store as an attribute.
if _isattribute(data):
if isinstance(data, np.longdouble):
data = np.double(data)
utils.simon("Casting data as double instead of longdouble.")
hf.attrs[attr] = data
# If data is a numpy array, create a dataset.
else:
if isinstance(data[0], np.longdouble):
data = np.double(data)
utils.simon("Casting data as double instead of longdouble.")
hf.create_dataset(attr, data=data)
def fastColumnWiseCorrcoef(O, P): #P = age_mtx = 1 x n, O = subjects_vertices_mtx = n x m
n = P.size
DO = O - (np.einsum('ij->j',O) / np.double(n))
P -= (np.einsum('i->',P) / np.double(n))
tmp = np.einsum('ij,ij->j',DO,DO)
tmp *= np.einsum('i,i->',P,P)
return np.dot(P, DO) / np.sqrt(tmp)
def estOE(d):
gt = d['gt']['graph']
gt = bfu.undersample(gt, 1)
e = gk.OCE(d['estimate'], gt)
N = np.double(len(gk.edgelist(gt))) +\
np.double(len(gk.bedgelist(gt)))
return (e['directed'][0] + e['bidirected'][0]) / N
def test_high_precision_split2(self):
C_I, C_F, C_l, k = \
mp.io._split_high_precision_number("C", np.double(1.01), 8)
assert C_I == 1
np.testing.assert_almost_equal(C_F, np.double(0.01), 6)
assert C_l == 0
assert k == "double"
def get_precision_recall_values(predictionImage, labelImage, threshLevels, doPlot):
# normalize images
predictionImage = predictionImage / np.double(np.max(predictionImage))
labelImage = labelImage > 0
precisionValues = np.zeros(threshLevels)
recallValues = np.zeros(threshLevels)
counter = 0
# get sufficient number of thresholds
threshValues = np.linspace(0.01,0.99,threshLevels)
for thresh in threshValues:
#threshold the prediction image
threshPrediction = predictionImage>thresh
truePositives = np.double(np.sum(np.logical_and(labelImage, threshPrediction)))
falsePositives = np.double(np.sum(np.logical_and(np.logical_not(labelImage), threshPrediction)))
falseNegatives = np.double(np.sum(np.logical_and(labelImage, np.logical_not(threshPrediction))))
precision = truePositives / (truePositives + falsePositives)
recall = truePositives / (truePositives + falseNegatives)
precisionValues[counter] = precision
recallValues[counter] = recall
counter += 1
if doPlot:
plt.figure('Precision vs Recall')
plt.plot(precisionValues, recallValues, 'o')
plt.xlabel("Precision")
plt.ylabel("Recall")
return (precisionValues, recallValues)
def threshold(frame, threshold = 0.5, normalized_threshold = True, threshold_type = cv2.THRESH_BINARY, debug = False):
"""
tresholding an image: the input type has to be either np.uint8 or np.float32
returns a uint8 image
"""
fmin = np.double( np.nanmin(frame) )
fmax = np.double( np.nanmax(frame) )
if debug:
print "fmin: ", fmin
print "fmax: ", fmax
floatframe = ( (1.0 - 0.0) / (fmax - fmin) * (np.double(frame) - fmin) ).astype(np.float32)
if debug:
print "floatframe: ", floatframe.dtype
print "min: ", np.nanmin(floatframe)
print "max: ", np.nanmax(floatframe)
if not normalized_threshold:
threshold = 1.0 / (fmax - fmin) * (threshold - fmin)
if debug:
print "normalized threshold: ", threshold
retval, t = cv2.threshold(floatframe, thresh = threshold, maxval = 255, type = threshold_type)
return t
def get_fn_likelihood(residuals, sigma_w, sigma_r, gamma=1.0):
like=0.0
# Arrays of zeros to be passed to the likelihood function
aa,bb,M = Wavelets.getDWT(residuals)
# Calculate the g(gamma) factor used in Carter & Winn...
if(gamma==1.0):
g_gamma=1.0/(2.0*np.log(2.0)) # (value assuming gamma=1)
else:
g_gamma=(2.0)-(2.0)**gamma
# log-Likelihood of the aproximation coefficients
sigmasq_S=(sigma_r**2)*g_gamma+(sigma_w)**2
tau_a = 1.0/sigmasq_S
like += normal_like( bb[0], 0.0 , tau_a )
k=long(0)
SS=range(M)
for ii in SS:
# log-Likelihood of the detail coefficients with m=i...
if(ii==0):
sigmasq_W=(sigma_r**2)*(2.0**(-gamma*np.double(1.0)))+(sigma_w)**2
tau=1.0/sigmasq_W
like += normal_like( bb[1], 0.0, tau )
else:
sigmasq_W=(sigma_r**2)*(2.0**(-gamma*np.double(ii+1)))+(sigma_w)**2
tau=1.0/sigmasq_W
for j in range(2**ii):
like += normal_like( aa[k], 0.0 , tau )
k=k+1
return like
def gen_backbone(CDFmat, alpha=0.05, S=None):
""" Returns a backbone network given a CDF matrix and significance value
and an optional similarity matrix for weights.
Finds all entries in the matrix s.t. 1-alpha < CDF matrix entry
"""
# do some input checking
if type(CDFmat) != N.matrix:
raise TypeError('gen_backbone: Invalid input type -- must be numpy.matrix')
# now find the size of this matrix
sz = CDFmat.shape
# check for correct dimensions
if sz[0] != sz[1]:
raise ValueError('gen_ecdf_matrix: Invalid input -- matrix is not square')
# now make sure the matrix is of doubles
CDFmat = N.double(CDFmat)
# convenience renaming
n = sz[0]
# now we need to find the entries
BBout = N.double(CDFmat > 1-alpha)
# add weights if desired
print type(S)
if S != None:
print '######## Adding weights to matrix...'
BBout = N.multiply(BBout,S)
else:
print '######## NO weights specified...'
return BBout
def split_numbers(number):
"""
Split high precision number(s) into doubles.
TODO: Consider the option of using a third number to specify shift.
Parameters
----------
number: long double
The input high precision number which is to be split
Returns
-------
number_I: double
First part of high precision number
number_F: double
Second part of high precision number
"""
if isinstance(number, collections.Iterable):
mods = [math.modf(n) for n in number]
number_F = [f for f,_ in mods]
number_I = [i for _,i in mods]
else:
number_F, number_I = math.modf(number)
return np.double(number_I), np.double(number_F)
def get_mappings(mesh,
rotations,
is_shift=np.zeros(3, dtype='intc'),
is_time_reversal=True,
qpoints=np.double([])):
"""
Return k-point map to the irreducible k-points and k-point grid points .
The symmetry is searched from the input rotation matrices in real space.
is_shift=[0, 0, 0] gives Gamma center mesh and the values 1 give
half mesh distance shifts.
"""
mapping = np.zeros(np.prod(mesh), dtype='intc')
rot_mapping = np.zeros(np.prod(mesh), dtype="intc")
mesh_points = np.zeros((np.prod(mesh), 3), dtype='intc')
qpoints = np.double(qpoints).copy()
if qpoints.shape == (3,):
qpoints = np.double([qpoints])
spg.stabilized_reciprocal_mesh(mesh_points,
mapping,
rot_mapping,
np.intc(mesh).copy(),
np.intc(is_shift),
is_time_reversal * 1,
np.intc(rotations).copy(),
np.double(qpoints))
return mapping, rot_mapping
def best_match(dominf, key, rng):
"""compute rms, pos, s and o such that |i'*s + o - rng| is minimal
where i' is a submatrix of domain that has shape = rng.shape and
upper left corner at pos .
"""
n, m = rng.shape
N = np.double(n) * np.double(m)
dom = dominf.doms[key]
c = np.sum(rng ** 2) / N
mean = np.sum(rng) / N
b = all_dot(rng, dom) / N
dom_a = dominf.get_mean(key, rng.shape[0])
dom_mean = dominf.get_meansq(key, rng.shape[0])
denom = (dom_a - dom_mean ** 2)
denom[abs(denom) < allmost0] = allmost0
s = (b - mean * dom_mean) / denom
o = (mean * dom_a - b * dom_mean) / denom
# mean square error
ms = (dom_a * s ** 2 + o ** 2 + c +
2 * s * o * dom_mean - 2 * s * b - 2 * o * mean)
# penalty for too big s:
ms[np.logical_not(abs(s) <= limit_s)] = np.inf
# best domain
index = np.unravel_index(np.argmin(ms), ms.shape)
rms = np.sqrt(np.max((0.0, ms[index])))
return (rms, index, s[index], o[index])
def jaccard_distance(u, v):
"""return jaccard distance"""
u = numpy.asarray(u)
v = numpy.asarray(v)
return (numpy.double(numpy.bitwise_and((u != v),
numpy.bitwise_or(u != 0, v != 0)).sum())
/ numpy.double(numpy.bitwise_or(u != 0, v != 0).sum()))
def load_snotel_list(snotel_list_file):
"""Load station name/location information from a text file
example file:
ID Site Name Latitude Longitude Elevation
05J04S PHANTOM VALLEY 40.38333 -105.83333 2752.4
05J05S WILD BASIN 40.20000 -105.60000 2913.9
05J06S DEADMAN HILL 40.80000 -105.76667 3115.1
"""
with open(snotel_list_file) as f:
for i in range(1):l=f.next()
outputlist=[]
for l in f:
try:
line=l.split()
name=line[0].strip()
lat=np.double(line[-3])
lon=np.double(line[-2])
outputlist.append(Bunch(name=name,lat=lat,lon=lon))
except Exception as e:
print(e)
print(l)
return outputlist
def handle_read(self):
try:
data = self.recv(1024)
if not data:
return
datalist = data.split(',')
timestamp = datalist[2]
for i in range( (len(datalist) - 3 )/4):
if str(self.order) == datalist[ 3 + i*4]:
x = np.double(datalist[4 + i*4])
y = np.double(datalist[5 + i*4])
""" create and pubish tranmerc """
gps_message = gpgga_tranmerc()
gps_message.time = timestamp
gps_message.northing = y/100
gps_message.easting = x/100
gps_message.fix = np.uint8(4)
gps_message.sat = np.uint8(6)
gps_message.hdop = np.double(1.0)
self.gps_pub.publish(gps_message)
except:
pass
def price(self, Sl, Su, K, scheme=PenaltyRannacherScheme, **kwargs):
"""
Price the payoff for prices in range [Sl, Su], and K increments, using
the given FD scheme, and possibility using exponential increments
(zspace) for the price range.
"""
Sl = np.double(Sl)
Su = np.double(Su)
K = np.int(K)
P = FDEModel.Value(self.V.T, self.N, Sl, Su, K)
S = P.S
ds = P.S[1] - P.S[0]
Sl = P.S * self.dS.fde_l()
# Terminal stock price and derivative value
P.C[-1] = C = self.V.coupon(self.V.T)
P.V[-1] = V = self.V.terminal(S) + C
P.I[-1] = I = V
# Discount price backwards
t = P.t
scheme = scheme(self.dS, self.dt, ds, S, **kwargs)
for i in range(self.N - 1, -1, -1):
# Discount previous derivative value
P.C[i] = C = self.V.coupon(t[i])
P.X[i] = X = self.V.default(t[i], Sl)
V, I = scheme(t[i], V, X, C, self.V.transient)
P.V[i] = V
P.I[i] = I
return P
def CG(Afun, B, x0, par=None, callback=None):
"""
Conjugate gradients solver.
Parameters
----------
Afun : Matrix, LinOper, or numpy.array of shape (n, n)
it stores the matrix data of linear system and provides a matrix by
vector multiplication
B : VecTri or numpy.array of shape (n,)
it stores a right-hand side of linear system
x0 : VecTri or numpy.array of shape (n,)
initial approximation of solution of linear system
par : dict
parameters of the method
callback :
Returns
-------
x : VecTri or numpy.array of shape (n,)
resulting unknown vector
res : dict
results
"""
if par is None:
par = dict()
if 'tol' not in list(par.keys()):
par['tol'] = 1e-6
if 'maxiter' not in list(par.keys()):
par['maxiter'] = int(1e3)
scal=get_scal(B, par)
res = dict()
xCG = x0
Ax = Afun(x0)
R = B - Ax
P = R
rr = scal(R,R)
res['kit'] = 0
res['norm_res'] = np.double(rr)**0.5 # /np.norm(E_N)
norm_res_log = []
norm_res_log.append(res['norm_res'])
while (res['norm_res'] > par['tol']) and (res['kit'] < par['maxiter']):
res['kit'] += 1 # number of iterations
AP = Afun(P)
alp = float(rr/scal(P,AP))
xCG = xCG + alp*P
R = R - alp*AP
rrnext = scal(R,R)
bet = rrnext/rr
rr = rrnext
P = R + bet*P
res['norm_res'] = np.double(rr)**0.5
norm_res_log.append(res['norm_res'])
if callback is not None:
callback(xCG)
if res['kit'] == 0:
res['norm_res'] = 0
return xCG, res
def apply_bc_A(self):
"""
account for boundary conditions at x = 0 & x = L
- additionally, the stiffness matrix [K] is modified to maintain
symmetry => positive definiteness
"""
# apply BC at x = 0
if self.BC_type[0] != 0:
# essential BC
self.A[0, 0:2] = np.array([1, 0]) # modifies first equation
self.A[1, 0] = 0 # modification to mainatin symmetry
elif self.BC_type[0] == 0:
# natural BC, Flux (Q*)
n = -1.0 # unit outward normal
dT = self.BC_0 / (self.kA * n)
# modifies [K] first equation
self.A[0, 0:2] = self.kA/np.double(self.h)**2 * np.array([1, -1])
# apply BC at x = L
if self.BC_type[1] != 0:
# essential BC
self.A[-1][-2:] = np.array([0, 1]) # modifies last equation
self.A[-2][-1] = 0 # modification to mainatin symmetry
elif self.BC_type[1] == 0:
# natural BC, flux (Q*)
n = 1.0 # unit outward normal
dT = self.BC_L / (self.kA * n)
self.A[-1][-2:] = self.kA/np.double(self.h)**2 *\
np.array([-1, 1]) # modifies last equation
return self.A
def matchesWithReference(data, expected_data, data_name, precision): # ok if exactly equal (including strings or lists of strings) try : if expected_data == data: return True except: pass # If numbers: try: double_data = np.array(np.double(data), ndmin=1) if precision is not None: error = np.abs( double_data-np.array(np.double(expected_data), ndmin=1) ) max_error_location = np.unravel_index(np.argmax(error), error.shape) max_error = error[max_error_location] if max_error < precision: return True print( "Reference quantity '"+data_name+"' does not match the data (required precision "+str(precision)+")") print( "Max error = "+str(max_error)+" at index "+str(max_error_location)) else: if np.all(double_data == np.double(expected_data)): return True print( "Reference quantity '"+data_name+"' does not match the data") except Exception as e: print( "Reference quantity '"+data_name+"': unable to compare to data") print( e ) return False
def genARbasis(numberFrequencyBins, sizeOfFourier, Fs, \
formantsRange=None, \
bwRange=None, \
numberOfAmpsPerPole=5, \
numberOfFreqPerPole=60, \
maxF0 = 1000.0):
if formantsRange is None:
formantsRange = {}
formantsRange[0] = [80.0, 1400.0]
formantsRange[1] = [300.0, 4000.0]
formantsRange[2] = [1100.0, 6000.0]
formantsRange[3] = [6100.0, 20000.0]
numberOfFormants = len(formantsRange)
if bwRange is None:
bwMin = maxF0
bwMax = np.maximum(0.1 * Fs, bwMin)
bwRange = np.arange(numberOfAmpsPerPole, dtype=np.double) \
* (bwMax - bwMin) / \
np.double(numberOfAmpsPerPole-1.0) + \
bwMin
freqRanges = np.zeros([numberOfFormants, numberOfFreqPerPole])
for n in range(numberOfFormants):
freqRanges[n] = np.arange(numberOfFreqPerPole) \
* (formantsRange[n][1] - formantsRange[n][0]) / \
np.double(numberOfFreqPerPole-1.0) + \
formantsRange[n][0]
totNbElements = numberOfFreqPerPole * \
numberOfFormants * numberOfAmpsPerPole
poleAmp = np.zeros(totNbElements)
poleFrq = np.zeros(totNbElements)
WGAMMA = np.zeros([numberFrequencyBins, totNbElements])
cplxExp = np.exp(-1j * 2.0 * np.pi * \
np.arange(numberFrequencyBins) / \
np.double(sizeOfFourier))
for n in range(numberOfFormants):
for w in range(numberOfFreqPerPole):
for a in range(numberOfAmpsPerPole):
elementNb = n * numberOfAmpsPerPole * numberOfFreqPerPole + \
w * numberOfAmpsPerPole + \
a
poleAmp[elementNb] = np.exp(-bwRange[a] / np.double(Fs))
poleFrq[elementNb] = freqRanges[n][w]
## pole = poleAmp[elementNb] * \
## np.exp(1j * 2.0 * np.pi * \
## poleFrq[elementNb] / np.double(Fs))
WGAMMA[:,elementNb] = 1 / \
np.abs(1 - \
2.0 * \
poleAmp[elementNb] * \
np.cos(2.0 * np.pi * poleFrq[elementNb] / \
np.double(Fs)) * cplxExp +
(poleAmp[elementNb] * cplxExp) ** 2\
) ** 2
return bwRange, freqRanges, poleAmp, poleFrq, WGAMMA
def data2AB(data, x0=None):
n = data.shape[0]
T = data.shape[1]
YY = np.dot(data[:, 1:], data[:, 1:].T)
XX = np.dot(data[:, :-1], data[:, :-1].T)
YX = np.dot(data[:, 1:], data[:, :-1].T)
model = VAR(data.T)
r = model.fit(1)
A = r.coefs[0,:,:]
# A = np.ones((n,n))
B = np.ones((n, n))
np.fill_diagonal(B, 0)
B[np.triu_indices(n)] = 0
K = np.int(scipy.sum(abs(B)))#abs(A)+abs(B)))
a_idx = np.where(A != 0)
b_idx = np.where(B != 0)
np.fill_diagonal(B, 1)
try:
s = x0.shape
x = x0
except AttributeError:
x = np.r_[A.flatten(), 0.1*scipy.randn(K)]
o = optimize.fmin_bfgs(nllf2, x,
args=(np.double(A), np.double(B),
YY, XX, YX, T, a_idx, b_idx),
gtol=1e-12, maxiter=500,
disp=False, full_output=True)
A, B = x2M(o[0], np.double(A), np.double(B), a_idx, b_idx)
B = B+B.T
return A, B
def calculate_gh_at_sigma_and_temp(self):
import anharmonic._phono3py as phono3c
if self._is_precondition:
out = self._collision_out[self._isigma, :, self._itemp]
out_reverse = np.where(self._frequencies>self._cutoff_frequency, 1 / out, 0)
self._z[self._isigma, :, self._itemp] = self._r[self._isigma, :, self._itemp] * out_reverse[..., np.newaxis]
else:
self._z[self._isigma, :, self._itemp] = self._r[self._isigma, :, self._itemp]
self._z[:, np.where(np.any(np.abs(self._qpoints) > self._pp._criteria, axis=1))] = 0
zr0 = np.zeros(3, dtype="double")
phono3c.phonon_3_multiply_dvector_gb3_dvector_gb3(zr0,
self._z_prev[self._isigma,:,self._itemp].copy(),
self._r_prev[self._isigma,:,self._itemp].copy(),
np.intc(self._irr_index_mapping).copy(),
np.intc(self._kpoint_operations[self._rot_mappings]),
np.double(np.linalg.inv(self._primitive.get_cell())).copy())
#Flexibly preconditioned CG: r(i+1)-r(i) instead of r(i+1)
r = self._r[self._isigma,:,self._itemp] - self._r_prev[self._isigma,:,self._itemp]
# r = self._r[self._isigma,:,self._itemp]
zr1 = np.zeros(3, dtype="double")
phono3c.phonon_3_multiply_dvector_gb3_dvector_gb3(zr1,
self._z[self._isigma,:,self._itemp].copy(),
r.copy(),
np.intc(self._irr_index_mapping).copy(),
np.intc(self._kpoint_operations[self._rot_mappings]),
np.double(np.linalg.inv(self._primitive.get_cell())).copy())
zr1_over_zr0 = np.where(np.abs(zr0>0), zr1/zr0, 0)
self._p[self._isigma,:,self._itemp] = self._z[self._isigma,:,self._itemp] +\
zr1_over_zr0 * self._p_prev[self._isigma,:,self._itemp]
self._p[:, np.where(np.any(np.abs(self._qpoints) > self._pp._criteria, axis=1))] = 0
def getCDF(temp_response):
"""
This function calculates voting score of response magnitude.
The score is calculated as the rank of sorted magnitude.
Parameters
----------
temp_response: numpy array
Input batch response matrix
"""
shape = temp_response.shape
nrow = shape[0]
ncol = shape[1]
len_hist = nrow*ncol
resp_arr = temp_response.reshape((len_hist,))
sorted_arr = np.sort(resp_arr)
cdf = np.zeros(shape, dtype=np.double)
for r in range(nrow):
for c in range(ncol):
index = np.where(sorted_arr==temp_response[r,c])
cdf[r,c] = np.double(index[0][0])/np.double(len_hist)
return cdf
def train_perceptron(self, n_epochs):
"""Trains the parser by running the averaged perceptron algorithm for n_epochs."""
self.weights = np.zeros(self.features.n_feats)
total = np.zeros(self.features.n_feats)
for epoch in range(n_epochs):
print "Epoch {0}".format(epoch+1)
n_mistakes = 0
n_tokens = 0
n_instances = 0
for instance in self.reader.train_instances:
feats = self.features.create_features(instance)
scores = self.features.compute_scores(feats, self.weights)
if self.projective:
heads_pred = self.decoder.parse_proj(scores)
else:
heads_pred = self.decoder.parse_nonproj(scores)
for m in range(np.size(heads_pred)):
if heads_pred[m] != instance.heads[m]: # mistake
for f in feats[instance.heads[m]][m]:
if f < 0:
continue
self.weights[f] += 1.0
for f in feats[heads_pred[m]][m]:
if f < 0:
continue
self.weights[f] -= 1.0
n_mistakes += 1
n_tokens += 1
n_instances += 1
print "Training accuracy: {0}".format(np.double(n_tokens-n_mistakes) / np.double(n_tokens))
total += self.weights
self.weights = total / np.double(n_epochs)
def pattern_distance_jaccard(a, b):
"""
Computes a distance measure for two binary patterns based on their
Jaccard-Needham distance, defined as
.. math::
d_J(a,b) = 1 - J(a,b) = \\frac{|a \\cup b| - |a \\cap b|}{|a \\cup b|}.
The similarity measure takes values on the closed interval [0, 1],
where a value of 1 is attained for disjoint, i.e. maximally dissimilar
patterns a and b and a value of 0 for the case of :math:`a=b`.
Parameters
----------
a : list or array, int or bool
Input pattern
b : list or array, int or bool
Input pattern
Returns
-------
dist : double
Jaccard distance between `a` and `b`.
"""
# Note: code taken from scipy. Duplicated as only numpy references wanted for base functionality
a = np.atleast_1d(a).astype(bool)
b = np.atleast_1d(b).astype(bool)
dist = (np.double(np.bitwise_and((a != b), np.bitwise_or(a != 0, b != 0)).sum())
/ np.double(np.bitwise_or(a != 0, b != 0).sum()))
return dist
def __init__(self, mean, stddev, minval, maxval):
self.memoized_moments = [1.0] # 0th moment
self.mean = np.double(mean)
self.stddev = np.double(stddev)
# NOTE(ringwalt): The formula doesn't handle infinite values.
self.minval = np.double(max(-10, minval))
self.maxval = np.double(min(10, maxval))
def env_reset():
global thita
global g_drl_th_val
global g_eql_th_val
global g_rnd_th_val
global thita_3db
global P_max_macro
global MC_power
global slots_slots
global sub
global A_m
global R_m
global num_U
global num_M
global M_x
global M_y
global M_cell_beam
global M_cell_associated_user_id
global U_x
global U_y
global U_association_macro
global U_neighbor
global U_neighbor_sector
global M_cell_region_x
global M_cell_region_y
global U_macro_distance
global U_macro_power
global M_cell_txblock_power
global P_macro_subband
global U_received_power_subband
global U_throughput_subband
global U_SINR_subband
global U_CQI_subband
global g_action
global g_cntr
global current_throughput
global previous_throughput
global g_counter
global k_m_n
global v_k_old
global u_k_old
global w_k_old
global v_k_new
global u_k_new
global w_k_new
k_m_n = []
v_k_old = []
u_k_old = []
w_k_old = []
v_k_new = []
u_k_new = []
w_k_new = []
g_counter = 0
current_throughput = []
previous_throughput = 0
g_drl_th_val = 0
g_eql_th_val = 0
g_rnd_th_val = 0
M_cell_beam = []
M_cell_associated_user_id = []
M_cell_region_x = []
M_cell_region_y = []
M_cell_txblock_power = []
U_x = []
U_y = []
U_association_macro = []
U_neighbor = []
U_neighbor_sector = []
U_macro_distance = []
U_macro_power = []
U_throughputt = []
P_macro_subband = []
U_received_power_subband = []
U_throughput_subband = []
U_SINR_subband = []
U_CQI_subband = []
itr = 0
for i in range(len(M_x)):
M_cell_beam.append(random_beam())
M_cell_associated_user_id.append([])
for j in range(num_U):
M_cell_associated_user_id[i].extend([itr + 1])
U_association_macro.append(i + 1)
itr = itr + 1
# [x, y] = random_allocation(M_x[i], M_y[i], 360, 0.2 * R_m, R_m, num_U)
# U_x.extend(x)
# U_y.extend(y)
# ***********************************************************************************************************************
# g_data_x.append(U_x)
# g_data_y.append(U_y)
if g_cntr < g_data_len:
U_x = data_set['arr_0'][g_cntr].tolist()
# print(U_x)
U_y = data_set['arr_1'][g_cntr].tolist()
else:
for i in range(len(M_x)):
[x, y] = random_allocation(M_x[i], M_y[i], 360, 0.2 * R_m, R_m,
num_U)
U_x.extend(x)
U_y.extend(y)
g_data_x.append(U_x)
g_data_y.append(U_y)
g_cntr = g_cntr + 1
print(g_cntr)
# ***********************************************************************************************************************
for i in range(len(U_x)):
U_neighbor.append([])
U_neighbor_sector.append([])
for i in range(len(M_x)):
M_cell_region_x.append([[], [], []])
M_cell_region_y.append([[], [], []])
for j in range(3):
[a, b] = find_points_in_angle(M_x[i], M_y[i], U_x, U_y,
M_cell_beam[i][j], i + 1, j + 1)
M_cell_region_x[i][j].extend(a)
M_cell_region_y[i][j].extend(b)
thita = np.zeros((len(U_x), len(M_x)))
for i in range(len(U_x)):
U_macro_distance.append([])
U_macro_power.append([])
for j in range(len(M_x)):
U_macro_distance[i].append(find_dis(U_x[i], U_y[i], M_x[j],
M_y[j]))
angl = M_cell_beam[j][U_neighbor_sector[i][j] - 1]
thita[i][j] = find_angl(M_x[j], M_y[j], U_x[i], U_y[i], angl)
a = transmit_power(thita[i][j], thita_3db, A_m,
power_watt_dbm(P_max_macro))
U_macro_power[i].append(pw_m_hata(a, U_macro_distance[i][j]))
for i in range(len(M_x)):
M_cell_txblock_power.append(
profile_power(sub, slots_slots, MC_power, P_max_macro))
P_macro_subband.append(M_cell_txblock_power[i])
generate_power_matrix_macro()
throughput()
temp_var = []
for i in range(num_M * num_U):
temp_var.extend(U_CQI_subband[i])
temp_var.append(
np.double(
U_macro_distance[i][U_association_macro[i] - 1] / R_m >= 0.5))
return np.array(temp_var)
from Smilei import *
import numpy as np
import matplotlib.pyplot as plt
from scipy.special import erf as erf
path = "temperature_isotropization1"
sim = Smilei(path)
density_electron = np.double(sim.namelist.Species["electron1"].charge_density)
coulomb_log = np.double(sim.namelist.Collisions[0].coulomb_log)
dt = np.double(sim.namelist.Main.timestep) / (2 * np.pi)
re_ = 2.8179403267e-15 # meters
wavelength = 1e-6 # meters
c = 3e8
coeff = (2. * np.pi / wavelength)**2 * re_ * c / (2. * np.sqrt(np.pi))
times = sim.ParticleDiagnostic(diagNumber=0).getAvailableTimesteps()
electrons0 = sim.ParticleDiagnostic(0, slice={"x": "all"}).get()
vx = electrons0["vx"]
electrons1 = sim.ParticleDiagnostic(1, slice={"x": "all"}).get()
vy = electrons1["vy"]
electrons2 = sim.ParticleDiagnostic(2, slice={"x": "all"}).get()
vz = electrons2["vz"]
e_Tpar = np.zeros(len(times))
e_Tperp = np.zeros(len(times))
fig = None
#fig = plt.figure(1)
if fig: fig.clf()
# the data.
# print('\nVisualizing Neural Network... \n')
#
# displayData(Theta1[:,1:])
# input('\nProgram paused. Press enter to continue.\n')
## ================= Part 10: Implement Predict =================
# After training the neural network, we would like to use it to predict
# the labels. You will now implement the "predict" function to use the
# neural network to predict the labels of the training set. This lets
# you compute the training set accuracy.
pred = (predict(Theta1, Theta2, X) + 1) % 10
print('\nTraining Set Accuracy: %f\n' %
(np.mean(np.double(pred == y.T)) * 100))
# rp = np.random.permutation(m)
#
# for i in range(m):
# # Display
# print('\nDisplaying Example Image\n')
# t = np.array([X[rp[i]]])
# displayData(t)
#
# pred = predict(Theta1, Theta2, t)
# print('\nNeural Network Prediction: %d (digit %d)\n'%(pred, (pred+1)%10))
#
# input('Program paused. Press enter to continue.\n')
def preprocess():
""" Input:
Although this function doesn't have any input, you are required to load
the MNIST data set from file 'mnist_all.mat'.
Output:
train_data: matrix of training set. Each row of train_data contains
feature vector of a image
train_label: vector of label corresponding to each image in the training
set
validation_data: matrix of training set. Each row of validation_data
contains feature vector of a image
validation_label: vector of label corresponding to each image in the
training set
test_data: matrix of training set. Each row of test_data contains
feature vector of a image
test_label: vector of label corresponding to each image in the testing
set
Some suggestions for preprocessing step:
- divide the original data set to training, validation and testing set
with corresponding labels
- convert original data set from integer to double by using double()
function
- normalize the data to [0, 1]
- feature selection"""
mat = loadmat('D:\Machine Learning\mnist_all.mat'
) #loads the MAT object as a Dictionary
#Pick a reasonable size for validation data
train_data = np.array([])
test_data = np.array([])
validation_data = np.array([])
validation_label = np.array([])
test_data = np.array([])
test_label = np.array([])
validation_label = np.zeros(shape=(1000, 1))
trainx = mat.get('train0')
test_data = mat.get('test0')
a = range(trainx.shape[0])
aperm = np.random.permutation(a)
validation_data = trainx[aperm[0:1000], :]
train_data = trainx[aperm[1000:], :]
train_label = np.zeros(shape=((trainx.shape[0] - 1000), 1))
test_label = np.zeros(shape=(test_data.shape[0], 1))
for i in range(1, 10):
trainx = mat.get('train' + str(i))
testx = mat.get('test' + str(i))
a = range(trainx.shape[0])
aperm = np.random.permutation(a)
validation_data = np.concatenate(
((validation_data, trainx[aperm[0:1000], :])), axis=0)
b = np.zeros(shape=(1000, 1))
b[:] = i
validation_label = np.concatenate((validation_label, b), axis=0)
train_data = np.concatenate((train_data, trainx[aperm[1000:], :]),
axis=0)
c = np.zeros(shape=((trainx.shape[0] - 1000), 1))
c[:] = i
train_label = np.concatenate((train_label, c), axis=0)
d = np.zeros(shape=((testx.shape[0]), 1))
d[:] = i
test_label = np.concatenate((test_label, d), axis=0)
test_data = np.concatenate((test_data, testx), axis=0)
train_data = np.double(train_data)
test_data = np.double(test_data)
validation_data = np.double(validation_data)
train_data /= 255.0
test_data /= 255.0
validation_data /= 255.0
return train_data, train_label, validation_data, validation_label, test_data, test_label
def stFeatureExtraction(signal, Fs, Win, Step):
"""
This function implements the shor-term windowing process. For each short-term window a set of features is extracted.
This results to a sequence of feature vectors, stored in a numpy matrix.
ARGUMENTS
signal: the input signal samples
Fs: the sampling freq (in Hz)
Win: the short-term window size (in samples)
Step: the short-term window step (in samples)
RETURNS
stFeatures: a numpy array (numOfFeatures x numOfShortTermWindows)
"""
Win = int(Win)
Step = int(Step)
# Signal normalization
signal = numpy.double(signal)
signal = signal / (2.0**15)
DC = signal.mean()
MAX = (numpy.abs(signal)).max()
signal = (signal - DC) / MAX
N = len(signal) # total number of samples
curPos = 0
countFrames = 0
nFFT = Win / 2
[fbank, freqs] = mfccInitFilterBanks(
Fs, nFFT
) # compute the triangular filter banks used in the mfcc calculation
nChroma, nFreqsPerChroma = stChromaFeaturesInit(nFFT, Fs)
numOfTimeSpectralFeatures = 8
numOfHarmonicFeatures = 0
nceps = 13
numOfChromaFeatures = 13
totalNumOfFeatures = numOfTimeSpectralFeatures + nceps + numOfHarmonicFeatures + numOfChromaFeatures
# totalNumOfFeatures = numOfTimeSpectralFeatures + nceps + numOfHarmonicFeatures
stFeatures = numpy.array([], dtype=numpy.float64)
while (curPos + Win - 1 <
N): # for each short-term window until the end of signal
countFrames += 1
x = signal[curPos:curPos + Win] # get current window
curPos = curPos + Step # update window position
X = abs(fft(x)) # get fft magnitude
X = X[0:nFFT] # normalize fft
X = X / len(X)
if countFrames == 1:
Xprev = X.copy() # keep previous fft mag (used in spectral flux)
curFV = numpy.zeros((totalNumOfFeatures, 1))
curFV[0] = stZCR(x) # zero crossing rate
curFV[1] = stEnergy(x) # short-term energy
curFV[2] = stEnergyEntropy(x) # short-term entropy of energy
[curFV[3], curFV[4]
] = stSpectralCentroidAndSpread(X, Fs) # spectral centroid and spread
curFV[5] = stSpectralEntropy(X) # spectral entropy
curFV[6] = stSpectralFlux(X, Xprev) # spectral flux
curFV[7] = stSpectralRollOff(X, 0.90, Fs) # spectral rolloff
curFV[numOfTimeSpectralFeatures:numOfTimeSpectralFeatures + nceps,
0] = stMFCC(X, fbank, nceps).copy() # MFCCs
chromaNames, chromaF = stChromaFeatures(X, Fs, nChroma,
nFreqsPerChroma)
curFV[numOfTimeSpectralFeatures + nceps:numOfTimeSpectralFeatures +
nceps + numOfChromaFeatures - 1] = chromaF
curFV[numOfTimeSpectralFeatures + nceps + numOfChromaFeatures -
1] = chromaF.std()
# curFV[numOfTimeSpectralFeatures+nceps+numOfChromaFeatures-1] = numpy.nonzero( chromaF > 2.0 * chromaF.mean() )[0].shape[0]
# temp = numpy.sort(chromaF[:,0])
# curFV[numOfTimeSpectralFeatures+numOfChromaFeatures] = temp[-1] / numpy.mean(temp[0:5])
# temp = numpy.sort(chromaF[:,0])
# if countFrames==10 or countFrames==30:
# A = int(temp[-1] / numpy.mean(temp[0:5]))/10
# for a in range(A):
# print("|"),
# print
# if countFrames==20:
# print numpy.nonzero(chromaF > 5*chromaF.mean())[0].shape[0]
#HR, curFV[numOfTimeSpectralFeatures+nceps] = stHarmonic(x, Fs)
# curFV[numOfTimeSpectralFeatures+nceps+1] = freq_from_autocorr(x, Fs)
if countFrames == 1:
stFeatures = curFV # initialize feature matrix (if first frame)
else:
stFeatures = numpy.concatenate((stFeatures, curFV),
1) # update feature matrix
Xprev = X.copy()
return numpy.array(stFeatures)
def stSpectogram(signal, Fs, Win, Step, PLOT=False):
"""
Short-term FFT mag for spectogram estimation:
Returns:
a numpy array (nFFT x numOfShortTermWindows)
ARGUMENTS:
signal: the input signal samples
Fs: the sampling freq (in Hz)
Win: the short-term window size (in samples)
Step: the short-term window step (in samples)
PLOT: flag, 1 if results are to be ploted
RETURNS:
"""
Win = int(Win)
Step = int(Step)
signal = numpy.double(signal)
signal = signal / (2.0**15)
DC = signal.mean()
MAX = (numpy.abs(signal)).max()
signal = (signal - DC) / (MAX - DC)
N = len(signal) # total number of signals
curPos = 0
countFrames = 0
nfft = int(Win / 2)
specgram = numpy.array([], dtype=numpy.float64)
while (curPos + Win - 1 < N):
countFrames += 1
x = signal[curPos:curPos + Win]
curPos = curPos + Step
X = abs(fft(x))
X = X[0:nfft]
X = X / len(X)
if countFrames == 1:
specgram = X**2
else:
specgram = numpy.vstack((specgram, X))
FreqAxis = [((f + 1) * Fs) / (2 * nfft) for f in range(specgram.shape[1])]
TimeAxis = [(t * Step) / Fs for t in range(specgram.shape[0])]
if (PLOT):
fig, ax = plt.subplots()
imgplot = plt.imshow(specgram.transpose()[::-1, :])
Fstep = int(nfft / 5.0)
FreqTicks = range(0, int(nfft) + Fstep, Fstep)
FreqTicksLabels = [
str(Fs / 2 - int((f * Fs) / (2 * nfft))) for f in FreqTicks
]
ax.set_yticks(FreqTicks)
ax.set_yticklabels(FreqTicksLabels)
TStep = countFrames / 3
TimeTicks = range(0, countFrames, TStep)
TimeTicksLabels = ['%.2f' % (float(t * Step) / Fs) for t in TimeTicks]
ax.set_xticks(TimeTicks)
ax.set_xticklabels(TimeTicksLabels)
ax.set_xlabel('time (secs)')
ax.set_ylabel('freq (Hz)')
imgplot.set_cmap('jet')
plt.colorbar()
plt.show()
return (specgram, TimeAxis, FreqAxis)
def stChromagram(signal, Fs, Win, Step, PLOT=False):
"""
Short-term FFT mag for spectogram estimation:
Returns:
a numpy array (nFFT x numOfShortTermWindows)
ARGUMENTS:
signal: the input signal samples
Fs: the sampling freq (in Hz)
Win: the short-term window size (in samples)
Step: the short-term window step (in samples)
PLOT: flag, 1 if results are to be ploted
RETURNS:
"""
Win = int(Win)
Step = int(Step)
signal = numpy.double(signal)
signal = signal / (2.0**15)
DC = signal.mean()
MAX = (numpy.abs(signal)).max()
signal = (signal - DC) / (MAX - DC)
N = len(signal) # total number of signals
curPos = 0
countFrames = 0
nfft = int(Win / 2)
nChroma, nFreqsPerChroma = stChromaFeaturesInit(nfft, Fs)
chromaGram = numpy.array([], dtype=numpy.float64)
while (curPos + Win - 1 < N):
countFrames += 1
x = signal[curPos:curPos + Win]
curPos = curPos + Step
X = abs(fft(x))
X = X[0:nfft]
X = X / len(X)
chromaNames, C = stChromaFeatures(X, Fs, nChroma, nFreqsPerChroma)
C = C[:, 0]
if countFrames == 1:
chromaGram = C.T
else:
chromaGram = numpy.vstack((chromaGram, C.T))
FreqAxis = chromaNames
TimeAxis = [(t * Step) / Fs for t in range(chromaGram.shape[0])]
if (PLOT):
fig, ax = plt.subplots()
chromaGramToPlot = chromaGram.transpose()[::-1, :]
Ratio = chromaGramToPlot.shape[1] / (3 * chromaGramToPlot.shape[0])
chromaGramToPlot = numpy.repeat(chromaGramToPlot, Ratio, axis=0)
imgplot = plt.imshow(chromaGramToPlot)
Fstep = int(nfft / 5.0)
# FreqTicks = range(0, int(nfft) + Fstep, Fstep)
# FreqTicksLabels = [str(Fs/2-int((f*Fs) / (2*nfft))) for f in FreqTicks]
ax.set_yticks(range(Ratio / 2, len(FreqAxis) * Ratio, Ratio))
ax.set_yticklabels(FreqAxis[::-1])
TStep = countFrames / 3
TimeTicks = range(0, countFrames, TStep)
TimeTicksLabels = ['%.2f' % (float(t * Step) / Fs) for t in TimeTicks]
ax.set_xticks(TimeTicks)
ax.set_xticklabels(TimeTicksLabels)
ax.set_xlabel('time (secs)')
imgplot.set_cmap('jet')
plt.colorbar()
plt.show()
return (chromaGram, TimeAxis, FreqAxis)
def run_tracker(p, net_type, net_base_path, model_name):
"""
run tracker, return bounding result and speed
"""
# load model # IF QUANTISED
net = SiamNet(net_type)
net.load_state_dict(torch.load(
os.path.join("/home/vision/orig_dp/siamtrackopt/FINN", net_base_path, model_name))['state_dict'])
net = net.to(device)
# evaluation mode
net.eval()
# load sequence
img_list, target_position, target_size, gt_list = load_sequence(p.seq_base_path, p.video)
# first frame
img_uint8 = cv2.imread(img_list[0])
img_uint8 = cv2.cvtColor(img_uint8, cv2.COLOR_BGR2RGB)
img_double = np.double(img_uint8) # uint8 to float
# compute avg for padding
avg_chans = np.mean(img_double, axis=(0, 1))
wc_z = target_size[1] + p.context_amount * sum(target_size)
hc_z = target_size[0] + p.context_amount * sum(target_size)
s_z = np.sqrt(wc_z * hc_z)
scale_z = p.examplar_size / s_z
# crop examplar z in the first frame
z_crop = get_subwindow_tracking(img_double, target_position, p.examplar_size, round(s_z), avg_chans)
z_crop = np.uint8(z_crop) # you need to convert it to uint8
# convert image to tensor
z_crop_tensor = 255.0 * F.to_tensor(z_crop).unsqueeze(0)
d_search = (p.instance_size - p.examplar_size) / 2
pad = d_search / scale_z
s_x = s_z + 2 * pad
# arbitrary scale saturation
min_s_x = p.scale_min * s_x
max_s_x = p.scale_max * s_x
# generate cosine window
if p.windowing == 'cosine':
window = np.outer(np.hanning(p.score_size * p.response_UP), np.hanning(p.score_size * p.response_UP))
elif p.windowing == 'uniform':
window = np.ones((p.score_size * p.response_UP, p.score_size * p.response_UP))
window = window / sum(sum(window))
# pyramid scale search
scales = p.scale_step**np.linspace(-np.ceil(p.num_scale/2), np.ceil(p.num_scale/2), p.num_scale)
# extract feature for examplar z
z_features = net.conv_features(Variable(z_crop_tensor).to(device))
z_features = z_features.repeat(p.num_scale, 1, 1, 1)
# do tracking
bboxes = np.zeros((len(img_list), 4), dtype=np.double) # save tracking result
stats = Stats()
recorder = Recorder(str(net_type), p.video)
start_time = datetime.datetime.now()
for i in tqdm(range(0, len(img_list))):
if i > 0:
# do detection
# currently, we only consider RGB images for tracking
img_uint8 = cv2.imread(img_list[i])
img_uint8 = cv2.cvtColor(img_uint8, cv2.COLOR_BGR2RGB)
img_double = np.double(img_uint8) # uint8 to float
scaled_instance = s_x * scales
scaled_target = np.zeros((2, scales.size), dtype = np.double)
scaled_target[0, :] = target_size[0] * scales
scaled_target[1, :] = target_size[1] * scales
# extract scaled crops for search region x at previous target position
x_crops = make_scale_pyramid(img_double, target_position, scaled_instance, p.instance_size, avg_chans, p)
# get features of search regions
x_crops_tensor = torch.FloatTensor(x_crops.shape[3], x_crops.shape[2], x_crops.shape[1], x_crops.shape[0])
# response_map = SiameseNet.get_response_map(z_features, x_crops)
for k in range(x_crops.shape[3]):
tmp_x_crop = x_crops[:, :, :, k]
tmp_x_crop = np.uint8(tmp_x_crop)
# numpy array to tensor
x_crops_tensor[k, :, :, :] = 255.0 * F.to_tensor(tmp_x_crop).unsqueeze(0)
# get features of search regions
x_features = net.conv_features(Variable(x_crops_tensor).to(device))
# evaluate the offline-trained network for exemplar x features
target_position, new_scale = tracker_eval(net, round(s_x), z_features, x_features, target_position, window, p)
# scale damping and saturation
s_x = max(min_s_x, min(max_s_x, (1 - p.scale_LR) * s_x + p.scale_LR * scaled_instance[int(new_scale)]))
target_size = (1 - p.scale_LR) * target_size + p.scale_LR * np.array([scaled_target[0, int(new_scale)], scaled_target[1, int(new_scale)]])
rect_position = np.array([target_position[1]-target_size[1]/2, target_position[0]-target_size[0]/2, target_size[1], target_size[0]])
if p.visualization:
if not recorder.is_initialized:
recorder.set_options(img_uint8)
save_info = {"net_type": str(net_type),
"video_name": p.video,
"save_to_file": p.save_to_file,
"save_to_video": p.save_to_video}
visualize_tracking_result(img_uint8, rect_position, 1, i, gt_list[i], save_info, stats, recorder)
# output bbox in the original frame coordinates
o_target_position = target_position
o_target_size = target_size
bboxes[i,:] = np.array([o_target_position[1]-o_target_size[1]/2, o_target_position[0]-o_target_size[0]/2, o_target_size[1], o_target_size[0]])
if p.visualization:
recorder.close()
# print(stats.meanIoU())
# print(stats.meanCenterError())
# print(stats.meanPrecision())
end_time = datetime.datetime.now()
fps = len(img_list)/max(1.0, (end_time-start_time).seconds)
return bboxes, fps
def _rho(beta, nxkx, N, K):
kappa = beta * K
rx = np.array([_rhoi(x, nxkx, beta) for x in nxkx])
return rx.prod() / rf(kappa, np.double(N))
shp_gpd = gpd.read_file(shp_path)
print("adapting to our needs..")
lunghezza = lines_length(shp_gpd['geometry'])
solo_remi=[1 if i=='Rio Blu' else 0 for i in shp_gpd['BARCHE_A_R']]
# normativa: to be done
# c'è anche la stazza, ma per il momento non ci interessa (10t)
vel_max=shp_gpd['VEL_MAX'][:]
vel_max_mp=shp_gpd['VEL_MAX_MP'][:]
larghezza=shp_gpd['LARGHEZZA_']
senso_unico=shp_gpd['ONEWAY'][:]
orario_senso_start=gpd.GeoSeries(np.zeros([larghezza.size], dtype=np.int8))
orario_senso_end=gpd.GeoSeries(np.ones([larghezza.size], dtype=np.int8)*24)
orario_chiuso_start=gpd.GeoSeries(np.ones([larghezza.size], dtype=np.int8)*24)
orario_chiuso_end=gpd.GeoSeries(np.ones([larghezza.size], dtype=np.int8)*24)
altezza=np.double(np.ones_like(solo_remi))*1000
lista_sensi_inversi=["DE SAN LUCA - ROSSINI", "DE PALAZZO - CANONICA", "DE LA FAVA","DE LA PIETA' - SANT'ANTONIN","DE SAN GIUSEPPE", "DE LA TETA - SAN GIOVANNI LATERANO RAMO BASSO", "DE SAN GIACOMO DALL'ORIO","DE SAN VIO"]
lista_limiti_sette=["GRANDE","DE CANNAREGIO"]
lista_no_info_per_non_dimenticare=["DE LE GALEAZZE", "SCUOLA GABELLI","DE LA VERONA - MENUO" ]
lista_limiti_laguna=["DE LA RANA", "MOLO B", "DI CAMPALTO","COA DI LATTE","MORTO - MAZZORBO","CARBONERA","ALTINO","MONTIRON","DE SANT'ANTONIO","DEL COLPO", "DI BOSSOLARO", "LA ROTTA","CAMPANA", "BOMBAE", "PORTOSECCO","DE LA CAVA","CODA REZIOL"]
vel_laguna=10
lista_limiti_centro=["SCUOLA GABELLI","A. CANAL","DE LA SACA DE LA MISERICORDIA","ARSENAL VECHIO"] # in realtà sono quasi tutti al lido
lista_divieti_0_24 = ["ARSENAL VECHIO"]################## caso particolare, gli altri hanno il divieto di transito nella colonna giusta!
# canali con limite per canoe e simili dalle 8 alle 15 lunedì-venerdì e dalle 8 alle 13 il sabato
lista_no_remetti = ["Canal Grande", "Cannaregio", "Giardini", "Greci - San Lorenzo", "- Santa Giustina - Sant’Antonin – Pietà", "Noale", "Novo", "Ca’ Foscari", "Santi Apostoli - Gesuiti"] # nomi da correggere
vel_centro=5
ponte_codes_list=[]
epsilon=0.001
for index,canal in shp_gpd.iterrows():
def triangulateCallback(self, p_left, p_right):
t = p_left.header.stamp.to_time()
if self.start == 0:
self.start = t
if len(p_left.polygon.points) == 0 or len(p_right.polygon.points) == 0:
if self.tracker:
if self.tracker.updateNotFound() >= 15:
rospy.loginfo("Lost!")
self.tracker = None
try:
data = "%.9f,%.3f,%.3f,%.3f,%d,%.3f,%.3f" % (
t - self.start, 0, 0, 0, 0, 0, 0)
self.writer.writerow(data.split(','))
except csv.Error as e:
sys.exit('File %s, line %d: %s' %
(self.file_name, self.writer.line_num, e))
else:
try:
data = "%.9f,%.3f,%.3f,%.3f,%d,%.3f,%.3f" % (
t - self.start, self.tracker.pos[0],
self.tracker.pos[1], self.tracker.pos[2],
self.tracker.isTracked, 0, 0)
self.writer.writerow(data.split(','))
except csv.Error as e:
sys.exit('File %s, line %d: %s' %
(self.file_name, self.writer.line_num, e))
else:
try:
data = "%.9f,%.3f,%.3f,%.3f,%d,%.3f,%.3f" % (
t - self.start, 0, 0, 0, 0, 0, 0)
self.writer.writerow(data.split(','))
except csv.Error as e:
sys.exit('File %s, line %d: %s' %
(self.file_name, self.writer.line_num, e))
return
psi_beta_list_left = []
psi_beta_list_right = []
for point in p_left.polygon.points:
u = point.x
v = point.y
x = (u - u01) / mu1
y = (v - v01) / mv1
phi = np.arctan2(y, x)
r = np.sqrt(x**2 + y**2)
p = coeffs1[:]
p.append(-r)
thetas = np.roots(p)
for theta in thetas:
if np.imag(theta) == 0:
if 0 < np.real(theta) < np.pi / 2:
theta = np.double(np.real(theta))
break
else:
print "Unable to find theta"
return None
u_cam = np.array([[np.sin(theta) * np.cos(phi)],
[np.sin(theta) * np.sin(phi)], [np.cos(theta)]])
rect_r = np.dot(R1, u_cam)
psi = np.arcsin(rect_r[0, 0])
beta = np.arctan2(rect_r[1, 0], rect_r[2, 0])
psi_beta_list_left.append((psi, beta, point.z))
for point in p_right.polygon.points:
u = point.x
v = point.y
x = (u - u02) / mu2
y = (v - v02) / mv2
phi = np.arctan2(y, x)
r = np.sqrt(x**2 + y**2)
p = coeffs1[:]
p.append(-r)
thetas = np.roots(p)
for theta in thetas:
if np.imag(theta) == 0:
if 0 < np.real(theta) < np.pi / 2:
theta = np.double(np.real(theta))
break
else:
print "Unable to find theta"
return None
u_cam = np.array([[np.sin(theta) * np.cos(phi)],
[np.sin(theta) * np.sin(phi)], [np.cos(theta)]])
rect_r = np.dot(R1, u_cam)
psi = np.arcsin(rect_r[0, 0])
beta = np.arctan2(rect_r[1, 0], rect_r[2, 0])
psi_beta_list_right.append((psi, beta, point.z))
psi_beta_list_left = sorted(psi_beta_list_left,
key=itemgetter(2),
reverse=True)
psi_beta_list_right = sorted(psi_beta_list_right,
key=itemgetter(2),
reverse=True)
i = 0
j = 0
found = False
while not found:
z1 = psi_beta_list_left[i][2]
z2 = psi_beta_list_right[j][2]
if z1 == 0 and z2 == 0:
rospy.loginfo("No match")
if self.tracker:
if self.tracker.updateNotFound() >= 15:
rospy.loginfo("Lost!")
self.tracker = None
break # 0 should not be matched with zero, just end
if z1 < z2: # right camera has better membership value
#rospy.loginfo("j = %d, %f", j, z2)
(psi2, beta2, z2) = psi_beta_list_right[j]
for (psi1, beta1, z1) in psi_beta_list_left:
if abs(beta1 - beta2) < 0.06: # On the same epipolar line
if psi1 <= psi2:
# In the stereo vision, psi1 must be more than psi2 for the same object
continue
rho = baseline * np.cos(psi2) / np.sin(psi1 - psi2)
if rho > 10:
# 15 meters away from the camera, that is impossible and would be a misdetection
continue
x_out = rho * np.sin(psi1)
y_out = rho * np.cos(psi1) * np.sin(beta1)
z_out = rho * np.cos(psi1) * np.cos(beta1)
if not self.tracker:
self.tracker = Tracker(x_out, y_out, z_out)
else:
self.tracker.predict()
if self.tracker.distance(
(x_out, y_out, z_out)) > 0.36:
continue
self.tracker.updateTrack((x_out, y_out, z_out))
self.broadcaster.sendTransform(
self.tracker.pos,
tf.transformations.quaternion_from_euler(0, 0, 0),
rospy.Time.now(), '/blimp', '/world')
print z1, z2
found = True
break
else:
# Arriving here means there is no match, increase index of right camera
j += 1
else: # left camera has better membership value
#rospy.loginfo("i = %d, %f", i, z1)
(psi1, beta1, z1) = psi_beta_list_left[i]
for (psi2, beta2, z2) in psi_beta_list_right:
if abs(beta1 - beta2) < 0.06: # On the same epipolar line
if psi1 <= psi2:
# In the stereo vision, psi1 must be more than psi2 for the same object
continue
rho = baseline * np.cos(psi2) / np.sin(psi1 - psi2)
if rho > 10:
# 15 meters away from the camera, that is impossible and would be a misdetection
continue
x_out = rho * np.sin(psi1)
y_out = rho * np.cos(psi1) * np.sin(beta1)
z_out = rho * np.cos(psi1) * np.cos(beta1)
if not self.tracker:
self.tracker = Tracker(x_out, y_out, z_out)
else:
self.tracker.predict()
if self.tracker.distance(
(x_out, y_out, z_out)) > 0.36:
continue
self.tracker.updateTrack((x_out, y_out, z_out))
self.broadcaster.sendTransform(
self.tracker.pos,
tf.transformations.quaternion_from_euler(0, 0, 0),
rospy.Time.now(), '/blimp', '/world')
print z1, z2
found = True
break
else:
# Arriving here means there is no match, increase index of right camera
i += 1
if i >= len(psi_beta_list_left) or j >= len(psi_beta_list_right):
rospy.loginfo("No match")
if self.tracker:
if self.tracker.updateNotFound() >= 15:
rospy.loginfo("Lost!")
self.tracker = None
break
if self.tracker:
try:
data = "%.9f,%.3f,%.3f,%.3f,%d,%.3f,%.3f" % (
t - self.start, self.tracker.pos[0], self.tracker.pos[1],
self.tracker.pos[2], self.tracker.isTracked, z1, z2)
self.writer.writerow(data.split(','))
except csv.Error as e:
sys.exit('File %s, line %d: %s' %
(self.file_name, self.writer.line_num, e))
else:
try:
data = "%.9f,%.3f,%.3f,%.3f,%d,%.3f,%.3f" % (t - self.start, 0,
0, 0, 0, 0, 0)
self.writer.writerow(data.split(','))
except csv.Error as e:
sys.exit('File %s, line %d: %s' %
(self.file_name, self.writer.line_num, e))
def main():
global lowpass_filter_b, lowpass_filter_a
global hz_ire_scale, minn
global f_deemp_b, f_deemp_a
global deemp_t1, deemp_t2
global Bcutr, Acutr
global Inner
global blocklen
outfile = sys.stdout #.buffer
audio_mode = 0
CAV = 0
byte_start = 0
byte_end = 0
f_seconds = False
optlist, cut_argv = getopt.getopt(sys.argv[1:], "d:D:hLCaAwSs:")
for o, a in optlist:
if o == "-d":
deemp_t1 = np.double(a)
if o == "-D":
deemp_t2 = np.double(a)
if o == "-a":
audio_mode = 1
blocklen = blocklen * 4
if o == "-L":
Inner = 1
if o == "-A":
CAV = 1
Inner = 1
if o == "-h":
# use full spec deemphasis filter - will result in overshoot, but higher freq resonse
f_deemp_b = [3.778720395899611e-01, -2.442559208200777e-01]
f_deemp_a = [1.000000000000000e+00, -8.663838812301168e-01]
if o == "-C":
Bcutr, Acutr = sps.butter(1,
[2.50 / (freq / 2), 3.26 / (freq / 2)],
btype='bandstop')
Bcutr, Acutr = sps.butter(1,
[2.68 / (freq / 2), 3.08 / (freq / 2)],
btype='bandstop')
if o == "-w":
hz_ire_scale = (9360000 - 8100000) / 100
minn = 8100000 + (hz_ire_scale * -60)
if o == "-S":
f_seconds = True
if o == "-s":
ia = int(a)
if ia == 0:
lowpass_filter_b, lowpass_filter_a = sps.butter(
5, (4.2 / (freq / 2)), 'low')
if ia == 1:
lowpass_filter_b, lowpass_filter_a = sps.butter(
5, (4.4 / (freq / 2)), 'low')
if ia == 2:
lowpass_filter_b, lowpass_filter_a = sps.butter(
6, (4.6 / (freq / 2)), 'low')
lowpass_filter_b, lowpass_filter_a = sps.butter(
6, (4.6 / (freq / 2)), 'low')
deemp_t1 = .825
deemp_t2 = 2.35
if ia == 3:
# high frequency response - and ringing. choose your poison ;)
lowpass_filter_b, lowpass_filter_a = sps.butter(
10, (5.0 / (freq / 2)), 'low')
lowpass_filter_b, lowpass_filter_a = sps.butter(
7, (5.0 / (freq / 2)), 'low')
if ia == 4:
lowpass_filter_b, lowpass_filter_a = sps.butter(
10, (5.3 / (freq / 2)), 'low')
lowpass_filter_b, lowpass_filter_a = sps.butter(
7, (5.3 / (freq / 2)), 'low')
if ia >= 30:
lpfreq = ia / 10.0
lowpass_filter_b, lowpass_filter_a = sps.butter(
5, (lpfreq / (freq / 2)), 'low')
# set up deemp filter
[tf_b, tf_a] = sps.zpk2tf(-deemp_t2 * (10**-8), -deemp_t1 * (10**-8),
deemp_t1 / deemp_t2)
[f_deemp_b, f_deemp_a] = sps.bilinear(tf_b, tf_a, 1 / (freq_hz / 2))
# test()
argc = len(cut_argv)
if argc >= 1:
infile = open(cut_argv[0], "rb")
else:
infile = sys.stdin
byte_start = 0
if (argc >= 2):
byte_start = float(cut_argv[1])
if (argc >= 3):
byte_end = float(cut_argv[2])
limit = 1
else:
limit = 0
if f_seconds:
byte_start *= (freq_hz * 2)
byte_end *= (freq_hz * 2)
else:
byte_end += byte_start
byte_end -= byte_start
byte_start = int(byte_start)
byte_end = int(byte_end)
if (byte_start > 0):
infile.seek(byte_start)
if CAV and byte_start > 11454654400:
CAV = 0
Inner = 0
# set up deemp filter
[tf_b, tf_a] = sps.zpk2tf(-deemp_t2 * (10**-8), -deemp_t1 * (10**-8),
deemp_t1 / deemp_t2)
[f_deemp_b, f_deemp_a] = sps.bilinear(tf_b, tf_a, 1 / (freq_hz / 2))
total = toread = blocklen
inbuf = infile.read(toread * 2)
indata = np.fromstring(inbuf, 'int16', toread) + 32768
total = 0
total_prevread = 0
total_read = 0
while (len(inbuf) > 0):
toread = blocklen - indata.size
if toread > 0:
inbuf = infile.read(toread * 2)
indata = np.append(
indata,
np.fromstring(inbuf, 'int16',
len(inbuf) // 2) + 32768)
if indata.size < blocklen:
exit()
if audio_mode:
output, osamp = process_audio(indata)
nread = osamp
outfile.write(output)
else:
output_16 = process_video(indata)
outfile.write(output_16)
nread = len(output_16)
total_pread = total_read
total_read += nread
if CAV:
if (total_read + byte_start) > 11454654400:
CAV = 0
Inner = 0
indata = indata[nread:len(indata)]
if limit == 1:
byte_end -= toread
if (byte_end < 0):
inbuf = ""
def main(args):
if args.info:
args.input.append(args.output)
df = pd.concat(
(star.parse_star(inp, augment=args.augment) for inp in args.input),
join="inner")
dfaux = None
if args.cls is not None:
df = star.select_classes(df, args.cls)
if args.info:
if star.is_particle_star(df) and star.Relion.CLASS in df.columns:
c = df[star.Relion.CLASS].value_counts()
print("%s particles in %d classes" %
("{:,}".format(df.shape[0]), len(c)))
print(" ".join([
'%d: %s (%.2f %%)' % (i, "{:,}".format(s), 100. * s / c.sum())
for i, s in iteritems(c.sort_index())
]))
elif star.is_particle_star(df):
print("%s particles" % "{:,}".format(df.shape[0]))
if star.Relion.MICROGRAPH_NAME in df.columns:
mgraphcnt = df[star.Relion.MICROGRAPH_NAME].value_counts()
print(
"%s micrographs, %s +/- %s particles per micrograph" %
("{:,}".format(len(mgraphcnt)), "{:,.3f}".format(
np.mean(mgraphcnt)), "{:,.3f}".format(np.std(mgraphcnt))))
try:
print("%f A/px (%sX magnification)" %
(star.calculate_apix(df), "{:,.0f}".format(
df[star.Relion.MAGNIFICATION][0])))
except KeyError:
pass
if len(df.columns.intersection(star.Relion.ORIGINS3D)) > 0:
print("Largest shift is %f pixels" % np.max(
np.abs(df[df.columns.intersection(
star.Relion.ORIGINS3D)].values)))
return 0
if args.drop_angles:
df.drop(star.Relion.ANGLES, axis=1, inplace=True, errors="ignore")
if args.drop_containing is not None:
containing_fields = [
f for q in args.drop_containing for f in df.columns if q in f
]
if args.invert:
containing_fields = df.columns.difference(containing_fields)
df.drop(containing_fields, axis=1, inplace=True, errors="ignore")
if args.offset_group is not None:
df[star.Relion.GROUPNUMBER] += args.offset_group
if args.restack is not None:
if not args.augment:
star.augment_star_ucsf(df, inplace=True)
star.set_original_fields(df, inplace=True)
df[star.UCSF.IMAGE_PATH] = args.restack
df[star.UCSF.IMAGE_INDEX] = np.arange(df.shape[0])
if args.subsample_micrographs is not None:
if args.bootstrap is not None:
print("Only particle sampling allows bootstrapping")
return 1
mgraphs = df[star.Relion.MICROGRAPH_NAME].unique()
if args.subsample_micrographs < 1:
args.subsample_micrographs = np.int(
max(np.round(args.subsample_micrographs * len(mgraphs)), 1))
else:
args.subsample_micrographs = np.int(args.subsample_micrographs)
ind = np.random.choice(len(mgraphs),
size=args.subsample_micrographs,
replace=False)
mask = df[star.Relion.MICROGRAPH_NAME].isin(mgraphs[ind])
if args.auxout is not None:
dfaux = df.loc[~mask]
df = df.loc[mask]
if args.subsample is not None and args.suffix == "":
if args.subsample < 1:
args.subsample = np.int(
max(np.round(args.subsample * df.shape[0]), 1))
else:
args.subsample = np.int(args.subsample)
ind = np.random.choice(df.shape[0], size=args.subsample, replace=False)
mask = df.index.isin(ind)
if args.auxout is not None:
dfaux = df.loc[~mask]
df = df.loc[mask]
if args.copy_angles is not None:
angle_star = star.parse_star(args.copy_angles, augment=args.augment)
df = star.smart_merge(df,
angle_star,
fields=star.Relion.ANGLES,
key=args.merge_key)
if args.copy_alignments is not None:
align_star = star.parse_star(args.copy_alignments,
augment=args.augment)
df = star.smart_merge(df,
align_star,
fields=star.Relion.ALIGNMENTS,
key=args.merge_key)
if args.copy_reconstruct_images is not None:
recon_star = star.parse_star(args.copy_reconstruct_images,
augment=args.augment)
df[star.Relion.RECONSTRUCT_IMAGE_NAME] = recon_star[
star.Relion.IMAGE_NAME]
if args.transform is not None:
if args.transform.count(",") == 2:
r = geom.euler2rot(
*np.deg2rad([np.double(s) for s in args.transform.split(",")]))
else:
r = np.array(json.loads(args.transform))
df = star.transform_star(df, r, inplace=True)
if args.invert_hand:
df[star.Relion.ANGLEROT] = -df[star.Relion.ANGLEROT]
df[star.Relion.ANGLETILT] = 180 - df[star.Relion.ANGLETILT]
if args.copy_paths is not None:
path_star = star.parse_star(args.copy_paths)
star.set_original_fields(df, inplace=True)
df[star.Relion.IMAGE_NAME] = path_star[star.Relion.IMAGE_NAME]
if args.copy_ctf is not None:
ctf_star = pd.concat((star.parse_star(inp, augment=args.augment)
for inp in glob.glob(args.copy_ctf)),
join="inner")
df = star.smart_merge(df,
ctf_star,
star.Relion.CTF_PARAMS,
key=args.merge_key)
if args.copy_micrograph_coordinates is not None:
coord_star = pd.concat(
(star.parse_star(inp, augment=args.augment)
for inp in glob.glob(args.copy_micrograph_coordinates)),
join="inner")
df = star.smart_merge(df,
coord_star,
fields=star.Relion.MICROGRAPH_COORDS,
key=args.merge_key)
if args.scale is not None:
star.scale_coordinates(df, args.scale, inplace=True)
star.scale_origins(df, args.scale, inplace=True)
star.scale_magnification(df, args.scale, inplace=True)
if args.scale_particles is not None:
star.scale_origins(df, args.scale_particles, inplace=True)
star.scale_magnification(df, args.scale_particles, inplace=True)
if args.scale_coordinates is not None:
star.scale_coordinates(df, args.scale_coordinates, inplace=True)
if args.scale_origins is not None:
star.scale_origins(df, args.scale_origins, inplace=True)
if args.scale_magnification is not None:
star.scale_magnification(df, args.scale_magnfication, inplace=True)
if args.recenter:
df = star.recenter(df, inplace=True)
if args.zero_origins:
df = star.zero_origins(df, inplace=True)
if args.pick:
df.drop(df.columns.difference(star.Relion.PICK_PARAMS),
axis=1,
inplace=True,
errors="ignore")
if args.subsample is not None and args.suffix != "":
if args.subsample < 1:
print("Specific integer sample size")
return 1
nsamplings = args.bootstrap if args.bootstrap is not None else df.shape[
0] / np.int(args.subsample)
inds = np.random.choice(df.shape[0],
size=(nsamplings, np.int(args.subsample)),
replace=args.bootstrap is not None)
for i, ind in enumerate(inds):
star.write_star(
os.path.join(
args.output,
os.path.basename(args.input[0])[:-5] + args.suffix +
"_%d" % (i + 1)), df.iloc[ind])
if args.to_micrographs:
gb = df.groupby(star.Relion.MICROGRAPH_NAME)
mu = gb.mean()
df = mu[[
c for c in star.Relion.CTF_PARAMS + star.Relion.MICROSCOPE_PARAMS +
[star.Relion.MICROGRAPH_NAME] if c in mu
]].reset_index()
if args.micrograph_range:
df.set_index(star.Relion.MICROGRAPH_NAME, inplace=True)
m, n = [int(tok) for tok in args.micrograph_range.split(",")]
mg = df.index.unique().sort_values()
outside = list(range(0, m)) + list(range(n, len(mg)))
dfaux = df.loc[mg[outside]].reset_index()
df = df.loc[mg[m:n]].reset_index()
if args.micrograph_path is not None:
df = star.replace_micrograph_path(df,
args.micrograph_path,
inplace=True)
if args.min_separation is not None:
gb = df.groupby(star.Relion.MICROGRAPH_NAME)
dupes = []
for n, g in gb:
nb = algo.query_connected(
g[star.Relion.COORDS],
args.min_separation / star.calculate_apix(df))
dupes.extend(g.index[~np.isnan(nb)])
dfaux = df.loc[dupes]
df.drop(dupes, inplace=True)
if args.merge_source is not None:
if args.merge_fields is not None:
if "," in args.merge_fields:
args.merge_fields = args.merge_fields.split(",")
else:
args.merge_fields = [args.merge_fields]
else:
print("Merge fields must be specified using --merge-fields")
return 1
if args.merge_key is not None:
if "," in args.merge_key:
args.merge_key = args.merge_key.split(",")
merge_star = star.parse_star(args.merge_source, augment=args.augment)
df = star.smart_merge(df,
merge_star,
fields=args.merge_fields,
key=args.merge_key)
if args.split_micrographs:
dfs = star.split_micrographs(df)
for mg in dfs:
star.write_star(
os.path.join(args.output,
os.path.basename(mg)[:-4]) + args.suffix, dfs[mg])
return 0
if args.auxout is not None and dfaux is not None:
star.write_star(args.auxout, dfaux, simplify=args.augment_output)
if args.output is not None:
star.write_star(args.output, df, simplify=args.augment_output)
return 0
def _rhoi(x, nxkx, beta):
return power(rf(beta, np.double(x)), nxkx[x])
np.short() np.intc() np.intp() np.int0() np.int_() np.longlong() np.ubyte() np.ushort() np.uintc() np.uintp() np.uint0() np.uint() np.ulonglong() np.half() np.single() np.double() np.float_() np.longdouble() np.longfloat() np.csingle() np.singlecomplex() np.cdouble() np.complex_() np.cfloat() np.clongdouble() np.clongfloat() np.longcomplex()
dNm = midRange[1] - midRange[0] highRange = np.round(np.interp(highband, F, np.arange(NTmax))) dNh = highRange[1] - highRange[0] dN = max(dNl, dNm, dNh) print 'NTmax', NTmax print 'NLmax', NLmax print 'NTleft', NTleft print 'Nread', Nread plt.hold(True) lastImgLine = 0 # Remember last line, which was filled in Rrgb array. for i in range(int(Nread)): if i == Nread - 1: R = np.double(dataset[(i) * NTmax:, 0:NLmax]) T = np.arange((i) * NTmax, dataset.shape[0]) / fp F = np.linspace(0, fp, num=NTleft) lowRange = np.round(np.interp(lowband, F, np.arange(NTleft))) dNl = lowRange[1] - lowRange[0] midRange = np.round(np.interp(midband, F, np.arange(NTleft))) dNm = midRange[1] - midRange[0] highRange = np.round(np.interp(highband, F, np.arange(NTleft))) dNh = highRange[1] - highRange[0] dN = max(dNl, dNm, dNh) else: R = dataset[(i) * NTmax:(i + 1) * NTmax, 0:NLmax] T = np.arange((i) * NTmax, (i + 1) * NTmax) / fp S = sp.fft(R, axis=0) Rl = sp.ifft(S[int(lowRange[0]): int(lowRange[1]), :], n=int(dN), axis = 0)
# Sample training estimation batch (for initial scale and batchnorm parameters)
beta0_list = []
for i, (sampler, embedder) in enumerate(zip(sampler_list, embedder_list)):
Tp_est, Tv_est, Ap_est, Av_est = sampler.sample(
args.est_batch_size[0],
numpy.inf,
args.fract_inside[0],
xvalset='train',
update_epoch=False) # get all neighbors for estimation
Tp_est = normalize_patches(args.patch_norm[0], Tp_est)
Ap_est = normalize_patches(args.patch_norm[0], Ap_est)
Y_est = numpy.double(Av_est == Tv_est)
#
# Estimate initial scale parameter
print('estimating initial scale parameter')
global scale_vars
# scale for the original method
l2n = id_net.get_l2n(Tp_est, Ap_est)
scale_vars = (l2n, Y_est)
beta0_list.append(
minimize(labfus, 0.001, method='Nelder-Mead', tol=1e-6).x)
embedder.beta0 = beta0_list[-1] # add beta0 to model
import numpy as np
from keras.layers import Input
from keras.layers import MaxPooling2D, UpSampling2D, Conv2D
from keras.models import Sequential
from tensorflow.examples.tutorials.mnist import input_data
import matplotlib.pyplot as plt
from keras.datasets import mnist
#load data
img_rows, img_cols = 28, 28
input_shape = [28, 28, 1]
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = np.double(np.reshape(x_train,
(len(x_train), img_rows, img_cols, 1))) / 255
x_test = np.double(np.reshape(x_test,
(len(x_test), img_rows, img_cols, 1))) / 255
# construct a convolutional stack autoencoder
auto_encoder = Sequential()
auto_encoder.add(
Conv2D(15, (3, 3),
activation='relu',
padding='same',
input_shape=input_shape)) # (?, 28, 28, 32)
auto_encoder.add(MaxPooling2D((2, 2), padding='same')) # (?, 14, 14, 32)
auto_encoder.add(
Conv2D(10, (3, 3),
activation='relu',
padding='same',
input_shape=input_shape)) # (?, 28, 28, 32)
auto_encoder.add(MaxPooling2D((2, 2), padding='same')) # (?, 14, 14, 32)
def domain2polygons(zlonc_target, zlatc_target, wk_proj="wgs84", data=None):
nzonal_target, nmerid_target = zlonc_target[:-1, :-1].shape
if data is not None:
nbands = data.shape[0]
# Reference projections
wgs84 = osr.SpatialReference()
wgs84.ImportFromProj4("+proj=longlat +ellps=WGS84 +datum=WGS84 +no_defs")
wgs84_pyproj = Proj(init="epsg:4326")
if wk_proj == "wgs84":
ref_proj = osr.SpatialReference()
ref_proj.ImportFromProj4(
"+proj=longlat +ellps=WGS84 +datum=WGS84 +no_defs")
ref_pyproj = Proj(init="epsg:4326")
else:
ref_proj = osr.SpatialReference()
ref_proj.ImportFromProj4(wk_proj)
ref_pyproj = Proj(wk_proj)
# Convert lon/lat domain coordinates to reference coordinate system
xc, yc = transform(wgs84_pyproj, ref_pyproj, zlonc_target, zlatc_target)
# define the layer
vector_grid_driver = ogr.GetDriverByName("MEMORY")
vector_grid_ds = vector_grid_driver.CreateDataSource("memData")
vector_grid_layer = vector_grid_ds.CreateLayer("grid",
ref_proj,
geom_type=ogr.wkbPolygon)
if data is not None:
for iband in range(nbands):
idField = ogr.FieldDefn("field_{}".format(iband), ogr.OFTReal)
vector_grid_layer.CreateField(idField)
featureDefn = vector_grid_layer.GetLayerDefn()
for i in range(nmerid_target):
for j in range(nzonal_target):
lon_tmp = [
xc[j, i],
xc[j, i + 1],
xc[j + 1, i + 1],
xc[j + 1, i],
xc[j, i],
]
if wk_proj == "wgs84":
lon_tmp = np.unwrap(lon_tmp, discont=180)
# Create ring
ring = ogr.Geometry(ogr.wkbLinearRing)
_ = ring.AddPoint(lon_tmp[0], yc[j, i])
_ = ring.AddPoint(lon_tmp[1], yc[j, i + 1])
_ = ring.AddPoint(lon_tmp[2], yc[j + 1, i + 1])
_ = ring.AddPoint(lon_tmp[3], yc[j + 1, i])
_ = ring.AddPoint(lon_tmp[4], yc[j, i])
# Create polygon
poly = ogr.Geometry(ogr.wkbPolygon)
poly.AddGeometry(ring)
# Add feature to layer
feature = ogr.Feature(featureDefn)
feature.SetGeometry(poly)
# Add data if any
if data is not None:
for iband in range(nbands):
feature.SetField("field_{}".format(iband),
np.double(data[iband, j, i]))
vector_grid_layer.CreateFeature(feature)
feature = None
return vector_grid_driver, vector_grid_ds, vector_grid_layer
def maxent_edge_sampling(model, theta_g, block, l, psi, beta, x_out):
"""
:param model: GNM and parameters
:type model: dict
:param theta_g: parameters for marginal distribution of network structure P(G)
:type theta_g: matrix
:param block: sample block from penultimate iteration of mKPGM
:type block: matrix
:param psi: edge types
:type psi: list
:param beta: fraction of edges of each type
:type beta: list
:param x_out: node attributes
:type x_out: list
:return: graphOut (output graph, contains num. of vertices and edge list)
:rtype: tuple
"""
# U = unique probabilities
# T = edge locations
# (3)
U, T = get_unique_prob_edge_location(model, theta_g, block, psi, x_out)
# N_e = 0
Nus = []
# for each unique prob. pi_u
for pi_u in U:
# (5) Draw num. edges to sample per unique prob.
n_u = np.random.binomial(len(T[pi_u]), pi_u)
Nus.append(n_u) # (6)
# Accum. total num. edges to be sampled
# N_e += n_u
N_e = sum(Nus)
# (7) Draw num. edges per edge-type to match rho_IN
gamma = list(np.random.multinomial(n=N_e, pvals=beta, size=1))
# n = Number of experiments (int)
# pvals = Probabilities of each of the p different outcomes (sequence of floats, length p)
# size = Output shape (int or tuple of ints, optional)
E_OUT = []
for i, u in enumerate(U):
# (9) Draw num. edges per edge type for pi_u
Y = list(
np.random.multinomial(n=Nus[i],
pvals=[np.double(g) / N_e for g in gamma[0]],
size=1))
# (10 - 14)
for j, p in enumerate(psi):
# (11) Sampling Y_j edges at random from T_uj possible locations
possible_edges = list(T[u][p])
random.shuffle(possible_edges)
edges = possible_edges[:Y[0][j]]
E_OUT.extend(edges) # (12)
gamma[0][j] -= Y[0][j] # (13)
N_e -= Y[0][j] # (14)
# vertices = len(block[0])
vertices = pow(model['b'], model['K'])
return vertices, E_OUT
dates = years * 10000 + months * 100 + days
co2file = options.co2file
if co2file != '':
co2reader = csv.reader(open(co2file), delimiter=',')
co2data = []
cnt = 0
for row in co2reader:
if cnt:
year, jday, co2 = row
d = datetime.datetime(int(year), 1,
1) + datetime.timedelta(int(jday) - 1)
co2data.append(
[d.year * 10000 + d.month * 100 + d.day,
double(co2)])
cnt += 1
co2data = array(co2data)
co2interp = interp(dates, co2data[:, 0],
co2data[:, 1]) # linear interpolate to dates
else:
co2interp = None # no co2 data
firstyear = years[0]
lastyear = years[-1]
yeartostop = lastyear - 30 + 1
cnt = 1
for i in range(firstyear, yeartostop + 1, 5):
idx = indices(dates, i, i + 49)
data = alldata[idx]
def preprocess():
""" Input:
Although this function doesn't have any input, you are required to load
the MNIST data set from file 'mnist_all.mat'.
Output:
train_data: matrix of training set. Each row of train_data contains
feature vector of a image
train_label: vector of label corresponding to each image in the training
set
validation_data: matrix of training set. Each row of validation_data
contains feature vector of a image
validation_label: vector of label corresponding to each image in the
training set
test_data: matrix of training set. Each row of test_data contains
feature vector of a image
test_label: vector of label corresponding to each image in the testing
set
Some suggestions for preprocessing step:
- feature selection"""
mat = loadmat('mnist_all.mat') # loads the MAT object as a Dictionary
# Pick a reasonable size for validation data
# ------------Initialize preprocess arrays----------------------#
train_preprocess = np.zeros(shape=(50000, 784))
validation_preprocess = np.zeros(shape=(10000, 784))
test_preprocess = np.zeros(shape=(10000, 784))
train_label_preprocess = np.zeros(shape=(50000,))
validation_label_preprocess = np.zeros(shape=(10000,))
test_label_preprocess = np.zeros(shape=(10000,))
# ------------Initialize flag variables----------------------#
train_len = 0
validation_len = 0
test_len = 0
train_label_len = 0
validation_label_len = 0
# ------------Start to split the data set into 6 arrays-----------#
for key in mat:
# -----------when the set is training set--------------------#
if "train" in key:
label = key[-1] # record the corresponding label
tup = mat.get(key)
sap = range(tup.shape[0])
tup_perm = np.random.permutation(sap)
tup_len = len(tup) # get the length of current training set
tag_len = tup_len - 1000 # defines the number of examples which will be added into the training set
# ---------------------adding data to training set-------------------------#
train_preprocess[train_len:train_len + tag_len] = tup[tup_perm[1000:], :]
train_len += tag_len
train_label_preprocess[train_label_len:train_label_len + tag_len] = label
train_label_len += tag_len
# ---------------------adding data to validation set-------------------------#
validation_preprocess[validation_len:validation_len + 1000] = tup[tup_perm[0:1000], :]
validation_len += 1000
validation_label_preprocess[validation_label_len:validation_label_len + 1000] = label
validation_label_len += 1000
# ---------------------adding data to test set-------------------------#
elif "test" in key:
label = key[-1]
tup = mat.get(key)
sap = range(tup.shape[0])
tup_perm = np.random.permutation(sap)
tup_len = len(tup)
test_label_preprocess[test_len:test_len + tup_len] = label
test_preprocess[test_len:test_len + tup_len] = tup[tup_perm]
test_len += tup_len
# ---------------------Shuffle,double and normalize-------------------------#
train_size = range(train_preprocess.shape[0])
train_perm = np.random.permutation(train_size)
train_data = train_preprocess[train_perm]
train_data = np.double(train_data)
train_data = train_data / 255.0
train_label = train_label_preprocess[train_perm]
validation_size = range(validation_preprocess.shape[0])
vali_perm = np.random.permutation(validation_size)
validation_data = validation_preprocess[vali_perm]
validation_data = np.double(validation_data)
validation_data = validation_data / 255.0
validation_label = validation_label_preprocess[vali_perm]
test_size = range(test_preprocess.shape[0])
test_perm = np.random.permutation(test_size)
test_data = test_preprocess[test_perm]
test_data = np.double(test_data)
test_data = test_data / 255.0
test_label = test_label_preprocess[test_perm]
# Feature selection
# Your code here.
tot_data = np.vstack((train_data,validation_data,test_data)) # combining train, validation and test
ftr_indices=np.all(tot_data == tot_data[0,:], axis = 0)
fltrd_data = tot_data[:,~ftr_indices] # removing columns which are similar to the first one by performing shift operations on false columns
tr_len = len(train_data)
va_len = len(validation_data)
tst_len = len(test_data)
train_data = fltrd_data[0:tr_len,:] # separating train data from the filtered data
validation_data = fltrd_data[tr_len: (tr_len + va_len),:] # separating validation data from the filtered data
test_data = fltrd_data[(tr_len + va_len): (tr_len + va_len + tst_len),:] # separating test data from the filtered data
print('preprocess done')
return train_data, train_label, validation_data, validation_label, test_data, test_label
## plt.rc('text', usetex=True)
plt.rc('image', cmap='jet')
plt.ion()
windowSizeInSamples = 2048
hopsize = 256
NFT = 2048
niter = 50
# TODO: also process these as options:
minF0 = 80
maxF0 = 500
Fs = 44100.0
maxFinFT = 8000.0
F = np.ceil(maxFinFT * NFT / np.double(Fs))
stepNotes = 16 # this is the number of F0s within one semitone
K = 10 # number of spectral shapes for the filter part
R = 0 # number of spectral shapes for the accompaniment
P = 30 # number of elements in dictionary of smooth filters
# number of chirped spectral shapes between each F0
# this feature should be further studied before
# we find a good way of doing that.
chirpPerF0 = 1
# Create the harmonic combs, for each F0 between minF0 and maxF0:
F0Table, WF0 = \
generate_WF0_chirped(minF0, maxF0, Fs, Nfft=NFT,
from scipy.spatial.transform import Slerp
from scipy import interpolate
curr_path = os.path.dirname(os.path.realpath(__file__))
sys.path.insert(0, os.path.join(curr_path,'../'))
from flightgoggles.utils import *
from flightgoggles.env import flightgoggles_env
parser = argparse.ArgumentParser(description='FlightGoggles log2video')
parser.add_argument('-f', "--file_path", help="assign log file path", default=os.path.join(curr_path,"example_log.csv"))
parser.add_argument('-o', "--output", help="assign output video name", default=os.path.join(curr_path,"test.avi"))
args = parser.parse_args()
print("Reading log file from {}".format(args.file_path))
data = pd.read_csv(args.file_path, sep=',', header=None).values[1:,:]
data = np.double(data)
save_path = os.path.join(curr_path,"./tmp")
if not os.path.exists(save_path):
os.makedirs(save_path)
env = flightgoggles_env(cfg_fgclient="FlightGogglesClient_debug_env.yaml")
# pos_curr = env.get_state("uav1")["position"]
# yaw_curr = env.get_state("uav1")["attitude_euler_angle"][2]
FPS_VIDEO = 60
# Interpolation
time_array = data[:,0]*1e-9
total_time = time_array[-1]
t_new = np.arange(time_array[0], total_time, 1.0/FPS_VIDEO)
print(debug_J)
print('training neural network...')
# 随机初始化参数
initialTheta1 = randInitializeWeights(hidden_layer_size, input_layer_size)
initialTheta2 = randInitializeWeights(num_labels, hidden_layer_size)
initialParams = np.r_[initialTheta1.flatten(), initialTheta2.flatten()]
lamb = 1
# 优化参数
tnc = minimize(costFunction,
initialParams,
args=(input_layer_size, hidden_layer_size, num_labels, X,
y_m, lamb),
jac=gradientVectorized,
method='TNC',
options={'maxiter': 500})
print(tnc)
nn_params = tnc.x
Theta1 = nn_params[:initialTheta1.size].reshape(initialTheta1.shape)
Theta2 = nn_params[initialTheta1.size:].reshape(initialTheta2.shape)
displayData(Theta1[:, 1:])
pre_y = predict(Theta1, Theta2, X)
print('Train Accuracy: ', np.mean(np.double(pre_y == y)) * 100)
i_map = 0
i_rprec = 0
i_p10 = 0
map_array = np.empty([50, 4])
rprec_array = np.empty([50, 4])
p10_array = np.empty([50, 4])
# file parsing
for line in open('measuresrun1ignore.txt', 'r'):
# create map matrix
if ' map ' in line:
line = line.split(" ")
map_array[i_map, 0] = np.double(line[23])
i_map += 1
# create Rprec matrix
if 'Rprec ' in line:
line = line.split(" ")
rprec_array[i_rprec, 0] = np.double(line[19])
i_rprec += 1
# create Precision at 10 Matrix
if ' P_10 ' in line:
line = line.split(" ")
p10_array[i_p10, 0] = np.double(line[22])
i_p10 += 1
i_map = 0
lat2 = 90. - latdelta * latidx
lon1 = -180. + londelta * (lonidx - 1)
lon2 = -180. + londelta * lonidx
if lat > lat1 or lat < lat2 or lon < lon1 or lon > lon2:
raise Exception('Not in cell!')
return latidx, lonidx
if len(sys.argv) != 6:
print 'Usage: trans_gidx.py latidx1 lonidx1 lat_delta1[,lon_delta1] lat_delta2[,lon_delta2] weathdir2'
sys.exit(1)
latidx = int(sys.argv[1])
lonidx = int(sys.argv[2])
delta1 = [double(d) / 60. for d in sys.argv[3].split(',')] # arcminutes -> degrees
delta2 = [double(d) / 60. for d in sys.argv[4].split(',')]
weathdir2 = sys.argv[5]
if len(delta1) == 1:
latdelta1 = londelta1 = delta1[0]
else:
latdelta1, londelta1 = delta1
if len(delta2) == 1:
latdelta2 = londelta2 = delta2[0]
else:
latdelta2, londelta2 = delta2
latrat, lonrat = latdelta2 / latdelta1, londelta2 / londelta1
latd = [0.5, 0, 1, 0, 1] # start at center
imlow = cv2.normalize(imlow.astype('float'), None, 0.0, 1.0,
cv2.NORM_MINMAX)
## laplacian = cv2.Laplacian(imlow,cv2.CV_64F)
## laplacian1 = np.absolute(laplacian)
## laplacian2 = np.uint8(laplacian1)
kernel = np.array([[-1, 0, 1]])
dst = cv2.filter2D(imlow, -1, kernel)
dst[dst < 0] = 0 # Where values are low
dst[dst > 1] = 1 # Where values are high
dst = dst * 255
dst = np.uint8(dst)
ret, th3 = cv2.threshold(dst, 20, 255, cv2.THRESH_BINARY)
lines = cv2.HoughLines(th3, 1, np.pi / 180, 1)
lines = np.double(lines)
condition11 = 1
condition12 = 1
if lines != []:
for rho, theta in lines[:, 0, :]:
if (theta < 80 * npi) & (theta > 0 * npi) & (condition11 == 1):
thetaL = theta
rhoL = rho
condition11 = 0
if (theta > 120 * npi) & (theta < 180 * npi) & (condition12 == 1):
thetaR = theta
rhoR = rho
condition12 = 0
if (condition11 == 0) & (condition12 == 0):
flag_betterspecanal = int(sys.argv[2])
if (flag_betterspecanal == 0):
print('Enter BlockSize: ')
block_size = int(input())
im = Image.open(img_name)
print('Read ' + img_name)
print('Image size: ', im.size)
# Display image object by PIL.
im.show(title='image')
# Import Image Data into Numpy array.
# The matrix x contains a 2-D array of 8-bit gray scale values.
x = np.array(im)
print('Data type: ', x.dtype)
# Display numpy array by matplotlib.
plt.imshow(x, cmap=plt.cm.gray)
plt.title('Image')
# Set colorbar location. [left, bottom, width, height].
cax = plt.axes([0.9, 0.15, 0.04, 0.7])
plt.colorbar(cax=cax)
plt.show()
x = np.double(x) / 255.0
if flag_betterspecanal == 0:
SpecAnal(x, 99, 99, block_size, base_name)
else:
BetterSpecAnal(x, base_name)
def preprocess():
""" Input:
Although this function doesn't have any input, you are required to load
the MNIST data set from file 'mnist_all.mat'.
Output:
train_data: matrix of training set. Each row of train_data contains
feature vector of a image
train_label: vector of label corresponding to each image in the training
set
validation_data: matrix of training set. Each row of validation_data
contains feature vector of a image
validation_label: vector of label corresponding to each image in the
training set
test_data: matrix of training set. Each row of test_data contains
feature vector of a image
test_label: vector of label corresponding to each image in the testing
set
Some suggestions for preprocessing step:
- feature selection"""
mat = loadmat('/Users/chenshihchia/PycharmProjects/MLassignment1/mnist_all.mat') # loads the MAT object as a Dictionary
# Pick a reasonable size for validation data
# ------------Initialize preprocess arrays----------------------#
train_preprocess = np.zeros(shape=(50000, 784))
validation_preprocess = np.zeros(shape=(10000, 784))
test_preprocess = np.zeros(shape=(10000, 784))
train_label_preprocess = np.zeros(shape=(50000,))
validation_label_preprocess = np.zeros(shape=(10000,))
test_label_preprocess = np.zeros(shape=(10000,))
# ------------Initialize flag variables----------------------#
train_len = 0
validation_len = 0
test_len = 0
train_label_len = 0
validation_label_len = 0
# ------------Start to split the data set into 6 arrays-----------#
for key in mat:
# -----------when the set is training set--------------------#
if "train" in key:
label = key[-1] # record the corresponding label
tup = mat.get(key)
sap = range(tup.shape[0])
tup_perm = np.random.permutation(sap)
tup_len = len(tup) # get the length of current training set
tag_len = tup_len - 1000 # defines the number of examples which will be added into the training set
# ---------------------adding data to training set-------------------------#
train_preprocess[train_len:train_len + tag_len] = tup[tup_perm[1000:], :]
train_len += tag_len
train_label_preprocess[train_label_len:train_label_len + tag_len] = label
train_label_len += tag_len
# ---------------------adding data to validation set-------------------------#
validation_preprocess[validation_len:validation_len + 1000] = tup[tup_perm[0:1000], :]
validation_len += 1000
validation_label_preprocess[validation_label_len:validation_label_len + 1000] = label
validation_label_len += 1000
# ---------------------adding data to test set-------------------------#
elif "test" in key:
label = key[-1]
tup = mat.get(key)
sap = range(tup.shape[0])
tup_perm = np.random.permutation(sap)
tup_len = len(tup)
test_label_preprocess[test_len:test_len + tup_len] = label
test_preprocess[test_len:test_len + tup_len] = tup[tup_perm]
test_len += tup_len
# ---------------------Shuffle,double and normalize-------------------------#
train_size = range(train_preprocess.shape[0])
train_perm = np.random.permutation(train_size)
train_data = train_preprocess[train_perm]
train_data = np.double(train_data)
train_data = train_data / 255.0
train_label = train_label_preprocess[train_perm]
validation_size = range(validation_preprocess.shape[0])
vali_perm = np.random.permutation(validation_size)
validation_data = validation_preprocess[vali_perm]
validation_data = np.double(validation_data)
validation_data = validation_data / 255.0
validation_label = validation_label_preprocess[vali_perm]
test_size = range(test_preprocess.shape[0])
test_perm = np.random.permutation(test_size)
test_data = test_preprocess[test_perm]
test_data = np.double(test_data)
test_data = test_data / 255.0
test_label = test_label_preprocess[test_perm]
# Feature selection
# Your code here.
selected_features = np.zeros(shape=(784, 1))
j = 0
all_data = np.concatenate((train_data, validation_data,test_data), axis=0)
all_data_vstack = np.array(np.vstack((all_data)))
all_data_reduced = np.all(all_data_vstack == all_data_vstack[0, :], axis=0)
for i in range(all_data_reduced.shape[0]):
if (bool(all_data_reduced[i])):
selected_features[j] = i
j = j+1
all_data = all_data[:, ~all_data_reduced]
train_data = all_data[0:train_data.shape[0],:]
validation_data = all_data[train_data.shape[0]:train_data.shape[0]+validation_data.shape[0],:]
test_data = all_data[train_data.shape[0]+validation_data.shape[0]:all_data.shape[0],:]
print('preprocess done')
return train_data, train_label, validation_data, validation_label, test_data, test_label, selected_features