def pvbBound(n):
a = multiply(2, n)
b = power(a, 50)
c = multiply(6, b)
d = divide(c, 0.05)
e = log(d)
f = divide(1.0, n)
return divide(1.0, n) + sqrt(divide(1.0, power(n, 2)) + multiply(f, e))
def ma_multiply_test():
XM=[[1,0,1,0],[0,1,0,0],[0,0,100, 0]]
Coins=[[1],[2],[3]]
from numpy import ma
print(ma.multiply(XM, Coins))
print(ma_multiply(XM, Coins))
Coins=[1,2,3,4]
print(ma.multiply(XM, Coins))
print(ma_multiply(XM, Coins))
def ma_multiply_test():
XM = [[1, 0, 1, 0], [0, 1, 0, 0], [0, 0, 100, 0]]
Coins = [[1], [2], [3]]
from numpy import ma
print(ma.multiply(XM, Coins))
print(ma_multiply(XM, Coins))
Coins = [1, 2, 3, 4]
print(ma.multiply(XM, Coins))
print(ma_multiply(XM, Coins))
def runLogisticRegression(self):
diff = 1
while diff > 0.01:
permutation = np.random.permutation(self.N)
newWeights = self.w.copy()
for i in permutation:
x, y = self.trainingData[i]
gradient = divide(
multiply(-1.0, multiply(x, y)),
(1.0 + exp(multiply(y, np.dot(transpose(self.w), x)))))
newWeights = subtract(newWeights,
multiply(self.learningRate, gradient))
self.epoch += 1
diff = norm(self.w - newWeights)
self.w = newWeights
def normalize(self, info: PoseNormalizationInfo, scale_factor: float = 1):
"""
Normalize the point to a fixed distance between two points
"""
mask = self.body.data.mask
transposed = self.body.zero_filled().points_perspective()
p1s = transposed[info.p1]
p2s = transposed[info.p2]
if transposed.shape[1] == 0:
p1s = p1s[0]
p2s = p2s[0]
else:
p1s = ma.concatenate(p1s)
p2s = ma.concatenate(p2s)
# try:
mean_distance = np.mean(distance_batch(p1s, p2s))
# except FloatingPointError:
# print(self.body.data)
# print(p1s)
# print(p2s)
scale = scale_factor / mean_distance # scale all points to dist/scale
if round(scale, 5) != 1:
self.body.data = ma.multiply(self.body.data, scale)
self.body.data = ma.array(self.body.data, mask=mask)
return self
def b_test():
from numpy import ma as ma
td = 0.33333333333333333
XM = [[-td, -td, 2 * td], [2 * td, 2 * td, -td], [-td, -td, -td]]
Coins = [1000000] * 3
print(ma.multiply(XM, Coins).T.dot(XM))
print(dot(switch_row_cols(matrix_multiply(XM, Coins)), XM))
def findEout(self, numSamples=1000):
e_out = 0
dataSamples = [self.createDataPoint() for _ in range(numSamples)]
for x, y in dataSamples:
e_out += log(1 +
exp(-1 * multiply(y, np.dot(transpose(self.w), x))))
e_out /= float(numSamples)
return e_out
def logicreg_mle_iteration(Y, X, w):
w_result = None
#Terminating control params
w_ctrl = 0.0001
G_ctrl = 0.00001
L_ctrl = 0.000001
#Initialize looping params
w_base = w
P_base = logicreg_func(X, w_base)
#Initialize monitoring params
iteration_round = 1
while True:
#Log-likelihood gradients
G = dot(transpose(X), subtract(Y, P_base))
#Terminating control 1: G ~ 0
if norm(G) < G_ctrl:
print 'G condition met, iteration_round:', iteration_round
w_result = w_base
break
else:
#Sample_diag for Hessian matrix computation
D = diagflat(multiply(P_base, subtract(1, P_base)))
#Hessian matrix H
H = multiply(-1, dot(dot(transpose(X), D), X))
#Each new param W is approximated base on the latest w and sample data set Y, X
w_new = subtract(w_base, dot(logicreg_matrix_inverse(H), G))
#Terminating control 2: w_new ~ w_base
if norm(subtract(w_new, w_base)) < w_ctrl:
print 'w condition met, iteration_round:', iteration_round
w_result = w_new
break
else:
P_new = logicreg_func(X, w_new)
#Terminating control 3: L_new ~ L_base
if abs(logicreg_log_likelihood_func(Y, P_new) - logicreg_log_likelihood_func(Y, P_base)) < L_ctrl:
print 'L condition met, iteration_round:', iteration_round
w_result = w_new
break
else:
#Prepare params for next loop
P_base = P_new
w_base = w_new
iteration_round += 1
continue
return w_result
def plotSpectrogram(self, f, startTime=0):
if self.filePath != '':
samplingFrequency, signalData = f
self.figureSpec.clear()
# Plot the signal read from wav file
ax = self.figureSpec.add_subplot(111)
nfft = int(self.nfftComboBox.currentText())
nover = int(self.noverlapComboBox.currentText())
if self.windowComboBox.currentText() == 'boxcar':
win = signal.get_window('boxcar', len(signalData))
elif self.windowComboBox.currentText() == 'triang':
win = signal.get_window('triang', len(signalData))
elif self.windowComboBox.currentText() == 'blackman':
win = signal.get_window('blackman', len(signalData))
elif self.windowComboBox.currentText() == 'hamming':
win = signal.get_window('hamming', len(signalData))
elif self.windowComboBox.currentText() == 'hann':
win = signal.get_window('hann', len(signalData))
elif self.windowComboBox.currentText() == 'bartlett':
win = signal.get_window('bartlett', len(signalData))
elif self.windowComboBox.currentText() == 'flattop':
win = signal.get_window('flattop', len(signalData))
elif self.windowComboBox.currentText() == 'parzen':
win = signal.get_window('parzen', len(signalData))
elif self.windowComboBox.currentText() == 'nuttall':
win = signal.get_window('nuttall', len(signalData))
elif self.windowComboBox.currentText() == 'taylor':
win = signal.get_window('taylor', len(signalData))
signalData = multiply(signalData, win)
ax.specgram(signalData,
NFFT=nfft,
noverlap=nover,
Fs=samplingFrequency,
cmap=self.colorMap)
ax.set_yscale(self.scaleComboBox.currentText()) #linear or symlog
if self.maxFreqComboBox.currentText() != 'Auto':
ax.set_ylim(
0, int(self.maxFreqComboBox.currentText().split(' ',
1)[0]))
ax.set_ylim(
int(self.minFreqComboBox.currentText().split(' ', 1)[0]), )
ax.set_title('Spectrogram', fontsize=12)
ax.set_xlabel('Time [s]', fontsize=8)
ax.set_ylabel('Frequency [Hz]', fontsize=8)
ax.tick_params(labelsize=7)
self.figureSpec.tight_layout(pad=0.3)
self.figureSpecCanvas.draw()
def b_test():
from numpy import ma as ma
td=0.33333333333333333
XM=[[-td, -td, 2*td],
[2*td, 2*td, -td],
[-td, -td, -td]]
Coins=[1000000]*3
print(ma.multiply(XM, Coins).T.dot(XM))
print(dot(switch_row_cols(matrix_multiply(XM, Coins)), XM))
def heat_equation_FE(n, m, e, umin, umax, theta, alpha, sigma, r):
dx = (umax - umin) / n # Price step
dt = theta / m # time step
vam = initial_option(n, r, m, e, umin, dt, dx, alpha, sigma)
F = dt / dx**2
if F >= 0.5:
print("unstable!!!")
# Implementing the explicit algorithm
for i in range(1, m, 1):
# Checks if early exercise is better for the American Option
vam[1:n - 1, i] = fmax(
vam[1:n - 1, i - 1] + multiply(
F, vam[2:n, i - 1] + vam[0:n - 2, i - 1] -
2 * vam[1:n - 1, i - 1]), vam[1:n - 1, 0])
# Reversal of the time components in the matrix as the solution of the Black Scholes equation was performed
# backwards
vam = fliplr(vam)
return vam
def applyDownscaling(self, emissivity_image_100m, mean_emissivity_100m):
#emissivity_image = self.calcEmissivitySobrino()
# ******** MODIS LST (1km) ***********
lst_image = Image(
self.modis_image.lst.split(".")[0] + '_subdivided_100m.tif')
modis_array = lst_image.getArray(masked=True,
lower_valid_range=7500,
upper_valid_range=65535)
# convertir à des températures de surface en Celsius
lst_metadata = lst_image.getMetadata()
# vérifier si le scale_factor est présent dans les métadonnées (c'est le cas pour AppEEARS, pas EarthData)
if 'scale_factor' in lst_metadata:
scale_factor = float(
lst_metadata['scale_factor']) # multiplier par 0.02
add_offset = float(lst_metadata['add_offset'])
else:
scale_factor = float(0.02)
add_offset = float(0)
# conversion en Kelvin, puis en Celsius
kelvin_array = np.add(np.multiply(modis_array, scale_factor),
add_offset)
lst_celsius_array = np.subtract(kelvin_array, 273.15)
# apply PBIM formula (T_high = T_low * emissivity_high / emissivity_avg) for each pixel
# ********** Émissivité (100m) *********
emissivity = Image(emissivity_image_100m).getArray(masked=False)
mean_emissivity = Image(mean_emissivity_100m).getArray(masked=False)
# PBIM formula
t_high = ma.divide(ma.multiply(lst_celsius_array, emissivity),
mean_emissivity)
Image(emissivity_image_100m).save_band(
t_high, r'secteur3/PBIM_100m_result.tif')
def forward_backward(evidence_sequence, prior):
"""Does the forwards-backwards-algorithm with the given evidence sequence and prior"""
t = len(evidence_sequence)
forward_messages = [] # forward messages
b = matrix([1, 1])
b = transpose(b) # B needs to be as columns for matrix operations.
smoothed_estimates = [0] * (t + 1) #
forward_messages.append(prior)
for i in range(1, t + 1):
forward_messages.append(forward(forward_messages[i - 1], evidence_sequence[i - 1]))
print("forwarded= %s" % forward_messages)
for j in range(t, 1, -1):
normal = multiply(forward_messages[j], b)
normal = normalize(transpose(normal)) # Normalizing requires rows.
smoothed_estimates[j] = normal
b = backward(b, evidence_sequence[j - 1])
print("Backward message: %s" % b)
return smoothed_estimates
def WeightedCov(self, votes_filled):
"""Weights are the number of coins people start with, so the aim of this
weighting is to count 1 vote for each of their coins -- e.g., guy with 10
coins effectively gets 10 votes, guy with 1 coin gets 1 vote, etc.
http://stats.stackexchange.com/questions/61225/correct-equation-for-weighted-unbiased-sample-covariance
"""
# Compute the weighted mean (of all voters) for each decision
weighted_mean = ma.average(votes_filled,
axis=0,
weights=self.rep_coins.squeeze())
# Each vote's difference from the mean of its decision (column)
mean_deviation = np.matrix(votes_filled - weighted_mean)
# Compute the unbiased weighted population covariance
# (for uniform weights, equal to np.cov(votes_filled.T, bias=1))
covariance_matrix = 1/float(np.sum(self.rep_coins)-1) * ma.multiply(mean_deviation, self.rep_coins).T.dot(mean_deviation)
return covariance_matrix, mean_deviation
def forward_backward(evidence_sequence, prior):
"""Does the forwards-backwards-algorithm with the given evidence sequence and prior"""
t = len(evidence_sequence)
forward_messages = [] # forward messages
b = matrix([1, 1])
b = transpose(b) # B needs to be as columns for matrix operations.
smoothed_estimates = [0] * (t + 1) #
forward_messages.append(prior)
for i in range(1, t + 1):
forward_messages.append(
forward(forward_messages[i - 1], evidence_sequence[i - 1]))
print("forwarded= %s" % forward_messages)
for j in range(t, 1, -1):
normal = multiply(forward_messages[j], b)
normal = normalize(transpose(normal)) # Normalizing requires rows.
smoothed_estimates[j] = normal
b = backward(b, evidence_sequence[j - 1])
print("Backward message: %s" % b)
return smoothed_estimates
def l2norm_weighted(values, overall_scale, term_weights):
"""Calculates scaled and weighted Euclidean distance.
Calculated distance is of form: scale * sqrt((a1*a)**2 + (b1*b)**2 + ...)
where a, b, ... are terms to be summed and a1, a2, ... are optional weights
for the terms.
Args:
values (tuple): Tuple containing the values.
overall_scale (float): Scale factor for the calculated
Euclidean distance.
term_weights (tuple): Weights for the terms. Must be single
float or a list of numbers (one per term).
Returns:
float: Scaled and weighted Euclidean distance.
TODO: Probably better use masked arrays instead of tuples.
"""
weighted_values = ma.multiply(values, term_weights)
return overall_scale * l2norm(*weighted_values)
def l2norm_weighted(values: tuple, overall_scale: float,
term_weights: tuple) -> ma.MaskedArray:
"""Calculates scaled and weighted Euclidean distance.
Calculated distance is of form: scale * sqrt((a1*a)**2 + (b1*b)**2 + ...)
where a, b, ... are terms to be summed and a1, a2, ... are optional weights
for the terms.
Args:
values: Tuple containing the values.
overall_scale: Scale factor for the calculated Euclidean distance.
term_weights: Weights for the terms. Must be single float or a list of numbers
(one per term).
Returns:
Scaled and weighted Euclidean distance.
TODO: Use masked arrays instead of tuples.
"""
generic_values = ma.array(values, dtype=object)
weighted_values = ma.multiply(generic_values, term_weights)
return overall_scale * l2norm(*weighted_values)
def test_testArithmetic(self):
# Test of basic arithmetic.
(x, y, a10, m1, m2, xm, ym, z, zm, xf, s) = self.d
a2d = array([[1, 2], [0, 4]])
a2dm = masked_array(a2d, [[0, 0], [1, 0]])
assert_(eq(a2d * a2d, a2d * a2dm))
assert_(eq(a2d + a2d, a2d + a2dm))
assert_(eq(a2d - a2d, a2d - a2dm))
for s in [(12,), (4, 3), (2, 6)]:
x = x.reshape(s)
y = y.reshape(s)
xm = xm.reshape(s)
ym = ym.reshape(s)
xf = xf.reshape(s)
assert_(eq(-x, -xm))
assert_(eq(x + y, xm + ym))
assert_(eq(x - y, xm - ym))
assert_(eq(x * y, xm * ym))
with np.errstate(divide='ignore', invalid='ignore'):
assert_(eq(x / y, xm / ym))
assert_(eq(a10 + y, a10 + ym))
assert_(eq(a10 - y, a10 - ym))
assert_(eq(a10 * y, a10 * ym))
with np.errstate(divide='ignore', invalid='ignore'):
assert_(eq(a10 / y, a10 / ym))
assert_(eq(x + a10, xm + a10))
assert_(eq(x - a10, xm - a10))
assert_(eq(x * a10, xm * a10))
assert_(eq(x / a10, xm / a10))
assert_(eq(x ** 2, xm ** 2))
assert_(eq(abs(x) ** 2.5, abs(xm) ** 2.5))
assert_(eq(x ** y, xm ** ym))
assert_(eq(np.add(x, y), add(xm, ym)))
assert_(eq(np.subtract(x, y), subtract(xm, ym)))
assert_(eq(np.multiply(x, y), multiply(xm, ym)))
with np.errstate(divide='ignore', invalid='ignore'):
assert_(eq(np.divide(x, y), divide(xm, ym)))
def test_testArithmetic(self):
# Test of basic arithmetic.
(x, y, a10, m1, m2, xm, ym, z, zm, xf, s) = self.d
a2d = array([[1, 2], [0, 4]])
a2dm = masked_array(a2d, [[0, 0], [1, 0]])
assert_(eq(a2d * a2d, a2d * a2dm))
assert_(eq(a2d + a2d, a2d + a2dm))
assert_(eq(a2d - a2d, a2d - a2dm))
for s in [(12,), (4, 3), (2, 6)]:
x = x.reshape(s)
y = y.reshape(s)
xm = xm.reshape(s)
ym = ym.reshape(s)
xf = xf.reshape(s)
assert_(eq(-x, -xm))
assert_(eq(x + y, xm + ym))
assert_(eq(x - y, xm - ym))
assert_(eq(x * y, xm * ym))
with np.errstate(divide='ignore', invalid='ignore'):
assert_(eq(x / y, xm / ym))
assert_(eq(a10 + y, a10 + ym))
assert_(eq(a10 - y, a10 - ym))
assert_(eq(a10 * y, a10 * ym))
with np.errstate(divide='ignore', invalid='ignore'):
assert_(eq(a10 / y, a10 / ym))
assert_(eq(x + a10, xm + a10))
assert_(eq(x - a10, xm - a10))
assert_(eq(x * a10, xm * a10))
assert_(eq(x / a10, xm / a10))
assert_(eq(x ** 2, xm ** 2))
assert_(eq(abs(x) ** 2.5, abs(xm) ** 2.5))
assert_(eq(x ** y, xm ** ym))
assert_(eq(np.add(x, y), add(xm, ym)))
assert_(eq(np.subtract(x, y), subtract(xm, ym)))
assert_(eq(np.multiply(x, y), multiply(xm, ym)))
with np.errstate(divide='ignore', invalid='ignore'):
assert_(eq(np.divide(x, y), divide(xm, ym)))
def lspolyfit(myTime,
data,
order,
mode,
thresholdMask=None,
firstNan=None,
fitColumnIndex=None,
fitColumnMask=None,
full=False,
alpha=0.95,
tryLowerOrders=False):
if thresholdMask is None:
# well use all data points regardless
thresholdMask = np.zeros(data.shape, dtype=np.int8)
firstNan = np.int64(np.ones(data.shape[1]) * len(myTime))
if mode == naming.FAST_CAMERA_MODE:
# fast mode does not use the first frame for fit
data = data[1:, :]
# time vector is shorter but must maintain the values
orgMyTime = myTime
myTime = myTime[1:]
else:
orgMyTime = myTime
# masked arrays can be multi dimensional
coef = np.zeros((order + 1, data.shape[1]))
coefMask = np.ones((order + 1, data.shape[1]))
if full:
# diagonal has same mask as coef
myDiag = np.zeros((order + 1, data.shape[1]))
lengthVec = np.zeros(data.shape[1])
if not tryLowerOrders:
# Vandermonde matrices for legth N of time series are
# just the first N lines of the full matrix
van = np.vander(myTime, order + 1)
# first list entry contains array with all columns
# that can fit full order
# unmask everything that can do full order
coefMask[:, fitColumnIndex[0]] = 0
# now we need to loop through all the possible lengths
# of time vectors
goodLengths = np.int8(np.unique(firstNan[fitColumnIndex[0]]))
# print(goodLengths)
for k in goodLengths:
# print('looping through all good lengths', k)
tmpMask = np.where(firstNan == k,
np.zeros(firstNan.shape, dtype=np.int8), 1)
# combine threshold and fitColumnMask
tmpMask = thresholdMask + tmpMask
tmpMask[tmpMask == 2] = 1
# mask everything that cannot do full order
yMa = ma.array(data, mask=tmpMask)
cov = np.linalg.inv(np.dot(van[:k].T, van[:k]))
coef = coef + ma.dot(cov, ma.dot(van[:k].T, yMa[:k])).data
if full:
tmpInd = np.where(firstNan == k)[0]
myDiag[:, tmpInd] = np.repeat(np.diag(cov)[:, None],
len(tmpInd),
axis=1)
lengthVec[tmpInd] = k
# masked arrays can be multi dimensional
coef = ma.array(coef, mask=coefMask)
if not full:
return coef
myDiag = ma.array(myDiag, mask=coefMask)
lengthVec = ma.array(lengthVec, mask=fitColumnMask[0])
dataMask = thresholdMask + fitColumnMask[0]
dataMask[dataMask == 2] = 1
yMa = ma.array(data, mask=dataMask)
# covariance matrix for parameters is (X.T X)-1 * sig2
# where sig2 is variance of noise
# use variance of residuals to estimate sig2
# n number of points, p degree of polynomial + 1
# sig2 = 1/(n-p)*sum(res_i^2)
# coeffsig_i = c*sqrt(sig2*diag((X.T X)-1)_ii)
# estimating variance for fit parameters
fitX = np.tile(myTime, (yMa.shape[1], 1))
fit = np.polyval(coef, fitX.T)
res = (yMa - fit)
if mode == naming.FAST_CAMERA_MODE:
# for fast mode redo the fit so it matches input data shape
# and add zeros to residuals
fitX = np.tile(orgMyTime, (yMa.shape[1], 1))
fit = np.polyval(coef, fitX.T)
res = res.insert(res, 0, axis=0)
yRes = yMa - ma.mean(yMa)
# lengthVec - order -1 is effective degree of freedom
effDf = lengthVec - (order + 1)
rSquared = 1 - (ma.sum(res**2, axis=0)) / (ma.sum(yRes**2, axis=0))
# r2Adj = 1 - (1-r2) * (lengthVec -1)/(effDf - 1)
sig2 = 1 / (effDf) * ma.sum(res**2, axis=0)
tValue = stats.t.ppf(alpha, effDf)
tValue = ma.array(tValue, mask=fitColumnMask[0])
coefSig = tValue * ma.sqrt(ma.multiply(myDiag, sig2))
# tSquared = ma.divide(coef**2, (coefSig/tValue)**2)
# significanHighestOrder = tSquared[0] < tValue**2
# print(tValue**2)
# print(tSquared[0])
# print(significanHighestOrder)
return coef, coefSig, fit, res, sig2, rSquared
# else try for all other orders
# Vandermonde matrices for legth N of time series are
# just the first N lines of the full matrix
# we need different matrices for different lower orders
van = []
for i in range(order, 0, -1):
print(i)
van.append(np.vander(myTime, i + 1))
print(van)
return van
def validationExterne():
""" Permet d'effectuer une validation externe entre l'image résultante de la réduction d'échelle et une image de
température de surface calculée à partir des bandes 10 et 11 de Landsat 8 (disponibles sur EarthData).
Les résultats de la validation externe sont des métriques de qualité en comparant les résultats de la réduction
d'échelle à la température de surface calculée à 100m. Ces résultats sont présentés dans la console par des
'print' (lignes 129 à 144).
"""
# Match prediction result extent
landsat_b10 = Image(
r'data/LC08_L1TP_014028_20200706_20200721_01_T1_B10.TIF')
landsat_b10.reprojectMatch(
r'data/MOD11_L2.clipped_test2.tif'.split(".")[0] +
'_subdivided_100m.tif', False)
landsat_b10.setNewFile(
landsat_b10.filename.replace(".TIF", "_reproject.tif"))
# Get TOA radiance
b10_array = landsat_b10.getArray(masked=True,
lower_valid_range=1,
upper_valid_range=65535)
b10_array_radiance = ma.add(ma.multiply(b10_array, 0.00033420), 0.10000)
# Get Brightness Temperature
b10_array_brightness_temp = (1321.0789 / (ma.log(
(774.8853 / b10_array_radiance) + 1))) - 273.15
# Get NDVI
landsat_b4 = Image(
r'data/LC08_L1TP_014028_20200706_20200721_01_T1_B4_reproject.tif')
b4_DN = landsat_b4.getArray(masked=True,
lower_valid_range=1,
upper_valid_range=65535)
b4 = np.add(np.multiply(b4_DN, float(0.00002)), float(-0.10))
landsat_b5 = Image(
r'data/LC08_L1TP_014028_20200706_20200721_01_T1_B5_reproject.tif')
b5_DN = landsat_b5.getArray(masked=True,
lower_valid_range=1,
upper_valid_range=65535)
b5 = np.add(np.multiply(b5_DN, float(0.00002)), float(-0.10))
ndvi = np.divide(np.subtract(b5, b4),
np.add(b5, b4),
where=((np.add(b5, b4)) != 0))
# Get proportion of vegetation
min_ndvi = ma.amin(ndvi)
max_ndvi = ma.amax(ndvi)
pv = ma.power(
ma.divide(ma.subtract(ndvi, min_ndvi),
(ma.subtract(max_ndvi, min_ndvi)),
where=(ma.subtract(max_ndvi, min_ndvi)) != 0), 2)
# Get emissivity
emissivity = 0.004 * pv + 0.986
# Get Landsat 8 LST
landsat_lst = b10_array_brightness_temp / (
1 +
(0.00115 * b10_array_brightness_temp / 1.4388) * ma.log(emissivity))
# Save LST image for visualization
landsat_b10.save_band(landsat_lst, r'data/landsat_lst.tif')
# Validation between both arrays
predicted_lst = ma.masked_invalid(
Image(r'data/MODIS_predit_100m.tif').getArray())
predicted_lst_with_residuals = ma.masked_invalid(
Image(r'data/MODIS_predit_100m_avec_residus.tif').getArray())
predicted_lst = ma.filled(predicted_lst, 0)
predicted_lst_with_residuals = ma.filled(predicted_lst_with_residuals, 0)
# Without residuals
print('Without residual correction')
print('Mean Absolute Error (MAE):',
metrics.mean_absolute_error(predicted_lst, landsat_lst))
print('Mean Squared Error:',
metrics.mean_squared_error(predicted_lst, landsat_lst))
print('Root Mean Squared Error:',
np.sqrt(metrics.mean_squared_error(predicted_lst, landsat_lst)),
"°C")
print(
'Accuracy:', 100 -
np.mean(100 * ((abs(predicted_lst - landsat_lst)) / landsat_lst)), "%")
print('Explained variance score (EVS):',
metrics.explained_variance_score(predicted_lst, landsat_lst))
# With residuals
print("\n")
print('With residual correction')
print(
'Mean Absolute Error (MAE):',
metrics.mean_absolute_error(predicted_lst_with_residuals, landsat_lst))
print(
'Mean Squared Error:',
metrics.mean_squared_error(predicted_lst_with_residuals, landsat_lst))
print(
'Root Mean Squared Error:',
np.sqrt(
metrics.mean_squared_error(predicted_lst_with_residuals,
landsat_lst)), "°C")
print(
'Accuracy:', 100 - np.mean(100 * (
(abs(predicted_lst_with_residuals - landsat_lst)) / landsat_lst)),
"%")
print(
'Explained variance score (EVS):',
metrics.explained_variance_score(predicted_lst_with_residuals,
landsat_lst))
def rpBound(n):
return sqrt(
divide(multiply(2, log(multiply(multiply(
2, n), power(n, 50)))), n)) + sqrt(
multiply(divide(2, n), log(divide(1, 0.05)))) + divide(1, n)
def vcBound(n):
return sqrt(
multiply(divide(8.0, n),
log(multiply(4, divide(power(multiply(2, n), 50), 0.05)))))
def devroyeBound(n):
# Ran into overflow error performing naive calculation. Had to decompose natural log components.
return divide(1, (n - 2.0)) + sqrt(
divide(1, power(n - 2.0, 2)) +
multiply(divide(1, multiply(2, (n - 2.0))),
log(4) + multiply(100, log(n)) - log(0.05)))
