def setTrainingData(self, state, factors, output, mode='All', samples=None):
'''
@param state Raster of the current state (classes) values.
@param factors List of the factor rasters (predicting variables).
@param output Raster that contains classes to predict.
@param mode Type of sampling method:
All Get all pixels
Normal Get samples. Count of samples in the data=samples.
Balanced Undersampling of major classes and/or oversampling of minor classes.
@samples Sample count of the training data (doesn't used in 'All' mode).
'''
if not self.logreg:
raise LRError('You must create a Logistic Regression model before!')
# Normalize factors before sampling:
for f in factors:
f.normalize(mode = 'mean')
sampler = Sampler(state, factors, output, ns=self.ns)
sampler.setTrainingData(state, factors, output, shuffle=False, mode=mode, samples=samples)
outputVecLen = sampler.outputVecLen
stateVecLen = sampler.stateVecLen
factorVectLen = sampler.factorVectLen
size = len(sampler.data)
self.data = sampler.data
def setTrainingData(self, state, factors, output, shuffle=True, mode='All', samples=None):
'''
@param state Raster of the current state (classes) values.
@param factors List of the factor rasters (predicting variables).
@param output Raster that contains classes to predict.
@param shuffle Perform random shuffle.
@param mode Type of sampling method:
All Get all pixels
Normal Get samples. Count of samples in the data=samples.
Balanced Undersampling of major classes and/or oversampling of minor classes.
@samples Sample count of the training data (doesn't used in 'All' mode).
'''
if not self.MLP:
raise MlpManagerError('You must create a MLP before!')
# Normalize factors before sampling:
for f in factors:
f.normalize(mode = 'mean')
sampler = Sampler(state, factors, output, self.ns)
sampler.setTrainingData(state, factors, output, shuffle, mode, samples)
outputVecLen = self.getOutputVectLen()
stateVecLen = sampler.stateVecLen
factorVectLen = sampler.factorVectLen
size = len(sampler.data)
self.data = np.zeros(size, dtype=[('state', float, stateVecLen), ('factors', float, factorVectLen), ('output', float, outputVecLen)])
self.data['state'] = sampler.data['state']
self.data['factors'] = sampler.data['factors']
self.data['output'] = [self.getOutputVector(sample['output']) for sample in sampler.data]
def setTrainingData(self):
state, factors, output, mode, samples = self.state, self.factors, self.output, self.mode, self.samples
if not self.logreg:
raise LRError(
'You must create a Logistic Regression model before!')
# Normalize factors before sampling:
for f in factors:
f.normalize(mode='mean')
self.sampler = Sampler(state, factors, output, ns=self.ns)
self.__propagateSamplerSignals()
self.sampler.setTrainingData(state,
output,
shuffle=False,
mode=mode,
samples=samples)
outputVecLen = self.sampler.outputVecLen
stateVecLen = self.sampler.stateVecLen
factorVectLen = self.sampler.factorVectLen
size = len(self.sampler.data)
self.data = self.sampler.data
self.catlist = np.unique(self.data['output'])
def test_get_state(self):
smp = Sampler(self.state, self.factors, self.output, ns=0)
assert_array_equal(smp.get_state(self.state, 1, 1), [0, 0])
assert_array_equal(smp.get_state(self.state, 0, 0), [0, 1])
smp = Sampler(self.state, self.factors, self.output, ns=1)
res = [
0,
1,
0,
0,
0,
1,
0,
1,
0,
0,
0,
1,
1,
0,
0,
1,
0,
0,
]
assert_array_equal(smp.get_state(self.state, 1, 1), res)
inputRast = Raster('../../examples/sites.tif')
inputRast.resetMask([0])
smp = Sampler(inputRast, self.factors, self.output, ns=0)
assert_array_equal(smp.get_state(self.state, 1, 1), [0])
assert_array_equal(smp.get_state(self.state, 0, 0), [1])
def test_cat2vect(self):
smp = Sampler(self.state, self.factors, self.output, ns=0)
assert_array_equal(smp.cat2vect(0), [1, 0])
assert_array_equal(smp.cat2vect(1), [0, 1])
assert_array_equal(smp.cat2vect(2), [0, 0])
inputRast = Raster('../../examples/sites.tif')
inputRast.resetMask([0])
smp = Sampler(inputRast, self.factors, self.output, ns=0)
assert_array_equal(smp.cat2vect(1), [1])
assert_array_equal(smp.cat2vect(2), [0])
def setTrainingData(self,
state,
factors,
output,
shuffle=True,
mode='All',
samples=None):
'''
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
@param output Raster that contains categories to predict.
@param shuffle Perform random shuffle.
@param mode Type of sampling method:
All Get all pixels
Random Get samples. Count of samples in the data=samples.
Stratified Undersampling of major categories and/or oversampling of minor categories.
@samples Sample count of the training data (doesn't used in 'All' mode).
'''
if not self.MLP:
raise MlpManagerError('You must create a MLP before!')
# Normalize factors before sampling:
for f in factors:
f.normalize(mode='mean')
self.sampler = Sampler(state, factors, output, self.ns)
self.sampler.setTrainingData(state=state,
output=output,
shuffle=shuffle,
mode=mode,
samples=samples)
outputVecLen = self.getOutputVectLen()
stateVecLen = self.sampler.stateVecLen
factorVectLen = self.sampler.factorVectLen
size = len(self.sampler.data)
self.data = np.zeros(size,
dtype=[('coords', float, 2),
('state', float, stateVecLen),
('factors', float, factorVectLen),
('output', float, outputVecLen)])
self.data['coords'] = self.sampler.data['coords']
self.data['state'] = self.sampler.data['state']
self.data['factors'] = self.sampler.data['factors']
self.data['output'] = [
self.getOutputVector(sample['output'])
for sample in self.sampler.data
]
def _predict(self, state, factors):
'''
Calculate output and confidence rasters using MLP model and input rasters
@param state Raster of the current state (classes) values.
@param factors List of the factor rasters (predicting variables).
'''
geodata = state.getGeodata()
rows, cols = geodata['ySize'], geodata['xSize']
for r in factors:
if not state.geoDataMatch(r):
raise MlpManagerError('Geometries of the input rasters are different!')
# Normalize factors before prediction:
for f in factors:
f.normalize(mode = 'mean')
predicted_band = np.zeros([rows, cols])
confidence_band = np.zeros([rows, cols])
sampler = Sampler(state, factors, ns=self.ns)
mask = state.getBand(1).mask.copy()
for i in xrange(rows):
for j in xrange(cols):
if not mask[i,j]:
input = sampler.get_inputs(state, factors, i,j)
if input != None:
out = self.getOutput(input)
# Get index of the biggest output value as the result
biggest = max(out)
res = list(out).index(biggest)
predicted_band[i, j] = self.classlist[res]
confidence = self.outputConfidence(out)
confidence_band[i, j] = confidence
else: # Input sample is incomplete => mask this pixel
mask[i, j] = True
predicted_bands = [np.ma.array(data = predicted_band, mask = mask)]
confidence_bands = [np.ma.array(data = confidence_band, mask = mask)]
self.prediction = Raster()
self.prediction.create(predicted_bands, geodata)
self.confidence = Raster()
self.confidence.create(confidence_bands, geodata)
def test_cat2vect(self):
smp = Sampler(self.state, self.factors, self.output, ns=0)
assert_array_equal(smp.cat2vect(0), [1, 0])
assert_array_equal(smp.cat2vect(1), [0, 1])
assert_array_equal(smp.cat2vect(2), [0, 0])
inputRast = Raster('../../examples/sites.tif')
inputRast.resetMask([0])
smp = Sampler(inputRast, self.factors, self.output, ns=0)
assert_array_equal(smp.cat2vect(1), [1])
assert_array_equal(smp.cat2vect(2), [0])
def setTrainingData(self):
state, factors, output, mode, samples = self.state, self.factors, self.output, self.mode, self.samples
if not self.logreg:
raise LRError('You must create a Logistic Regression model before!')
# Normalize factors before sampling:
for f in factors:
f.normalize(mode = 'mean')
self.sampler = Sampler(state, factors, output, ns=self.ns)
self.__propagateSamplerSignals()
self.sampler.setTrainingData(state, output, shuffle=False, mode=mode, samples=samples)
outputVecLen = self.sampler.outputVecLen
stateVecLen = self.sampler.stateVecLen
factorVectLen = self.sampler.factorVectLen
size = len(self.sampler.data)
self.data = self.sampler.data
self.catlist = np.unique(self.data['output'])
def test_get_state(self):
smp = Sampler(self.state, self.factors, self.output, ns=0)
assert_array_equal(smp.get_state(self.state, 1,1), [0, 0])
assert_array_equal(smp.get_state(self.state, 0,0), [0, 1])
smp = Sampler(self.state, self.factors, self.output, ns=1)
res = [ 0, 1, 0, 0, 0, 1,
0, 1, 0, 0, 0, 1,
1, 0, 0, 1, 0, 0,
]
assert_array_equal(smp.get_state(self.state, 1,1), res)
inputRast = Raster('../../examples/sites.tif')
inputRast.resetMask([0])
smp = Sampler(inputRast, self.factors, self.output, ns=0)
assert_array_equal(smp.get_state(self.state, 1,1), [0])
assert_array_equal(smp.get_state(self.state, 0,0), [1])
class MlpManager(QObject):
'''This class gets the data extracted from the UI and
pass it to multi-layer perceptron, then gets and stores the result.
'''
updateGraph = pyqtSignal(float, float) # Train error, val. error
updateMinValErr = pyqtSignal(float) # Min validation error
updateDeltaRMS = pyqtSignal(
float) # Delta of RMS: min(valError) - currentValError
updateKappa = pyqtSignal(float) # Kappa value
processFinished = pyqtSignal()
processInterrupted = pyqtSignal()
logMessage = pyqtSignal(str)
errorReport = pyqtSignal(str)
rangeChanged = pyqtSignal(str, int)
updateProgress = pyqtSignal()
def __init__(self, ns=0, MLP=None):
QObject.__init__(self)
self.MLP = MLP
self.interrupted = False
self.layers = None
if self.MLP:
self.layers = self.getMlpTopology()
self.ns = ns # Neighbourhood size of training rasters.
self.data = None # Training data
self.catlist = None # List of unique output values of the output raster
self.train_error = None # Error on training set
self.val_error = None # Error on validation set
self.minValError = None # The minimum error that is achieved on the validation set
self.valKappa = 0 # Kappa on on the validation set
self.sampler = None # Sampler
# Results of the MLP prediction
self.prediction = None # Raster of the MLP prediction results
self.confidence = None # Raster of the MLP results confidence (1 = the maximum confidence, 0 = the least confidence)
self.transitionPotentials = None # Dictionary of transition potencial maps: {category1: map1, category2: map2, ...}
# Outputs of the activation function for small and big numbers
self.sigmax, self.sigmin = sigmoid(100), sigmoid(
-100) # Max and Min of the sigmoid function
self.sigrange = self.sigmax - self.sigmin # Range of the sigmoid
def computeMlpError(self, sample):
'''Get MLP error on the sample'''
input = np.hstack((sample['state'], sample['factors']))
out = self.getOutput(input)
err = ((sample['output'] - out)**2).sum() / len(out)
return err
def computePerformance(self, train_indexes, val_ind):
'''Check errors of training and validation sets
@param train_indexes Tuple that contains indexes of the first and last elements of the training set.
@param val_ind Tuple that contains indexes of the first and last elements of the validation set.
'''
train_error = 0
train_sampl = train_indexes[1] - train_indexes[
0] # Count of training samples
for i in range(train_indexes[0], train_indexes[1]):
train_error = train_error + self.computeMlpError(
sample=self.data[i])
self.setTrainError(train_error / train_sampl)
if val_ind:
val_error = 0
val_sampl = val_ind[1] - val_ind[0]
answers = np.ma.zeros(val_sampl)
out = np.ma.zeros(val_sampl)
for i in xrange(val_ind[0], val_ind[1]):
sample = self.data[i]
val_error = val_error + self.computeMlpError(
sample=self.data[i])
input = np.hstack((sample['state'], sample['factors']))
output = self.getOutput(input)
out[i - val_ind[0]] = self.outCategory(output)
answers[i - val_ind[0]] = self.outCategory(sample['output'])
self.setValError(val_error / val_sampl)
depCoef = DependenceCoef(out, answers, expand=True)
self.valKappa = depCoef.kappa(mode=None)
def copyWeights(self):
'''Deep copy of the MLP weights'''
return copy.deepcopy(self.MLP.weights)
def createMlp(self, state, factors, output, hidden_layers):
'''
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
@param hidden_layers List of neuron counts in hidden layers.
@param ns Neighbourhood size.
'''
if output.getBandsCount() != 1:
raise MlpManagerError('Output layer must have one band!')
input_neurons = 0
for raster in factors:
input_neurons = input_neurons + raster.getNeighbourhoodSize(
self.ns)
# state raster contains categories. We need use n-1 dummy variables (where n = number of categories)
input_neurons = input_neurons + (len(state.getBandGradation(1)) -
1) * state.getNeighbourhoodSize(
self.ns)
# Output category's (neuron) list and count
self.catlist = output.getBandGradation(1)
categories = len(self.catlist)
# set neuron counts in the MLP layers
self.layers = hidden_layers
self.layers.insert(0, input_neurons)
self.layers.append(categories)
self.MLP = MLP(*self.layers)
def getConfidence(self):
return self.confidence
def getInputVectLen(self):
'''Length of input data vector of the MLP'''
shape = self.getMlpTopology()
return shape[0]
def getOutput(self, input_vector):
out = self.MLP.propagate_forward(input_vector)
return out
def getOutputVectLen(self):
'''Length of input data vector of the MLP'''
shape = self.getMlpTopology()
return shape[-1]
def getOutputVector(self, val):
'''Convert a number val into vector,
for example, let self.catlist = [1, 3, 4] then
if val = 1, result = [ 1, -1, -1]
if val = 3, result = [-1, 1, -1]
if val = 4, result = [-1, -1, 1]
where -1 is minimum of the sigmoid, 1 is max of the sigmoid
'''
size = self.getOutputVectLen()
res = np.ones(size) * (self.sigmin)
ind = np.where(self.catlist == val)
res[ind] = self.sigmax
return res
def getMinValError(self):
return self.minValError
def getMlpTopology(self):
return self.MLP.shape
def getKappa(self):
return self.valKappa
def getPrediction(self, state, factors, calcTransitions=False):
self._predict(state, factors, calcTransitions)
return self.prediction
def getTrainError(self):
return self.train_error
def getTransitionPotentials(self):
return self.transitionPotentials
def getValError(self):
return self.val_error
def outCategory(self, out_vector):
# Get index of the biggest output value as the result
biggest = max(out_vector)
res = list(out_vector).index(biggest)
res = self.catlist[res]
return res
def outputConfidence(self, output, scale=True):
'''
Return confidence (difference between 2 biggest values) of the MLP output.
@param output: The confidence
@param scale: If True, then scale the confidence to int [0, 1, ..., 100] percent
'''
out_scl = self.scaleOutput(output, percent=scale)
out_scl.sort()
return out_scl[-1] - out_scl[-2]
def outputTransitions(self, output, scale=True):
'''
Return transition potencial of the outputs scaled to [0,1] or 1-100
@param output: The output of MLP
@param scale: If True, then scale the transitions to int ([0, 1, ..., 100]) percent
'''
out_scl = self.scaleOutput(output, percent=scale)
result = {}
for r, v in enumerate(out_scl):
cat = self.catlist[r]
result[cat] = v
return result
def scaleOutput(self, output, percent=True):
'''
Scale the output to range [0,1] or 1-100
@param output: Output of a MLP
@param percent: If True, then scale the output to int [0, 1, ..., 100] percent
'''
res = 1.0 * (output - self.sigmin) / self.sigrange
if percent:
res = [int(100 * x) for x in res]
return res
def _predict(self, state, factors, calcTransitions=False):
'''
Calculate output and confidence rasters using MLP model and input rasters
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
'''
try:
self.rangeChanged.emit(self.tr("Initialize model %p%"), 1)
geodata = state.getGeodata()
rows, cols = geodata['ySize'], geodata['xSize']
for r in factors:
if not state.geoDataMatch(r):
raise MlpManagerError(
'Geometries of the input rasters are different!')
self.transitionPotentials = None # Reset tr.potentials if they exist
# Normalize factors before prediction:
for f in factors:
f.normalize(mode='mean')
predicted_band = np.zeros([rows, cols], dtype=np.uint8)
confidence_band = np.zeros([rows, cols], dtype=np.uint8)
if calcTransitions:
self.transitionPotentials = {}
for cat in self.catlist:
self.transitionPotentials[cat] = np.zeros([rows, cols],
dtype=np.uint8)
self.sampler = Sampler(state, factors, ns=self.ns)
mask = state.getBand(1).mask.copy()
if mask.shape == ():
mask = np.zeros([rows, cols], dtype=np.bool)
self.updateProgress.emit()
self.rangeChanged.emit(self.tr("Prediction %p%"), rows)
for i in xrange(rows):
for j in xrange(cols):
if not mask[i, j]:
input = self.sampler.get_inputs(state, i, j)
if input != None:
out = self.getOutput(input)
res = self.outCategory(out)
predicted_band[i, j] = res
confidence = self.outputConfidence(out)
confidence_band[i, j] = confidence
if calcTransitions:
potentials = self.outputTransitions(out)
for cat in self.catlist:
map = self.transitionPotentials[cat]
map[i, j] = potentials[cat]
else: # Input sample is incomplete => mask this pixel
mask[i, j] = True
self.updateProgress.emit()
predicted_bands = [
np.ma.array(data=predicted_band, mask=mask, dtype=np.uint8)
]
confidence_bands = [
np.ma.array(data=confidence_band, mask=mask, dtype=np.uint8)
]
self.prediction = Raster()
self.prediction.create(predicted_bands, geodata)
self.confidence = Raster()
self.confidence.create(confidence_bands, geodata)
if calcTransitions:
for cat in self.catlist:
band = [
np.ma.array(data=self.transitionPotentials[cat],
mask=mask,
dtype=np.uint8)
]
self.transitionPotentials[cat] = Raster()
self.transitionPotentials[cat].create(band, geodata)
except MemoryError:
self.errorReport.emit(
self.tr("The system out of memory during ANN prediction"))
raise
except:
self.errorReport.emit(
self.tr("An unknown error occurs during ANN prediction"))
raise
def readMlp(self):
pass
def resetErrors(self):
self.val_error = np.finfo(np.float).max
self.train_error = np.finfo(np.float).max
def resetMlp(self):
self.MLP.reset()
self.resetErrors()
def saveMlp(self):
pass
def saveSamples(self, fileName):
self.sampler.saveSamples(fileName)
def setMlpWeights(self, w):
'''Set weights of the MLP'''
self.MLP.weights = w
def setTrainingData(self,
state,
factors,
output,
shuffle=True,
mode='All',
samples=None):
'''
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
@param output Raster that contains categories to predict.
@param shuffle Perform random shuffle.
@param mode Type of sampling method:
All Get all pixels
Random Get samples. Count of samples in the data=samples.
Stratified Undersampling of major categories and/or oversampling of minor categories.
@samples Sample count of the training data (doesn't used in 'All' mode).
'''
if not self.MLP:
raise MlpManagerError('You must create a MLP before!')
# Normalize factors before sampling:
for f in factors:
f.normalize(mode='mean')
self.sampler = Sampler(state, factors, output, self.ns)
self.sampler.setTrainingData(state=state,
output=output,
shuffle=shuffle,
mode=mode,
samples=samples)
outputVecLen = self.getOutputVectLen()
stateVecLen = self.sampler.stateVecLen
factorVectLen = self.sampler.factorVectLen
size = len(self.sampler.data)
self.data = np.zeros(size,
dtype=[('coords', float, 2),
('state', float, stateVecLen),
('factors', float, factorVectLen),
('output', float, outputVecLen)])
self.data['coords'] = self.sampler.data['coords']
self.data['state'] = self.sampler.data['state']
self.data['factors'] = self.sampler.data['factors']
self.data['output'] = [
self.getOutputVector(sample['output'])
for sample in self.sampler.data
]
def setTrainError(self, error):
self.train_error = error
def setValError(self, error):
self.val_error = error
def setEpochs(self, epochs):
self.epochs = epochs
def setValPercent(self, value=20):
self.valPercent = value
def setLRate(self, value=0.1):
self.lrate = value
def setMomentum(self, value=0.01):
self.momentum = value
def setContinueTrain(self, value=False):
self.continueTrain = value
def startTrain(self):
self.train(self.epochs, self.valPercent, self.lrate, self.momentum,
self.continueTrain)
def stopTrain(self):
self.interrupted = True
def train(self,
epochs,
valPercent=20,
lrate=0.1,
momentum=0.01,
continue_train=False):
'''Perform the training procedure on the MLP and save the best neural net
@param epoch Max iteration count.
@param valPercent Percent of the validation set.
@param lrate Learning rate.
@param momentum Learning momentum.
@param continue_train If False then it is new training cycle, reset weights training and validation error. If True, then continue training.
'''
try:
samples_count = len(self.data)
val_sampl_count = samples_count * valPercent / 100
apply_validation = True if val_sampl_count > 0 else False # Use or not use validation set
train_sampl_count = samples_count - val_sampl_count
# Set first train_sampl_count as training set, the other as validation set
train_indexes = (0, train_sampl_count)
val_indexes = (train_sampl_count,
samples_count) if apply_validation else None
if not continue_train: self.resetMlp()
self.minValError = self.getValError(
) # The minimum error that is achieved on the validation set
last_train_err = self.getTrainError()
best_weights = self.copyWeights(
) # The MLP weights when minimum error that is achieved on the validation set
self.rangeChanged.emit(self.tr("Train model %p%"), epochs)
for epoch in range(epochs):
self.trainEpoch(train_indexes, lrate, momentum)
self.computePerformance(train_indexes, val_indexes)
self.updateGraph.emit(self.getTrainError(), self.getValError())
self.updateDeltaRMS.emit(self.getMinValError() -
self.getValError())
self.updateKappa.emit(self.getKappa())
QCoreApplication.processEvents()
if self.interrupted:
self.processInterrupted.emit()
break
last_train_err = self.getTrainError()
self.setTrainError(last_train_err)
if apply_validation and (self.getValError() <
self.getMinValError()):
self.minValError = self.getValError()
best_weights = self.copyWeights()
self.updateMinValErr.emit(self.getMinValError())
self.updateProgress.emit()
self.setMlpWeights(best_weights)
except MemoryError:
self.errorReport.emit(
self.tr("The system out of memory during ANN training"))
raise
except:
self.errorReport.emit(
self.tr("An unknown error occurs during ANN trainig"))
raise
finally:
self.processFinished.emit()
def trainEpoch(self, train_indexes, lrate=0.1, momentum=0.01):
'''Perform a training epoch on the MLP
@param train_ind Tuple of the min&max indexes of training samples in the samples data.
@param val_ind Tuple of the min&max indexes of validation samples in the samples data.
@param lrate Learning rate.
@param momentum Learning momentum.
'''
train_sampl = train_indexes[1] - train_indexes[0]
for i in range(train_sampl):
n = np.random.randint(*train_indexes)
sample = self.data[n]
input = np.hstack((sample['state'], sample['factors']))
self.getOutput(input) # Forward propagation
self.MLP.propagate_backward(sample['output'], lrate, momentum)
def test_setTrainingData(self):
smp = Sampler(self.output, self.factors, self.output, ns=0)
smp.setTrainingData(self.output, self.factors, self.output, shuffle=False)
data = np.array(
[
(1.0, 1.0, 1.0),
(2.0, 1.0, 2.0),
(1.0, 3.0, 1.0),
(1.0, 3.0, 1.0),
(2.0, 2.0, 2.0),
(1.0, 1.0, 1.0),
(0.0, 0.0, 0.0),
(1.0, 3.0, 1.0),
(2.0, 1.0, 2.0),
],
dtype=[('state', float, (1,)), ('factors', float, (1,)), ('output', float, 1)]
)
for i in range(len(data)):
assert_array_equal(data[i]['factors'], smp.data[i]['factors'])
assert_array_equal(data[i]['output'], smp.data[i]['output'])
assert_array_equal(data[i]['state'], smp.data[i]['state'])
# two factor_rasters
smp = Sampler(self.output, self.factors2, self.output, ns=1)
smp.setTrainingData(self.output, self.factors2, self.output)
data = np.array(
[
([1,2,1, 1,2,1, 0,1,2],
[ 1., 1., 3., 3., 2., 1., 0., 3., 1.,
1., 1., 3., 3., 2., 1., 0., 3., 1.],
2.0)
],
dtype=[('state', float, (9,)), ('factors', float, (18,)), ('output', float, 1)]
)
assert_array_equal(data[0]['factors'], smp.data[0]['factors'])
assert_array_equal(data[0]['output'], smp.data[0]['output'])
assert_array_equal(data[0]['state'], smp.data[0]['state'])
# Multiband factors
smp = Sampler(self.output, self.factors3, self.output, ns=1)
smp.setTrainingData(self.output, self.factors3, self.output)
data = np.array(
[
([1,2,1, 1,2,1, 0,1,2],
[ 1., 2., 1., 1., 2., 1., 0., 1., 2.,
1., 1., 3., 3., 2., 1., 0., 3., 1.],
2.0)
],
dtype=[('state', float, (9,)), ('factors', float, (18,)), ('output', float, 1)]
)
assert_array_equal(data[0]['factors'], smp.data[0]['factors'])
assert_array_equal(data[0]['output'], smp.data[0]['output'])
assert_array_equal(data[0]['state'], smp.data[0]['state'])
# Several factor bands, several factor rasters
smp = Sampler(self.output, self.factors4, self.output, ns=1)
smp.setTrainingData(self.output, self.factors4, self.output)
data = np.array(
[
([1,2,1, 1,2,1, 0,1,2],
[ 1., 2., 1., 1., 2., 1., 0., 1., 2.,
1., 1., 3., 3., 2., 1., 0., 3., 1.,
1., 1., 3., 3., 2., 1., 0., 3., 1.],
2.0)
],
dtype=[('state', float, (9,)), ('factors', float, (27,)), ('output', float, 1)]
)
assert_array_equal(data[0]['factors'], smp.data[0]['factors'])
assert_array_equal(data[0]['output'], smp.data[0]['output'])
assert_array_equal(data[0]['state'], smp.data[0]['state'])
# Mode = Normal
# As the Multiband factors example, but 10 samples:
data = np.array(
[
([1,2,1, 1,2,1, 0,1,2],
[ 1., 2., 1., 1., 2., 1., 0., 1., 2.,
1., 1., 3., 3., 2., 1., 0., 3., 1.],
2.0)
],
dtype=[('state', float, (9,)), ('factors', float, (18,)), ('output', float, 1)]
)
smp = Sampler(self.output, self.factors3, self.output, ns=1)
smp.setTrainingData(self.output, self.factors3, self.output, mode='Normal', samples=10)
for i in range(10):
assert_array_equal(data[0]['factors'], smp.data[i]['factors'])
assert_array_equal(data[0]['output'], smp.data[i]['output'])
assert_array_equal(data[0]['state'], smp.data[i]['state'])
# Mode = Balanced
smp = Sampler(self.output, self.factors, self.output, ns=0)
smp.setTrainingData(self.output, self.factors, self.output, mode='Balanced', samples=15)
out = smp.data['output']
out.sort()
self.assertEqual(out[0], 0)
self.assertEqual(out[4], 0)
self.assertEqual(out[5], 1)
self.assertEqual(out[9], 1)
self.assertEqual(out[10], 2)
class LR(QObject):
"""
Implements Logistic Regression model definition and calibration
(maximum liklihood parameter estimation).
"""
rangeChanged = pyqtSignal(str, int)
updateProgress = pyqtSignal()
processFinished = pyqtSignal()
samplingFinished = pyqtSignal()
finished = pyqtSignal()
logMessage = pyqtSignal(str)
errorReport = pyqtSignal(str)
def __init__(self, ns=0, logreg=None):
QObject.__init__(self)
if logreg:
self.logreg = logreg
else:
self.logreg = mlr.MLR()
self.state = None
self.factors = None
self.output = None
self.mode = "All"
self.samples = None
self.catlist = None
self.ns = ns # Neighbourhood size of training rasters.
self.data = None # Training data
self.maxiter = 100 # Maximum of fitting iterations
self.sampler = None # Sampler
# Results of the LR prediction
self.prediction = None # Raster of the LR prediction results
self.confidence = None # Raster of the LR results confidence (1 = the maximum confidence, 0 = the least confidence)
self.Kappa = 0 # Kappa value
self.pseudoR = 0 # Pseudo R-squared (Count) (http://www.ats.ucla.edu/stat/mult_pkg/faq/general/Psuedo_RSquareds.htm)
self.transitionPotentials = None # Dictionary of transition potencial maps: {category1: map1, category2: map2, ...}
def getCoef(self):
return self.logreg.get_weights().T
def getConfidence(self):
return self.confidence
def getIntercept(self):
return self.logreg.get_intercept()
def getKappa(self):
return self.Kappa
def getStdErrIntercept(self):
X = np.column_stack( (self.data['state'], self.data['factors']) )
return self.logreg.get_stderr_intercept(X)
def getStdErrWeights(self):
X = np.column_stack( (self.data['state'], self.data['factors']) )
return self.logreg.get_stderr_weights(X).T
def get_PvalIntercept(self):
X = np.column_stack( (self.data['state'], self.data['factors']) )
return self.logreg.get_pval_intercept(X)
def get_PvalWeights(self):
X = np.column_stack( (self.data['state'], self.data['factors']) )
return self.logreg.get_pval_weights(X).T
def getPrediction(self, state, factors, calcTransitions=False):
self._predict(state, factors, calcTransitions)
return self.prediction
def getPseudoR(self):
return self.pseudoR
def getTransitionPotentials(self):
return self.transitionPotentials
def _outputConfidence(self, input):
'''
Return confidence (difference between 2 biggest probabilities) of the LR output.
1 = the maximum confidence, 0 = the least confidence
'''
out_scl = self.logreg.predict_proba(input)[0]
# Calculate the confidence:
out_scl.sort()
return int(100 * (out_scl[-1] - out_scl[-2]) )
def outputTransitions(self, input):
'''
Return transition potential of the outputs
'''
out_scl = self.logreg.predict_proba(input)[0]
out_scl = [int(100 * x) for x in out_scl]
result = {}
for r, v in enumerate(out_scl):
cat = self.catlist[r]
result[cat] = v
return result
def _predict(self, state, factors, calcTransitions=False):
'''
Calculate output and confidence rasters using LR model and input rasters
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
'''
try:
self.rangeChanged.emit(self.tr("Initialize model %p%"), 1)
geodata = state.getGeodata()
rows, cols = geodata['ySize'], geodata['xSize']
for r in factors:
if not state.geoDataMatch(r):
raise LRError('Geometries of the input rasters are different!')
self.transitionPotentials = None # Reset tr.potentials if they exist
# Normalize factors before prediction:
for f in factors:
f.normalize(mode = 'mean')
predicted_band = np.zeros([rows, cols], dtype=np.uint8)
confidence_band = np.zeros([rows, cols], dtype=np.uint8)
if calcTransitions:
self.transitionPotentials = {}
for cat in self.catlist:
self.transitionPotentials[cat] = np.zeros([rows, cols], dtype=np.uint8)
self.sampler = Sampler(state, factors, ns=self.ns)
mask = state.getBand(1).mask.copy()
if mask.shape == ():
mask = np.zeros([rows, cols], dtype=np.bool)
self.updateProgress.emit()
self.rangeChanged.emit(self.tr("Prediction %p%"), rows)
for i in xrange(rows):
for j in xrange(cols):
if not mask[i,j]:
input = self.sampler.get_inputs(state, i,j)
if input != None:
input = np.array([input])
out = self.logreg.predict(input)
predicted_band[i,j] = out
confidence = self._outputConfidence(input)
confidence_band[i, j] = confidence
if calcTransitions:
potentials = self.outputTransitions(input)
for cat in self.catlist:
map = self.transitionPotentials[cat]
map[i, j] = potentials[cat]
else: # Input sample is incomplete => mask this pixel
mask[i, j] = True
self.updateProgress.emit()
predicted_bands = [np.ma.array(data = predicted_band, mask = mask, dtype=np.uint8)]
confidence_bands = [np.ma.array(data = confidence_band, mask = mask, dtype=np.uint8)]
self.prediction = Raster()
self.prediction.create(predicted_bands, geodata)
self.confidence = Raster()
self.confidence.create(confidence_bands, geodata)
if calcTransitions:
for cat in self.catlist:
band = [np.ma.array(data=self.transitionPotentials[cat], mask=mask, dtype=np.uint8)]
self.transitionPotentials[cat] = Raster()
self.transitionPotentials[cat].create(band, geodata)
except MemoryError:
self.errorReport.emit(self.tr("The system out of memory during LR prediction"))
raise
except:
self.errorReport.emit(self.tr("An unknown error occurs during LR prediction"))
raise
finally:
self.processFinished.emit()
def __propagateSamplerSignals(self):
self.sampler.rangeChanged.connect(self.__samplerProgressRangeChanged)
self.sampler.updateProgress.connect(self.__samplerProgressChanged)
self.sampler.samplingFinished.connect(self.__samplerFinished)
def __samplerFinished(self):
self.sampler.rangeChanged.disconnect(self.__samplerProgressRangeChanged)
self.sampler.updateProgress.disconnect(self.__samplerProgressChanged)
self.sampler.samplingFinished.disconnect(self.__samplerFinished)
self.samplingFinished.emit()
def __samplerProgressRangeChanged(self, message, maxValue):
self.rangeChanged.emit(message, maxValue)
def __samplerProgressChanged(self):
self.updateProgress.emit()
def save(self):
pass
def saveSamples(self, fileName):
self.sampler.saveSamples(fileName)
def setMaxIter(self, maxiter):
self.maxiter = maxiter
def setTrainingData(self):
state, factors, output, mode, samples = self.state, self.factors, self.output, self.mode, self.samples
if not self.logreg:
raise LRError('You must create a Logistic Regression model before!')
# Normalize factors before sampling:
for f in factors:
f.normalize(mode = 'mean')
self.sampler = Sampler(state, factors, output, ns=self.ns)
self.__propagateSamplerSignals()
self.sampler.setTrainingData(state, output, shuffle=False, mode=mode, samples=samples)
outputVecLen = self.sampler.outputVecLen
stateVecLen = self.sampler.stateVecLen
factorVectLen = self.sampler.factorVectLen
size = len(self.sampler.data)
self.data = self.sampler.data
self.catlist = np.unique(self.data['output'])
def train(self):
X = np.column_stack( (self.data['state'], self.data['factors']) )
Y = self.data['output']
self.labelCodes = np.unique(Y)
self.logreg.fit(X, Y, maxiter=self.maxiter)
out = self.logreg.predict(X)
depCoef = DependenceCoef(np.ma.array(out), np.ma.array(Y), expand=True)
self.Kappa = depCoef.kappa(mode=None)
self.pseudoR = depCoef.correctness(percent = False)
def setState(self, state):
self.state = state
def setFactors(self, factors):
self.factors = factors
def setOutput(self, output):
self.output = output
def setMode(self, mode):
self.mode = mode
def setSamples(self, samples):
self.samples = samples
def startTrain(self):
try:
self.setTrainingData()
self.train()
except MemoryError:
self.errorReport.emit(self.tr("The system out of memory during LR training"))
raise
except:
self.errorReport.emit(self.tr("An unknown error occurs during LR trainig"))
raise
finally:
self.finished.emit()
def _predict(self, state, factors, calcTransitions=False):
'''
Calculate output and confidence rasters using LR model and input rasters
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
'''
try:
self.rangeChanged.emit(self.tr("Initialize model %p%"), 1)
geodata = state.getGeodata()
rows, cols = geodata['ySize'], geodata['xSize']
for r in factors:
if not state.geoDataMatch(r):
raise LRError('Geometries of the input rasters are different!')
self.transitionPotentials = None # Reset tr.potentials if they exist
# Normalize factors before prediction:
for f in factors:
f.normalize(mode = 'mean')
predicted_band = np.zeros([rows, cols], dtype=np.uint8)
confidence_band = np.zeros([rows, cols], dtype=np.uint8)
if calcTransitions:
self.transitionPotentials = {}
for cat in self.catlist:
self.transitionPotentials[cat] = np.zeros([rows, cols], dtype=np.uint8)
self.sampler = Sampler(state, factors, ns=self.ns)
mask = state.getBand(1).mask.copy()
if mask.shape == ():
mask = np.zeros([rows, cols], dtype=np.bool)
self.updateProgress.emit()
self.rangeChanged.emit(self.tr("Prediction %p%"), rows)
for i in xrange(rows):
for j in xrange(cols):
if not mask[i,j]:
input = self.sampler.get_inputs(state, i,j)
if input != None:
input = np.array([input])
out = self.logreg.predict(input)
predicted_band[i,j] = out
confidence = self._outputConfidence(input)
confidence_band[i, j] = confidence
if calcTransitions:
potentials = self.outputTransitions(input)
for cat in self.catlist:
map = self.transitionPotentials[cat]
map[i, j] = potentials[cat]
else: # Input sample is incomplete => mask this pixel
mask[i, j] = True
self.updateProgress.emit()
predicted_bands = [np.ma.array(data = predicted_band, mask = mask, dtype=np.uint8)]
confidence_bands = [np.ma.array(data = confidence_band, mask = mask, dtype=np.uint8)]
self.prediction = Raster()
self.prediction.create(predicted_bands, geodata)
self.confidence = Raster()
self.confidence.create(confidence_bands, geodata)
if calcTransitions:
for cat in self.catlist:
band = [np.ma.array(data=self.transitionPotentials[cat], mask=mask, dtype=np.uint8)]
self.transitionPotentials[cat] = Raster()
self.transitionPotentials[cat].create(band, geodata)
except MemoryError:
self.errorReport.emit(self.tr("The system out of memory during LR prediction"))
raise
except:
self.errorReport.emit(self.tr("An unknown error occurs during LR prediction"))
raise
finally:
self.processFinished.emit()
def test_setTrainingData(self):
smp = Sampler(self.state, self.factors, self.output, ns=0)
smp.setTrainingData(self.state, self.output, shuffle=False)
data = np.array([
([0, 3], [0, 1], 1.0, 1.0),
([1, 3], [0, 0], 1.0, 2.0),
([2, 3], [0, 1], 3.0, 1.0),
([0, 2], [0, 1], 3.0, 1.0),
([1, 2], [0, 0], 2.0, 2.0),
([2, 2], [0, 1], 1.0, 1.0),
([0, 1], [1, 0], 0.0, 0.0),
([1, 1], [0, 1], 3.0, 1.0),
([2, 1], [0, 0], 1.0, 2.0),
],
dtype=[('coords', float, 2), ('state', float, (2, )),
('factors', float, (1, )),
('output', float, 1)])
for i in range(len(data)):
assert_array_equal(data[i]['coords'], smp.data[i]['coords'])
assert_array_equal(data[i]['factors'], smp.data[i]['factors'])
assert_array_equal(data[i]['output'], smp.data[i]['output'])
assert_array_equal(data[i]['state'], smp.data[i]['state'])
# two factor_rasters
smp = Sampler(self.state, self.factors2, self.output, ns=1)
smp.setTrainingData(self.state, self.output)
# Check all except coords
data = np.array(
[( #State categories: [1,2,1, 1,2,1, 0,1,2],
[0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0],
# Factors:
[
1., 1., 3., 3., 2., 1., 0., 3., 1., 1., 1., 3., 3., 2., 1.,
0., 3., 1.
],
# Output:
2.0)],
dtype=[('state', float, (18, )), ('factors', float, (18, )),
('output', float, 1)])
assert_array_equal(data[0]['factors'], smp.data[0]['factors'])
assert_array_equal(data[0]['output'], smp.data[0]['output'])
assert_array_equal(data[0]['state'], smp.data[0]['state'])
# Multiband factors
smp = Sampler(self.state, self.factors3, self.output, ns=1)
smp.setTrainingData(self.state, self.output)
# Check all except coords
data = np.array(
[([0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0], [
1., 2., 1., 1., 2., 1., 0., 1., 2., 1., 1., 3., 3., 2., 1., 0.,
3., 1.
], 2.0)],
dtype=[('state', float, (18, )), ('factors', float, (18, )),
('output', float, 1)])
assert_array_equal(data[0]['factors'], smp.data[0]['factors'])
assert_array_equal(data[0]['output'], smp.data[0]['output'])
assert_array_equal(data[0]['state'], smp.data[0]['state'])
# Several factor bands, several factor rasters
smp = Sampler(self.state, self.factors4, self.output, ns=1)
smp.setTrainingData(self.state, self.output)
# Check all except coords
data = np.array([
([0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0], [
1., 2., 1., 1., 2., 1., 0., 1., 2., 1., 1., 3., 3., 2., 1., 0.,
3., 1., 1., 1., 3., 3., 2., 1., 0., 3., 1.
], 2.0)
],
dtype=[('state', float, (18, )),
('factors', float, (27, )),
('output', float, 1)])
assert_array_equal(data[0]['factors'], smp.data[0]['factors'])
assert_array_equal(data[0]['output'], smp.data[0]['output'])
assert_array_equal(data[0]['state'], smp.data[0]['state'])
# Mode = Random
# As the Multiband factors example, but 10 samples:
data = np.array(
[([0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0], [
1., 2., 1., 1., 2., 1., 0., 1., 2., 1., 1., 3., 3., 2., 1., 0.,
3., 1.
], 2.0)],
dtype=[('state', float, (18, )), ('factors', float, (18, )),
('output', float, 1)])
smp = Sampler(self.state, self.factors3, self.output, ns=1)
smp.setTrainingData(self.state, self.output, mode='Random', samples=10)
for i in range(10):
assert_array_equal(data[0]['factors'], smp.data[i]['factors'])
assert_array_equal(data[0]['output'], smp.data[i]['output'])
assert_array_equal(data[0]['state'], smp.data[i]['state'])
# Mode = Stratified
smp = Sampler(self.state, self.factors, self.output, ns=0)
smp.setTrainingData(self.state,
self.output,
mode='Stratified',
samples=15)
out = smp.data['output']
out.sort()
self.assertEqual(out[0], 0)
self.assertEqual(out[4], 0)
self.assertEqual(out[5], 1)
self.assertEqual(out[9], 1)
self.assertEqual(out[10], 2)
def _predict(self, state, factors, calcTransitions=False):
'''
Calculate output and confidence rasters using LR model and input rasters
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
'''
try:
self.rangeChanged.emit(self.tr("Initialize model %p%"), 1)
geodata = state.getGeodata()
rows, cols = geodata['ySize'], geodata['xSize']
for r in factors:
if not state.geoDataMatch(r):
raise LRError(
'Geometries of the input rasters are different!')
self.transitionPotentials = None # Reset tr.potentials if they exist
# Normalize factors before prediction:
for f in factors:
f.normalize(mode='mean')
predicted_band = np.zeros([rows, cols], dtype=np.uint8)
confidence_band = np.zeros([rows, cols], dtype=np.uint8)
if calcTransitions:
self.transitionPotentials = {}
for cat in self.catlist:
self.transitionPotentials[cat] = np.zeros([rows, cols],
dtype=np.uint8)
self.sampler = Sampler(state, factors, ns=self.ns)
mask = state.getBand(1).mask.copy()
if mask.shape == ():
mask = np.zeros([rows, cols], dtype=np.bool)
self.updateProgress.emit()
self.rangeChanged.emit(self.tr("Prediction %p%"), rows)
for i in xrange(rows):
for j in xrange(cols):
if not mask[i, j]:
input = self.sampler.get_inputs(state, i, j)
if input != None:
input = np.array([input])
out = self.logreg.predict(input)
predicted_band[i, j] = out
confidence = self._outputConfidence(input)
confidence_band[i, j] = confidence
if calcTransitions:
potentials = self.outputTransitions(input)
for cat in self.catlist:
map = self.transitionPotentials[cat]
map[i, j] = potentials[cat]
else: # Input sample is incomplete => mask this pixel
mask[i, j] = True
self.updateProgress.emit()
predicted_bands = [
np.ma.array(data=predicted_band, mask=mask, dtype=np.uint8)
]
confidence_bands = [
np.ma.array(data=confidence_band, mask=mask, dtype=np.uint8)
]
self.prediction = Raster()
self.prediction.create(predicted_bands, geodata)
self.confidence = Raster()
self.confidence.create(confidence_bands, geodata)
if calcTransitions:
for cat in self.catlist:
band = [
np.ma.array(data=self.transitionPotentials[cat],
mask=mask,
dtype=np.uint8)
]
self.transitionPotentials[cat] = Raster()
self.transitionPotentials[cat].create(band, geodata)
except MemoryError:
self.errorReport.emit(
self.tr("The system out of memory during LR prediction"))
raise
except:
self.errorReport.emit(
self.tr("An unknown error occurs during LR prediction"))
raise
finally:
self.processFinished.emit()
class LR(QObject):
"""
Implements Logistic Regression model definition and calibration
(maximum liklihood parameter estimation).
"""
rangeChanged = pyqtSignal(str, int)
updateProgress = pyqtSignal()
processFinished = pyqtSignal()
samplingFinished = pyqtSignal()
finished = pyqtSignal()
logMessage = pyqtSignal(str)
errorReport = pyqtSignal(str)
def __init__(self, ns=0, logreg=None):
QObject.__init__(self)
if logreg:
self.logreg = logreg
else:
self.logreg = mlr.MLR()
self.state = None
self.factors = None
self.output = None
self.mode = "All"
self.samples = None
self.catlist = None
self.ns = ns # Neighbourhood size of training rasters.
self.data = None # Training data
self.maxiter = 100 # Maximum of fitting iterations
self.sampler = None # Sampler
# Results of the LR prediction
self.prediction = None # Raster of the LR prediction results
self.confidence = None # Raster of the LR results confidence (1 = the maximum confidence, 0 = the least confidence)
self.Kappa = 0 # Kappa value
self.pseudoR = 0 # Pseudo R-squared (Count) (http://www.ats.ucla.edu/stat/mult_pkg/faq/general/Psuedo_RSquareds.htm)
self.transitionPotentials = None # Dictionary of transition potencial maps: {category1: map1, category2: map2, ...}
def getCoef(self):
return self.logreg.get_weights().T
def getConfidence(self):
return self.confidence
def getIntercept(self):
return self.logreg.get_intercept()
def getKappa(self):
return self.Kappa
def getStdErrIntercept(self):
X = np.column_stack((self.data['state'], self.data['factors']))
return self.logreg.get_stderr_intercept(X)
def getStdErrWeights(self):
X = np.column_stack((self.data['state'], self.data['factors']))
return self.logreg.get_stderr_weights(X).T
def get_PvalIntercept(self):
X = np.column_stack((self.data['state'], self.data['factors']))
return self.logreg.get_pval_intercept(X)
def get_PvalWeights(self):
X = np.column_stack((self.data['state'], self.data['factors']))
return self.logreg.get_pval_weights(X).T
def getPrediction(self, state, factors, calcTransitions=False):
self._predict(state, factors, calcTransitions)
return self.prediction
def getPseudoR(self):
return self.pseudoR
def getTransitionPotentials(self):
return self.transitionPotentials
def _outputConfidence(self, input):
'''
Return confidence (difference between 2 biggest probabilities) of the LR output.
1 = the maximum confidence, 0 = the least confidence
'''
out_scl = self.logreg.predict_proba(input)[0]
# Calculate the confidence:
out_scl.sort()
return int(100 * (out_scl[-1] - out_scl[-2]))
def outputTransitions(self, input):
'''
Return transition potential of the outputs
'''
out_scl = self.logreg.predict_proba(input)[0]
out_scl = [int(100 * x) for x in out_scl]
result = {}
for r, v in enumerate(out_scl):
cat = self.catlist[r]
result[cat] = v
return result
def _predict(self, state, factors, calcTransitions=False):
'''
Calculate output and confidence rasters using LR model and input rasters
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
'''
try:
self.rangeChanged.emit(self.tr("Initialize model %p%"), 1)
geodata = state.getGeodata()
rows, cols = geodata['ySize'], geodata['xSize']
for r in factors:
if not state.geoDataMatch(r):
raise LRError(
'Geometries of the input rasters are different!')
self.transitionPotentials = None # Reset tr.potentials if they exist
# Normalize factors before prediction:
for f in factors:
f.normalize(mode='mean')
predicted_band = np.zeros([rows, cols], dtype=np.uint8)
confidence_band = np.zeros([rows, cols], dtype=np.uint8)
if calcTransitions:
self.transitionPotentials = {}
for cat in self.catlist:
self.transitionPotentials[cat] = np.zeros([rows, cols],
dtype=np.uint8)
self.sampler = Sampler(state, factors, ns=self.ns)
mask = state.getBand(1).mask.copy()
if mask.shape == ():
mask = np.zeros([rows, cols], dtype=np.bool)
self.updateProgress.emit()
self.rangeChanged.emit(self.tr("Prediction %p%"), rows)
for i in xrange(rows):
for j in xrange(cols):
if not mask[i, j]:
input = self.sampler.get_inputs(state, i, j)
if input != None:
input = np.array([input])
out = self.logreg.predict(input)
predicted_band[i, j] = out
confidence = self._outputConfidence(input)
confidence_band[i, j] = confidence
if calcTransitions:
potentials = self.outputTransitions(input)
for cat in self.catlist:
map = self.transitionPotentials[cat]
map[i, j] = potentials[cat]
else: # Input sample is incomplete => mask this pixel
mask[i, j] = True
self.updateProgress.emit()
predicted_bands = [
np.ma.array(data=predicted_band, mask=mask, dtype=np.uint8)
]
confidence_bands = [
np.ma.array(data=confidence_band, mask=mask, dtype=np.uint8)
]
self.prediction = Raster()
self.prediction.create(predicted_bands, geodata)
self.confidence = Raster()
self.confidence.create(confidence_bands, geodata)
if calcTransitions:
for cat in self.catlist:
band = [
np.ma.array(data=self.transitionPotentials[cat],
mask=mask,
dtype=np.uint8)
]
self.transitionPotentials[cat] = Raster()
self.transitionPotentials[cat].create(band, geodata)
except MemoryError:
self.errorReport.emit(
self.tr("The system out of memory during LR prediction"))
raise
except:
self.errorReport.emit(
self.tr("An unknown error occurs during LR prediction"))
raise
finally:
self.processFinished.emit()
def __propagateSamplerSignals(self):
self.sampler.rangeChanged.connect(self.__samplerProgressRangeChanged)
self.sampler.updateProgress.connect(self.__samplerProgressChanged)
self.sampler.samplingFinished.connect(self.__samplerFinished)
def __samplerFinished(self):
self.sampler.rangeChanged.disconnect(
self.__samplerProgressRangeChanged)
self.sampler.updateProgress.disconnect(self.__samplerProgressChanged)
self.sampler.samplingFinished.disconnect(self.__samplerFinished)
self.samplingFinished.emit()
def __samplerProgressRangeChanged(self, message, maxValue):
self.rangeChanged.emit(message, maxValue)
def __samplerProgressChanged(self):
self.updateProgress.emit()
def save(self):
pass
def saveSamples(self, fileName):
self.sampler.saveSamples(fileName)
def setMaxIter(self, maxiter):
self.maxiter = maxiter
def setTrainingData(self):
state, factors, output, mode, samples = self.state, self.factors, self.output, self.mode, self.samples
if not self.logreg:
raise LRError(
'You must create a Logistic Regression model before!')
# Normalize factors before sampling:
for f in factors:
f.normalize(mode='mean')
self.sampler = Sampler(state, factors, output, ns=self.ns)
self.__propagateSamplerSignals()
self.sampler.setTrainingData(state,
output,
shuffle=False,
mode=mode,
samples=samples)
outputVecLen = self.sampler.outputVecLen
stateVecLen = self.sampler.stateVecLen
factorVectLen = self.sampler.factorVectLen
size = len(self.sampler.data)
self.data = self.sampler.data
self.catlist = np.unique(self.data['output'])
def train(self):
X = np.column_stack((self.data['state'], self.data['factors']))
Y = self.data['output']
self.labelCodes = np.unique(Y)
self.logreg.fit(X, Y, maxiter=self.maxiter)
out = self.logreg.predict(X)
depCoef = DependenceCoef(np.ma.array(out), np.ma.array(Y), expand=True)
self.Kappa = depCoef.kappa(mode=None)
self.pseudoR = depCoef.correctness(percent=False)
def setState(self, state):
self.state = state
def setFactors(self, factors):
self.factors = factors
def setOutput(self, output):
self.output = output
def setMode(self, mode):
self.mode = mode
def setSamples(self, samples):
self.samples = samples
def startTrain(self):
try:
self.setTrainingData()
self.train()
except MemoryError:
self.errorReport.emit(
self.tr("The system out of memory during LR training"))
raise
except:
self.errorReport.emit(
self.tr("An unknown error occurs during LR trainig"))
raise
finally:
self.finished.emit()
def test_setTrainingData(self):
smp = Sampler(self.state, self.factors, self.output, ns=0)
smp.setTrainingData(self.state, self.output, shuffle=False)
data = np.array(
[
([0,3], [0, 1], 1.0, 1.0),
([1,3], [0, 0], 1.0, 2.0),
([2,3], [0, 1], 3.0, 1.0),
([0,2], [0, 1], 3.0, 1.0),
([1,2], [0, 0], 2.0, 2.0),
([2,2], [0, 1], 1.0, 1.0),
([0,1], [1, 0], 0.0, 0.0),
([1,1], [0, 1], 3.0, 1.0),
([2,1], [0, 0], 1.0, 2.0),
],
dtype=[('coords', float, 2), ('state', float, (2,)), ('factors', float, (1,)), ('output', float, 1)]
)
for i in range(len(data)):
assert_array_equal(data[i]['coords'], smp.data[i]['coords'])
assert_array_equal(data[i]['factors'], smp.data[i]['factors'])
assert_array_equal(data[i]['output'], smp.data[i]['output'])
assert_array_equal(data[i]['state'], smp.data[i]['state'])
# two factor_rasters
smp = Sampler(self.state, self.factors2, self.output, ns=1)
smp.setTrainingData(self.state, self.output)
# Check all except coords
data = np.array(
[
(#State categories: [1,2,1, 1,2,1, 0,1,2],
[0,1, 0,0, 0,1, 0,1, 0,0, 0,1, 1,0, 0,1, 0,0],
# Factors:
[ 1., 1., 3., 3., 2., 1., 0., 3., 1.,
1., 1., 3., 3., 2., 1., 0., 3., 1.],
# Output:
2.0)
],
dtype=[('state', float, (18,)), ('factors', float, (18,)), ('output', float, 1)]
)
assert_array_equal(data[0]['factors'], smp.data[0]['factors'])
assert_array_equal(data[0]['output'], smp.data[0]['output'])
assert_array_equal(data[0]['state'], smp.data[0]['state'])
# Multiband factors
smp = Sampler(self.state, self.factors3, self.output, ns=1)
smp.setTrainingData(self.state, self.output)
# Check all except coords
data = np.array(
[
([0,1, 0,0, 0,1, 0,1, 0,0, 0,1, 1,0, 0,1, 0,0],
[ 1., 2., 1., 1., 2., 1., 0., 1., 2.,
1., 1., 3., 3., 2., 1., 0., 3., 1.],
2.0)
],
dtype=[('state', float, (18,)), ('factors', float, (18,)), ('output', float, 1)]
)
assert_array_equal(data[0]['factors'], smp.data[0]['factors'])
assert_array_equal(data[0]['output'], smp.data[0]['output'])
assert_array_equal(data[0]['state'], smp.data[0]['state'])
# Several factor bands, several factor rasters
smp = Sampler(self.state, self.factors4, self.output, ns=1)
smp.setTrainingData(self.state, self.output)
# Check all except coords
data = np.array(
[
([0,1, 0,0, 0,1, 0,1, 0,0, 0,1, 1,0, 0,1, 0,0],
[ 1., 2., 1., 1., 2., 1., 0., 1., 2.,
1., 1., 3., 3., 2., 1., 0., 3., 1.,
1., 1., 3., 3., 2., 1., 0., 3., 1.],
2.0)
],
dtype=[('state', float, (18,)), ('factors', float, (27,)), ('output', float, 1)]
)
assert_array_equal(data[0]['factors'], smp.data[0]['factors'])
assert_array_equal(data[0]['output'], smp.data[0]['output'])
assert_array_equal(data[0]['state'], smp.data[0]['state'])
# Mode = Random
# As the Multiband factors example, but 10 samples:
data = np.array(
[
([0,1, 0,0, 0,1, 0,1, 0,0, 0,1, 1,0, 0,1, 0,0],
[ 1., 2., 1., 1., 2., 1., 0., 1., 2.,
1., 1., 3., 3., 2., 1., 0., 3., 1.],
2.0)
],
dtype=[('state', float, (18,)), ('factors', float, (18,)), ('output', float, 1)]
)
smp = Sampler(self.state, self.factors3, self.output, ns=1)
smp.setTrainingData(self.state, self.output, mode='Random', samples=10)
for i in range(10):
assert_array_equal(data[0]['factors'], smp.data[i]['factors'])
assert_array_equal(data[0]['output'], smp.data[i]['output'])
assert_array_equal(data[0]['state'], smp.data[i]['state'])
# Mode = Stratified
smp = Sampler(self.state, self.factors, self.output, ns=0)
smp.setTrainingData(self.state, self.output, mode='Stratified', samples=15)
out = smp.data['output']
out.sort()
self.assertEqual(out[0], 0)
self.assertEqual(out[4], 0)
self.assertEqual(out[5], 1)
self.assertEqual(out[9], 1)
self.assertEqual(out[10], 2)
class MlpManager(QObject):
'''This class gets the data extracted from the UI and
pass it to multi-layer perceptron, then gets and stores the result.
'''
updateGraph = pyqtSignal(float, float) # Train error, val. error
updateMinValErr = pyqtSignal(float) # Min validation error
updateDeltaRMS = pyqtSignal(float) # Delta of RMS: min(valError) - currentValError
updateKappa = pyqtSignal(float) # Kappa value
processFinished = pyqtSignal()
processInterrupted = pyqtSignal()
logMessage = pyqtSignal(str)
errorReport = pyqtSignal(str)
rangeChanged = pyqtSignal(str, int)
updateProgress = pyqtSignal()
def __init__(self, ns=0, MLP=None):
QObject.__init__(self)
self.MLP = MLP
self.interrupted = False
self.layers = None
if self.MLP:
self.layers = self.getMlpTopology()
self.ns = ns # Neighbourhood size of training rasters.
self.data = None # Training data
self.catlist = None # List of unique output values of the output raster
self.train_error = None # Error on training set
self.val_error = None # Error on validation set
self.minValError = None # The minimum error that is achieved on the validation set
self.valKappa = 0 # Kappa on on the validation set
self.sampler = None # Sampler
# Results of the MLP prediction
self.prediction = None # Raster of the MLP prediction results
self.confidence = None # Raster of the MLP results confidence (1 = the maximum confidence, 0 = the least confidence)
self.transitionPotentials = None # Dictionary of transition potencial maps: {category1: map1, category2: map2, ...}
# Outputs of the activation function for small and big numbers
self.sigmax, self.sigmin = sigmoid(100), sigmoid(-100) # Max and Min of the sigmoid function
self.sigrange = self.sigmax - self.sigmin # Range of the sigmoid
def computeMlpError(self, sample):
'''Get MLP error on the sample'''
input = np.hstack( (sample['state'], sample['factors']) )
out = self.getOutput( input )
err = ((sample['output'] - out)**2).sum()/len(out)
return err
def computePerformance(self, train_indexes, val_ind):
'''Check errors of training and validation sets
@param train_indexes Tuple that contains indexes of the first and last elements of the training set.
@param val_ind Tuple that contains indexes of the first and last elements of the validation set.
'''
train_error = 0
train_sampl = train_indexes[1] - train_indexes[0] # Count of training samples
for i in range(train_indexes[0], train_indexes[1]):
train_error = train_error + self.computeMlpError(sample = self.data[i])
self.setTrainError(train_error/train_sampl)
if val_ind:
val_error = 0
val_sampl = val_ind[1] - val_ind[0]
answers = np.ma.zeros(val_sampl)
out = np.ma.zeros(val_sampl)
for i in xrange(val_ind[0], val_ind[1]):
sample = self.data[i]
val_error = val_error + self.computeMlpError(sample = self.data[i])
input = np.hstack( (sample['state'],sample['factors']) )
output = self.getOutput(input)
out[i-val_ind[0]] = self.outCategory(output)
answers[i-val_ind[0]] = self.outCategory(sample['output'])
self.setValError(val_error/val_sampl)
depCoef = DependenceCoef(out, answers, expand=True)
self.valKappa = depCoef.kappa(mode=None)
def copyWeights(self):
'''Deep copy of the MLP weights'''
return copy.deepcopy(self.MLP.weights)
def createMlp(self, state, factors, output, hidden_layers):
'''
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
@param hidden_layers List of neuron counts in hidden layers.
@param ns Neighbourhood size.
'''
if output.getBandsCount() != 1:
raise MlpManagerError('Output layer must have one band!')
input_neurons = 0
for raster in factors:
input_neurons = input_neurons+ raster.getNeighbourhoodSize(self.ns)
# state raster contains categories. We need use n-1 dummy variables (where n = number of categories)
input_neurons = input_neurons + (len(state.getBandGradation(1))-1) * state.getNeighbourhoodSize(self.ns)
# Output category's (neuron) list and count
self.catlist = output.getBandGradation(1)
categories = len(self.catlist)
# set neuron counts in the MLP layers
self.layers = hidden_layers
self.layers.insert(0, input_neurons)
self.layers.append(categories)
self.MLP = MLP(*self.layers)
def getConfidence(self):
return self.confidence
def getInputVectLen(self):
'''Length of input data vector of the MLP'''
shape = self.getMlpTopology()
return shape[0]
def getOutput(self, input_vector):
out = self.MLP.propagate_forward( input_vector )
return out
def getOutputVectLen(self):
'''Length of input data vector of the MLP'''
shape = self.getMlpTopology()
return shape[-1]
def getOutputVector(self, val):
'''Convert a number val into vector,
for example, let self.catlist = [1, 3, 4] then
if val = 1, result = [ 1, -1, -1]
if val = 3, result = [-1, 1, -1]
if val = 4, result = [-1, -1, 1]
where -1 is minimum of the sigmoid, 1 is max of the sigmoid
'''
size = self.getOutputVectLen()
res = np.ones(size) * (self.sigmin)
ind = np.where(self.catlist==val)
res[ind] = self.sigmax
return res
def getMinValError(self):
return self.minValError
def getMlpTopology(self):
return self.MLP.shape
def getKappa(self):
return self.valKappa
def getPrediction(self, state, factors, calcTransitions=False):
self._predict(state, factors, calcTransitions)
return self.prediction
def getTrainError(self):
return self.train_error
def getTransitionPotentials(self):
return self.transitionPotentials
def getValError(self):
return self.val_error
def outCategory(self, out_vector):
# Get index of the biggest output value as the result
biggest = max(out_vector)
res = list(out_vector).index(biggest)
res = self.catlist[res]
return res
def outputConfidence(self, output, scale=True):
'''
Return confidence (difference between 2 biggest values) of the MLP output.
@param output: The confidence
@param scale: If True, then scale the confidence to int [0, 1, ..., 100] percent
'''
out_scl = self.scaleOutput(output, percent=scale)
out_scl.sort()
return out_scl[-1] - out_scl[-2]
def outputTransitions(self, output, scale=True):
'''
Return transition potencial of the outputs scaled to [0,1] or 1-100
@param output: The output of MLP
@param scale: If True, then scale the transitions to int ([0, 1, ..., 100]) percent
'''
out_scl = self.scaleOutput(output, percent=scale)
result = {}
for r, v in enumerate(out_scl):
cat = self.catlist[r]
result[cat] = v
return result
def scaleOutput(self, output, percent=True):
'''
Scale the output to range [0,1] or 1-100
@param output: Output of a MLP
@param percent: If True, then scale the output to int [0, 1, ..., 100] percent
'''
res = 1.0 * (output - self.sigmin) / self.sigrange
if percent:
res = [ int(100 * x) for x in res]
return res
def _predict(self, state, factors, calcTransitions=False):
'''
Calculate output and confidence rasters using MLP model and input rasters
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
'''
try:
self.rangeChanged.emit(self.tr("Initialize model %p%"), 1)
geodata = state.getGeodata()
rows, cols = geodata['ySize'], geodata['xSize']
for r in factors:
if not state.geoDataMatch(r):
raise MlpManagerError('Geometries of the input rasters are different!')
self.transitionPotentials = None # Reset tr.potentials if they exist
# Normalize factors before prediction:
for f in factors:
f.normalize(mode = 'mean')
predicted_band = np.zeros([rows, cols], dtype=np.uint8)
confidence_band = np.zeros([rows, cols], dtype=np.uint8)
if calcTransitions:
self.transitionPotentials = {}
for cat in self.catlist:
self.transitionPotentials[cat] = np.zeros([rows, cols], dtype=np.uint8)
self.sampler = Sampler(state, factors, ns=self.ns)
mask = state.getBand(1).mask.copy()
if mask.shape == ():
mask = np.zeros([rows, cols], dtype=np.bool)
self.updateProgress.emit()
self.rangeChanged.emit(self.tr("Prediction %p%"), rows)
for i in xrange(rows):
for j in xrange(cols):
if not mask[i,j]:
input = self.sampler.get_inputs(state, i,j)
if input != None:
out = self.getOutput(input)
res = self.outCategory(out)
predicted_band[i, j] = res
confidence = self.outputConfidence(out)
confidence_band[i, j] = confidence
if calcTransitions:
potentials = self.outputTransitions(out)
for cat in self.catlist:
map = self.transitionPotentials[cat]
map[i, j] = potentials[cat]
else: # Input sample is incomplete => mask this pixel
mask[i, j] = True
self.updateProgress.emit()
predicted_bands = [np.ma.array(data = predicted_band, mask = mask, dtype=np.uint8)]
confidence_bands = [np.ma.array(data = confidence_band, mask = mask, dtype=np.uint8)]
self.prediction = Raster()
self.prediction.create(predicted_bands, geodata)
self.confidence = Raster()
self.confidence.create(confidence_bands, geodata)
if calcTransitions:
for cat in self.catlist:
band = [np.ma.array(data=self.transitionPotentials[cat], mask=mask, dtype=np.uint8)]
self.transitionPotentials[cat] = Raster()
self.transitionPotentials[cat].create(band, geodata)
except MemoryError:
self.errorReport.emit(self.tr("The system out of memory during ANN prediction"))
raise
except:
self.errorReport.emit(self.tr("An unknown error occurs during ANN prediction"))
raise
def readMlp(self):
pass
def resetErrors(self):
self.val_error = np.finfo(np.float).max
self.train_error = np.finfo(np.float).max
def resetMlp(self):
self.MLP.reset()
self.resetErrors()
def saveMlp(self):
pass
def saveSamples(self, fileName):
self.sampler.saveSamples(fileName)
def setMlpWeights(self, w):
'''Set weights of the MLP'''
self.MLP.weights = w
def setTrainingData(self, state, factors, output, shuffle=True, mode='All', samples=None):
'''
@param state Raster of the current state (categories) values.
@param factors List of the factor rasters (predicting variables).
@param output Raster that contains categories to predict.
@param shuffle Perform random shuffle.
@param mode Type of sampling method:
All Get all pixels
Random Get samples. Count of samples in the data=samples.
Stratified Undersampling of major categories and/or oversampling of minor categories.
@samples Sample count of the training data (doesn't used in 'All' mode).
'''
if not self.MLP:
raise MlpManagerError('You must create a MLP before!')
# Normalize factors before sampling:
for f in factors:
f.normalize(mode = 'mean')
self.sampler = Sampler(state, factors, output, self.ns)
self.sampler.setTrainingData(state=state, output=output, shuffle=shuffle, mode=mode, samples=samples)
outputVecLen = self.getOutputVectLen()
stateVecLen = self.sampler.stateVecLen
factorVectLen = self.sampler.factorVectLen
size = len(self.sampler.data)
self.data = np.zeros(size, dtype=[('coords', float, 2), ('state', float, stateVecLen), ('factors', float, factorVectLen), ('output', float, outputVecLen)])
self.data['coords'] = self.sampler.data['coords']
self.data['state'] = self.sampler.data['state']
self.data['factors'] = self.sampler.data['factors']
self.data['output'] = [self.getOutputVector(sample['output']) for sample in self.sampler.data]
def setTrainError(self, error):
self.train_error = error
def setValError(self, error):
self.val_error = error
def setEpochs(self, epochs):
self.epochs = epochs
def setValPercent(self, value=20):
self.valPercent = value
def setLRate(self, value=0.1):
self.lrate = value
def setMomentum(self, value=0.01):
self.momentum = value
def setContinueTrain(self, value=False):
self.continueTrain = value
def startTrain(self):
self.train(self.epochs, self.valPercent, self.lrate, self.momentum, self.continueTrain)
def stopTrain(self):
self.interrupted = True
def train(self, epochs, valPercent=20, lrate=0.1, momentum=0.01, continue_train=False):
'''Perform the training procedure on the MLP and save the best neural net
@param epoch Max iteration count.
@param valPercent Percent of the validation set.
@param lrate Learning rate.
@param momentum Learning momentum.
@param continue_train If False then it is new training cycle, reset weights training and validation error. If True, then continue training.
'''
try:
samples_count = len(self.data)
val_sampl_count = samples_count*valPercent/100
apply_validation = True if val_sampl_count>0 else False # Use or not use validation set
train_sampl_count = samples_count - val_sampl_count
# Set first train_sampl_count as training set, the other as validation set
train_indexes = (0, train_sampl_count)
val_indexes = (train_sampl_count, samples_count) if apply_validation else None
if not continue_train: self.resetMlp()
self.minValError = self.getValError() # The minimum error that is achieved on the validation set
last_train_err = self.getTrainError()
best_weights = self.copyWeights() # The MLP weights when minimum error that is achieved on the validation set
self.rangeChanged.emit(self.tr("Train model %p%"), epochs)
for epoch in range(epochs):
self.trainEpoch(train_indexes, lrate, momentum)
self.computePerformance(train_indexes, val_indexes)
self.updateGraph.emit(self.getTrainError(), self.getValError())
self.updateDeltaRMS.emit(self.getMinValError() - self.getValError())
self.updateKappa.emit(self.getKappa())
QCoreApplication.processEvents()
if self.interrupted:
self.processInterrupted.emit()
break
last_train_err = self.getTrainError()
self.setTrainError(last_train_err)
if apply_validation and (self.getValError() < self.getMinValError()):
self.minValError = self.getValError()
best_weights = self.copyWeights()
self.updateMinValErr.emit(self.getMinValError())
self.updateProgress.emit()
self.setMlpWeights(best_weights)
except MemoryError:
self.errorReport.emit(self.tr("The system out of memory during ANN training"))
raise
except:
self.errorReport.emit(self.tr("An unknown error occurs during ANN trainig"))
raise
finally:
self.processFinished.emit()
def trainEpoch(self, train_indexes, lrate=0.1, momentum=0.01):
'''Perform a training epoch on the MLP
@param train_ind Tuple of the min&max indexes of training samples in the samples data.
@param val_ind Tuple of the min&max indexes of validation samples in the samples data.
@param lrate Learning rate.
@param momentum Learning momentum.
'''
train_sampl = train_indexes[1] - train_indexes[0]
for i in range(train_sampl):
n = np.random.randint( *train_indexes )
sample = self.data[n]
input = np.hstack( (sample['state'],sample['factors']) )
self.getOutput( input ) # Forward propagation
self.MLP.propagate_backward( sample['output'], lrate, momentum )
