def test(self, fImgs, fLbls, fName):
'''Neural network testing after training.
INPUT : Images set, labels set (None for autoencoders),
name for save.
OUTPUT : Nothing'''
if PREPROCESSING:
fImgs, _key = ld.normalization(fName, fImgs)
print "Testing the neural networks..."
_out = self.propagation(fImgs.T)
_cost = self.error(_out[-1], fImgs.T) / len(fImgs)
print "Cost {0}\n".format(_cost)
# Save output in order to have a testset for next layers
if fName is not None:
self.save_output(fName, "test", fImgs)
# Displaying the results
if PREPROCESSING:
_decrypt = np.dot(_out[-1],_key.T)
dy.display(fName, "out", fImgs, _decrypt.T)
else:
dy.display(fName, "out", fImgs, _out[-1].T)
# Approximated vision of first hidden layer neurons
dy.display(fName, "neurons", self.neurons_vision())
def test(self, fImgs, fLbls, fName):
'''Neural network testing after training.
INPUT : Images set, labels set (None for autoencoders),
name for save.
OUTPUT : Nothing'''
if PREPROCESSING:
fImgs, _key = ld.normalization(fName, fImgs)
print "Testing the neural networks..."
_out = self.propagation(fImgs.T)
_cost = self.error(_out[-1], fImgs.T) / len(fImgs)
print "Cost {0}\n".format(_cost)
# Save output in order to have a testset for next layers
if fName is not None:
self.save_output(fName, "test", fImgs)
# Displaying the results
if PREPROCESSING:
_decrypt = np.dot(_out[-1], _key.T)
dy.display(fName, "out", fImgs, _decrypt.T)
else:
dy.display(fName, "out", fImgs, _out[-1].T)
# Approximated vision of first hidden layer neurons
dy.display(fName, "neurons", self.neurons_vision())
def train(self, fImgs, fLbls, fIterations, fName):
'''Training algorithm. Can evolved according to your need.
INPUT : Images set, labels set (None for autoencoders),
number of iterations before stopping, name for save
OUTPUT : Nothing'''
if PREPROCESSING:
fImgs, _key = ld.normalization(fName, fImgs)
print "Training...\n"
_gcost = []
_gtime = []
_done = fIterations
for i in xrange(fIterations):
_gtime.append(tm.time())
_gcost.append(0)
for j in xrange(self.mCycle):
_trn, _tst = self.cross_validation(j, fImgs)
for k in xrange(len(_trn) / self.mBatchSize):
if DEBUG:
print "Learning rates :", self.mEpsilon
print "Momentums :", self.mMomentum
# Inputs and labels batch
_in = self.build_batch(k, _trn)
# Activation propagation
_out = self.propagation(_in, DROPOUT)
# Local error for each layer
_err = self.layer_error(_out, _in, SPARSITY)
# Gradient for stochastic gradient descent
_wGrad, _bGrad = self.gradient(_err, _out)
# Gradient checking
if GRAD_CHECK:
print "Gradient checking ..."
self.gradient_checking(_in,_in,_wGrad,_bGrad)
# Adapt learning rate
if (i > 0 or j > 0 or k > 0) and ANGLE_DRIVEN:
self.angle_driven_approach(_wGrad)
# Weight variations
self.variations(_wGrad)
# Update weights and biases
self.update(_bGrad)
# Adapt learning rate
if AVG_GRADIENT:
self.average_gradient_approach(_wGrad)
# Evaluate the network
_cost = self.evaluate(_tst)
_gcost[i] += _cost
if DEBUG:
print "Cost :", _cost
# Iteration information
_gtime[i] = tm.time() - _gtime[i]
print "Iteration {0} in {1}s".format(i, _gtime[i])
# Global cost for one cycle
_gcost[i] /= self.mCycle
print "Cost of iteration : {0}".format(_gcost[i])
# Parameters
print "Epsilon {0} Momentum {1}\n".format(self.mEpsilon,
self.mMomentum)
# Stop condition
if i > 0 and abs(_gcost[i-1] - _gcost[i]) < 0.001:
_done = i + 1
break
elif self.mStop:
_done = i + 1
break
dy.plot(xrange(_done), _gcost, fName, "_cost.png")
dy.plot(xrange(_done), _gtime, fName, "_time.png")
if fName is not None:
self.save_output(fName, "train", fImgs)
def train(self, fImgs, fLbls, fIterations, fName):
'''Training algorithm. Can evolved according to your need.
INPUT : Images set, labels set (None for autoencoders),
number of iterations before stopping, name for save
OUTPUT : Nothing'''
if PREPROCESSING:
fImgs, _key = ld.normalization(fName, fImgs)
print "Training...\n"
_gcost = []
_gtime = []
_done = fIterations
for i in xrange(fIterations):
_gtime.append(tm.time())
_gcost.append(0)
for j in xrange(self.mCycle):
_trn, _tst = self.cross_validation(j, fImgs)
for k in xrange(len(_trn) / self.mBatchSize):
if DEBUG:
print "Learning rates :", self.mEpsilon
print "Momentums :", self.mMomentum
# Inputs and labels batch
_in = self.build_batch(k, _trn)
# Activation propagation
_out = self.propagation(_in, DROPOUT)
# Local error for each layer
_err = self.layer_error(_out, _in, SPARSITY)
# Gradient for stochastic gradient descent
_wGrad, _bGrad = self.gradient(_err, _out)
# Gradient checking
if GRAD_CHECK:
print "Gradient checking ..."
self.gradient_checking(_in, _in, _wGrad, _bGrad)
# Adapt learning rate
if (i > 0 or j > 0 or k > 0) and ANGLE_DRIVEN:
self.angle_driven_approach(_wGrad)
# Weight variations
self.variations(_wGrad)
# Update weights and biases
self.update(_bGrad)
# Adapt learning rate
if AVG_GRADIENT:
self.average_gradient_approach(_wGrad)
# Evaluate the network
_cost = self.evaluate(_tst)
_gcost[i] += _cost
if DEBUG:
print "Cost :", _cost
# Iteration information
_gtime[i] = tm.time() - _gtime[i]
print "Iteration {0} in {1}s".format(i, _gtime[i])
# Global cost for one cycle
_gcost[i] /= self.mCycle
print "Cost of iteration : {0}".format(_gcost[i])
# Parameters
print "Epsilon {0} Momentum {1}\n".format(self.mEpsilon,
self.mMomentum)
# Stop condition
if i > 0 and abs(_gcost[i - 1] - _gcost[i]) < 0.001:
_done = i + 1
break
elif self.mStop:
_done = i + 1
break
dy.plot(xrange(_done), _gcost, fName, "_cost.png")
dy.plot(xrange(_done), _gtime, fName, "_time.png")
if fName is not None:
self.save_output(fName, "train", fImgs)