def train_LR(self, lr):
trainer.train_LR(self, lr)
dataloader = DataLoader()
datasets =dataloader.load_shared_data()
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
params=np.empty((28*28)*10+10);
climin.initialize.randomize_normal(params,0,1)
params = params/(28*28)
lr.setParams(params);
x=lr.x
y=lr.y
cost = (
lr.negative_log_likelihood(y)
+ self.L1_lambda * lr.L1
+ self.L2_lambda * lr.L2_sqr
)
g_W = T.grad(cost=cost, wrt=lr.W)
g_b = T.grad(cost=cost, wrt=lr.b)
g_W_model = theano.function(
inputs=[x,y],
outputs=g_W
)
g_b_model = theano.function(
inputs=[x,y],
outputs=g_b
)
batch_size = self.batch_size
index = T.lscalar()
test_err_model = theano.function(
inputs=[index],
outputs=lr.zeroOneLoss(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]
}
)
train_err_model = theano.function(
inputs=[index],
outputs=lr.zeroOneLoss(y),
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]
}
)
validate_err_model = theano.function(
inputs=[index],
outputs=lr.zeroOneLoss(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]
})
# compute number of minibatches for training, validation and testing
batch_size = self.batch_size;
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
def train_error():
train_losses = [train_err_model(i)
for i in xrange(n_train_batches)]
this_train_losses = np.mean(train_losses)
return this_train_losses;
def validate_error():
validation_losses = [validate_err_model(i)
for i in xrange(n_valid_batches)]
this_validation_loss = np.mean(validation_losses)
return this_validation_loss;
def test_error():
test_losses = [test_err_model(i)
for i in xrange(n_test_batches)]
this_test_loss = np.mean(test_losses)
return this_test_loss;
def d_loss_wrt_pars(parameters, inpt, targets):
lr.setParams(parameters)
gwValue = g_W_model(inpt,targets)
gbValue = g_b_model(inpt,targets)
return np.concatenate([gwValue.flatten(),gbValue])
args = ((i, {}) for i in climin.util.iter_minibatches([train_set_x.eval(), train_set_y.eval()], self.batch_size, [0, 0]))
opt = climin.rmsprop.RmsProp(params, d_loss_wrt_pars, step_rate=self.learning_rate,decay=self.decay, momentum=self.momentum, args=args)
validation_frequency = n_train_batches
directory=check_create_observations_dir()
self.output_directory = directory
bestValidationLoss = np.Inf;
for info in opt:
if info['n_iter'] % validation_frequency ==0:
epoch_no = info['n_iter']/n_train_batches
train_err=train_error()
validation_err = validate_error()
test_err = test_error()
self.add_train_data(epoch_no, train_err, validation_err, test_err)
if epoch_no % 10 ==0:
repfields_path=os.path.join(directory,"repFields"+str(epoch_no).zfill(3)+'.png')
W_vals=lr.W.get_value(borrow=True)
display(W_vals,repfields_path)
if epoch_no >= self.n_epochs:
break
if validation_err < bestValidationLoss:
bestValidationLoss = validation_err
#
# if validation_err *0.95 > bestValidationLoss:
#
# print "Best Validation Error : %f Validation err:%f " %(bestValidationLoss,validation_err)
# break;
if epoch_no > 15 and train_err < 0.9* validation_err:
break
print "Iteration no: %d Validation error = %f" %(epoch_no,validation_err*100)
trainer.save_errors(self, directory)
repfields_final_path=os.path.join(directory,"repFields.png")
W_vals=lr.W.get_value(borrow=True)
display(W_vals,repfields_final_path)
def train_LR(self, lr):
trainer.train_LR(self, lr)
dataloader = DataLoader()
datasets =dataloader.load_shared_data()
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# compute number of minibatches for training, validation and testing
batch_size = self.batch_size;
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
index = T.lscalar() # index to a [mini]batch
x = lr.x # the data is presented as rasterized images
y = lr.y
cost = (
lr.negative_log_likelihood(y)
+ self.L1_lambda * lr.L1
+ self.L2_lambda * lr.L2_sqr
)
test_model = theano.function(
inputs=[index],
outputs=lr.zeroOneLoss(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]
}
)
train_err_model = theano.function(
inputs=[index],
outputs=lr.zeroOneLoss(y),
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]
}
)
validate_model = theano.function(
inputs=[index],
outputs=lr.zeroOneLoss(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]
})
g_W = T.grad(cost=cost, wrt=lr.W)
g_b = T.grad(cost=cost, wrt=lr.b)
updates = [(lr.W, lr.W - self.learning_rate * g_W),
(lr.b, lr.b - self.learning_rate * g_b)]
train_model = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size]
}
)
def train_error():
train_losses = [train_err_model(i)
for i in xrange(n_train_batches)]
this_train_losses = numpy.mean(train_losses)
return this_train_losses;
def validate_error():
validation_losses = [validate_model(i)
for i in xrange(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
return this_validation_loss;
def test_error():
test_losses = [test_model(i)
for i in xrange(n_test_batches)]
this_test_loss = numpy.mean(test_losses)
return this_test_loss;
print '... training'
#Train in mini batches
minEpochs = 4
validationFrequency = n_train_batches;
iteration = 0;
bestValidationLoss = numpy.Inf;
max_epoch_reached = False
directory=check_create_observations_dir()
self.output_directory = directory
while not max_epoch_reached :
iteration = iteration + 1;
epochNo = (iteration / n_train_batches)
batchId= iteration % n_train_batches;
currentCost=train_model(batchId)
if iteration % validationFrequency == 0:
validation_err = validate_error()
test_err = test_error()
train_err = train_error()
print "Epoch no: %d Validation Loss = %f" %(epochNo,validation_err*100)
self.add_train_data(epochNo, train_err, validation_err, test_err)
if epochNo %5 ==0:
W_vals=lr.W.get_value(borrow=True)
repfields_path=os.path.join(directory,"repFields"+str(epochNo).zfill(3)+'.png')
display(W_vals,repfields_path)
if validation_err < bestValidationLoss:
bestValidationLoss = validation_err
if epochNo > minEpochs and validation_err *0.995 > bestValidationLoss:
#print "------------------------Validation Loss = %f" %(validationLoss*100)
break;
if epochNo >= self.n_epochs:
max_epoch_reached = True
testLoss=test_error()
print "iteration %d complete. Cost = %f Best Validation Loss = %f Test Loss = %f" %(iteration,currentCost,bestValidationLoss *100,testLoss *100)
trainer.save_errors(self, directory)
repfields_final_path=os.path.join(directory,"repFields.png")
W_vals=lr.W.get_value(borrow=True)
display(W_vals,repfields_final_path)
def train_NN(self, nn):
trainer.train_NN(self, nn)
data_loader = DataLoader()
datasets = data_loader.load_shared_data()
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# compute number of minibatches for training, validation and testing
batch_size = self.batch_size;
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
index = T.lscalar() # index to a [mini]batch
x = nn.input # the data is presented as rasterized images
y = T.ivector('y')
cost = (nn.negative_log_likelihood_dropout(y)
+ self.L1_lambda * nn.L1
+ self.L2_lambda * nn.L2
)
train_err_model = theano.function(
inputs=[index],
outputs=nn.errors(y),
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]
}
)
test_err_model = theano.function(
inputs=[index],
outputs=nn.errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]
}
)
validate_err_model = theano.function(
inputs=[index],
outputs=nn.errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]
})
gparams = [T.grad(cost, param) for param in nn.params]
updates = [
(param, param - self.learning_rate * gparam)
for param, gparam in zip(nn.params, gparams)
]
train_model = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size]
}
)
def validate():
validation_losses = [validate_err_model(i)
for i in xrange(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
return this_validation_loss;
def test():
test_losses = [test_err_model(i)
for i in xrange(n_test_batches)]
this_test_loss = numpy.mean(test_losses)
return this_test_loss;
def train():
train_losses = [train_err_model(i)
for i in xrange(n_train_batches)]
this_train_loss = numpy.mean(train_losses)
return this_train_loss;
print '... training'
#Train in mini batches
minEpochs = 4
validationFrequency = n_train_batches;
iteration = 0;
bestValidationLoss = numpy.Inf;
directory=check_create_observations_dir()
self.output_directory = directory
max_epoch_reached = False
while not max_epoch_reached :
iteration = iteration + 1;
epochNo = (iteration / n_train_batches) + 1
batchId= iteration % n_train_batches;
currentCost=train_model(batchId)
#print "Cost = %f" %(currentCost)
if iteration % validationFrequency == 0:
validation_err = validate()
train_err = train()
test_err = test()
self.add_train_data(epochNo, train_err, validation_err, test_err)
print "Epoch no: %d Validation Loss = %f" %(epochNo,validation_err*100)
if validation_err < bestValidationLoss:
bestValidationLoss = validation_err
if epochNo > minEpochs and validation_err *self.early_stopping_threshold > bestValidationLoss:
#print "------------------------Validation Loss = %f" %(validationLoss*100)
break;
if epochNo >= self.n_epochs:
max_epoch_reached = True
testLoss=test()
trainer.save_errors(self, directory)
repfields_final_path=os.path.join(directory,"repFields.png")
W_vals=nn.W1.get_value(borrow=True)
display(W_vals,repfields_final_path)
print "iteration %d complete. Cost = %f Best Validation Loss = %f Test Loss = %f" %(iteration,currentCost,bestValidationLoss *100,testLoss *100)
def train_NN(self, nn):
trainer.train_NN(self, nn)
data_loader = DataLoader()
datasets = data_loader.load_shared_data()
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# compute number of minibatches for training, validation and testing
batch_size = self.batch_size;
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
index = T.lscalar() # index to a [mini]batch
x = nn.input # the data is presented as rasterized images
y = T.ivector('y')
cost = (
nn.negative_log_likelihood_dropout(y)
+ self.L1_lambda * nn.L1
+ self.L2_lambda * nn.L2
)
test_err_model = theano.function(
inputs=[index],
outputs=nn.errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]
}
)
validate_err_model = theano.function(
inputs=[index],
outputs=nn.errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]
})
gparams = [T.grad(cost, param) for param in nn.params];
s =nn.params
# grad_model = theano.function(
# inputs=[index],
# outputs=gparams(y),
# givens={
# x: valid_set_x[index * batch_size:(index + 1) * batch_size],
# y: valid_set_y[index * batch_size:(index + 1) * batch_size]
# })
# mean_square_t = [theano.shared(
# value=np.zeros(
# (param.shape),
# dtype=theano.config.floatX
# ), borrow=True
# ) for param in mlp.params]
l=[]
mean_square_t=l
mlp_params = []
#mlp_params = theano.printing.Print('this is a very important value')(mlp_params)
for param in nn.params:
p=param.get_value(borrow = True)
obj=theano.shared(
value=np.ones(
(p.shape),
dtype=theano.config.floatX # @UndefinedVariable
),
name = "xysdfa" ,
borrow=True
)
#mlp_params.append(theano.printing.Print('asdfas')(param))
mlp_params.append(param)
#obj=theano.shared(1)
mean_square_t.append(obj)
#mean_square_t = [np.zeros(param.shape) for param in mlp.params]
#mean_square_t_minus_1 = np.zeros(mlp.params.shape);
#mean_square_update = [(r, self.decay * r + (1 - self.decay) * g**2) for r, g in zip(mean_square_t, gparams)]
new_mean_square_t =[self.decay * mt + (1-self.decay) * gp**2 for gp,mt in zip (gparams,mean_square_t)]
mean_square_update =[(t,t_plus_1) for t,t_plus_1 in zip(mean_square_t,new_mean_square_t)]
param_update =[(param, param - ( self.learning_rate*gp/(T.sqrt(mt+1e-8)))) for param, gp,mt in zip(mlp_params, gparams,new_mean_square_t) ]
#mt+1e-8
#r_t_minus_1 = np.zeros(mlp.params.shape);
#r_t = np.zeros(mlp.params.shape);
# updates = [
# (param, param - self.learning_rate * gparam)
# for param, gparam in zip(mlp.params, gparams)
# ]
train_model = theano.function(
inputs=[index],
outputs=cost,
#mode='DebugMode',
updates=mean_square_update + param_update,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size]
}
)
train_model_print = theano.function(
inputs=[index],
outputs=cost,
mode='DebugMode',
updates=mean_square_update + param_update,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size]
}
)
train_err_model = theano.function(
inputs=[index],
outputs=nn.errors(y),
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]
}
)
def validate():
validation_losses = [validate_err_model(i)
for i in xrange(n_valid_batches)]
this_validation_loss = np.mean(validation_losses)
return this_validation_loss;
def test():
test_losses = [test_err_model(i)
for i in xrange(n_test_batches)]
this_test_loss = np.mean(test_losses)
return this_test_loss;
def train():
train_losses = [train_err_model(i)
for i in xrange(n_train_batches)]
this_train_loss = np.mean(train_losses)
return this_train_loss;
print '... training'
bestValidationLoss = np.inf
minEpochs = 4
validationFrequency = n_train_batches;
iteration = 0;
directory=check_create_observations_dir()
self.output_directory = directory
max_epoch_reached = False
while not max_epoch_reached :
iteration = iteration + 1;
epochNo = (iteration / n_train_batches) + 1
batchId= iteration % n_train_batches;
currentCost=train_model(batchId)
#print "Cost = %f" %(currentCost)
if iteration % validationFrequency == 0:
validation_err = validate()
train_err = train()
test_err = test()
self.add_train_data(epochNo, train_err, validation_err, test_err)
print "Epoch no: %d Validation Loss = %f" %(epochNo,validation_err*100)
if validation_err < bestValidationLoss:
bestValidationLoss = validation_err
if epochNo > minEpochs and validation_err *self.early_stopping_threshold > bestValidationLoss:
#print "------------------------Validation Loss = %f" %(validationLoss*100)
break;
if epochNo >= self.n_epochs:
max_epoch_reached = True
testLoss=test()
trainer.save_errors(self, directory)
repfields_final_path=os.path.join(directory,"repFields.png")
W_vals=nn.W1.get_value(borrow=True)
display(W_vals,repfields_final_path)
print "iteration %d complete. Cost = %f Best Validation Loss = %f Test Loss = %f" %(iteration,currentCost,bestValidationLoss *100,testLoss *100)