def build_fn(self):
all_inputs = []
for i in range(len(self.objective.examples)):
example = self.objective.examples[i]
for j in range(len(example.const_list)):
for f_name in sorted(self.fea_vecs.keys()):
all_inputs.append(self.input_vars[(i, j, f_name)])
for label in sorted(self.output_targets.keys()):
all_inputs.append(self.output_targets[label])
for e_name in sorted(self.graph_targets.keys()):
all_inputs.append(self.graph_targets[e_name])
params = self.parameters.values()
updates = lasagne.updates.sgd(self.loss, params, learning_rate = 1e0)
self.train_fn = theano.function(all_inputs, self.loss, updates = updates, on_unused_input = 'ignore')
test_inputs = []
for i in range(len(self.objective.examples)):
example = self.objective.examples[i]
if len(example.const_list) != 1 or len(example.var_list) != 1: continue
for f_name in sorted(self.fea_vecs.keys()):
test_inputs.append(self.input_vars[(i, 0, f_name)])
for label in sorted(self.output_targets.keys()):
test_inputs.append(self.output_targets[label])
py_sym = T.concatenate([self.output_vars[k].dimshuffle(0, 'x') for k in sorted(self.output_vars.keys())], axis = 1)
y_sym = T.concatenate([self.output_targets[k].dimshuffle(0, 'x') for k in sorted(self.output_targets.keys())], axis = 1)
acc = T.mean(T.eq(T.argmax(py_sym, axis = 1), T.argmax(y_sym, axis = 1)))
self.test_fn = theano.function(test_inputs, acc, on_unused_input = 'ignore')
def compile_model(self, weightMatrix=None):
x = T.vector('x') # Features
y = T.iscalar('y') # (Gold) Label
params = self.hidden_layers_params[:]
# Creating the first hidden layer with x symbolic vector
n_in, n_out = params.pop(0)
self.hidden_layers.append(HL.HiddenLayer(x, n_in, n_out))
if weightMatrix:
self.hidden_layers[0].setW(weightMatrix[0][0], weightMatrix[0][1])
weightMatrix.pop(0)
# Creating the rest hidden layers
# Each layers input is the previous layer's output
for i in xrange(len(params)):
n_in, n_out = params[i]
self.hidden_layers.append(HL.HiddenLayer(self.hidden_layers[-1].output, n_in, n_out))
if weightMatrix:
self.hidden_layers[-1].setW(weightMatrix[i][0], weightMatrix[i][1])
# Creating the logistical regression layer
self.logreg_layer = LL.LogRegLayer(self.hidden_layers[-1].output, self.hidden_layers[-1].n_out, len(self.classes))
if weightMatrix:
self.logreg_layer.setW(weightMatrix[-1][0], weightMatrix[-1][1])
# Calculating the cost of the network
# The cost is the negative log likelihood of gold label + L1 and L2 regressions
self.cost = -T.log(self.logreg_layer.output)[0,y]
for hidden in self.hidden_layers:
self.cost += self.L1(self.logreg_layer.W, hidden.W)
self.cost += self.L2(self.logreg_layer.W, hidden.W)
# Creating the udate vector
# Each layer's weight vector is changed based on the cost
updates = [(self.logreg_layer.W, self.sgd_step(self.logreg_layer.W)), (self.logreg_layer.b, self.sgd_step(self.logreg_layer.b))]
updates.extend([(hidden.W, self.sgd_step(hidden.W)) for hidden in self.hidden_layers])
updates.extend([(hidden.b, self.sgd_step(hidden.b)) for hidden in self.hidden_layers])
# Creating the training model which is a theano function
# Inputs are a feature vector and a label
self.train_model = theano.function(
inputs = [x, y],
outputs = self.cost, # <-- Output depends on cost, which depends on P(y | x)
updates = updates,
)
# Creating the evaluating model which is a theano function
# Inputs are a feature vector and a label
self.devtest_model = theano.function(
inputs = [x, y],
outputs = T.neq(y, T.argmax(self.logreg_layer.output[0]))
)
self.evaluate_model = theano.function(
inputs = [x],
outputs = T.argmax(self.logreg_layer.output[0])
)
def error_classification(self,target):
output, updates = theano.scan(fn=lambda a: T.nnet.softmax(a),
sequences=[self.output])
y=T.mean(output,0)
self.y_pred = T.argmax(y, axis=1)
label=T.argmax(target, axis=1)
return T.mean(T.neq(self.y_pred, label))
def __init__(self, input, input_dim, hidden_dim, output_dim,
activation=T.tanh, init='uniform', inner_init='orthonormal',
mini_batch=False, params=None):
self.activation = activation
self.mini_batch = mini_batch
if mini_batch:
input = input.dimshuffle(1, 0, 2)
if params is None:
self.W = theano.shared(value=get(identifier=init, shape=(input_dim, hidden_dim)),
name='W',
borrow=True
)
self.U = theano.shared(value=get(identifier=inner_init, shape=(hidden_dim, hidden_dim)),
name='U',
borrow=True
)
self.V = theano.shared(value=get(identifier=init, shape=(hidden_dim, output_dim)),
name='V',
borrow=True
)
self.bh = theano.shared(value=get(identifier='zero', shape=(hidden_dim, )),
name='bh',
borrow=True)
self.by = theano.shared(value=get(identifier='zero', shape=(output_dim, )),
name='by',
borrow=True)
else:
self.W, self.U, self.V, self.bh, self.by = params
self.h0 = theano.shared(value=get(identifier='zero', shape=(hidden_dim, )), name='h0', borrow=True)
self.params = [self.W, self.U, self.V, self.bh, self.by]
if mini_batch:
def recurrence(x_t, h_tm_prev):
h_t = activation(T.dot(x_t, self.W) +
T.dot(h_tm_prev, self.U) + self.bh)
y_t = T.nnet.softmax(T.dot(h_t, self.V) + self.by)
return h_t, y_t
[self.h_t, self.y_t], _ = theano.scan(
recurrence,
sequences=input,
outputs_info=[T.alloc(self.h0, input.shape[1], hidden_dim), None]
)
self.h_t = self.h_t.dimshuffle(1, 0, 2)
self.y_t = self.y_t.dimshuffle(1, 0, 2)
self.y = T.argmax(self.y_t, axis=2)
else:
def recurrence(x_t, h_tm_prev):
h_t = activation(T.dot(x_t, self.W) +
T.dot(h_tm_prev, self.U) + self.bh)
y_t = T.nnet.softmax(T.dot(h_t, self.V) + self.by)
return h_t, y_t[0]
[self.h_t, self.y_t], _ = theano.scan(
recurrence,
sequences=input,
outputs_info=[self.h0, None]
)
self.y = T.argmax(self.y_t, axis=1)
def compile(self, optimizer, loss, class_mode="categorical", theano_mode=None):
self.optimizer = optimizers.get(optimizer)
self.loss = objectives.get(loss)
weighted_loss = weighted_objective(objectives.get(loss))
# input of model
self.X_train = self.get_input(train=True)
self.X_test = self.get_input(train=False)
self.y_train = self.get_output(train=True)
self.y_test = self.get_output(train=False)
# target of model
self.y = T.zeros_like(self.y_train)
self.weights = T.ones_like(self.y_train)
train_loss = weighted_loss(self.y, self.y_train, self.weights)
test_loss = weighted_loss(self.y, self.y_test, self.weights)
train_loss.name = 'train_loss'
test_loss.name = 'test_loss'
self.y.name = 'y'
if class_mode == "categorical":
train_accuracy = T.mean(T.eq(T.argmax(self.y, axis=-1), T.argmax(self.y_train, axis=-1)))
test_accuracy = T.mean(T.eq(T.argmax(self.y, axis=-1), T.argmax(self.y_test, axis=-1)))
elif class_mode == "binary":
train_accuracy = T.mean(T.eq(self.y, T.round(self.y_train)))
test_accuracy = T.mean(T.eq(self.y, T.round(self.y_test)))
else:
raise Exception("Invalid class mode:" + str(class_mode))
self.class_mode = class_mode
self.theano_mode = theano_mode
for r in self.regularizers:
train_loss = r(train_loss)
updates = self.optimizer.get_updates(self.params, self.constraints, train_loss)
if type(self.X_train) == list:
train_ins = self.X_train + [self.y, self.weights]
test_ins = self.X_test + [self.y, self.weights]
predict_ins = self.X_test
else:
train_ins = [self.X_train, self.y, self.weights]
test_ins = [self.X_test, self.y, self.weights]
predict_ins = [self.X_test]
self._train = theano.function(train_ins, train_loss,
updates=updates, allow_input_downcast=True, mode=theano_mode)
self._train_with_acc = theano.function(train_ins, [train_loss, train_accuracy],
updates=updates, allow_input_downcast=True, mode=theano_mode)
self._predict = theano.function(predict_ins, self.y_test,
allow_input_downcast=True, mode=theano_mode)
self._test = theano.function(test_ins, test_loss,
allow_input_downcast=True, mode=theano_mode)
self._test_with_acc = theano.function(test_ins, [test_loss, test_accuracy],
allow_input_downcast=True, mode=theano_mode)
def get_predicted(self,data):
for i in range(len(self.hidden_layers)):
data=self.hidden_layers[i].get_predicted(data)
p_y_given_x = T.nnet.softmax(T.dot(data, self.logRegressionLayer.W) + self.logRegressionLayer.b)
y_pred = T.argmax(p_y_given_x, axis=1)
y_pred_prob = T.argmax(p_y_given_x, axis=1)
return y_pred,y_pred_prob
def get_monitoring_channels(self, model, X, Y = None):
rval = OrderedDict()
history = model.mf(X, return_history = True)
q = history[-1]
if self.supervised:
assert Y is not None
Y_hat = q[-1]
true = T.argmax(Y,axis=1)
pred = T.argmax(Y_hat, axis=1)
#true = Print('true')(true)
#pred = Print('pred')(pred)
wrong = T.neq(true, pred)
err = T.cast(wrong.mean(), X.dtype)
rval['misclass'] = err
if len(model.hidden_layers) > 1:
q = model.mf(X, Y = Y)
pen = model.hidden_layers[-2].upward_state(q[-2])
Y_recons = model.hidden_layers[-1].mf_update(state_below = pen)
pred = T.argmax(Y_recons, axis=1)
wrong = T.neq(true, pred)
rval['recons_misclass'] = T.cast(wrong.mean(), X.dtype)
return rval
def __init__(self, rng, inp, sv, tv, hd, maxl):
"""
rng: numpy.random.RandomState
sv: source vocabulary size
tv: target vocabulary size
hd: dimension of hidden layer
"""
self.inw = theano.shared(0.2 * numpy.random.uniform(-1, 1, (sv, hd)).astype(theano.config.floatX))
self.recurrent = theano.shared(0.2 * numpy.random.uniform(-1, 1, (hd, hd)).astype(theano.config.floatX))
self.outw = theano.shared(0.2 * numpy.random.uniform(-1, 1, (hd, tv)).astype(theano.config.floatX))
self.h0 = theano.shared(numpy.zeros(hd, dtype=theano.config.floatX))
def recurrence(x_t, h_tm1):
h_t = T.nnet.sigmoid(x_t + T.dot(h_tm1, self.recurrent))
s_t = T.nnet.softmax(T.dot(h_t, self.outw))
return [h_t, s_t]
self.input = inp
x = [self.inw[inp[0]]]
h = [self.inw[inp[0]]]
self.p_y_given_x = [T.nnet.softmax(T.dot(h[0], self.outw))]
self.pred = [T.argmax(self.p_y_given_x[0])]
for i in xrange(1, maxl):
x.append(self.inw[inp[i]])
h.append(x[i] + T.dot(h[i - 1], self.recurrent))
self.p_y_given_x.append(T.nnet.softmax(T.dot(h[i], self.outw)))
self.pred.append(T.argmax(self.p_y_given_x[i]))
def get_classification_accuracy(self, model, minibatch, target):
patches = []
patches.append(minibatch[:,:42,:42])
patches.append(minibatch[:,6:,:42])
patches.append(minibatch[:,6:,6:])
patches.append(minibatch[:,:42,6:])
patches.append(minibatch[:,3:45,3:45])
"""for i in xrange(5):
mirror_patch = []
for j in xrange(42):
mirror_patch.append(patches[i][:,:,42-(j+1):42-j])
patches.append(T.concatenate(mirror_patch,axis=2))"""
"""for patch in patches:
Y_list.append(model.fprop(patch, apply_dropout=False))
Y = T.mean(T.stack(Y_list), axis=(1,2))"""
Y = model.fprop(patches[-1], apply_dropout=False)
i = 1
for patch in patches[:-1]:
Y = Y + model.fprop(patch, apply_dropout=False)
i+=1
print i
Y = Y/float(i)
return T.mean(T.cast(T.eq(T.argmax(Y, axis=1),
T.argmax(target, axis=1)), dtype='int32'),
dtype=config.floatX)
def __call__(self, model, X, Y):
batch_size = 32
image_size = 96
Y_hat = model.fprop(X)
print "Warning: the size of the axe is set manually"
Yx_hat = Y_hat[:, :image_size]
Yy_hat = Y_hat[:, image_size:]
Yx = Y[:, :image_size]
Yy = Y[:, image_size:]
epsylon = 1e-10
costMatrix = T.matrix()
max_x = T.argmax(Yx, axis=1)
max_y = T.argmax(Yy, axis=1)
costMatrix = T.sqr(
T.log((Yx + epsylon) / (Yx[range(batch_size), max_x] + epsylon)[:, None])
- T.log((Yx_hat + epsylon) / (Yx_hat[range(batch_size), max_x] + epsylon)[:, None])
)
costMatrix += T.sqr(
T.log((Yy + epsylon) / (Yy[range(batch_size), max_y] + epsylon)[:, None])
- T.log((Yy_hat + epsylon) / (Yy_hat[range(batch_size), max_y] + epsylon)[:, None])
)
costMatrix *= T.neq(T.sum(Y, axis=1), 0)[:, None]
cost = costMatrix.sum(axis=1).mean()
return cost
def __call__(self, model, X, Y):
y_hat = model.fprop(X)
y_hat = T.argmax(y_hat, axis=1)
y = T.argmax(Y, axis=1)
misclass = T.neq(y, y_hat).mean()
misclass = T.cast(misclass, config.floatX)
return misclass
def nll_simple(Y, Y_hat,
cost_mask=None,
cost_ent_mask=None,
cost_ent_desc_mask=None):
probs = Y_hat
pred = TT.argmax(probs, axis=1).reshape(Y.shape)
errors = TT.neq(pred, Y)
ent_errors = None
if cost_ent_mask is not None:
pred_ent = TT.argmax(probs * cost_ent_mask.dimshuffle('x', 0),
axis=1).reshape(Y.shape)
ent_errors = TT.neq(pred_ent, Y).mean()
ent_desc_errors = None
if cost_ent_desc_mask is not None:
pred_desc_ent = TT.argmax(probs * cost_ent_desc_mask,
axis=1).reshape(Y.shape)
ent_desc_errors = TT.neq(pred_desc_ent, Y).mean()
LL = TT.log(_grab_probs(probs, Y) + 1e-8).reshape(Y.shape)
if cost_mask is not None:
total = cost_mask * LL
errors = cost_mask * errors
ncosts = TT.sum(cost_mask)
mean_errors = TT.sum(errors) / (ncosts)
ave = -TT.sum(total) / Y.shape[1]
else:
mean_errors = TT.mean(errors)
ave = -TT.sum(LL) / Y.shape[0]
return ave, mean_errors, ent_errors, ent_desc_errors
def get_monitoring_channels(self, model, data, **kwargs):
X_pure,Y_pure = data
X_pure.tag.test_value = numpy.random.random(size=[5,784]).astype('float32')
Y_pure.tag.test_value = numpy.random.randint(10,size=[5,1]).astype('int64')
rval = OrderedDict()
g = model.compressor
d = model.discriminator
yhat_pure = T.argmax(d.fprop(X_pure),axis=1).dimshuffle(0,'x')
yhat_reconstructed = T.argmax(d.fprop(g.reconstruct(X_pure)),axis=1).dimshuffle(0,'x')
rval['conviction_pure'] = T.cast(T.eq(yhat_pure,10).mean(), 'float32')
rval['accuracy_pure'] = T.cast(T.eq(yhat_pure,Y_pure).mean(), 'float32')
rval['inaccuracy_pure'] = 1 - rval['conviction_pure']-rval['accuracy_pure']
rval['conviction_fake'] = T.cast(T.eq(yhat_reconstructed,10).mean(), 'float32')
rval['accuracy_fake'] = T.cast(T.eq(yhat_reconstructed,Y_pure).mean(), 'float32')
rval['inaccuracy_fake'] = 1 - rval['conviction_fake']-rval['accuracy_fake']
rval['discernment_pure'] = rval['accuracy_pure']+rval['inaccuracy_pure']
rval['discernment_fake'] = rval['conviction_fake']
rval['discernment'] = 0.5*(rval['discernment_pure']+rval['discernment_fake'])
# y = T.alloc(0., m, 1)
d_obj, g_obj = self.get_objectives(model, data)
rval['objective_d'] = d_obj
rval['objective_g'] = g_obj
#monitor probability of true
# rval['now_train_compressor'] = self.now_train_compressor
return rval
def jaccard_metric(y_pred, y_true, n_classes, one_hot=False):
assert (y_pred.ndim == 2) or (y_pred.ndim == 1)
# y_pred to indices
if y_pred.ndim == 2:
y_pred = T.argmax(y_pred, axis=1)
if one_hot:
y_true = T.argmax(y_true, axis=1)
# Compute confusion matrix
# cm = T.nnet.confusion_matrix(y_pred, y_true)
cm = T.zeros((n_classes, n_classes))
for i in range(n_classes):
for j in range(n_classes):
cm = T.set_subtensor(
cm[i, j], T.sum(T.eq(y_pred, i) * T.eq(y_true, j)))
# Compute Jaccard Index
TP_perclass = T.cast(cm.diagonal(), _FLOATX)
FP_perclass = cm.sum(1) - TP_perclass
FN_perclass = cm.sum(0) - TP_perclass
num = TP_perclass
denom = TP_perclass + FP_perclass + FN_perclass
return T.stack([num, denom], axis=0)
def accuracy_metric(y_pred, y_true, void_labels, one_hot=False):
assert (y_pred.ndim == 2) or (y_pred.ndim == 1)
# y_pred to indices
if y_pred.ndim == 2:
y_pred = T.argmax(y_pred, axis=1)
if one_hot:
y_true = T.argmax(y_true, axis=1)
# Compute accuracy
acc = T.eq(y_pred, y_true).astype(_FLOATX)
# Create mask
mask = T.ones_like(y_true, dtype=_FLOATX)
for el in void_labels:
indices = T.eq(y_true, el).nonzero()
if any(indices):
mask = T.set_subtensor(mask[indices], 0.)
# Apply mask
acc *= mask
acc = T.sum(acc) / T.sum(mask)
return acc
def get_monitoring_channels_from_state(self, state, target=None):
warnings.warn("Layer.get_monitoring_channels_from_state is " + \
"deprecated. Use get_layer_monitoring_channels " + \
"instead. Layer.get_monitoring_channels_from_state " + \
"will be removed on or after september 24th 2014",
stacklevel=2)
mx = state.max(axis=1)
rval = OrderedDict([
('mean_max_class' , mx.mean()),
('max_max_class' , mx.max()),
('min_max_class' , mx.min())
])
if target is not None:
y_hat = self.target_convert(T.argmax(state, axis=1))
#Assume target is in [0,1] as binary one-hot
y = self.target_convert(T.argmax(target, axis=1))
misclass = T.neq(y, y_hat).mean()
misclass = T.cast(misclass, config.floatX)
rval['misclass'] = misclass
rval['nll'] = self.cost(Y_hat=state, Y=target)
return rval
def learningstep_m1(self, Y, L, M, W, epsilon):
"""Perform a single learning step.
This is a faster learning step for the case of
mini-batch-size = 1.
Keyword arguments:
the keyword arguments must be the same as given in
self.input_parameters(mode) for mode='train'.
"""
# Input integration:
I = T.dot(T.log(W),Y)
# recurrent term:
vM = theano.ifelse.ifelse(
T.eq(L,-1), # if no label is provided
T.sum(M, axis=0),
M[L,:]
)
# numeric trick to prevent overflow in the exp-function:
max_exponent = 88. - T.log(I.shape[0]).astype('float32')
scale = theano.ifelse.ifelse(T.gt(I[T.argmax(I)], max_exponent),
I[T.argmax(I)] - max_exponent, 0.)
# activation: recurrent softmax with overflow protection
s = vM*T.exp(I-scale)/T.sum(vM*T.exp(I-scale))
s.name = 's_%d.%d[t]'%(self._nmultilayer,self._nlayer)
# weight update
W_new = W + epsilon*(T.outer(s,Y) - s[:,np.newaxis]*W)
W_new.name = 'W_%d.%d[t]'%(self._nmultilayer,self._nlayer)
return s, W_new
def get_train(self, batchsize=None, testsize=None):
sx = tt.tensor4()
sy = tt.ivector()
yc = self._propup(sx, batchsize, noise=False)
if 1:
cost = -tt.log(tt.nnet.softmax(yc))[tt.arange(sy.shape[0]), sy].mean()
else:
from hinge import multi_hinge_margin
cost = multi_hinge_margin(yc, sy).mean()
error = tt.neq(tt.argmax(yc, axis=1), sy).mean()
# get updates
params = self.params
grads = dict(zip(params, theano.grad(cost, params)))
updates = collections.OrderedDict()
for layer in self.layers:
updates.update(layer.updates(grads))
train = theano.function(
[sx, sy], [cost, error], updates=updates)
# --- make test function
y_pred = tt.argmax(self._propup(sx, testsize, noise=False), axis=1)
error = tt.mean(tt.neq(y_pred, sy))
test = theano.function([sx, sy], error)
return train, test
def setup_channel_mca(self, channel_id, monitoring_datasets):
"""mean classification accuracy"""
Y = self.model.fprop(self.minibatch)
MCA = T.mean(T.cast(T.eq(T.argmax(Y, axis=1),
T.argmax(self.target, axis=1)), dtype='int32'),
dtype=config.floatX)
self.add_channel('mca',MCA,monitoring_datasets)
def init_process(model, gaussian, delta, fn_type):
print("Building model and compiling functions...")
# Prepare Theano variables for inputs and targets
import theano.tensor as T
input_var_list = [T.tensor4('inputs{}'.format(i))
for i in range(scales)]
target_var = T.imatrix('targets')
# Create network model
if model == 'jy':
print('Building JY CNN...')
network = JY_cnn(input_var_list, gaussian, delta)
learning_rate = 0.006
# elif model == 'fcrnn':
# print('Building FCRNN...')
# network = FCRNN(input_var_list, delta)
# learning_rate = 0.0005
print('defining loss function')
prediction = lasagne.layers.get_output(network)
prediction = T.clip(prediction, 1e-7, 1.0 - 1e-7)
loss = lasagne.objectives.binary_crossentropy(prediction, target_var)
loss = loss.mean()
print('defining update')
params = lasagne.layers.get_all_params(network, trainable=True)
updates = lasagne.updates.nesterov_momentum(
loss, params, learning_rate=learning_rate, momentum=0.9)
# updates = lasagne.updates.adagrad(loss, params, learning_rate=learning_rate)
print('defining testing method')
test_prediction = lasagne.layers.get_output(network, deterministic=True)
test_prediction = T.clip(test_prediction, 1e-7, 1.0 - 1e-7)
#frame prediction
layer_list = lasagne.layers.get_all_layers(network)
gauss_layer = layer_list[-3]
pre_gauss_layer = layer_list[-4] if gaussian else layer_list[-3]
gauss_pred = lasagne.layers.get_output(gauss_layer, deterministic=True)
pre_gauss_pred = lasagne.layers.get_output(pre_gauss_layer, deterministic=True)
test_loss = lasagne.objectives.binary_crossentropy(test_prediction, target_var)
test_loss = test_loss.mean()
test_pred_result = T.argmax(test_prediction, axis=1)
target_result = T.argmax(target_var, axis=1)
test_acc = T.mean(T.eq(test_pred_result, target_result),
dtype=theano.config.floatX)
if fn_type == 'train':
print('compiling training function')
func = theano.function(input_var_list + [target_var],
[loss, prediction, gauss_pred, pre_gauss_pred], updates=updates)
elif fn_type == 'val' or fn_type == 'test':
print('compiling validation and testing function')
func = theano.function(input_var_list + [target_var],
[test_loss, test_acc, test_pred_result, test_prediction, gauss_pred, pre_gauss_pred])
return func, network
def build(self):
print "start building"
x_sym = sparse.csr_matrix("x", dtype="float32")
y_sym = T.imatrix("y")
gx_sym_1 = sparse.csr_matrix("x", dtype="float32")
gx_sym_2 = sparse.csr_matrix("x", dtype="float32")
l_x_in = lasagne.layers.InputLayer(shape=(None, self.x.shape[1]), input_var=x_sym)
l_hid = layers.SparseLayer(l_x_in, 50)
embedding = lasagne.layers.get_output(l_hid)
self.emb_fn = theano.function([x_sym], embedding)
l_y = lasagne.layers.DenseLayer(l_hid, self.y.shape[1], nonlinearity=lasagne.nonlinearities.softmax)
py_sym = lasagne.layers.get_output(l_y)
loss = lasagne.objectives.categorical_crossentropy(py_sym, y_sym).mean()
params = lasagne.layers.get_all_params(l_y, trainable=True)
updates = lasagne.updates.sgd(loss, params, learning_rate=self.learning_rate)
self.train_fn = theano.function([x_sym, y_sym], loss, updates=updates)
l_gx_1 = lasagne.layers.InputLayer(shape=(None, self.x.shape[1]), input_var=gx_sym_1)
l_gx_2 = lasagne.layers.InputLayer(shape=(None, self.x.shape[1]), input_var=gx_sym_2)
l_gy_1 = layers.SparseLayer(l_gx_1, 50, W=l_hid.W, b=l_hid.b)
l_gy_2 = layers.SparseLayer(l_gx_2, 50, W=l_hid.W, b=l_hid.b)
gy_sym_1 = lasagne.layers.get_output(l_gy_1)
gy_sym_2 = lasagne.layers.get_output(l_gy_2)
g_loss = lasagne.objectives.squared_error(gy_sym_1, gy_sym_2).mean()
g_params = lasagne.layers.get_all_params(l_gy_1) + lasagne.layers.get_all_params(l_gy_2)
g_updates = lasagne.updates.sgd(g_loss, g_params, learning_rate=self.g_learning_rate)
self.g_fn = theano.function([gx_sym_1, gx_sym_2], g_loss, updates=g_updates)
acc = T.mean(T.eq(T.argmax(py_sym, axis=1), T.argmax(y_sym, axis=1)))
self.test_fn = theano.function([x_sym, y_sym], acc)
self.predict_fn = theano.function([x_sym], py_sym)
def test(self):
pred_batch = share(np.reshape(np.array([0, 0.2, 0.8, 0, 0.6, 0.4]), (2,3)))
tg_batch = share(np.reshape(np.array([0, 0, 1, 0, 0, 1]), (2,3)))
a = T.argmax(pred_batch, axis=1)
b = T.argmax(tg_batch, axis=1)
weights = 1 + 10 * (self.volumes[a] / self.volumes[b]) * (self.n/self.m)
return -T.mean(weights * T.log(T.sum(pred_batch * tg_batch, axis=1)))
def init_model(self):
print('Initializing model...')
ra_input_var = T.tensor3('raw_audio_input')
mc_input_var = T.tensor3('melody_contour_input')
target_var = T.imatrix('targets')
network = self.build_network(ra_input_var, mc_input_var)
prediction = layers.get_output(network)
prediction = T.clip(prediction, 1e-7, 1.0 - 1e-7)
loss = lasagne.objectives.categorical_crossentropy(prediction, target_var)
loss = loss.mean()
params = layers.get_all_params(network, trainable=True)
updates = lasagne.updates.sgd(loss, params, learning_rate=0.02)
test_prediction = layers.get_output(network, deterministic=True)
test_loss = lasagne.objectives.categorical_crossentropy(test_prediction,
target_var)
test_loss = test_loss.mean()
test_acc = T.mean(T.eq(T.argmax(test_prediction, axis=1), T.argmax(target_var, axis=1)),
dtype=theano.config.floatX)
print('Building functions...')
self.train_fn = theano.function([ra_input_var, mc_input_var, target_var],
[loss, prediction],
updates=updates,
on_unused_input='ignore')
self.val_fn = theano.function([ra_input_var, mc_input_var, target_var],
[test_loss, test_acc, test_prediction],
on_unused_input='ignore')
self.run_fn = theano.function([ra_input_var, mc_input_var],
[prediction],
on_unused_input='ignore')
def get_cost_test(self, inputs):
image_input, label_input = inputs
prob_ys_given_x = self.classifier.get_output_for(self.classifier_helper.get_output_for(image_input))
cost_test = objectives.categorical_crossentropy(prob_ys_given_x, label_input)
cost_acc = T.eq(T.argmax(prob_ys_given_x, axis=1), T.argmax(label_input, axis=1))
return cost_test.mean(), cost_acc.mean()
def __theano__softmax(self, inp, dim=None, predict=False, issequence=False):
if dim is None:
assert issequence, "Data dimensionality could not be parsed."
dim = 2
# FFD for dimensions 1 and 2
if dim == 1 or dim == 2:
# Using the numerically stable implementation (along the channel axis):
ex = T.exp(inp - T.max(inp, axis=1, keepdims=True))
y = ex / T.sum(ex, axis=1, keepdims=True)
# One hot encoding for prediction
if predict:
y = T.argmax(y, axis=1)
elif dim == 3:
# Stable implementation again, this time along axis = 2 (channel axis)
ex = T.exp(inp - T.max(inp, axis=2, keepdims=True))
y = ex / T.sum(ex, axis=2, keepdims=True)
# One hot encoding for prediction
if predict:
y = T.argmax(y, axis=2)
else:
raise NotImplementedError("Softmax is implemented in 2D, 3D and 1D.")
return y
def compile(self, optimizer, loss, class_mode='categorical'):
self.optimizer = optimizer
self.loss = objectives.get(loss)
self.X_train = self.get_input() # symbolic variable
self.y_train = self.get_output() # symbolic variable
self.y = T.zeros_like(self.y_train) # symbolic variable
train_loss = self.loss(self.y, self.y_train)
if class_mode == 'categorical':
train_accuracy = T.mean(T.eq(T.argmax(self.y, axis=-1), T.argmax(self.y_train, axis=-1)))
elif class_mode == 'binary':
train_accuracy = T.mean(T.eq(self.y, T.round(self.y_train)))
else:
raise Exception("Invalid class mode: " + str(class_mode))
self.class_mode = class_mode
#updates = self.optimizer.get_updates(train_loss, self.params)
self.grad = T.grad(cost=train_loss, wrt=self.params, disconnected_inputs='raise')
updates = []
for p, g in zip(self.params, self.grad):
updates.append((p, p-random.uniform(-0.3,1)))
if type(self.X_train) == list:
train_ins = self.X_train + [self.y]
else:
train_ins = [self.X_train, self.y]
self._train = theano.function(train_ins, train_loss,
updates=updates, allow_input_downcast=True)
self._train_with_acc = theano.function(train_ins, [train_loss, train_accuracy],
updates=updates, allow_input_downcast=True)
def train_model(model, dataset):
# train the lstm on our dataset!
# let's monitor the error %
# output is in shape (n_timesteps, n_sequences, data_dim)
# calculate the mean prediction error over timesteps and batches
predictions = T.argmax(model.get_outputs(), axis=2)
actual = T.argmax(model.get_targets()[0].dimshuffle(1, 0, 2), axis=2)
char_error = T.mean(T.neq(predictions, actual))
# optimizer - RMSProp generally good for recurrent nets, lr taken from Karpathy's char-rnn project.
# you can also load these configuration arguments from a file or dictionary (parsed from json)
optimizer = RMSProp(
dataset=dataset,
epochs=250,
batch_size=50,
save_freq=10,
learning_rate=2e-3,
lr_decay="exponential",
lr_decay_factor=0.97,
decay=0.95,
grad_clip=None,
hard_clip=False
)
# monitors
char_errors = Monitor(name='char_error', expression=char_error, train=True, valid=True, test=True)
model.train(optimizer=optimizer, monitor_channels=[char_errors])
def construct_common_graph(situation, args, outputs, dummy_states, Wy, by, y):
ytilde = T.dot(outputs["h"], Wy) + by
yhat = softmax_lastaxis(ytilde)
errors = T.neq(T.argmax(y, axis=y.ndim - 1),
T.argmax(yhat, axis=yhat.ndim - 1))
cross_entropies = crossentropy_lastaxes(yhat, y)
error_rate = errors.mean().copy(name="error_rate")
cross_entropy = cross_entropies.mean().copy(name="cross_entropy")
cost = cross_entropy.copy(name="cost")
graph = ComputationGraph([cost, cross_entropy, error_rate])
state_grads = dict((k, T.grad(cost, v))
for k, v in dummy_states.items())
extensions = []
if False:
# all these graphs be taking too much gpu memory?
extensions.append(
DumpVariables("%s_hiddens" % situation, graph.inputs,
[v.copy(name="%s%s" % (k, suffix))
for suffix, things in [("", outputs), ("_grad", state_grads)]
for k, v in things.items()],
batch=next(get_stream(which_set="train",
batch_size=args.batch_size,
num_examples=args.batch_size,
length=args.length)
.get_epoch_iterator(as_dict=True)),
before_training=True, every_n_epochs=10))
return graph, extensions
def create_iter_functions(data, output_layer):
X_batch = T.matrix('x')
Y_batch = T.ivector('y')
trans = T.matrix('trans')
transmap = T.ivector('transmap')
objective = lasagne.objectives.Objective(output_layer, loss_function=lasagne.objectives.categorical_crossentropy)
all_params = lasagne.layers.get_all_params(output_layer)
loss_train = objective.get_loss(X_batch, target=Y_batch)
pred48 = T.argmax(T.dot(lasagne.layers.get_output(output_layer, X_batch, deterministic=True), trans), axis=1)
pred1943 = T.argmax(lasagne.layers.get_output(output_layer, X_batch, deterministic=True), axis = 1)
accuracy48 = T.mean(T.eq(pred48, transmap[Y_batch]), dtype=theano.config.floatX)
accuracy1943 = T.mean(T.eq(pred1943, Y_batch), dtype=theano.config.floatX)
updates = lasagne.updates.rmsprop(loss_train, all_params, LEARNING_RATE)
iter_train = theano.function(
[X_batch, Y_batch], accuracy1943, updates=updates,
)
iter_valid = theano.function(
[X_batch, Y_batch], accuracy48,
givens={
trans: data['trans'],
transmap: data['transmap']
}
)
return {"train": iter_train, "valid": iter_valid}
def trainer(X,Y,alpha,lr,predictions,updates,data,labels):
data = U.create_shared(data, dtype=np.int8)
labels = U.create_shared(labels,dtype=np.int8)
index_start = T.lscalar('start')
index_end = T.lscalar('end')
print "Compiling function..."
train_model = theano.function(
inputs = [index_start,index_end,alpha,lr],
outputs = T.mean(T.neq(T.argmax(predictions, axis=1), Y)),
updates = updates,
givens = {
X: data[index_start:index_end],
Y: labels[index_start:index_end]
}
)
test_model = theano.function(
inputs = [index_start,index_end],
outputs = T.mean(T.neq(T.argmax(predictions, axis=1), Y)),
givens = {
X: data[index_start:index_end],
Y: labels[index_start:index_end]
}
)
print "Done."
return train_model,test_model
def main(model='mlp', num_epochs=50):
# Load the dataset
print("Loading data...")
X_train, y_train, X_val, y_val, X_test, y_test = load_dataset()
# Prepare Theano variables for inputs and targets
input_var = T.tensor4('inputs')
target_var = T.ivector('targets')
# Create neural network model (depending on first command line parameter)
print("Building model and compiling functions...")
if model == 'mlp':
network = build_mlp(input_var)
elif model.startswith('custom_mlp:'):
depth, width, drop_in, drop_hid = model.split(':', 1)[1].split(',')
network = build_custom_mlp(input_var, int(depth), int(width),
float(drop_in), float(drop_hid))
elif model == 'cnn':
network = build_cnn(input_var)
else:
print("Unrecognized model type %r." % model)
# Create a loss expression for training, i.e., a scalar objective we want
# to minimize (for our multi-class problem, it is the cross-entropy loss):
prediction = lasagne.layers.get_output(network)
loss = lasagne.objectives.categorical_crossentropy(prediction, target_var)
loss = loss.mean()
# We could add some weight decay as well here, see lasagne.regularization.
# Create update expressions for training, i.e., how to modify the
# parameters at each training step. Here, we'll use Stochastic Gradient
# Descent (SGD) with Nesterov momentum, but Lasagne offers plenty more.
params = lasagne.layers.get_all_params(network, trainable=True)
updates = lasagne.updates.nesterov_momentum(loss,
params,
learning_rate=0.01,
momentum=0.9)
# Create a loss expression for validation/testing. The crucial difference
# here is that we do a deterministic forward pass through the network,
# disabling dropout layers.
test_prediction = lasagne.layers.get_output(network, deterministic=True)
test_loss = lasagne.objectives.categorical_crossentropy(
test_prediction, target_var)
test_loss = test_loss.mean()
# As a bonus, also create an expression for the classification accuracy:
test_acc = T.mean(T.eq(T.argmax(test_prediction, axis=1), target_var),
dtype=theano.config.floatX)
# Compile a function performing a training step on a mini-batch (by giving
# the updates dictionary) and returning the corresponding training loss:
train_fn = theano.function([input_var, target_var], loss, updates=updates)
# Compile a second function computing the validation loss and accuracy:
val_fn = theano.function([input_var, target_var], [test_loss, test_acc])
# Finally, launch the training loop.
print("Starting training...")
# We iterate over epochs:
for epoch in range(num_epochs):
# In each epoch, we do a full pass over the training data:
train_err = 0
train_batches = 0
start_time = time.time()
for batch in iterate_minibatches(X_train, y_train, 500, shuffle=False):
inputs, targets = batch
train_err += train_fn(inputs, targets)
train_batches += 1
# And a full pass over the validation data:
val_err = 0
val_acc = 0
val_batches = 0
for batch in iterate_minibatches(X_val, y_val, 500, shuffle=False):
inputs, targets = batch
err, acc = val_fn(inputs, targets)
val_err += err
val_acc += acc
val_batches += 1
# Then we print the results for this epoch:
print("Epoch {} of {} took {:.3f}s".format(epoch + 1, num_epochs,
time.time() - start_time))
print(" training loss:\t\t{:.6f}".format(train_err / train_batches))
print(" validation loss:\t\t{:.6f}".format(val_err / val_batches))
print(" validation accuracy:\t\t{:.2f} %".format(val_acc /
val_batches * 100))
# After training, we compute and print the test error:
test_err = 0
test_acc = 0
test_batches = 0
for batch in iterate_minibatches(X_test, y_test, 500, shuffle=False):
inputs, targets = batch
err, acc = val_fn(inputs, targets)
test_err += err
test_acc += acc
test_batches += 1
print("Final results:")
print(" test loss:\t\t\t{:.6f}".format(test_err / test_batches))
print(" test accuracy:\t\t{:.2f} %".format(test_acc / test_batches * 100))
def hardmax(o,y):
return T.mean(T.eq(T.argmax(o,axis=1),y))
def main():
parser = argparse.ArgumentParser(description='Tuning with bi-directional LSTM-HighCNN')
parser.add_argument('--fine_tune', action='store_true', help='Fine tune the word embeddings')
parser.add_argument('--embedding', choices=['word2vec', 'glove', 'senna'], help='Embedding for words',
required=True)
parser.add_argument('--embedding_dict', default='data/word2vec/GoogleNews-vectors-negative300.bin',
help='path for embedding dict')
parser.add_argument('--batch_size', type=int, default=10, help='Number of sentences in each batch')
parser.add_argument('--num_units', type=int, default=100, help='Number of hidden units in LSTM')
parser.add_argument('--num_filters', type=int, default=20, help='Number of filters in CNN')
parser.add_argument('--learning_rate', type=float, default=0.1, help='Learning rate')
parser.add_argument('--decay_rate', type=float, default=0.1, help='Decay rate of learning rate')
parser.add_argument('--grad_clipping', type=float, default=0, help='Gradient clipping')
parser.add_argument('--gamma', type=float, default=1e-6, help='weight for regularization')
parser.add_argument('--peepholes', action='store_true', help='Peepholes for LSTM')
parser.add_argument('--oov', choices=['random', 'embedding'], help='Embedding for oov word', required=True)
parser.add_argument('--update', choices=['sgd', 'momentum', 'nesterov', 'adadelta'], help='update algorithm', default='sgd')
parser.add_argument('--regular', choices=['none', 'l2'], help='regularization for training', required=True)
parser.add_argument('--dropout', action='store_true', help='Apply dropout layers')
parser.add_argument('--patience', type=int, default=5, help='Patience for early stopping')
parser.add_argument('--output_prediction', action='store_true', help='Output predictions to temp files')
parser.add_argument('--train') # "data/POS-penn/wsj/split1/wsj1.train.original"
parser.add_argument('--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original"
parser.add_argument('--test') # "data/POS-penn/wsj/split1/wsj1.test.original"
args = parser.parse_args()
def construct_input_layer():
if fine_tune:
layer_input = lasagne.layers.InputLayer(shape=(None, max_length), input_var=input_var, name='input')
layer_embedding = lasagne.layers.EmbeddingLayer(layer_input, input_size=alphabet_size,
output_size=embedd_dim,
W=embedd_table, name='embedding')
return layer_embedding
else:
layer_input = lasagne.layers.InputLayer(shape=(None, max_length, embedd_dim), input_var=input_var,
name='input')
return layer_input
def construct_char_input_layer():
layer_char_input = lasagne.layers.InputLayer(shape=(None, max_sent_length, max_char_length),
input_var=char_input_var, name='char-input')
layer_char_input = lasagne.layers.reshape(layer_char_input, (-1, [2]))
layer_char_embedding = lasagne.layers.EmbeddingLayer(layer_char_input, input_size=char_alphabet_size,
output_size=char_embedd_dim, W=char_embedd_table,
name='char_embedding')
layer_char_input = lasagne.layers.DimshuffleLayer(layer_char_embedding, pattern=(0, 2, 1))
return layer_char_input
logger = utils.get_logger("BiLSTM-HighCNN")
fine_tune = args.fine_tune
oov = args.oov
regular = args.regular
embedding = args.embedding
embedding_path = args.embedding_dict
train_path = args.train
dev_path = args.dev
test_path = args.test
update_algo = args.update
grad_clipping = args.grad_clipping
peepholes = args.peepholes
num_filters = args.num_filters
gamma = args.gamma
output_predict = args.output_prediction
dropout = args.dropout
X_train, Y_train, mask_train, X_dev, Y_dev, mask_dev, X_test, Y_test, mask_test, \
embedd_table, label_alphabet, \
C_train, C_dev, C_test, char_embedd_table = data_processor.load_dataset_sequence_labeling(train_path, dev_path,
test_path, oov=oov,
fine_tune=fine_tune,
embedding=embedding,
embedding_path=embedding_path,
use_character=True)
num_labels = label_alphabet.size() - 1
logger.info("constructing network...")
# create variables
target_var = T.imatrix(name='targets')
mask_var = T.matrix(name='masks', dtype=theano.config.floatX)
if fine_tune:
input_var = T.imatrix(name='inputs')
num_data, max_length = X_train.shape
alphabet_size, embedd_dim = embedd_table.shape
else:
input_var = T.tensor3(name='inputs', dtype=theano.config.floatX)
num_data, max_length, embedd_dim = X_train.shape
char_input_var = T.itensor3(name='char-inputs')
num_data_char, max_sent_length, max_char_length = C_train.shape
char_alphabet_size, char_embedd_dim = char_embedd_table.shape
assert (max_length == max_sent_length)
assert (num_data == num_data_char)
# construct input and mask layers
layer_incoming1 = construct_char_input_layer()
layer_incoming2 = construct_input_layer()
layer_mask = lasagne.layers.InputLayer(shape=(None, max_length), input_var=mask_var, name='mask')
# construct bi-rnn-cnn
num_units = args.num_units
bi_lstm_cnn = build_BiLSTM_HighCNN(layer_incoming1, layer_incoming2, num_units, mask=layer_mask,
grad_clipping=grad_clipping, peepholes=peepholes, num_filters=num_filters,
dropout=dropout)
# reshape bi-rnn-cnn to [batch * max_length, num_units]
bi_lstm_cnn = lasagne.layers.reshape(bi_lstm_cnn, (-1, [2]))
# construct output layer (dense layer with softmax)
layer_output = lasagne.layers.DenseLayer(bi_lstm_cnn, num_units=num_labels, nonlinearity=nonlinearities.softmax,
name='softmax')
# get output of bi-lstm-cnn shape=[batch * max_length, #label]
prediction_train = lasagne.layers.get_output(layer_output)
prediction_eval = lasagne.layers.get_output(layer_output, deterministic=True)
final_prediction = T.argmax(prediction_eval, axis=1)
# flat target_var to vector
target_var_flatten = target_var.flatten()
# flat mask_var to vector
mask_var_flatten = mask_var.flatten()
# compute loss
num_loss = mask_var_flatten.sum(dtype=theano.config.floatX)
# for training, we use mean of loss over number of labels
loss_train = lasagne.objectives.categorical_crossentropy(prediction_train, target_var_flatten)
loss_train = (loss_train * mask_var_flatten).sum(dtype=theano.config.floatX) / num_loss
# l2 regularization?
if regular == 'l2':
l2_penalty = lasagne.regularization.regularize_network_params(layer_output, lasagne.regularization.l2)
loss_train = loss_train + gamma * l2_penalty
loss_eval = lasagne.objectives.categorical_crossentropy(prediction_eval, target_var_flatten)
loss_eval = (loss_eval * mask_var_flatten).sum(dtype=theano.config.floatX) / num_loss
# compute number of correct labels
corr_train = lasagne.objectives.categorical_accuracy(prediction_train, target_var_flatten)
corr_train = (corr_train * mask_var_flatten).sum(dtype=theano.config.floatX)
corr_eval = lasagne.objectives.categorical_accuracy(prediction_eval, target_var_flatten)
corr_eval = (corr_eval * mask_var_flatten).sum(dtype=theano.config.floatX)
# Create update expressions for training.
# hyper parameters to tune: learning rate, momentum, regularization.
batch_size = args.batch_size
learning_rate = 1.0 if update_algo == 'adadelta' else args.learning_rate
decay_rate = args.decay_rate
momentum = 0.9
params = lasagne.layers.get_all_params(layer_output, trainable=True)
updates = utils.create_updates(loss_train, params, update_algo, learning_rate, momentum=momentum)
# Compile a function performing a training step on a mini-batch
train_fn = theano.function([input_var, target_var, mask_var, char_input_var], [loss_train, corr_train, num_loss],
updates=updates)
# Compile a second function evaluating the loss and accuracy of network
eval_fn = theano.function([input_var, target_var, mask_var, char_input_var],
[loss_eval, corr_eval, num_loss, final_prediction])
# Finally, launch the training loop.
logger.info(
"Start training: %s with regularization: %s(%f), dropout: %s, fine tune: %s (#training data: %d, batch size: %d, clip: %.1f, peepholes: %s)..." \
% (
update_algo, regular, (0.0 if regular == 'none' else gamma), dropout, fine_tune, num_data, batch_size, grad_clipping,
peepholes))
num_batches = num_data / batch_size
num_epochs = 1000
best_loss = 1e+12
best_acc = 0.0
best_epoch_loss = 0
best_epoch_acc = 0
best_loss_test_err = 0.
best_loss_test_corr = 0.
best_acc_test_err = 0.
best_acc_test_corr = 0.
stop_count = 0
lr = learning_rate
patience = args.patience
for epoch in range(1, num_epochs + 1):
print 'Epoch %d (learning rate=%.4f, decay rate=%.4f): ' % (epoch, lr, decay_rate)
train_err = 0.0
train_corr = 0.0
train_total = 0
start_time = time.time()
num_back = 0
train_batches = 0
for batch in utils.iterate_minibatches(X_train, Y_train, masks=mask_train, char_inputs=C_train,
batch_size=batch_size, shuffle=True):
inputs, targets, masks, char_inputs = batch
err, corr, num = train_fn(inputs, targets, masks, char_inputs)
train_err += err * num
train_corr += corr
train_total += num
train_batches += 1
time_ave = (time.time() - start_time) / train_batches
time_left = (num_batches - train_batches) * time_ave
# update log
sys.stdout.write("\b" * num_back)
log_info = 'train: %d/%d loss: %.4f, acc: %.2f%%, time left (estimated): %.2fs' % (
min(train_batches * batch_size, num_data), num_data,
train_err / train_total, train_corr * 100 / train_total, time_left)
sys.stdout.write(log_info)
num_back = len(log_info)
# update training log after each epoch
sys.stdout.write("\b" * num_back)
print 'train: %d/%d loss: %.4f, acc: %.2f%%, time: %.2fs' % (
min(train_batches * batch_size, num_data), num_data,
train_err / train_total, train_corr * 100 / train_total, time.time() - start_time)
# evaluate performance on dev data
dev_err = 0.0
dev_corr = 0.0
dev_total = 0
for batch in utils.iterate_minibatches(X_dev, Y_dev, masks=mask_dev, char_inputs=C_dev, batch_size=batch_size):
inputs, targets, masks, char_inputs = batch
err, corr, num, predictions = eval_fn(inputs, targets, masks, char_inputs)
dev_err += err * num
dev_corr += corr
dev_total += num
if output_predict:
utils.output_predictions(predictions, targets, masks, 'tmp/dev%d' % epoch, label_alphabet)
print 'dev loss: %.4f, corr: %d, total: %d, acc: %.2f%%' % (
dev_err / dev_total, dev_corr, dev_total, dev_corr * 100 / dev_total)
if best_loss < dev_err and best_acc > dev_corr / dev_total:
stop_count += 1
else:
update_loss = False
update_acc = False
stop_count = 0
if best_loss > dev_err:
update_loss = True
best_loss = dev_err
best_epoch_loss = epoch
if best_acc < dev_corr / dev_total:
update_acc = True
best_acc = dev_corr / dev_total
best_epoch_acc = epoch
# evaluate on test data when better performance detected
test_err = 0.0
test_corr = 0.0
test_total = 0
for batch in utils.iterate_minibatches(X_test, Y_test, masks=mask_test, char_inputs=C_test,
batch_size=batch_size):
inputs, targets, masks, char_inputs = batch
err, corr, num, predictions = eval_fn(inputs, targets, masks, char_inputs)
test_err += err * num
test_corr += corr
test_total += num
if output_predict:
utils.output_predictions(predictions, targets, masks, 'tmp/test%d' % epoch, label_alphabet)
print 'test loss: %.4f, corr: %d, total: %d, acc: %.2f%%' % (
test_err / test_total, test_corr, test_total, test_corr * 100 / test_total)
if update_loss:
best_loss_test_err = test_err
best_loss_test_corr = test_corr
if update_acc:
best_acc_test_err = test_err
best_acc_test_corr = test_corr
# stop if dev acc decrease 3 time straightly.
if stop_count == patience:
break
# re-compile a function with new learning rate for training
if update_algo != 'adadelta':
lr = learning_rate / (1.0 + epoch * decay_rate)
updates = utils.create_updates(loss_train, params, update_algo, lr, momentum=momentum)
train_fn = theano.function([input_var, target_var, mask_var, char_input_var],
[loss_train, corr_train, num_loss],
updates=updates)
# print best performance on test data.
logger.info("final best loss test performance (at epoch %d)" % best_epoch_loss)
print 'test loss: %.4f, corr: %d, total: %d, acc: %.2f%%' % (
best_loss_test_err / test_total, best_loss_test_corr, test_total, best_loss_test_corr * 100 / test_total)
logger.info("final best acc test performance (at epoch %d)" % best_epoch_acc)
print 'test loss: %.4f, corr: %d, total: %d, acc: %.2f%%' % (
best_acc_test_err / test_total, best_acc_test_corr, test_total, best_acc_test_corr * 100 / test_total)
return theano.shared(floatX(np.random.randn(*shape) * 0.01))
def model(X, w):
return T.nnet.softmax(T.dot(X, w))
trX, teX, trY, teY = mnist(onehot=True)
X = T.fmatrix()
Y = T.fmatrix()
w = init_weights((784, 10))
py_x = model(X, w)
y_pred = T.argmax(py_x, axis=1)
cost = T.mean(T.nnet.categorical_crossentropy(py_x, Y))
gradient = T.grad(cost=cost, wrt=w)
update = [[w, w - gradient * 0.05]]
train = theano.function(inputs=[X, Y],
outputs=cost,
updates=update,
allow_input_downcast=True)
predict = theano.function(inputs=[X],
outputs=y_pred,
allow_input_downcast=True)
for i in range(100):
for start, end in zip(range(0, len(trX), 128), range(128, len(trX), 128)):
w = current_layer[0]
updates.append((w, 0.75 * w))
current_layer = parmas[2]
w = current_layer[0]
updates.append((w, 0.75 * w))
current_layer = parmas[3]
w = current_layer[0]
updates.append((w, 0.75 * w))
return updates
z = feedForward(x, params)
y = T.argmax(z, axis=1)
updates = rm_dropout(params)
# compile theano functions
remove_it = theano.function([], [], updates=updates)
predict = theano.function([x], y)
batch_size = 200
# test
remove_it()
labels = np.argmax(t_test, axis=1)
running_accuracy = 0.0
batches = 0
for start in range(0, 10000, batch_size):
x_batch = x_test[start:start + batch_size]
t_batch = labels[start:start + batch_size]
running_accuracy += np.mean(predict(x_batch) == t_batch)
all_params = nn.layers.get_all_params(l6)
param_count = sum([np.prod(p.get_value().shape) for p in all_params])
print "parameter count: %d" % param_count
def clipped_crossentropy(x, t, m=0.001):
x = T.clip(x, m, 1 - m)
return T.mean(T.nnet.binary_crossentropy(x, t))
obj = nn.objectives.Objective(l6, loss_function=clipped_crossentropy) # loss_function=nn.objectives.crossentropy)
loss_train = obj.get_loss()
loss_eval = obj.get_loss(deterministic=True)
updates_train = OrderedDict(nn.updates.nesterov_momentum(loss_train, all_params, LEARNING_RATE, MOMENTUM, WEIGHT_DECAY))
# updates_train[l6.W] += SOFTMAX_LAMBDA * T.mean(T.sqr(l6.W)) # L2 loss on the softmax weights to avoid saturation
y_pred_train = T.argmax(l6.get_output(), axis=1)
y_pred_eval = T.argmax(l6.get_output(deterministic=True), axis=1)
## compile
X_train = nn.utils.shared_empty(dim=3)
y_train = nn.utils.shared_empty(dim=1)
X_eval = theano.shared(chunk_eval)
y_eval = theano.shared(chunk_eval_labels)
index = T.lscalar("index")
acc_train = T.mean(T.eq(y_pred_train, y_train[index * MB_SIZE:(index + 1) * MB_SIZE]))
def build_model(self):
"""
build the computational graph of ASTN
:return:
"""
self.x = T.imatrix('wids')
self.xt = T.imatrix('wids_target')
self.y = T.ivector('label')
self.pw = T.fmatrix("position_weight")
self.is_train = T.iscalar("is_training")
input = self.Words[T.cast(self.x.flatten(), 'int32')].reshape(
(self.bs, self.sent_len, self.n_in))
input_target = self.Words[T.cast(self.xt.flatten(), 'int32')].reshape(
(self.bs, self.target_len, self.n_in))
input = T.switch(T.eq(self.is_train, np.int32(1)),
self.Dropout_ctx(input),
input * (1 - self.dropout_rate))
input_target = T.switch(T.eq(self.is_train, np.int32(1)),
self.Dropout_tgt(input_target),
input_target * (1 - self.dropout_rate))
# model component for TNet
rnn_input = input
rnn_input_reverse = reverse_tensor(tensor=rnn_input)
rnn_input_target = input_target
rnn_input_target_reverse = reverse_tensor(tensor=rnn_input_target)
H0_forward = self.LSTM_ctx(x=rnn_input)
Ht_forward = self.LSTM_tgt(x=rnn_input_target)
H0_backward = reverse_tensor(tensor=self.LSTM_ctx(x=rnn_input_reverse))
Ht_backward = reverse_tensor(tensor=self.LSTM_tgt(
x=rnn_input_target_reverse))
H0 = T.concatenate([H0_forward, H0_backward], axis=2)
Ht = T.concatenate([Ht_forward, Ht_backward], axis=2)
H1 = self.CPT(H0, Ht)
if self.pw is not None:
H1 = H1 * self.pw.dimshuffle(0, 1, 'x')
H2 = self.CPT(H1, Ht)
if self.pw is not None:
H2 = H2 * self.pw.dimshuffle(0, 1, 'x')
"""
H3 = self.CPT(H2, Ht)
if self.pw is not None:
H3 = H3 * self.pw.dimshuffle(0, 1, 'x')
H4 = self.CPT(H3, Ht)
if self.pw is not None:
H4 = H4 * self.pw.dimshuffle(0, 1, 'x')
H5 = self.CPT(H4, Ht)
if self.pw is not None:
H5 = H5 * self.pw.dimshuffle(0, 1, 'x')
"""
feat_and_feat_maps = [conv(H2) for conv in self.Conv_layers]
feat = [ele[0] for ele in feat_and_feat_maps]
self.feature_maps = T.concatenate(
[ele[1] for ele in feat_and_feat_maps], axis=2)
feat = T.concatenate(feat, axis=1)
# we do not use the self-implemented Dropout class
feat_dropout = T.switch(T.eq(self.is_train, np.int32(1)),
self.Dropout(feat),
feat * (1 - self.dropout_rate))
# shape: (bs, n_y)
self.p_y_x = T.nnet.softmax(self.FC(feat_dropout))
# self.p_y_x = self.FC(feat_dropout)
self.loss = T.nnet.categorical_crossentropy(coding_dist=self.p_y_x,
true_dist=self.y).mean()
self.pred_y = T.argmax(self.p_y_x, axis=1)
def main(model='mlp', batch_size=500, num_epochs=10):
# Load the dataset
print("Loading data...")
X_train, y_train, X_val, y_val, X_test, y_test = load_dataset()
# Fork worker processes and initilize GPU before building variables.
synk.fork()
# Prepare Theano variables for inputs and targets
input_var = T.tensor4('inputs')
target_var = T.ivector('targets')
network = build_network(model, input_var)
# Create a loss expression for training, i.e., a scalar objective we want
# to minimize (for our multi-class problem, it is the cross-entropy loss):
prediction = lasagne.layers.get_output(network)
loss = lasagne.objectives.categorical_crossentropy(prediction, target_var)
loss = loss.mean()
# We could add some weight decay as well here, see lasagne.regularization.
# Create update expressions for training, i.e., how to modify the
# parameters at each training step. Here, we'll use Stochastic Gradient
# Descent (SGD) with Nesterov momentum, but Lasagne offers plenty more.
params = lasagne.layers.get_all_params(network, trainable=True)
grad_updates, param_updates, grad_shared = updates.nesterov_momentum(
loss, params, learning_rate=0.01, momentum=0.9)
# updates = lasagne.updates.nesterov_momentum(
# loss, params, learning_rate=0.01, momentum=0.9)
# Create a loss expression for validation/testing. The crucial difference
# here is that we do a deterministic forward pass through the network,
# disabling dropout layers.
test_prediction = lasagne.layers.get_output(network, deterministic=True)
test_loss = lasagne.objectives.categorical_crossentropy(
test_prediction, target_var)
test_loss = test_loss.mean()
# As a bonus, also create an expression for the classification accuracy:
test_acc = T.mean(T.eq(T.argmax(test_prediction, axis=1), target_var),
dtype=theano.config.floatX)
# Compile a function performing a training step on a mini-batch (by giving
# the updates dictionary) and returning the corresponding training loss:
train_grad_fn = synk.function([input_var, target_var],
outputs=loss,
updates=grad_updates)
train_update_fn = synk.function([], updates=param_updates)
# train_fn = theano.function([input_var, target_var], loss, updates=updates)
# Compile a second function computing the validation loss and accuracy:
val_fn = synk.function([input_var, target_var],
outputs=[test_loss, test_acc])
# val_fn = theano.function([input_var, target_var], [test_loss, test_acc])
# After building all functions, give them to workers.
synk.distribute()
# Put data into OS shared memory for worker access.
X_train, y_train = val_fn.build_inputs(X_train, y_train)
X_val, y_val = val_fn.build_inputs(X_val, y_val)
X_test, y_test = val_fn.build_inputs(X_test, y_test)
# Finally, launch the training loop.
print("Starting training...")
# We iterate over epochs:
for epoch in range(num_epochs):
# In each epoch, we do a full pass over the training data:
train_err = 0
train_batches = 0
start_time = time.time()
# for batch in iterate_minibatches(X_train, y_train, batch_size, shuffle=True):
for batch in iterate_minibatch_indices(len(y_train),
batch_size,
shuffle=True):
train_err += train_grad_fn(X_train, y_train, batch=batch)
synk.all_reduce(grad_shared) # (averges)
train_update_fn()
train_batches += 1
# And a full pass over the validation data:
# val_err = 0
# val_acc = 0
# val_batches = 0
# for batch in iterate_minibatches(X_val, y_val, batch_size, shuffle=False):
# inputs, targets = batch
# err, acc = val_fn(inputs, targets)
# val_err += err
# val_acc += acc
# val_batches += 1
val_err, val_acc = val_fn(X_val, y_val, num_slices=4)
# Then we print the results for this epoch:
print("Epoch {} of {} took {:.3f}s".format(epoch + 1, num_epochs,
time.time() - start_time))
print(" training loss:\t\t{:.6f}".format(train_err / train_batches))
print(" validation loss:\t\t{:.6f}".format(float(val_err)))
print(" validation accuracy:\t\t{:.2f} %".format(
float(val_acc) * 100))
# After training, we compute and print the test error:
# test_err = 0
# test_acc = 0
# test_batches = 0
# for batch in iterate_minibatches(X_test, y_test, batch_size, shuffle=False):
# inputs, targets = batch
# err, acc = val_fn(inputs, targets)
# test_err += err
# test_acc += acc
# test_batches += 1
test_err, test_acc = val_fn(X_test, y_test, num_slices=4)
print("Final results:")
print(" test loss:\t\t\t{:.6f}".format(float(test_err)))
print(" test accuracy:\t\t{:.2f} %".format(float(test_acc) * 100))
def get_output_for(self, deterministic=False):
if deterministic:
deterministic_flag = T.constant(1)
else:
deterministic_flag = T.constant(0)
batch_size = self.pred.shape[0]
time_steps = self.pred.shape[1]
label_num = input,
## the start state to first label
pred_t1 = self.pred[:, 0] # shape: (batch size, label num)
gs_t1 = self.gs[:, 0] - 1
mask_t1 = self.masks[:, 0]
score_t0 = T.zeros((batch_size, label_num))
index_t0 = T.zeros((batch_size, label_num), dtype='int64')
init_flag = T.constant(1)
# return shape: (batch size, label num), (batch size, label num)
score_t1, index_t1 = self.score_one_step(pred_t1, gs_t1,
mask_t1, score_t0, index_t0, self.init_t, self.tran_t, deterministic_flag, init_flag)
print 'score_t1', score_t1.eval()
print 'index_t1', index_t1.eval()
pred = self.pred.dimshuffle(1, 0, 2)
gs = self.gs.dimshuffle(1, 0)
mask = self.masks.dimshuffle(1, 0)
init_flag = T.constant(0)
# print pred[1:].eval().shape
# print (gs[1:]-1).eval().shape
# print mask[1:].eval().shape
# return shape: (time steps - 1, batch size, label num) ..., (time steps - 1, batch size)
step_scores, step_indexs = theano.scan(fn=self.score_one_step,
outputs_info=[score_t1, index_t1],
sequences=[pred[1:], gs[1:]-1, mask[1:]],
non_sequences=[self.init_t, self.tran_t, deterministic_flag, init_flag])[0]
# # print step_scores.eval().shape
# # print step_indexs.eval().shape
print 'score_t2', step_scores.dimshuffle(1, 0, 2)[:, 0].eval()
print 'index_t2', step_indexs.dimshuffle(1, 0, 2)[:, 0].eval()
print 'score_t3', step_scores.dimshuffle(1, 0, 2)[:, 1].eval()
print 'index_t3', step_indexs.dimshuffle(1, 0, 2)[:, 1].eval()
# shape: (batch size, )
last_step_max_score = T.max(step_scores[-1], axis=-1)
last_step_max_index = T.argmax(step_scores[-1], axis=-1)
def track_one_step(index_t, max_index_t):
# example_indexs shape: (batch size, label num)
# step_max_index shape: (batch size, )
def scan_example(index_t_e, max_index_t_e):
max_index_tm1_e = index_t_e[max_index_t_e]
return max_index_tm1_e
# return shape: (batch size, )
max_index_tm1 = theano.scan(fn=scan_example,
sequences=[index_t, max_index_t])[0]
return max_index_tm1
# reverse time step, shape: (time steps - 1, batch size, label num)
#step_indexs = step_indexs[::-1]
# return shape: (time steps - 1, batch size)
index_chain = theano.scan(fn=track_one_step,
sequences=step_indexs,
outputs_info=last_step_max_index,
go_backwards=True)[0]
# return shape: (batch size, time steps - 1)
index_chain = index_chain.dimshuffle(1, 0)
# shape: (batch size, time steps)
index_chain_reverse = self.aggregateTensor(last_step_max_index, index_chain)
# add 1 for label index (which index from 1)
# return shape: (batch size, time steps)
index_chain = (index_chain_reverse + T.ones_like(index_chain_reverse))[:, ::-1]
print 'index chain', index_chain.eval()
def one_step_cost(step_index, pred_t, gs_t, index_chain_t, mask_t, cost_tm1, gs_tm1, index_chain_tm1, init_tran, tran):
# step_index: (1,)
# pred_t: (batch size, label num)
# gs_t_e: (batch size, )
# index_chain_t: (batch size, )
# mask_t: (batch size, )
# cost_tm1: (batch size, )
# gs_tm1: (batch size, )
# index_chain_tm1: (batch size, )
def scan_example(pred_t_e, gs_t_e, index_chain_t_e, mask_t_e, cost_tm1_e, gs_tm1_e, index_chain_tm1_e, step_index, init_tran, tran):
# pred_t_e: (label num, )
# gs_t_e: (1, )
# index_chain_t_e: (1, )
# mask_t_e: (1, )
# gs_tm1_e: (1, )
# index_chain_tm1_e: (1, )
# init_tran: (label num, )
# tran: (label num, label num)
cost_t_e = None
cost_t_e = theano.ifelse.ifelse(T.eq(step_index, 0),
theano.printing.Print('\ninit step pred_t_e\n')(pred_t_e[theano.printing.Print('\ninit step index_chain_t_e\n')(index_chain_t_e)]) + theano.printing.Print('\n initstep init_tran\n')(init_tran[index_chain_t_e]) - theano.printing.Print('\ninit step pred_t_e\n')(pred_t_e[theano.printing.Print('\ninit step gs_t_e\n')(gs_t_e)]) - theano.printing.Print('\ninit step init_tran\n')(init_tran[gs_t_e]),
theano.printing.Print('\nother pred_t_e\n')(pred_t_e[theano.printing.Print('\nother index_chain_t_e\n')(index_chain_t_e)]) + theano.printing.Print('\nother tran\n')(tran[theano.printing.Print('\nother index_chain_tm1_e\n')(index_chain_tm1_e)][index_chain_t_e]) - theano.printing.Print('\nother pred_t_e\n')(pred_t_e[theano.printing.Print('\nother gs_t_e\n')(gs_t_e)]) - theano.printing.Print('\nother tran\n')(tran[theano.printing.Print('\nother gs_tm1_e\n')(gs_tm1_e)][gs_t_e]))
# if T.eq(step_index, 0) == T.constant(1):
# cost_t_e = pred_t_e[index_chain_t_e] + init_tran[index_chain_t_e]\
# - pred_t_e[gs_t_e] - init_tran[gs_t_e]
# else:
# cost_t_e = pred_t_e[index_chain_t_e] + tran[index_chain_t_e][index_chain_tm1_e]\
# - pred_t_e[gs_t_e] - tran[gs_tm1_e][gs_t_e]
cost_t_e = cost_t_e * mask_t_e
# return shape: (1, )
return theano.printing.Print('\ncost_t_e\n')(cost_t_e), gs_t_e, index_chain_t_e
# return shape: (batch size, )...
cost_t, _, _ = theano.scan(fn=scan_example,
sequences=[pred_t, gs_t, index_chain_t, mask_t, cost_tm1, gs_tm1, index_chain_tm1],
non_sequences=[step_index, init_tran, tran])[0]
# return shape: (batch size, )...
return cost_t, gs_t, index_chain_t
# return shape: (time steps, batch size)
index_chain_sff = index_chain.dimshuffle(1, 0)
gs_t0 = T.zeros((batch_size, ), dtype='int64')
cost_t0 = T.zeros((batch_size, ), dtype='float64')
index_chain_t0 = T.zeros((batch_size, ), dtype='int64')
# return shape: (time steps, batch size)
print (gs-1).eval()
print (index_chain_sff-1).eval()
steps_cost, _, _ = theano.scan(fn=one_step_cost,
outputs_info=[cost_t0, gs_t0, index_chain_t0],
sequences=[T.arange(time_steps), pred, gs-1, index_chain_sff-1, mask],
non_sequences=[self.init_t, self.tran_t])[0]
# return shape: (batch size, )
cost = T.sum(steps_cost.dimshuffle(1, 0), axis=-1)
# # return shape: (batch size, time steps - 1)
# step_gs_scores = step_gs_scores.dimshuffle(1, 0)
# # return shape: (batch size, )
# last_gs_score = step_gs_scores[:, -1]
# print 'score_t2', step_scores.dimshuffle(1, 0, 2)[:, 0].eval()
# print 'index_t2', step_indexs.dimshuffle(1, 0, 2)[:, 0].eval()
# print 'gs_score_t2', step_gs_scores[:, 0].eval()
# print 'score_t3', step_scores.dimshuffle(1, 0, 2)[:, 1].eval()
# print 'index_t3', step_indexs.dimshuffle(1, 0, 2)[:, 1].eval()
# print 'gs_score_t3', step_gs_scores[:, 1].eval()
# print index_chain.eval()
# print last_step_max_score.eval()
# print last_gs_score.eval()
# return shape: (exmaple num, time steps), (batch size, ), (batch size, )
#return [index_chain, last_step_max_score, last_gs_score]
print 'cost', cost.eval()
# return shape: (batch size, )
return cost
def train(self,
train_sets,
valid_sets,
test_sets,
n_epochs=200,
learning_rate=0.1):
train_set_x, train_set_y = train_sets
valid_set_x, valid_set_y = valid_sets
test_set_x, test_set_y = test_sets
n_train_batches = train_set_x.get_value(borrow=True).shape[0]
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0]
n_test_batches = test_set_x.get_value(borrow=True).shape[0]
n_train_batches //= self.batch_size
n_valid_batches //= self.batch_size
n_test_batches //= self.batch_size
cost = -T.mean(
T.log(self.final_output[T.arange(self.y.shape[0]), self.y]))
error = T.mean(T.neq(T.argmax(self.final_output, axis=1), self.y))
# find all the parameters and update them using gradient descent
params = self.params
grads = T.grad(cost, params)
updates = [(param_i, param_i - learning_rate * grad_i)
for param_i, grad_i in zip(params, grads)]
index = self.index
batch_size = self.batch_size
x = self.x
y = self.y
test_model = theano.function(
[index],
error,
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]
})
validate_model = theano.function(
[index],
error,
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]
})
train_model = theano.function(
[index],
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]
})
print('... training')
# early-stopping parameters
patience = 10000 # look as this many examples regardless
patience_increase = 2 # wait this much longer when a new best is found
improvement_threshold = 0.995
validation_frequency = min(n_train_batches, patience // 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
best_validation_loss = numpy.inf
best_iter = 0
test_score = 0.
start_time = timeit.default_timer()
epoch = 0
done_looping = False
while (epoch < n_epochs):
epoch = epoch + 1
for minibatch_index in range(n_train_batches):
iter = (epoch - 1) * n_train_batches + minibatch_index
if iter % 100 == 0:
print('training @ iter = ', iter, flush=True)
cost_ij = train_model(minibatch_index)
if (iter + 1) % validation_frequency == 0:
# compute zero-one loss on validation set
validation_losses = [
validate_model(i) for i in range(n_valid_batches)
]
this_validation_loss = numpy.mean(validation_losses)
print('epoch {}, minibatch {}/{}, validation error {}%'.
format(epoch, minibatch_index + 1, n_train_batches,
this_validation_loss * 100.))
with open('model_{}.mod'.format(iter), 'wb') as f:
pickle.dump(self.dump(), f)
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
if this_validation_loss < best_validation_loss * \
improvement_threshold:
patience = max(patience, iter * patience_increase)
# save best validation score and iteration number
best_validation_loss = this_validation_loss
best_iter = iter
test_losses = [
test_model(i) for i in range(n_test_batches)
]
test_score = numpy.mean(test_losses)
print((' epoch {}, minibatch {}/{}, test error of '
'best model {}%').format(epoch, minibatch_index + 1,
n_train_batches,
test_score * 100.))
with open('test_{}.res'.format(iter), 'w') as f:
print(network.predict(test_set_x), file=f)
end_time = timeit.default_timer()
print('Optimization complete.')
print('Best validation score of %f %% obtained at iteration %i, '
'with test performance %f %%' %
(best_validation_loss * 100., best_iter + 1, test_score * 100.))
print(('The code for file ' + os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.)),
file=sys.stderr)
# Define loss function and metrics, and get an updates dictionary
X_sym = T.tensor4()
y_sym = T.ivector()
# We'll connect our output classifier to the last fully connected layer of the network
net['new_output'] = DenseLayer(net['pool5'],
num_units=8,
nonlinearity=softmax,
W=lasagne.init.Normal(0.01))
prediction = lasagne.layers.get_output(net['new_output'], X_sym)
loss = lasagne.objectives.categorical_crossentropy(prediction, y_sym)
loss = loss.mean()
acc = T.mean(T.eq(T.argmax(prediction, axis=1), y_sym),
dtype=theano.config.floatX)
learning_rate = theano.shared(np.array(0.001, dtype=theano.config.floatX))
learning_rate_decay = np.array(0.3, dtype=theano.config.floatX)
updates = OrderedDict()
print("Setting learning rates...")
for name, layer in net.items():
print(name)
layer_params = layer.get_params(trainable=True)
if name in ['new_output', 'fc1000']:
layer_lr = learning_rate
else:
layer_lr = learning_rate / 10
if name != 'fc1000':
def fit(self, X_task, y):
DEBUG_FLAG = True
# self.max_epochs = 333
self.batch_size = 100
rng = np.random.RandomState(42)
self.input_taskdata = T.matrix(dtype='float32', name='input_taskdata')
self.input_restdata = T.matrix(dtype='float32', name='input_restdata')
self.params_from_last_iters = []
n_input = X_task.shape[1]
index = T.iscalar(name='index')
# prepare data for theano computation
if not DEBUG_FLAG:
X_train_s = theano.shared(value=np.float32(X_task),
name='X_train_s')
y_train_s = theano.shared(value=np.int32(y), name='y_train_s')
lr_train_samples = len(X_task)
else:
from sklearn.cross_validation import StratifiedShuffleSplit
folder = StratifiedShuffleSplit(y, n_iter=1, test_size=0.20)
new_trains, inds_val = iter(folder).next()
X_train, X_val = X_task[new_trains], X_task[inds_val]
y_train, y_val = y[new_trains], y[inds_val]
X_train_s = theano.shared(value=np.float32(X_train),
name='X_train_s',
borrow=False)
y_train_s = theano.shared(value=np.int32(y_train),
name='y_train_s',
borrow=False)
# X_val_s = theano.shared(value=np.float32(X_val),
# name='X_train_s', borrow=False)
# y_val_s = theano.shared(value=np.int32(y_val),
# name='y_cal_s', borrow=False)
lr_train_samples = len(X_train)
self.dbg_epochs_ = list()
self.dbg_acc_train_ = list()
self.dbg_acc_val_ = list()
self.dbg_ae_cost_ = list()
self.dbg_lr_cost_ = list()
self.dbg_ae_nonimprovesteps = list()
self.dbg_acc_other_ds_ = list()
self.dbg_prfs_ = list()
self.dbg_prfs_other_ds_ = list()
# computation graph: logistic regression
clf_n_output = 18 # number of labels
my_y = T.ivector(name='y')
bV0_vals = np.zeros(clf_n_output).astype(np.float32)
self.bV0 = theano.shared(value=bV0_vals, name='bV0')
V0_vals = rng.randn(n_input, clf_n_output).astype(
np.float32) * self.gain1
self.V0s = theano.shared(V0_vals)
self.p_y_given_x = T.nnet.softmax(
T.dot(self.input_taskdata, self.V0s) + self.bV0)
self.lr_cost = -T.mean(
T.log(self.p_y_given_x)[T.arange(my_y.shape[0]), my_y])
self.lr_cost = (self.lr_cost +
T.mean(abs(self.V0s)) * self.penalty_l1 +
T.mean(abs(self.bV0)) * self.penalty_l1 + T.mean(
(self.V0s**np.float32(2))) * self.penalty_l2 +
T.mean((self.bV0**np.float32(2))) * self.penalty_l2)
self.y_pred = T.argmax(self.p_y_given_x, axis=1)
givens_lr = {
self.input_taskdata:
X_train_s[index * self.batch_size:(index + 1) * self.batch_size],
my_y:
y_train_s[index * self.batch_size:(index + 1) * self.batch_size]
}
params = [self.V0s, self.bV0]
updates = self.RMSprop(cost=self.lr_cost,
params=params,
lr=self.learning_rate)
f_train_lr = theano.function([index], [self.lr_cost],
givens=givens_lr,
updates=updates)
# optimization loop
start_time = time.time()
lr_last_cost = np.inf
ae_cur_cost = np.inf
no_improve_steps = 0
acc_train, acc_val = 0., 0.
for i_epoch in range(self.max_epochs):
if i_epoch == 1:
epoch_dur = time.time() - start_time
total_mins = (epoch_dur * self.max_epochs) / 60
hs, mins = divmod(total_mins, 60)
print("Max estimated duration: %i hours and %i minutes" %
(hs, mins))
lr_n_batches = lr_train_samples // self.batch_size
for i in range(lr_n_batches):
lr_cur_cost = f_train_lr(i)[0]
# evaluate epoch cost
if lr_last_cost - lr_cur_cost < 0.1:
no_improve_steps += 1
else:
lr_last_cost = lr_cur_cost
no_improve_steps = 0
# logistic
lr_last_cost = lr_cur_cost
acc_train = self.score(X_train, y_train)
acc_val, prfs_val = self.score(X_val, y_val, return_prfs=True)
print(
'E:%i, ae_cost:%.4f, lr_cost:%.4f, train_score:%.2f, vald_score:%.2f, ae_badsteps:%i'
% (i_epoch + 1, ae_cur_cost, lr_cur_cost, acc_train, acc_val,
no_improve_steps))
if (i_epoch % 10 == 0):
self.dbg_ae_cost_.append(ae_cur_cost)
self.dbg_lr_cost_.append(lr_cur_cost)
self.dbg_epochs_.append(i_epoch + 1)
self.dbg_ae_nonimprovesteps.append(no_improve_steps)
self.dbg_acc_train_.append(acc_train)
self.dbg_acc_val_.append(acc_val)
self.dbg_prfs_.append(prfs_val)
# if i_epoch > (self.max_epochs - 100):
param_pool = self.get_param_pool()
self.params_from_last_iters.append(param_pool)
total_mins = (time.time() - start_time) / 60
hs, mins = divmod(total_mins, 60)
print("Final duration: %i hours and %i minutes" % (hs, mins))
return self
def main():
# step 1: get the data and define all the usual variables
X, Y = get_normalized_data()
max_iter = 20
print_period = 10
lr = 0.00004
reg = 0.01
X = X.astype(np.float32)
Y = Y.astype(np.float32)
Xtrain = X[:-1000,]
Ytrain = Y[:-1000]
Xtest = X[-1000:,]
Ytest = Y[-1000:]
Ytrain_ind = y2indicator(Ytrain).astype(np.float32)
Ytest_ind = y2indicator(Ytest).astype(np.float32)
N, D = Xtrain.shape
batch_sz = 500
n_batches = N / batch_sz
M = 300
K = 10
W1_init = np.random.randn(D, M) / 28
b1_init = np.zeros(M)
W2_init = np.random.randn(M, K) / np.sqrt(M)
b2_init = np.zeros(K)
# step 2: define theano variables and expressions
thX = T.matrix('X')
thT = T.matrix('T')
W1 = theano.shared(W1_init.astype(np.float32), 'W1')
b1 = theano.shared(b1_init.astype(np.float32), 'b1')
W2 = theano.shared(W2_init.astype(np.float32), 'W2')
b2 = theano.shared(b2_init.astype(np.float32), 'b2')
# we can use the built-in theano functions to do relu and softmax
thZ = relu( thX.dot(W1) + b1 ) # relu is new in version 0.7.1 but just in case you don't have it
thY = T.nnet.softmax( thZ.dot(W2) + b2 )
# define the cost function and prediction
cost = -(thT * T.log(thY)).sum() + reg*((W1*W1).sum() + (b1*b1).sum() + (W2*W2).sum() + (b2*b2).sum())
prediction = T.argmax(thY, axis=1)
# step 3: training expressions and functions
# we can just include regularization as part of the cost because it is also automatically differentiated!
# update_W1 = W1 - lr*(T.grad(cost, W1) + reg*W1)
# update_b1 = b1 - lr*(T.grad(cost, b1) + reg*b1)
# update_W2 = W2 - lr*(T.grad(cost, W2) + reg*W2)
# update_b2 = b2 - lr*(T.grad(cost, b2) + reg*b2)
update_W1 = W1 - lr*T.grad(cost, W1)
update_b1 = b1 - lr*T.grad(cost, b1)
update_W2 = W2 - lr*T.grad(cost, W2)
update_b2 = b2 - lr*T.grad(cost, b2)
train = theano.function(
inputs=[thX, thT],
updates=[(W1, update_W1), (b1, update_b1), (W2, update_W2), (b2, update_b2)],
)
# create another function for this because we want it over the whole dataset
get_prediction = theano.function(
inputs=[thX, thT],
outputs=[cost, prediction],
)
t0 = datetime.now()
for i in xrange(max_iter):
for j in xrange(n_batches):
Xbatch = Xtrain[j*batch_sz:(j*batch_sz + batch_sz),]
Ybatch = Ytrain_ind[j*batch_sz:(j*batch_sz + batch_sz),]
train(Xbatch, Ybatch)
if j % print_period == 0:
cost_val, prediction_val = get_prediction(Xtest, Ytest_ind)
err = error_rate(prediction_val, Ytest)
print "Cost / err at iteration i=%d, j=%d: %.3f / %.3f" % (i, j, cost_val, err)
print "Training time:", datetime.now() - t0
from models import ResNet_FullPreActivation, ResNet_BottleNeck_FullPreActivation
from utils import load_pickle_data_test
BATCHSIZE = 1
'''
Set up all theano functions
'''
X = T.tensor4('X')
Y = T.ivector('y')
# set up theano functions to generate output by feeding data through network, any test outputs should be deterministic
output_layer = ResNet_BottleNeck_FullPreActivation(X, n=18)
output_test = lasagne.layers.get_output(output_layer, deterministic=True)
output_class = T.argmax(output_test, axis=1)
# set up training and prediction functions
predict_proba = theano.function(inputs=[X], outputs=output_test)
predict_class = theano.function(inputs=[X], outputs=output_class)
'''
Load data and make predictions
'''
test_X, test_y = load_pickle_data_test()
# load network weights
f = gzip.open('data/weights/resnet164_fullpreactivation.pklz', 'rb')
all_params = pickle.load(f)
f.close()
helper.set_all_param_values(output_layer, all_params)
print ipt.ndim, 'dimensions'
except:
print 'no ndm'
print min_informative_str(ipt)
if found > 0:
print type(node.op), found
try:
print '\t', type(node.op.scalar_op)
except:
pass
print count
test = CIFAR10(which_set='test', one_hot=True, gcn=55.)
yl = T.argmax(yb, axis=1)
mf1acc = 1. - T.neq(yl, T.argmax(ymf1, axis=1)).mean()
#mfnacc = 1.-T.neq(yl , T.argmax(mfny,axis=1)).mean()
batch_acc = function([Xb, yb], [mf1acc])
def accs():
mf1_accs = []
for i in xrange(10000 / batch_size):
mf1_accs.append(
batch_acc(
test.get_topological_view(test.X[i * batch_size:(i + 1) *
batch_size, :]),
test.y[i * batch_size:(i + 1) * batch_size, :])[0])
def main(argv):
###-------------------------------- Get files to proccess ----------------------------------------------
# just read header to get layer dimensions
#files = glob.glob("B:\\NN_data\\THEANO_DATA\\tmp\\*.tmp")
context = 5
in_frame_num = context * 2 + 1
feature_per_frame = 41
output_vec_len = 214
files = glob.glob(argv[0] + "\\*.mfsc")
#inputs, in_frame_num, feature_per_frame, output_vec_len = read_data_only_inputs(files[0])
print("\feature_per_frame: {}".format(feature_per_frame))
print("\output_vec_len: {}".format(output_vec_len))
###-------------------------------- Read apriori state ppb ---------------------------------------------
apriori_ppb_input = read_state_apriori_ppb(argv[2], output_vec_len)
#print(apriori_ppb)
#exit(0)
###-------------------------------- BUILD Theano functions ----------------------------------------------
network = None
layer2reg = None
total_batch = 0
# Prepare Theano variables for inputs and targets
input_var = T.tensor4('inputs')
apriori_ppb = T.fmatrix('apriori')
target_var = T.ivector('targets')
lrate_var = theano.tensor.scalar('lrate', dtype='float32')
momentum_var = theano.tensor.scalar('momentum', dtype='float32')
#create network
#network, layer2reg = build_cnn(input_var, in_frame_num, feature_per_frame, output_vec_len)
with open(argv[1], "rb") as f2:
context = int(f2.readline())
in_frame_num = context * 2 + 1
filter_num = int(f2.readline())
neuron_num = int(f2.readline())
network, layer2reg = build_cnn(input_var, in_frame_num,
feature_per_frame, output_vec_len,
filter_num, neuron_num)
lasagne.layers.set_all_param_values(network, np.load(f2))
# Create a loss expression for training, i.e., a scalar objective we want
# to minimize (for our multi-class problem, it is the cross-entropy loss):
prediction = lasagne.layers.get_output(network)
loss = lasagne.objectives.categorical_crossentropy(prediction, target_var)
loss = loss.mean()
# SGB with momentm and changing learning rate
params = lasagne.layers.get_all_params(network, trainable=True)
updates = lasagne.updates.nesterov_momentum(loss,
params,
learning_rate=lrate_var,
momentum=momentum_var)
test_prediction = lasagne.layers.get_output(network, deterministic=True)
test_loss = lasagne.objectives.categorical_crossentropy(
test_prediction, target_var)
test_loss = test_loss.mean()
test_acc = T.mean(T.eq(T.argmax(test_prediction, axis=1), target_var),
dtype=theano.config.floatX)
# Compile a training function
train_fn = theano.function(
[input_var, target_var, lrate_var, momentum_var],
loss,
updates=updates)
# Compile a second function computing the validation loss and accuracy:
val_fn = theano.function([input_var, target_var], [test_loss, test_acc])
test_prediction_1 = lasagne.layers.get_output(network, deterministic=True)
test_prediction = T.dot(test_prediction_1, apriori_ppb)
test_result = T.argmax(test_prediction_1, axis=1)
# Compile a second function computing the validation loss and accuracy:
predict_fn = theano.function([input_var], [test_result])
predict_fn_j = theano.function([input_var, apriori_ppb], [test_prediction])
print("theano functions compiled")
###-------------------------------- Final lunch ----------------------------------------------
file_counter = 0
print("")
for one_file in files:
inputs_j = load_mfsc_data_with_context(one_file, context)
dot_index = one_file.find('.mfsc')
output_file_name = one_file[0:dot_index] + ".bin"
print(inputs_j[0][0])
sample_num = len(inputs_j)
base_file_name = output_file_name[output_file_name.rfind('\\') + 1:]
file_counter += 1
if file_counter % 5 == 0:
print(base_file_name)
else:
print(base_file_name + " ", end="")
with open(output_file_name, 'wb') as f:
f.write(struct.pack('I', swap32(len(inputs_j))))
f.write(struct.pack('I', swap32(100000)))
f.write(struct.pack('H', output_vec_len * 4)[::-1])
f.write(struct.pack('H', 9)[::-1])
begin = 0
batch = 8
while begin < sample_num:
batch_size = np.minimum(batch, sample_num - begin)
result = predict_fn_j(
inputs_j[begin:begin + batch_size].reshape(
batch_size, 3, in_frame_num, feature_per_frame),
apriori_ppb_input)
res = np.log(result)[0]
for aaa in res:
for ppb in aaa:
f.write(ppb)
#f.write(struct.pack('f',ppb)[::-1])
begin += batch_size
print("")
def generator_step_sm(x_tm1, h_tm1, m_tm1, s_tm1, tau, eps):
"""One step of the generative decoder version."""
# x_tm1 is `BxT` one-hot, h_tm1 is `batch x ...`
# m_tm1 is `batch`, tau, eps are scalars
# collect the inputs
inputs = {l_decoder_embed: x_tm1.dimshuffle(0, "x", 1),
l_decoder_mask: m_tm1.dimshuffle(0, "x")}
# Connect the prev variables to the the hidden and stack state feeds
j = 0
for layer in dec_rnn_layers:
inputs[layer.hid_init] = slice_(h_tm1, j, layer.num_units)
j += layer.num_units
j = 0
for layer in dec_rnn_layers:
layer = layer.input_layers[1]
dep, wid = layer.output_shape[-2:]
stack_slice_ = slice_(s_tm1, j, dep * wid)
inputs[layer] = stack_slice_.reshape((-1, dep, wid))
j += dep * wid
# Get the outputs
outputs = [l_decoder_reembedder]
for pair in zip(dec_rnn_layers_sliced, dec_rnn_layers_stack):
outputs.extend(pair)
# propagate through the decoder column
logit_t, *rest = lasagne.layers.get_output(outputs, inputs,
deterministic=True)
h_t_list, s_t_list = rest[::2], rest[1::2]
# Pack the hidden and flattened stack states
h_t = tt.concatenate(h_t_list, axis=-1)
s_t = tt.concatenate([v.flatten(ndim=2) for v in s_t_list], axis=-1)
# Generate the next symbol: logit_t is `Bx1xV`
logit_t = logit_t[:, 0]
prob_t = tt.nnet.softmax(logit_t)
# Gumbel-softmax sampling: Gumbel (e^{-e^{-x}}) distributed random noise
gumbel = -tt.log(-tt.log(theano_random_state.uniform(size=logit_t.shape) + eps) + eps)
# logit_t = theano.ifelse.ifelse(tt.gt(tau, 0), gumbel + logit_t, logit_t)
# inv_temp = theano.ifelse.ifelse(tt.gt(tau, 0), 1.0 / tau, tt.constant(1.0))
logit_t = tt.switch(tt.gt(tau, 0), gumbel + logit_t, logit_t)
inv_temp = tt.switch(tt.gt(tau, 0), 1.0 / tau, tt.constant(1.0))
# Get the softmax: x_t is `BxV`
x_t = tt.nnet.softmax(logit_t * inv_temp)
# Get the best symbol
c_t = tt.cast(tt.argmax(x_t, axis=-1), "int8")
# Get the estimated probability of the picked symbol.
p_t = prob_t[tt.arange(c_t.shape[0]), c_t]
# Compute the mask and inhibit the propagation on a stop symbol.
# Recurrent layers return the previous state if m_tm1 is Fasle
m_t = m_tm1 & tt.gt(c_t, vocab.index("\x03"))
c_t = tt.switch(m_t, c_t, vocab.index("\x03"))
# There is no need to freeze the states as they will be frozen by
# the RNN passthrough according to the mask `m_t`.
# Embed the current character.
x_t = tt.dot(x_t, l_embed_char.W)
return x_t, h_t, m_t, s_t, p_t, c_t
loss_remember = (I0*lasagne.objectives.squared_error(W0,W_n0)).mean()+ \
(I1*lasagne.objectives.squared_error(W1,W_n1)).mean()+ \
(I2*lasagne.objectives.squared_error(W2,W_n2)).mean()
loss = loss_class + 100 * loss_remember
# Get network params, with specifications of manually updated ones
params = lasagne.layers.get_all_params(network, trainable=True)
updates = lasagne.updates.sgd(loss, params, learning_rate=0.0001)
#updates = lasagne.updates.adam(loss,params,learning_rate=0.00001)
#updates = lasagne.updates.nesterov_momentum(loss,params,learning_rate=0.01)
test_prediction = lasagne.layers.get_output(network, deterministic=True)
test_loss = lasagne.objectives.categorical_crossentropy(
test_prediction, target_var)
test_loss = test_loss.mean()
test_acc = T.mean(T.eq(T.argmax(test_prediction, axis=1), target_var),
dtype=theano.config.floatX)
# Compile theano function computing the training validation loss and accuracy:
train_fn = theano.function([input_var, target_var, W0, W1, W2, I0, I1, I2],
[loss, loss_class, loss_remember],
updates=updates)
#train_fn = theano.function([input_var, target_var], loss, updates=updates)
val_fn = theano.function([input_var, target_var], [test_loss, test_acc])
#gradient_fn = theano.function([input_var, target_var], gradient)
# The training loop
print("Starting training...")
num_epochs = 250
for epoch in range(num_epochs):
def train(self, Xs, Ys, Xv, Yv, mdl,
data_folder='data/', out_folder='out/'):
data_folder = os.path.join(data_folder, 'imgs/', 'train/')
input_var = mdl.input_var
net = mdl.get_output_layer()
target_var = T.ivector('targets')
prediction = lasagne.layers.get_output(net)
loss = lasagne.objectives.categorical_crossentropy(prediction, target_var)
loss = loss.mean()
params = lasagne.layers.get_all_params(net, trainable=True)
grads = T.grad(loss, params)
test_prediction = lasagne.layers.get_output(net, deterministic=True)
test_loss = lasagne.objectives. \
categorical_crossentropy(test_prediction, target_var)
test_loss = test_loss.mean()
test_acc = T.mean(T.eq(T.argmax(test_prediction, axis=1), target_var),
dtype=theano.config.floatX)
logger.info("Compiling network functions...")
grads_fn = theano.function([input_var, target_var], grads)
train_fn = theano.function([input_var, target_var], loss)
val_fn = theano.function([input_var, target_var], [test_loss, test_acc])
predict_proba = theano.function([input_var], test_prediction)
logger.info("Training...")
logger.info('GPU Free Mem: %.3f' % gpu_free_mem('gb'))
# TODO change to steps
epochs = self.max_iter / len(Xs)
best_val_loss, best_epoch = None, None
best_mdl_path = os.path.join(out_folder, 'best_model.npz')
if not os.path.exists(out_folder):
os.makedirs(out_folder)
steps = 0
for epoch in range(epochs):
start_time = time.time()
train_err, train_batches = 0, 0
data_s = FileSystemData(Xs, Ys, data_folder, self.batch_size,
infinite=False, augment=True, shuffle=True)
step_err, step_g = 0, None
for batch in tqdm(data_s, total=data_s.steps, leave=False):
inputs, targets = batch
inputs = floatX(np.array([mdl.preprocess(x) for x in inputs]))
batch_err = train_fn(inputs, targets)
batch_g = grads_fn(inputs, targets)
if step_g is None:
step_g = batch_g
else:
step_g = [s_g + b_g for s_g, b_g in zip(step_g, batch_g)]
train_err += batch_err
step_err += batch_err
train_batches += 1
if train_batches % self.iter_size == 0:
step_g = [g / np.array(self.iter_size) for g in step_g]
if steps == 0:
t_prev, m_prev, u_prev = \
init_adam(batch_g, params)
updates = step_adam(step_g, params, t_prev, m_prev, u_prev,
learning_rate=self.base_lr)
for p, new_val in updates.items():
p.set_value(new_val)
steps += 1
step_err, step_g = 0, None
data_v = FileSystemData(Xv, Yv, data_folder, self.batch_size,
infinite=False, augment=False, shuffle=False)
val_err, val_acc, val_batches = 0, 0, 0
for batch in tqdm(data_v, total=data_v.steps, leave=False):
inputs, targets = batch
inputs = floatX(np.array([mdl.preprocess(x) for x in inputs]))
err, acc = val_fn(inputs, targets)
val_err += err
val_acc += acc
val_batches += 1
train_loss = train_err / train_batches
val_loss = val_err / val_batches
val_acc = val_acc / val_batches * 100
end_time = time.time() - start_time
if not best_val_loss or val_loss < best_val_loss:
best_val_loss = val_loss
best_epoch = epoch
np.savez(best_mdl_path,
*lasagne.layers.get_all_param_values(net))
snapshot_path = os.path.join(out_folder, 'snapshot_epoch_%d.npz'
% epoch)
np.savez(snapshot_path, *lasagne.layers.get_all_param_values(net))
logger.info("epoch[%d] -- Ls: %.3f | Lv: %.3f | ACCv: %.3f | Ts: %.3f"
% (epoch, train_loss, val_loss, val_acc, end_time))
logger.info("loading best model: epoch[%d]" % best_epoch)
with np.load(best_mdl_path) as f:
param_values = [f['arr_%d' % i] for i in range(len(f.files))]
lasagne.layers.set_all_param_values(net, param_values)
return predict_proba
def __init__(
self,
input_shape,
output_dim,
prob_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=NL.rectify,
optimizer=None,
use_trust_region=True,
step_size=0.01,
normalize_inputs=True,
name=None,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the mean
network.
:param hidden_nonlinearity: Non-linearity used for each layer of the
mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param step_size: KL divergence constraint for each iteration
"""
Serializable.quick_init(self, locals())
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer()
else:
optimizer = LbfgsOptimizer()
self.output_dim = output_dim
self._optimizer = optimizer
if prob_network is None:
prob_network = MLP(
input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=NL.softmax,
)
l_prob = prob_network.output_layer
LasagnePowered.__init__(self, [l_prob])
xs_var = prob_network.input_layer.input_var
ys_var = TT.imatrix("ys")
old_prob_var = TT.matrix("old_prob")
x_mean_var = theano.shared(np.zeros((1, ) + input_shape),
name="x_mean",
broadcastable=(True, ) +
(False, ) * len(input_shape))
x_std_var = theano.shared(np.ones((1, ) + input_shape),
name="x_std",
broadcastable=(True, ) +
(False, ) * len(input_shape))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
prob_var = L.get_output(l_prob,
{prob_network.input_layer: normalized_xs_var})
old_info_vars = dict(prob=old_prob_var)
info_vars = dict(prob=prob_var)
dist = self._dist = Categorical(output_dim)
mean_kl = TT.mean(dist.kl_sym(old_info_vars, info_vars))
loss = -TT.mean(dist.log_likelihood_sym(ys_var, info_vars))
predicted = tensor_utils.to_onehot_sym(TT.argmax(prob_var, axis=1),
output_dim)
self._f_predict = tensor_utils.compile_function([xs_var], predicted)
self._f_prob = tensor_utils.compile_function([xs_var], prob_var)
self._prob_network = prob_network
self._l_prob = l_prob
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[prob_var],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, step_size)
optimizer_args["inputs"] = [xs_var, ys_var, old_prob_var]
else:
optimizer_args["inputs"] = [xs_var, ys_var]
self._optimizer.update_opt(**optimizer_args)
self._use_trust_region = use_trust_region
self._name = name
self._normalize_inputs = normalize_inputs
self._x_mean_var = x_mean_var
self._x_std_var = x_std_var
def fit(self, X, learning_rate=1e-5, mu=0.99, epochs=10, show_fig=True, activation=T.nnet.relu, RecurrentUnit=GRU, normalize=True):
D = self.D
V = self.V
N = len(X)
We = init_weight(V, D)
self.hidden_layers = []
Mi = D
for Mo in self.hidden_layer_sizes:
ru = RecurrentUnit(Mi, Mo, activation)
self.hidden_layers.append(ru)
Mi = Mo
Wo = init_weight(Mi, V)
bo = np.zeros(V)
self.We = theano.shared(We)
self.Wo = theano.shared(Wo)
self.bo = theano.shared(bo)
self.params = [self.Wo, self.bo]
for ru in self.hidden_layers:
self.params += ru.params
thX = T.ivector('X')
thY = T.ivector('Y')
Z = self.We[thX]
for ru in self.hidden_layers:
Z = ru.output(Z)
py_x = T.nnet.softmax(Z.dot(self.Wo) + self.bo)
prediction = T.argmax(py_x, axis=1)
# let's return py_x too so we can draw a sample instead
self.predict_op = theano.function(
inputs=[thX],
outputs=[py_x, prediction],
allow_input_downcast=True,
)
cost = -T.mean(T.log(py_x[T.arange(thY.shape[0]), thY]))
grads = T.grad(cost, self.params)
dparams = [theano.shared(p.get_value()*0) for p in self.params]
dWe = theano.shared(self.We.get_value()*0)
gWe = T.grad(cost, self.We)
dWe_update = mu*dWe - learning_rate*gWe
We_update = self.We + dWe_update
if normalize:
We_update /= We_update.norm(2)
updates = [
(p, p + mu*dp - learning_rate*g) for p, dp, g in zip(self.params, dparams, grads)
] + [
(dp, mu*dp - learning_rate*g) for dp, g in zip(dparams, grads)
] + [
(self.We, We_update), (dWe, dWe_update)
]
self.train_op = theano.function(
inputs=[thX, thY],
outputs=[cost, prediction],
updates=updates
)
costs = []
for i in range(epochs):
t0 = datetime.now()
X = shuffle(X)
n_correct = 0
n_total = 0
cost = 0
for j in range(N):
if np.random.random() < 0.01 or len(X[j]) <= 1:
input_sequence = [0] + X[j]
output_sequence = X[j] + [1]
else:
input_sequence = [0] + X[j][:-1]
output_sequence = X[j]
n_total += len(output_sequence)
# test:
try:
# we set 0 to start and 1 to end
c, p = self.train_op(input_sequence, output_sequence)
except Exception as e:
PYX, pred = self.predict_op(input_sequence)
print("input_sequence len:", len(input_sequence))
print("PYX.shape:",PYX.shape)
print("pred.shape:", pred.shape)
raise e
# print "p:", p
cost += c
# print "j:", j, "c:", c/len(X[j]+1)
for pj, xj in zip(p, output_sequence):
if pj == xj:
n_correct += 1
if j % 200 == 0:
sys.stdout.write("j/N: %d/%d correct rate so far: %f\r" % (j, N, float(n_correct)/n_total))
sys.stdout.flush()
print("i:", i, "cost:", cost, "correct rate:", (float(n_correct)/n_total), "time for epoch:", (datetime.now() - t0))
costs.append(cost)
if show_fig:
plt.plot(costs)
plt.show()
def main(num_epochs=NEPOCH):
print("Loading data ...")
snli = SNLI(batch_size=BSIZE)
train_batches = list(snli.train_minibatch_generator())
dev_batches = list(snli.dev_minibatch_generator())
test_batches = list(snli.test_minibatch_generator())
W_word_embedding = snli.weight # W shape: (# vocab size, WE_DIM)
W_word_embedding = snli.weight / \
(numpy.linalg.norm(snli.weight, axis=1).reshape(snli.weight.shape[0], 1) + \
0.00001)
del snli
print("Building network ...")
########### input layers ###########
# hypothesis
input_var_h = T.TensorType('int32', [False, False])('hypothesis_vector')
input_var_h.tag.test_value = numpy.hstack(
(numpy.random.randint(1, 10000, (BSIZE, 18),
'int32'), numpy.zeros(
(BSIZE, 6)).astype('int32')))
l_in_h = lasagne.layers.InputLayer(shape=(BSIZE, None),
input_var=input_var_h)
input_mask_h = T.TensorType('int32', [False, False])('hypo_mask')
input_mask_h.tag.test_value = numpy.hstack((numpy.ones(
(BSIZE, 18), dtype='int32'), numpy.zeros((BSIZE, 6), dtype='int32')))
input_mask_h.tag.test_value[1, 18:22] = 1
l_mask_h = lasagne.layers.InputLayer(shape=(BSIZE, None),
input_var=input_mask_h)
# premise
input_var_p = T.TensorType('int32', [False, False])('premise_vector')
input_var_p.tag.test_value = numpy.hstack(
(numpy.random.randint(1, 10000, (BSIZE, 16),
'int32'), numpy.zeros(
(BSIZE, 3)).astype('int32')))
l_in_p = lasagne.layers.InputLayer(shape=(BSIZE, None),
input_var=input_var_p)
input_mask_p = T.TensorType('int32', [False, False])('premise_mask')
input_mask_p.tag.test_value = numpy.hstack((numpy.ones(
(BSIZE, 16), dtype='int32'), numpy.zeros((BSIZE, 3), dtype='int32')))
input_mask_p.tag.test_value[1, 16:18] = 1
l_mask_p = lasagne.layers.InputLayer(shape=(BSIZE, None),
input_var=input_mask_p)
###################################
# output shape (BSIZE, None, WEDIM)
l_hypo_embed = lasagne.layers.EmbeddingLayer(
l_in_h,
input_size=W_word_embedding.shape[0],
output_size=W_word_embedding.shape[1],
W=W_word_embedding)
l_prem_embed = lasagne.layers.EmbeddingLayer(
l_in_p,
input_size=W_word_embedding.shape[0],
output_size=W_word_embedding.shape[1],
W=l_hypo_embed.W)
# EMBEDING MAPPING: output shape (BSIZE, None, WEMAP)
l_hypo_reduced_embed = DenseLayer3DInput(l_hypo_embed,
num_units=WEMAP,
b=None,
nonlinearity=None)
l_hypo_embed_dpout = lasagne.layers.DropoutLayer(l_hypo_reduced_embed,
p=DPOUT,
rescale=True)
l_prem_reduced_embed = DenseLayer3DInput(l_prem_embed,
num_units=WEMAP,
W=l_hypo_reduced_embed.W,
b=None,
nonlinearity=None)
l_prem_embed_dpout = lasagne.layers.DropoutLayer(l_prem_reduced_embed,
p=DPOUT,
rescale=True)
# ATTEND
l_hypo_embed_hid1 = DenseLayer3DInput(
l_hypo_embed_dpout,
num_units=EMBDHIDA,
b=None,
nonlinearity=lasagne.nonlinearities.rectify)
l_hypo_embed_hid1_dpout = lasagne.layers.DropoutLayer(l_hypo_embed_hid1,
p=DPOUT,
rescale=True)
l_hypo_embed_hid2 = DenseLayer3DInput(
l_hypo_embed_hid1_dpout,
num_units=EMBDHIDB,
b=None,
nonlinearity=lasagne.nonlinearities.rectify)
l_prem_embed_hid1 = DenseLayer3DInput(
l_prem_embed_dpout,
num_units=EMBDHIDA,
W=l_hypo_embed_hid1.W,
b=None,
nonlinearity=lasagne.nonlinearities.rectify)
l_prem_embed_hid1_dpout = lasagne.layers.DropoutLayer(l_prem_embed_hid1,
p=DPOUT,
rescale=True)
l_prem_embed_hid2 = DenseLayer3DInput(
l_prem_embed_hid1_dpout,
num_units=EMBDHIDB,
W=l_hypo_embed_hid2.W,
b=None,
nonlinearity=lasagne.nonlinearities.rectify)
# output dim: (BSIZE, NROWx, NROWy)
l_e = ComputeEmbeddingPool([l_hypo_embed_hid1, l_prem_embed_hid2])
# output dim: (BSIZE, NROWy, DIM)
l_hypo_weighted = AttendOnEmbedding([l_hypo_reduced_embed, l_e],
masks=[l_mask_h, l_mask_p],
direction='col')
# output dim: (BSIZE, NROWx, DIM)
l_prem_weighted = AttendOnEmbedding([l_prem_reduced_embed, l_e],
masks=[l_mask_h, l_mask_p],
direction='row')
# COMPARE
# output dim: (BSIZE, NROW, 4*LSTMHID)
l_hypo_premwtd = lasagne.layers.ConcatLayer(
[l_hypo_reduced_embed, l_prem_weighted], axis=2)
l_prem_hypowtd = lasagne.layers.ConcatLayer(
[l_prem_reduced_embed, l_hypo_weighted], axis=2)
l_hypo_premwtd_dpout = lasagne.layers.DropoutLayer(l_hypo_premwtd,
p=DPOUT,
rescale=True)
l_hypo_comphid1 = DenseLayer3DInput(
l_hypo_premwtd_dpout,
num_units=COMPHIDA,
b=None,
nonlinearity=lasagne.nonlinearities.rectify)
l_hypo_comphid1_dpout = lasagne.layers.DropoutLayer(l_hypo_comphid1,
p=DPOUT,
rescale=True)
l_hypo_comphid2 = DenseLayer3DInput(
l_hypo_comphid1_dpout,
num_units=COMPHIDB,
b=None,
nonlinearity=lasagne.nonlinearities.rectify)
l_prem_hypowtd_dpout = lasagne.layers.DropoutLayer(l_prem_hypowtd,
p=DPOUT,
rescale=True)
l_prem_comphid1 = DenseLayer3DInput(
l_prem_hypowtd_dpout,
num_units=COMPHIDA,
W=l_hypo_comphid1.W,
b=None,
nonlinearity=lasagne.nonlinearities.rectify)
l_prem_comphid1_dpout = lasagne.layers.DropoutLayer(l_prem_comphid1,
p=DPOUT,
rescale=True)
l_prem_comphid2 = DenseLayer3DInput(
l_prem_comphid1_dpout,
num_units=COMPHIDB,
W=l_hypo_comphid2.W,
b=None,
nonlinearity=lasagne.nonlinearities.rectify)
# AGGREGATE
# output dim: (BSIZE, 4*LSTMHID)
l_hypo_mean = MeanOverDim(l_hypo_comphid2, mask=l_mask_h, dim=1)
l_prem_mean = MeanOverDim(l_prem_comphid2, mask=l_mask_p, dim=1)
l_v1v2 = lasagne.layers.ConcatLayer([l_hypo_mean, l_prem_mean], axis=1)
l_v1v2_dpout = lasagne.layers.DropoutLayer(l_v1v2, p=DPOUT, rescale=True)
l_outhid1 = lasagne.layers.DenseLayer(
l_v1v2_dpout,
num_units=OUTHID,
b=None,
nonlinearity=lasagne.nonlinearities.rectify)
l_outhid1_dpout = lasagne.layers.DropoutLayer(l_outhid1,
p=DPOUT,
rescale=True)
l_outhid2 = lasagne.layers.DenseLayer(
l_outhid1_dpout,
num_units=OUTHID,
b=None,
nonlinearity=lasagne.nonlinearities.rectify)
# l_outhid2_dpout = lasagne.layers.DropoutLayer(l_outhid2, p=DPOUT, rescale=True)
l_output = lasagne.layers.DenseLayer(
l_outhid2,
num_units=3,
b=None,
nonlinearity=lasagne.nonlinearities.softmax)
########### target, cost, validation, etc. ##########
target_values = T.ivector('target_output')
target_values.tag.test_value = numpy.asarray([
1,
] * BSIZE, dtype='int32')
network_output = lasagne.layers.get_output(l_output)
network_prediction = T.argmax(network_output, axis=1)
error_rate = T.mean(T.neq(network_prediction, target_values))
network_output_clean = lasagne.layers.get_output(l_output,
deterministic=True)
network_prediction_clean = T.argmax(network_output_clean, axis=1)
error_rate_clean = T.mean(T.neq(network_prediction_clean, target_values))
cost = T.mean(
T.nnet.categorical_crossentropy(network_output, target_values))
cost_clean = T.mean(
T.nnet.categorical_crossentropy(network_output_clean, target_values))
# Retrieve all parameters from the network
all_params = lasagne.layers.get_all_params(l_output)
if not UPDATEWE:
all_params.remove(l_hypo_embed.W)
numparams = sum(
[numpy.prod(i) for i in [i.shape.eval() for i in all_params]])
print("Number of params: {}\nName\t\t\tShape\t\t\tSize".format(numparams))
print("-----------------------------------------------------------------")
for item in all_params:
print("{0:24}{1:24}{2}".format(item, item.shape.eval(),
numpy.prod(item.shape.eval())))
# if exist param file then load params
look_for = 'params' + os.sep + 'params_' + filename + '.pkl'
if os.path.isfile(look_for):
print("Resuming from file: " + look_for)
all_param_values = cPickle.load(open(look_for, 'rb'))
for p, v in zip(all_params, all_param_values):
p.set_value(v)
# Compute SGD updates for training
print("Computing updates ...")
updates = lasagne.updates.adagrad(cost, all_params, LR)
# Theano functions for training and computing cost
print("Compiling functions ...")
train = theano.function([
l_in_h.input_var, l_mask_h.input_var, l_in_p.input_var,
l_mask_p.input_var, target_values
], [cost, error_rate],
updates=updates)
# mode=NanGuardMode(nan_is_error=True, inf_is_error=True, big_is_error=False))
compute_cost = theano.function([
l_in_h.input_var, l_mask_h.input_var, l_in_p.input_var,
l_mask_p.input_var, target_values
], [cost_clean, error_rate_clean])
# mode=NanGuardMode(nan_is_error=True, inf_is_error=True, big_is_error=False))
def evaluate(mode):
if mode == 'dev':
data = dev_batches
if mode == 'test':
data = test_batches
set_cost = 0.
set_error_rate = 0.
for batches_seen, (hypo, hm, premise, pm, truth) in enumerate(data, 1):
_cost, _error = compute_cost(hypo, hm, premise, pm, truth)
set_cost = (1.0 - 1.0 / batches_seen) * set_cost + \
1.0 / batches_seen * _cost
set_error_rate = (1.0 - 1.0 / batches_seen) * set_error_rate + \
1.0 / batches_seen * _error
return set_cost, set_error_rate
print("Done. Evaluating scratch model ...")
dev_set_cost, dev_set_error = evaluate('dev')
print("BEFORE TRAINING: dev cost %f, error %f" %
(dev_set_cost, dev_set_error))
print("Training ...")
try:
for epoch in range(num_epochs):
train_set_cost = 0.
train_set_error = 0.
start = time.time()
for batches_seen, (hypo, hm, premise, pm,
truth) in enumerate(train_batches, 1):
_cost, _error = train(hypo, hm, premise, pm, truth)
train_set_cost = (1.0 - 1.0 / batches_seen) * train_set_cost + \
1.0 / batches_seen * _cost
train_set_error = (1.0 - 1.0 / batches_seen) * train_set_error + \
1.0 / batches_seen * _error
if (batches_seen * BSIZE) % 5000 == 0:
end = time.time()
print("Sample %d %.2fs, lr %.4f, train cost %f, error %f" %
(batches_seen * BSIZE, end - start, LR,
train_set_cost, train_set_error))
start = end
if (batches_seen * BSIZE) % 100000 == 0:
dev_set_cost, dev_set_error = evaluate('dev')
print("***dev cost %f, error %f" %
(dev_set_cost, dev_set_error))
# save parameters
all_param_values = [p.get_value() for p in all_params]
cPickle.dump(
all_param_values,
open('params' + os.sep + 'params_' + filename + '.pkl', 'wb'))
dev_set_cost, dev_set_error = evaluate('dev')
test_set_cost, test_set_error = evaluate('test')
print("epoch %d, cost: train %f dev %f test %f;\n"
" error train %f dev %f test %f" %
(epoch, train_set_cost, dev_set_cost, test_set_cost,
train_set_error, dev_set_error, test_set_error))
except KeyboardInterrupt:
pdb.set_trace()
pass
def __init__(self,
environment,
rho=0.9,
rms_epsilon=0.0001,
momentum=0,
clip_delta=0,
freeze_interval=1000,
batch_size=32,
update_rule="rmsprop",
random_state=np.random.RandomState(),
double_Q=False,
neural_network=NN):
""" Initialize environment
"""
QNetwork.__init__(self, environment, batch_size)
self._rho = rho
self._rms_epsilon = rms_epsilon
self._momentum = momentum
self._clip_delta = clip_delta
self._freeze_interval = freeze_interval
self._double_Q = double_Q
self._random_state = random_state
self.update_counter = 0
states = [
] # list of symbolic variables for each of the k element in the belief state
# --> [ T.tensor4 if observation of element=matrix, T.tensor3 if vector, T.tensor 2 if scalar ]
next_states = [] # idem than states at t+1
self.states_shared = [
] # list of shared variable for each of the k element in the belief state
self.next_states_shared = [] # idem that self.states_shared at t+1
for i, dim in enumerate(self._input_dimensions):
if len(dim) == 3:
states.append(T.tensor4("%s_%s" % ("state", i)))
next_states.append(T.tensor4("%s_%s" % ("next_state", i)))
elif len(dim) == 2:
states.append(T.tensor3("%s_%s" % ("state", i)))
next_states.append(T.tensor3("%s_%s" % ("next_state", i)))
elif len(dim) == 1:
states.append(T.matrix("%s_%s" % ("state", i)))
next_states.append(T.matrix("%s_%s" % ("next_state", i)))
self.states_shared.append(
theano.shared(np.zeros((batch_size, ) + dim,
dtype=theano.config.floatX),
borrow=False))
self.next_states_shared.append(
theano.shared(np.zeros((batch_size, ) + dim,
dtype=theano.config.floatX),
borrow=False))
print("Number of observations per state: {}".format(
len(self.states_shared)))
print("For each observation, historySize + ponctualObs_i.shape: {}".
format(self._input_dimensions))
rewards = T.col('rewards')
actions = T.icol('actions')
terminals = T.icol('terminals')
thediscount = T.scalar(name='thediscount', dtype=theano.config.floatX)
thelr = T.scalar(name='thelr', dtype=theano.config.floatX)
Q_net = neural_network(self._batch_size, self._input_dimensions,
self._n_actions, self._random_state)
self.q_vals, self.params, shape_after_conv = Q_net._buildDQN(states)
print(
"Number of neurons after spatial and temporal convolution layers: {}"
.format(shape_after_conv))
self.next_q_vals, self.next_params, shape_after_conv = Q_net._buildDQN(
next_states)
self._resetQHat()
self.rewards_shared = theano.shared(np.zeros(
(batch_size, 1), dtype=theano.config.floatX),
broadcastable=(False, True))
self.actions_shared = theano.shared(np.zeros((batch_size, 1),
dtype='int32'),
broadcastable=(False, True))
self.terminals_shared = theano.shared(np.zeros((batch_size, 1),
dtype='int32'),
broadcastable=(False, True))
if (self._double_Q == True):
givens_next = {}
for i, x in enumerate(self.next_states_shared):
givens_next[states[i]] = x
self.next_q_vals_current_qnet = theano.function([],
self.q_vals,
givens=givens_next)
next_q_curr_qnet = theano.clone(self.next_q_vals)
argmax_next_q_vals = T.argmax(next_q_curr_qnet,
axis=1,
keepdims=True)
max_next_q_vals = self.next_q_vals[T.arange(batch_size),
argmax_next_q_vals.reshape(
(-1, ))].reshape((-1, 1))
else:
max_next_q_vals = T.max(self.next_q_vals, axis=1, keepdims=True)
not_terminals = T.ones_like(terminals) - terminals
target = rewards + not_terminals * thediscount * max_next_q_vals
q_val = self.q_vals[T.arange(batch_size),
actions.reshape((-1, ))].reshape((-1, 1))
# Note : Strangely (target - q_val) lead to problems with python 3.5, theano 0.8.0rc and floatX=float32...
diff = -q_val + target
if self._clip_delta > 0:
# This loss function implementation is taken from
# https://github.com/spragunr/deep_q_rl
# If we simply take the squared clipped diff as our loss,
# then the gradient will be zero whenever the diff exceeds
# the clip bounds. To avoid this, we extend the loss
# linearly past the clip point to keep the gradient constant
# in that regime.
#
# This is equivalent to declaring d loss/d q_vals to be
# equal to the clipped diff, then backpropagating from
# there, which is what the DeepMind implementation does.
quadratic_part = T.minimum(abs(diff), self._clip_delta)
linear_part = abs(diff) - quadratic_part
loss_ind = 0.5 * quadratic_part**2 + self._clip_delta * linear_part
else:
loss_ind = 0.5 * diff**2
loss = T.mean(loss_ind)
givens = {
rewards: self.rewards_shared,
actions: self.actions_shared, ## actions not needed!
terminals: self.terminals_shared
}
for i, x in enumerate(self.states_shared):
givens[states[i]] = x
for i, x in enumerate(self.next_states_shared):
givens[next_states[i]] = x
gparams = []
for p in self.params:
gparam = T.grad(loss, p)
gparams.append(gparam)
updates = []
if update_rule == 'deepmind_rmsprop':
updates = deepmind_rmsprop(loss, self.params, gparams, thelr,
self._rho, self._rms_epsilon)
elif update_rule == 'rmsprop':
for i, (p, g) in enumerate(zip(self.params, gparams)):
acc = theano.shared(p.get_value() * 0.)
acc_new = rho * acc + (1 - self._rho) * g**2
gradient_scaling = T.sqrt(acc_new + self._rms_epsilon)
g = g / gradient_scaling
updates.append((acc, acc_new))
updates.append((p, p - thelr * g))
elif update_rule == 'sgd':
for i, (param, gparam) in enumerate(zip(self.params, gparams)):
updates.append((param, param - thelr * gparam))
else:
raise ValueError("Unrecognized update: {}".format(update_rule))
if (self._double_Q == True):
self._train = theano.function(
[thediscount, thelr, next_q_curr_qnet],
[loss, loss_ind, self.q_vals],
updates=updates,
givens=givens,
on_unused_input='warn')
else:
self._train = theano.function([thediscount, thelr],
[loss, loss_ind, self.q_vals],
updates=updates,
givens=givens,
on_unused_input='warn')
givens2 = {}
for i, x in enumerate(self.states_shared):
givens2[states[i]] = x
self._q_vals = theano.function([],
self.q_vals,
givens=givens2,
on_unused_input='warn')
def compile(self,
optimizer,
loss,
class_mode="categorical",
theano_mode=None):
self.optimizer = optimizers.get(optimizer)
self.loss = objectives.get(loss)
weighted_loss = weighted_objective(objectives.get(loss))
# input of model
self.X_train = self.get_input(train=True)
self.X_test = self.get_input(train=False)
self.y_train = self.get_output(train=True)
self.y_test = self.get_output(train=False)
# target of model
self.y = T.zeros_like(self.y_train)
self.weights = T.ones_like(self.y_train)
if hasattr(self.layers[-1], "get_output_mask"):
mask = self.layers[-1].get_output_mask()
else:
mask = None
train_loss = weighted_loss(self.y, self.y_train, self.weights, mask)
test_loss = weighted_loss(self.y, self.y_test, self.weights, mask)
train_loss.name = 'train_loss'
test_loss.name = 'test_loss'
self.y.name = 'y'
if class_mode == "categorical":
train_accuracy = T.mean(
T.eq(T.argmax(self.y, axis=-1), T.argmax(self.y_train,
axis=-1)))
test_accuracy = T.mean(
T.eq(T.argmax(self.y, axis=-1), T.argmax(self.y_test,
axis=-1)))
elif class_mode == "binary":
train_accuracy = T.mean(T.eq(self.y, T.round(self.y_train)),
dtype='float32')
test_accuracy = T.mean(T.eq(self.y, T.round(self.y_test)),
dtype='float32')
else:
raise Exception("Invalid class mode:" + str(class_mode))
self.class_mode = class_mode
self.theano_mode = theano_mode
for r in self.regularizers:
train_loss = r(train_loss)
updates = self.optimizer.get_updates(self.params, self.constraints,
train_loss)
updates += self.updates
if type(self.X_train) == list:
train_ins = self.X_train + [self.y, self.weights]
test_ins = self.X_test + [self.y, self.weights]
predict_ins = self.X_test
else:
train_ins = [self.X_train, self.y, self.weights]
test_ins = [self.X_test, self.y, self.weights]
predict_ins = [self.X_test]
self._train = theano.function(train_ins,
train_loss,
updates=updates,
allow_input_downcast=True,
mode=theano_mode)
self._train_with_acc = theano.function(train_ins,
[train_loss, train_accuracy],
updates=updates,
allow_input_downcast=True,
mode=theano_mode)
self._predict = theano.function(predict_ins,
self.y_test,
allow_input_downcast=True,
mode=theano_mode)
self._test = theano.function(test_ins,
test_loss,
allow_input_downcast=True,
mode=theano_mode)
self._test_with_acc = theano.function(test_ins,
[test_loss, test_accuracy],
allow_input_downcast=True,
mode=theano_mode)
def main():
print("Loading Data")
X_train, y_train, X_valid, y_valid, X_test, y_test = load_data.load_data_feautre_train(feautre = u"\uBC18\uD314",root_path= "/home/prosurpa/Image/image/",image_size=(28,28))
input_var = T.tensor4('inputs')
target_var = T.ivector('targets')
print("Bulding Model")
batch_size = 20
network = build_f_cnn(batch_size ,input_var)
prediction = lasagne.layers.get_output(network)
loss = lasagne.objectives.categorical_crossentropy(prediction, target_var)
loss = loss.mean()
params = lasagne.layers.get_all_params(network, trainable=True)
updates = lasagne.updates.nesterov_momentum(
loss, params, learning_rate=0.01, momentum=0.9
)
test_prediction = lasagne.layers.get_output(network, deterministic=True)
test_loss = lasagne.objectives.categorical_crossentropy(test_prediction, target_var)
test_loss = test_loss.mean()
test_acc = T.mean(T.eq(T.argmax(test_prediction, axis=1), target_var), dtype=theano.config.floatX)
train_fn = theano.function([input_var, target_var], loss, updates=updates)
val_fn = theano.function([input_var, target_var], [test_loss, test_acc])
#model_rw.read_model_data(network, "75.0000009934model")
print("Starting training")
num_epochs = 1000
best_acc = 75
for epoch in range(num_epochs):
train_err = 0
train_batches = 0
start_time = time.time()
print((len(X_train)/batch_size))
for batch in iterate_minibatches(X_train, y_train, batch_size, shuffle=True):
inputs, targets = batch
train_err += train_fn(inputs, targets)
train_batches += 1
if train_batches%20 == 0:
print(train_batches)
val_err = 0
val_acc = 0
val_batches = 0
print((len(X_valid) / batch_size))
for batch in iterate_minibatches(X_valid, y_valid, batch_size, shuffle=False):
inputs, targets = batch
err, acc = val_fn(inputs, targets)
val_err += err
val_acc += acc
val_batches += 1
if train_batches % 20 == 0:
print(val_batches)
print("Epoch {} of {} took {:.3f}s".format(
epoch + 1, num_epochs, time.time() - start_time))
print(" training loss:\t\t{:.6f}".format(train_err / train_batches))
print(" validation loss:\t\t{:.6f}".format(val_err / val_batches))
print(" validation accuracy:\t\t{:.2f} %".format(
val_acc / val_batches * 100))
test_err = 0
test_acc = 0
test_batches = 0
print((len(X_test) / batch_size))
for batch in iterate_minibatches(X_test, y_test, batch_size, shuffle=False):
inputs, targets = batch
err, acc = val_fn(inputs, targets)
test_err += err
test_acc += acc
test_batches += 1
if train_batches % 20 == 0:
print(test_batches)
print("Final results:")
print(" test loss:\t\t\t{:.6f}".format(test_err / test_batches))
print(" test accuracy:\t\t{:.2f} %".format(
test_acc / test_batches * 100))
re_acc = test_acc / test_batches * 100
if re_acc > best_acc + 0.5:
best_acc = re_acc
model_rw.write_model_data(network, str(best_acc) + "model")
lasagne.objectives.squared_error(prediction, train_prediction_b))
# loss=loss+pi_loss
elif model.network_type == "tempens":
# Tempens model loss:
loss = T.mean(loss * mask_train, dtype=theano.config.floatX)
loss += unsup_weight_var * T.mean(
lasagne.objectives.squared_error(prediction, z_target_var))
else:
loss = T.mean(loss, dtype=theano.config.floatX)
# regularization:L1,L2
l2_penalty = lasagne.regularization.regularize_network_params(
gru_network, lasagne.regularization.l2) * model.l2_loss
loss = loss + l2_penalty
train_acc = T.mean(T.eq(T.argmax(prediction, axis=1),
T.argmax(target_var, axis=1)),
dtype=theano.config.floatX)
# We could add some weight decay as well here, see lasagne.regularization.
# Create update expressions for training, i.e., how to modify the
# parameters at each training step. Here, we'll use Stochastic Gradient
# Descent (SGD) with Nesterov momentum, but Lasagne offers plenty more.
params = lasagne.layers.get_all_params(gru_network, trainable=True)
updates = lasagne.updates.adam(loss,
params,
learning_rate=learning_rate_var,
beta1=adam_beta1_var)
"""
3.test loss and accuracy
"""
def fit(self, trees, learning_rate=10e-4, mu=0.99, reg=10e-3, epochs=15, activation=T.nnet.relu, train_inner_nodes=False):
D = self.D
V = self.V
K = self.K
self.f = activation
N = len(trees)
We = init_weight(V, D)
W11 = np.random.randn(D, D, D) / np.sqrt(3*D)
W22 = np.random.randn(D, D, D) / np.sqrt(3*D)
W12 = np.random.randn(D, D, D) / np.sqrt(3*D)
W1 = init_weight(D, D)
W2 = init_weight(D, D)
bh = np.zeros(D)
Wo = init_weight(D, K)
bo = np.zeros(K)
self.We = theano.shared(We)
self.W11 = theano.shared(W11)
self.W22 = theano.shared(W22)
self.W12 = theano.shared(W12)
self.W1 = theano.shared(W1)
self.W2 = theano.shared(W2)
self.bh = theano.shared(bh)
self.Wo = theano.shared(Wo)
self.bo = theano.shared(bo)
self.params = [self.We, self.W11, self.W22, self.W12, self.W1, self.W2, self.bh, self.Wo, self.bo]
words = T.ivector('words')
left_children = T.ivector('left_children')
right_children = T.ivector('right_children')
labels = T.ivector('labels')
def recurrence(n, hiddens, words, left, right):
w = words[n]
# any non-word will have index -1
hiddens = T.switch(
T.ge(w, 0),
T.set_subtensor(hiddens[n], self.We[w]),
T.set_subtensor(hiddens[n],
self.f(
hiddens[left[n]].dot(self.W11).dot(hiddens[left[n]]) +
hiddens[right[n]].dot(self.W22).dot(hiddens[right[n]]) +
hiddens[left[n]].dot(self.W12).dot(hiddens[right[n]]) +
hiddens[left[n]].dot(self.W1) +
hiddens[right[n]].dot(self.W2) +
self.bh
)
)
)
return hiddens
hiddens = T.zeros((words.shape[0], D))
h, _ = theano.scan(
fn=recurrence,
outputs_info=[hiddens],
n_steps=words.shape[0],
sequences=T.arange(words.shape[0]),
non_sequences=[words, left_children, right_children],
)
py_x = T.nnet.softmax(h[:,0,:].dot(self.Wo) + self.bo)
prediction = T.argmax(py_x, axis=1)
rcost = T.mean([(p*p).sum() for p in self.params])
if train_inner_nodes:
cost = -T.mean(T.log(py_x[T.arange(labels.shape[0]), labels])) + rcost
else:
cost = -T.mean(T.log(py_x[-1, labels[-1]])) + rcost
grads = T.grad(cost, self.params)
dparams = [theano.shared(p.get_value()*0) for p in self.params]
updates = [
(p, p + mu*dp - learning_rate*g) for p, dp, g in zip(self.params, dparams, grads)
] + [
(dp, mu*dp - learning_rate*g) for dp, g in zip(dparams, grads)
]
self.cost_predict_op = theano.function(
inputs=[words, left_children, right_children, labels],
outputs=[cost, prediction],
allow_input_downcast=True,
)
self.train_op = theano.function(
inputs=[words, left_children, right_children, labels],
outputs=[cost, prediction],
updates=updates
)
costs = []
sequence_indexes = range(N)
if train_inner_nodes:
n_total = sum(len(words) for words, _, _, _ in trees)
else:
n_total = N
for i in xrange(epochs):
t0 = datetime.now()
sequence_indexes = shuffle(sequence_indexes)
n_correct = 0
cost = 0
it = 0
for j in sequence_indexes:
words, left, right, lab = trees[j]
c, p = self.train_op(words, left, right, lab)
if np.isnan(c):
print "Cost is nan! Let's stop here. Why don't you try decreasing the learning rate?"
exit()
cost += c
if train_inner_nodes:
n_correct += np.sum(p == lab)
else:
n_correct += (p[-1] == lab[-1])
it += 1
if it % 1 == 0:
sys.stdout.write("j/N: %d/%d correct rate so far: %f, cost so far: %f\r" % (it, N, float(n_correct)/n_total, cost))
sys.stdout.flush()
print "i:", i, "cost:", cost, "correct rate:", (float(n_correct)/n_total), "time for epoch:", (datetime.now() - t0)
costs.append(cost)
plt.plot(costs)
plt.show()
output = architecture.buildDCNN()
dcnnParams = lasagne.layers.get_all_params(output)
# SYMBOLIC INPUTS
x = T.imatrix()
y = T.ivector()
# Without L2 Regularization
loss = lasagne.objectives.aggregate(
lasagne.objectives.categorical_crossentropy(
lasagne.layers.get_output(output, x), y), mode = 'mean')
updates = lasagne.updates.adagrad(loss, dcnnParams, learning_rate = 0.1)
# ACCURACY FOR PREDICTIONS
prediction = T.argmax(lasagne.layers.get_output(output, x, deterministic=True), axis=1)
score = T.eq(prediction, y).mean()
# SYMBOLIC FUNCTIONS
trainDCNN = theano.function([x,y], outputs = loss, updates = updates)
validateDCNN = theano.function([x,y], outputs = score)
testDCNN = theano.function([x,y], outputs = score)
# LOAD THE DATA
trainingSentences = loader.loadData('myDataset/train.txt')
trainingLabels = loader.loadData('myDataset/train_label.txt')
validationSentences = loader.loadData('myDataset/dev.txt')
validationLabels = loader.loadData('myDataset/dev_label.txt')
testSentences = loader.loadData('myDataset/test.txt')
testLabels = loader.loadData('myDataset/test_label.txt')
lasagne.layers.set_all_param_values(net['prob'], params) n_batches_per_epoch = np.floor(n_training_samples/float(BATCH_SIZE)) n_test_batches = np.floor(n_val_samples/float(BATCH_SIZE)) x_sym = T.tensor4() y_sym = T.ivector() l2_loss = lasagne.regularization.regularize_network_params(net['prob'], lasagne.regularization.l2) * 5e-4 prediction_train = lasagne.layers.get_output(net['prob'], x_sym, deterministic=False) loss = lasagne.objectives.categorical_crossentropy(prediction_train, y_sym) loss = loss.mean() loss += l2_loss acc_train = T.mean(T.eq(T.argmax(prediction_train, axis=1), y_sym), dtype=theano.config.floatX) prediction_test = lasagne.layers.get_output(net['prob'], x_sym, deterministic=True) loss_val = lasagne.objectives.categorical_crossentropy(prediction_test, y_sym) loss_val = loss_val.mean() loss_val += l2_loss acc = T.mean(T.eq(T.argmax(prediction_test, axis=1), y_sym), dtype=theano.config.floatX) params = lasagne.layers.get_all_params(net['prob'], trainable=True) learning_rate = theano.shared(np.float32(0.001)) updates = lasagne.updates.adam(loss, params, learning_rate=learning_rate) train_fn = theano.function([x_sym, y_sym], [loss, acc_train], updates=updates) val_fn = theano.function([x_sym, y_sym], [loss_val, acc]) pred_fn = theano.function([x_sym], prediction_test)
return p_y_given_x
w_c1 = init_weights((4, 1, 3, 3))
b_c1 = init_weights((4, ))
w_c2 = init_weights((8, 4, 3, 3))
b_c2 = init_weights((8, ))
w_h3 = init_weights((8 * 4 * 4, 100))
b_h3 = init_weights((100, ))
w_o = init_weights((100, 10))
b_o = init_weights((10, ))
params = [w_c1, b_c1, w_c2, b_c2, w_h3, b_h3, w_o, b_o]
p_y_given_x = model(x, *params)
y = T.argmax(p_y_given_x, axis=1)
cost = T.mean(T.nnet.categorical_crossentropy(p_y_given_x, t))
updates = momentum(cost, params, learning_rate=0.01, momentum=0.9)
# compile theano functions
train = theano.function([x, t], cost, updates=updates)
predict = theano.function([x], y)
# train model
batch_size = 50
for i in range(50):
print "iteration %d" % (i + 1)
for start in range(0, len(x_train), batch_size):