def predict_proba(self, pairs, features):
features = np.array([feature+[1] for feature in features])
probs = [0.0 for i in xrange(len(features))]
for i in xrange(len(features)):
feature = features[i]
if pairs[i][self.field_for_model_num] not in self.val_map_model:
lr_val = logisticVal(self.global_para, feature)
else:
model_idx = self.val_map_model[pairs[i][self.field_for_model_num]]
lr_val = logisticVal(self.model_para[model_idx], feature)
probs[i] = lr_val
return probs
def predict(self, pairs, features):
features = np.array([feature+[1] for feature in features])
targets = [0 for i in xrange(len(features))]
for i in xrange(len(features)):
feature = features[i]
if pairs[i][self.field_for_model_num] not in self.val_map_model:
lr_val = logisticVal(self.global_para, feature)
else:
model_idx = self.val_map_model[pairs[i][self.field_for_model_num]]
lr_val = logisticVal(self.model_para[model_idx], feature)
if lr_val > 0.5:
targets[i] = 1
return targets
def predict_proba(self, pairs, features):
features = np.array([feature + [1] for feature in features])
probs = [0.0 for i in xrange(len(features))]
for i in xrange(len(features)):
feature = features[i]
if pairs[i][self.field_for_model_num] not in self.val_map_model:
lr_val = logisticVal(self.global_para, feature)
else:
model_idx = self.val_map_model[pairs[i][
self.field_for_model_num]]
lr_val = logisticVal(self.model_para[model_idx], feature)
probs[i] = lr_val
return probs
def predict(self, pairs, features):
features = np.array([feature + [1] for feature in features])
targets = [0 for i in xrange(len(features))]
for i in xrange(len(features)):
feature = features[i]
if pairs[i][self.field_for_model_num] not in self.val_map_model:
lr_val = logisticVal(self.global_para, feature)
else:
model_idx = self.val_map_model[pairs[i][
self.field_for_model_num]]
lr_val = logisticVal(self.model_para[model_idx], feature)
if lr_val > 0.5:
targets[i] = 1
return targets
def stochasticGraidentDescent(self, feature, target, model_idx):
lr_val = logisticVal(self.model_para[model_idx], feature)
avg_model_para = np.mean(self.model_para, 0)
self.model_para[model_idx] = self.model_para[model_idx]\
+ self.lr*(target*feature-lr_val*feature
- self.C*self.model_para[model_idx]
- self.eta*(self.model_para[model_idx]-avg_model_para))
def stochasticGraidentDescent(self, feature, target, model_idx):
lr_val = logisticVal(self.model_para[model_idx], feature)
avg_model_para = np.mean(self.model_para, 0)
self.model_para[model_idx] = self.model_para[model_idx]\
+ self.lr*(target*feature-lr_val*feature
- self.C*self.model_para[model_idx]
- self.eta*(self.model_para[model_idx]-avg_model_para))
def calLossVal(self):
correct_num = 0
for i in xrange(self.nInstance):
feature = self.features[i]
target = self.targets[i]
if self.pairs[i][self.field_for_model_num] in self.val_map_model:
model_idx = self.val_map_model[self.pairs[i][self.field_for_model_num]]
lr_val = logisticVal(self.model_para[model_idx], feature)
if np.abs(lr_val-0.5)*target>0:
correct_num += 1
return 1-1.0*correct_num/self.nInstance
def calLossVal(self):
correct_num = 0
for i in xrange(self.nInstance):
feature = self.features[i]
target = self.targets[i]
if self.pairs[i][self.field_for_model_num] in self.val_map_model:
model_idx = self.val_map_model[self.pairs[i][
self.field_for_model_num]]
lr_val = logisticVal(self.model_para[model_idx], feature)
if np.abs(lr_val - 0.5) * target > 0:
correct_num += 1
return 1 - 1.0 * correct_num / self.nInstance
def calLikelihood(self):
#pred_prob = np.array([0.0 for i in xrange(self.nInstance)])
loglikelihood = 0.0
for i in xrange(self.nInstance):
feature = self.features[i]
target = self.targets[i]
if self.pairs[i][self.field_for_model_num] in self.val_map_model:
model_idx = self.val_map_model[self.pairs[i][self.field_for_model_num]]
pred_prob = logisticVal(self.model_para[model_idx], feature)
if target == 1:
loglikelihood += np.log(pred_prob)
else:
loglikelihood += np.log(1-pred_prob)
#loglikelihood = np.dot(np.log(pred_prob).reshape(1, self.nInstance), self.targets)\
# + np.dot(np.log(1-pred_prob).reshape(1, self.nInstance), 1-self.targets)
return loglikelihood
def calLikelihood(self):
#pred_prob = np.array([0.0 for i in xrange(self.nInstance)])
loglikelihood = 0.0
for i in xrange(self.nInstance):
feature = self.features[i]
target = self.targets[i]
if self.pairs[i][self.field_for_model_num] in self.val_map_model:
model_idx = self.val_map_model[self.pairs[i][
self.field_for_model_num]]
pred_prob = logisticVal(self.model_para[model_idx], feature)
if target == 1:
loglikelihood += np.log(pred_prob)
else:
loglikelihood += np.log(1 - pred_prob)
#loglikelihood = np.dot(np.log(pred_prob).reshape(1, self.nInstance), self.targets)\
# + np.dot(np.log(1-pred_prob).reshape(1, self.nInstance), 1-self.targets)
return loglikelihood