def model_bin_train(self, data_row, truth, num_bins=2):
#TODO add epsilon
model = {}
cutoffsc = [[] for _ in range(len(data_row[0]))]
dmat = np.matrix(data_row)
drange = dmat.max() - dmat.min()
bin_size = float(drange) / num_bins
data_col = hw3.transpose_array(data_row)
for j in range(len(data_col)):
#cutoffsc.append([min(data_col)[0] + bin_size * i for i in range(num_bins)])
mu = np.asarray(data_col[j]).mean()
low_mu = np.asarray([data_col[j][i] for i in range(len(data_col[j])) if data_col[j][i] < mu]).mean()
high_mu = np.asarray([data_col[j][i] for i in range(len(data_col[j])) if data_col[j][i] > mu]).mean()
if num_bins == 4:
cutoffsc[j] = [min(data_col)[0], low_mu, mu, high_mu]
else:
cutoffsc[j] = [min(data_col)[0], (low_mu - min(data_col)[0])/2, mu, (high_mu-mu)/2, high_mu, (max(data_col)[0]-high_mu)/2]
cutoffs = [dmat.min() + bin_size * i for i in range(num_bins)]
#epsilon = float(alpha * 1) / len(covar_matrix)
for label in [0,1]:
# transpose to go by column
sub_data = hw3.transpose_array(hw3.get_sub_at_value(data_row, truth, label))
model[label] = hw3.bins_per_column(sub_data, cutoffs)
model[label] = hw3.bins_per_column_by_col(sub_data, cutoffsc)
# probability of bin given label
self.y_prob = float(sum(truth))/len(truth)
self.cutoffs = cutoffsc
return model
def model_gaussian_rand_var_train(self, data, truth):
mus = {}
std_dev = {}
for label in [0,1]:
sub_data = hw3.get_sub_at_value(data, truth, label)
mus[label] = hw3.get_mus(sub_data)
std_dev[label] = hw3.get_std_dev(sub_data)
self.y_prob = float(sum(truth))/len(truth)
return [mus, std_dev, float(sum(truth))/len(truth)]
def model_average_train(self, data_row, truth):
""" return [prob_over_given_1, prob_over_given_0, prob_y1]
prob_over_give_x = col1[mu, var, proabality], colx[mu, var, prob] ...
"""
mus = hw3.get_mus(data_row)
is_not_spam = hw3.get_sub_at_value(data_row, truth, 0)
is_spam = hw3.get_sub_at_value(data_row, truth, 1)
prob_over = get_prob_over(data_row, mus)
prob_over_given_1 = get_prob_over(is_spam, mus)
prob_over_given_0 = get_prob_over(is_not_spam, mus)
l0 = len(prob_over_given_0)
l1 = len(prob_over_given_1)
if l1 != l0:
addx = abs(l1-l0)
fake_row = [0 for _ in range(addx)]
if l1 > l0:
prob_over_given_0 = fake_row
else:
prob_over_given_1 = fake_row
prob_y1 = float(sum(truth))/len(truth)
self.y_prob = prob_y1
return [prob_over_given_1, prob_over_given_0, prob_over, prob_y1]
def initialize(self, data, k=2):
# start with k = 2 and std_dev = 1
self.k = k
self.labels = [ki for ki in range(self.k)]
models = [EMModel() for _ in range(self.k)]
mucheat = mu_cheat(hw3.transpose_array(data), k)
for ki in range(self.k):
#models[ki].random_mus(data)
models[ki].mu = mucheat[ki]
self.labels = self.assign_labels(data, models)
#self.labels = self.assign_labels2(data, model)
self.prevent_empty(data)
for ki in range(self.k):
sub_data = hw3.get_sub_at_value(data, self.labels, ki)
#models[ki].sigma = hw3.get_covar(sub_data)
models[ki].sigma = hw3.get_covar(data)
#models[ki].weight = float(len(sub_data)) / len(data)
models[ki].weight = .5
models[ki].likelihood = self.expectation(data, models[ki]) # multivarate_normal
self.models = models
