def uncertainty_gs(probs, likelyhoods, credal_size): sorted_index = np.argsort(likelyhoods, kind='stable') l = likelyhoods[sorted_index] p = probs[:,sorted_index] gs_total = [] gs_epist = [] gs_ale = [] for level in range(credal_size-1): p_cut = p[:,0:level+2] # get the level cut probs based on sorted likelyhood # computing levi (set14) for level cut p_cut and appeinding to the unc array entropy = -p_cut*np.ma.log2(p_cut) entropy = entropy.filled(0) entropy_sum = np.sum(entropy, axis=2) s_max = np.max(entropy_sum, axis=1) s_min = np.min(entropy_sum, axis=1) gh = set_gh(p_cut) total = s_max e = gh a = total - e gs_total.append(total) gs_epist.append(e) gs_ale.append(a) gs_total = np.mean(np.array(gs_total), axis=0) gs_epist = np.mean(np.array(gs_epist), axis=0) gs_ale = np.mean(np.array(gs_ale), axis=0) return gs_total, gs_epist, gs_ale
def uncertainty_ent_bays(probs, likelihoods): # three dimentianl array with d1 as datapoints, (d2) the rows as samples and (d3) the columns as probability for each class
p = np.array(probs)
# print("prob\n", probs)
# print("likelihoods in bays", likelihoods)
entropy = -p*np.ma.log2(p)
entropy = entropy.filled(0)
# print("entropy\n", entropy)
a = np.sum(entropy, axis=2)
al = a * likelihoods
a = np.sum(al, axis=1)
given_axis = 1
dim_array = np.ones((1,probs.ndim),int).ravel()
dim_array[given_axis] = -1
b_reshaped = likelihoods.reshape(dim_array)
mult_out = probs*b_reshaped
p_m = np.sum(mult_out, axis=1)
# p_m = np.mean(p, axis=1) #* likelihoods
total = -np.sum(p_m*np.ma.log2(p_m), axis=1)
total = total.filled(0)
e = total - a
return total, e, a
def uncertainty_set15(probs, bootstrap_size=0, sampling_size=0, credal_size=0):
if bootstrap_size > 0:
p = [] #np.array(probs)
for data_point in probs:
d_p = []
for sampling_seed in range(bootstrap_size):
d_p.append(resample(data_point, random_state=sampling_seed))
p.append(np.array(d_p))
p = np.array(p)
p = np.mean(p, axis=2)
if sampling_size > 0:
p = []
for sample_index in range(sampling_size):
# number_of_samples = int(probs.shape[1] / sampling_size)
# print("number_of_samples ", number_of_samples)
sampled_index = np.random.choice(probs.shape[1], credal_size)
p.append(probs[:,sampled_index,:])
p = np.array(p)
p = np.mean(p, axis=2)
p = p.transpose([1,0,2])
else:
p = probs
entropy = -p*np.ma.log2(p)
entropy = entropy.filled(0)
entropy_sum = np.sum(entropy, axis=2)
s_min = np.min(entropy_sum, axis=1)
s_max = np.max(entropy_sum, axis=1)
total = s_max
e = s_max - s_min
a = total - e
return total, e, a
def uncertainty_ent(probs): # three dimentianl array with d1 as datapoints, (d2) the rows as samples and (d3) the columns as probability for each class p = np.array(probs) entropy = -p*np.ma.log2(p) entropy = entropy.filled(0) a = np.sum(entropy, axis=1) a = np.sum(a, axis=1) / entropy.shape[1] p_m = np.mean(p, axis=1) total = -np.sum(p_m*np.ma.log2(p_m), axis=1) total = total.filled(0) e = total - a return total, e, a # now it should be correct
def uncertainty_ent_levi(probs, credal_size=30): # three dimentianl array with d1 as datapoints, (d2) the rows as samples and (d3) the columns as probability for each class p = [] #np.array(probs) for data_point in probs: d_p = [] for sampling_seed in range(credal_size): d_p.append(resample(data_point, random_state=sampling_seed)) p.append(np.array(d_p)) p = np.array(p) p = np.mean(p, axis=2) entropy = -p*np.ma.log10(p) entropy = entropy.filled(0) a = np.sum(entropy, axis=1) a = np.sum(a, axis=1) / entropy.shape[1] p_m = np.mean(p, axis=1) total = -np.sum(p_m*np.ma.log10(p_m), axis=1) total = total.filled(0) e = total - a return total, e, a # now it should be correct
def uncertainty_set17(probs, bootstrap_size=0, sampling_size=0, credal_size=0, log=False):
if bootstrap_size > 0:
p = [] #np.array(probs)
for data_point in probs:
d_p = []
for sampling_seed in range(bootstrap_size):
d_p.append(resample(data_point, random_state=sampling_seed))
p.append(np.array(d_p))
p = np.array(p)
p = np.mean(p, axis=2)
if sampling_size > 0:
p = []
for sample_index in range(sampling_size):
# number_of_samples = int(probs.shape[1] / sampling_size)
sampled_index = np.random.choice(probs.shape[1], credal_size)
p.append(probs[:,sampled_index,:])
p = np.array(p)
p = np.mean(p, axis=2)
p = p.transpose([1,0,2])
else:
p = probs
if log:
print("------------------------------------set14 prob after averaging each ensemble")
print("Set14 p \n" , p)
print(p.shape)
# entropy = -p*np.log2(p)
entropy = -p*np.ma.log2(p)
entropy = entropy.filled(0)
p_m = np.mean(p, axis=1)
total = -np.sum(p_m*np.ma.log2(p_m), axis=1)
total = total.filled(0)
entropy_sum = np.sum(entropy, axis=2)
s_max = np.max(entropy_sum, axis=1)
gh = set_gh(p)
e = gh
a = s_max
total = a + e
return total, e, a
def uncertainty_ent_standard(probs): # for tree p = np.array(probs) entropy = -p*np.ma.log10(p) entropy = entropy.filled(0) total = np.sum(entropy, axis=1) return total, total, total # now it should be correct