1
0
Fork 0
Old engine for Continuous Time Bayesian Networks. Superseded by reCTBN. 🐍 https://github.com/madlabunimib/PyCTBN
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
This repo is archived. You can view files and clone it, but cannot push or open issues/pull-requests.
PyCTBN/main_package/classes/parameters_estimator.py

391 lines
20 KiB

import os
import time as tm
from line_profiler import LineProfiler
import numba as nb
import numpy as np
import network_graph as ng
import sample_path as sp
import amalgamated_cims as acims
class ParametersEstimator:
def __init__(self, sample_path, net_graph):
self.sample_path = sample_path
self.net_graph = net_graph
self.scalar_indexes_converter = self.net_graph.scalar_indexing_structure
self.columns_filtering_structure = self.net_graph.filtering_structure
self.transition_scalar_index_converter = self.net_graph.transition_scalar_indexing_structure
self.transition_filtering = self.net_graph.transition_filtering
self.amalgamated_cims_struct = None
def init_amalgamated_cims_struct(self):
self.amalgamated_cims_struct = acims.AmalgamatedCims(self.net_graph.get_states_number_of_all_nodes_sorted(),
self.net_graph.get_nodes(),
self.net_graph.get_ordered_by_indx_parents_values_for_all_nodes())
def parameters_estimation(self):
print("Starting computing")
t0 = tm.time()
for trajectory in self.sample_path.trajectories:
#tr_length = trajectory.size()
self.parameters_estimation_single_trajectory(trajectory.get_trajectory())
#print("Finished Trajectory number", indx)
t1 = tm.time() - t0
print("Elapsed Time ", t1)
def parameters_estimation_single_trajectory(self, trajectory):
row_length = trajectory.shape[1]
for indx, row in enumerate(trajectory[:-1]):
self.compute_sufficient_statistics_for_trajectory(trajectory.times, trajectory.actual_trajectory, trajectory.transitions, row_length)
def compute_sufficient_statistics_for_trajectory(self, times, traj_values, traj_transitions, row_length):
#time = self.compute_time_delta(current_row, next_row)
#time = current_row[0]
print(times)
print(traj_values)
print(traj_transitions)
for row in traj_transitions:
time = times[0]
for indx in range(0, row_length):
if row[indx] == 1:
which_node = indx
transition = [which_node, (traj_values[indx - 1], traj_values[indx])]
which_matrix = self.which_matrix_to_update(row, which_node)
which_element = transition[1]
self.amalgamated_cims_struct.update_state_transition_for_matrix(which_node, which_matrix, which_element)
which_element = transition[1][0]
self.amalgamated_cims_struct.update_state_residence_time_for_matrix(which_node, which_matrix,
which_element,
time)
else:
which_node = indx
which_matrix = self.which_matrix_to_update(row, which_node)
which_element = row[indx]
self.amalgamated_cims_struct.update_state_residence_time_for_matrix(
which_node, which_matrix, which_element, time)
def which_matrix_to_update(self, current_row, node_indx):
#print(type(self.fancy_indexing_structure[node_indx]))
return tuple(current_row.take(self.fancy_indexing_structure[node_indx]))
#return tuple(ParametersEstimator.taker(current_row, self.fancy_indexing_structure[node_indx]))
def parameters_estimation_for_variable_multiple_parents(self, node_indx, times, transitions ,variable_values, parents_values):
#print(times)
#print(variable_values)
#print(parents_values)
#print("Starting computing")
#t0 = tm.time()
for indx, row in enumerate(variable_values):
time = times[indx]
which_matrix = tuple(parents_values[indx]) # questo è un vettore
current_state = variable_values[indx]
if transitions[indx] == 1:
prev_state = variable_values[indx - 1]
transition = [node_indx, (prev_state, current_state)]
#which_node = transition[0]
which_element = transition[1]
self.amalgamated_cims_struct.update_state_transition_for_matrix(node_indx, which_matrix, which_element)
#which_element = current_state
self.amalgamated_cims_struct.update_state_residence_time_for_matrix(node_indx, which_matrix,
current_state,
time)
def parameters_estimation_for_variable_single_parent(self, node_indx, times, transitions, variable_values,
parents_values):
for indx, row in enumerate(variable_values):
time = times[indx]
which_matrix = parents_values[indx] # Avendo un solo parent questo è uno scalare
current_state = variable_values[indx]
#which_matrix = ParametersEstimator.taker(parents_values, indx)
# print(which_matrix.dtype)
if transitions[indx] == 1:
prev_state = variable_values[indx - 1]
transition = [node_indx, (prev_state, current_state)]
which_element = transition[1]
self.amalgamated_cims_struct.update_state_transition_for_matrix(node_indx, which_matrix,
which_element)
which_element = current_state
self.amalgamated_cims_struct.update_state_residence_time_for_matrix(node_indx, which_matrix,
which_element,time)
def parameters_estimation_for_variable_no_parent(self, node_indx, times, transitions,variable_values):
for indx, row in enumerate(variable_values):
time = times[indx]
which_matrix = 0
current_state = variable_values[indx]
"""if transitions[indx] == 1:
prev_state = variable_values[indx - 1]
#current_state = variable_values[indx]
transition = [node_indx, (prev_state, current_state)]
which_element = transition[1]
self.amalgamated_cims_struct.update_state_transition_for_matrix(node_indx, which_matrix,
which_element)"""
which_element = current_state
self.amalgamated_cims_struct.update_state_residence_time_for_matrix(node_indx, which_matrix,
which_element,
time)
def parameters_estimation_for_variable_no_parent_in_place(self, node_indx, times, transitions, variable_values):
state_trans_matrix = np.zeros(shape=(3,3), dtype=np.int)
state_res_time_array = np.zeros(shape=(3), dtype=np.float)
for indx, row in enumerate(variable_values):
time = times[indx]
#which_matrix = 0
current_state = variable_values[indx]
if transitions[indx] == 1:
prev_state = variable_values[indx - 1]
#current_state = variable_values[indx]
transition = [node_indx, (prev_state, current_state)]
which_element = transition[1]
#self.amalgamated_cims_struct.update_state_transition_for_matrix(node_indx, which_matrix,
#which_element)
state_trans_matrix[which_element] += 1
which_element = current_state
#self.amalgamated_cims_struct.update_state_residence_time_for_matrix(node_indx, which_matrix,
#which_element,
#time)
state_res_time_array[which_element] += time
def parameters_estimation_for_variable_single_parent_in_place(self, node_indx, times, transitions, variable_values,
parents_values,values_tuple):
state_res_time_dim = values_tuple[1:]
state_trans_matricies = np.zeros(shape=27, dtype=np.int)
state_res_time_array = np.zeros(shape=9, dtype=np.float)
state_transition_indx = np.array(values_tuple, dtype=np.int)
for indx, row in enumerate(variable_values):
time = times[indx]
#which_matrix = np.ravel_multi_index(parents_values[indx], ) # Avendo un solo parent questo è uno scalare
#current_state = variable_values[indx]
#which_matrix = ParametersEstimator.taker(parents_values, indx)
state_transition_indx[0] = parents_values[indx]
state_transition_indx[1] = variable_values[indx]
# print(which_matrix.dtype)
if transitions[indx] == 1:
state_transition_indx[2] = variable_values[indx - 1]
#transition = [node_indx, (prev_state, current_state)]
#which_element = transition[1]
#self.amalgamated_cims_struct.update_state_transition_for_matrix(node_indx, which_matrix,
#which_element)
scalar_indx = np.ravel_multi_index(state_transition_indx, values_tuple)
print("State Transition", scalar_indx)
state_trans_matricies[scalar_indx] += 1
scalar_indx = np.ravel_multi_index(state_transition_indx[:-1], state_res_time_dim)
print("Res Time",scalar_indx)
state_res_time_array[scalar_indx] += time
#which_element = current_state
#self.amalgamated_cims_struct.update_state_residence_time_for_matrix(node_indx, which_matrix,
#which_element,time)
#t1 = tm.time() - t0
#print("Elapsed Time ", t1)
def compute_parameters(self):
for node_indx, set_of_cims in enumerate(self.amalgamated_cims_struct.sets_of_cims):
self.compute_state_res_time_for_node(node_indx, self.sample_path.trajectories[0].get_times(),
self.sample_path.trajectories[0].get_trajectory(),
self.columns_filtering_structure[node_indx],
self.scalar_indexes_converter[node_indx],
set_of_cims.state_residence_times)
self.compute_state_transitions_for_a_node(node_indx,
self.sample_path.trajectories[0].get_complete_trajectory(),
self.transition_filtering[node_indx],
self.transition_scalar_index_converter[node_indx],
set_of_cims.transition_matrices)
set_of_cims.build_cims(set_of_cims.state_residence_times, set_of_cims.transition_matrices)
def compute_state_res_time_for_node(self, node_indx, times, trajectory, cols_filter, scalar_indexes_struct, T):
#print(times)
#print(trajectory)
#print(cols_filter)
#print(scalar_indexes_struct)
#print(T)
T[:] = np.bincount(np.sum(trajectory[:, cols_filter] * scalar_indexes_struct / scalar_indexes_struct[0], axis=1)
.astype(np.int), \
times,
minlength=scalar_indexes_struct[-1]).reshape(-1, T.shape[1])
#print("Done This NODE", T)
def compute_state_residence_time_for_all_nodes(self):
for node_indx, set_of_cims in enumerate(self.amalgamated_cims_struct.sets_of_cims):
self.compute_state_res_time_for_node(node_indx, self.sample_path.trajectories[0].get_times(),
self.sample_path.trajectories[0].get_trajectory(), self.columns_filtering_structure[node_indx],
self.scalar_indexes_converter[node_indx], set_of_cims.state_residence_times)
def compute_state_transitions_for_a_node(self, node_indx, trajectory, cols_filter, scalar_indexing, M):
#print(node_indx)
#print(trajectory)
#print(cols_filter)
#print(scalar_indexing)
#print(M)
diag_indices = np.array([x * M.shape[1] + x % M.shape[1] for x in range(M.shape[0] * M.shape[1])],
dtype=np.int64)
trj_tmp = trajectory[trajectory[:, int(trajectory.shape[1] / 2) + node_indx].astype(np.int) >= 0]
#print(trj_tmp)
#print("Summing", np.sum(trj_tmp[:, cols_filter] * scalar_indexing / scalar_indexing[0], axis=1).astype(np.int))
#print(M.shape[1])
#print(M.shape[2])
M[:] = np.bincount(np.sum(trj_tmp[:, cols_filter] * scalar_indexing / scalar_indexing[0], axis=1).astype(np.int),
minlength=scalar_indexing[-1]).reshape(-1, M.shape[1], M.shape[2])
M_raveled = M.ravel()
M_raveled[diag_indices] = 0
#print(M_raveled)
M_raveled[diag_indices] = np.sum(M, axis=2).ravel()
#print(M_raveled)
#print(M)
def compute_state_transitions_for_all_nodes(self):
for node_indx, set_of_cims in enumerate(self.amalgamated_cims_struct.sets_of_cims):
self.compute_state_transitions_for_a_node(node_indx, self.sample_path.trajectories[0].get_complete_trajectory(),
self.transition_filtering[node_indx],
self.transition_scalar_index_converter[node_indx], set_of_cims.transition_matrices)
# Simple Test #
os.getcwd()
os.chdir('..')
path = os.getcwd() + '/data'
s1 = sp.SamplePath(path)
s1.build_trajectories()
s1.build_structure()
g1 = ng.NetworkGraph(s1.structure)
g1.init_graph()
pe = ParametersEstimator(s1, g1)
pe.init_amalgamated_cims_struct()
#print(pe.amalgamated_cims_struct.get_set_of_cims(0).get_cims_number())
#print(pe.amalgamated_cims_struct.get_set_of_cims(1).get_cims_number())
#print(pe.amalgamated_cims_struct.get_set_of_cims(2).get_cims_number())
#print(np.shape(s1.trajectories[0].transitions)[0])
#print(pe.columns_filtering_structure)
#print(pe.scalar_indexes_converter)
#print(pe.amalgamated_cims_struct.sets_of_cims[1].state_residence_times)
#print(pe.amalgamated_cims_struct.sets_of_cims[2].state_residence_times)
#print(pe.amalgamated_cims_struct.sets_of_cims[2].transition_matrices)
#print(pe.amalgamated_cims_struct.sets_of_cims[1].transition_matrices)
#print(pe.amalgamated_cims_struct.sets_of_cims[0].transition_matrices)
#pe.compute_state_transitions_for_all_nodes()
lp = LineProfiler()
"""pe.compute_state_residence_time_for_all_nodes()
#pe.parameters_estimation_for_variable(0, pe.sample_path.trajectories[0].get_trajectory()[:, 0],
# pe.sample_path.trajectories[0].get_trajectory()[:, 1], [])
#pe.parameters_estimation_single_trajectory(pe.sample_path.trajectories[0].get_trajectory())
#pe.parameters_estimation()
#lp.add_function(pe.compute_sufficient_statistics_for_row) # add additional function to profile
#lp_wrapper = lp(pe.parameters_estimation_single_trajectory)
#lp_wrapper = lp(pe.parameters_estimation)
#lp_wrapper(pe.sample_path.trajectories[0].get_trajectory())
#lp.print_stats()
#lp_wrapper = lp(pe.parameters_estimation_for_variable)
#lp_wrapper(2, pe.sample_path.trajectories[0].get_times(),
#pe.sample_path.trajectories[0].get_trajectory()[:, 2],
#pe.sample_path.trajectories[0].get_trajectory()[:, [0,1]])
lp_wrapper = lp(pe.parameters_estimation_for_variable_single_parent)
lp_wrapper(1, pe.sample_path.trajectories[0].get_times(),
pe.sample_path.trajectories[0].get_trajectory()[:, 1],
pe.sample_path.trajectories[0].get_trajectory()[:, 2])
lp.print_stats()
#print( pe.sample_path.trajectories[0].get_trajectory()[:, [1,2]])
for matrix in pe.amalgamated_cims_struct.get_set_of_cims(1).actual_cims:
print(matrix.state_residence_times)
print(matrix.state_transition_matrix)
matrix.compute_cim_coefficients()
print(matrix.cim)"""
"""lp_wrapper = lp(pe.parameters_estimation_for_variable_no_parent_in_place)
lp_wrapper(0, pe.sample_path.trajectories[0].get_times(), pe.sample_path.trajectories[0].transitions[:, 0],
pe.sample_path.trajectories[0].get_trajectory()[:, 0] )
lp.print_stats()
lp_wrapper = lp(pe.parameters_estimation_for_variable_single_parent)
lp_wrapper(1, pe.sample_path.trajectories[0].get_times(), pe.sample_path.trajectories[0].transitions[:, 1],
pe.sample_path.trajectories[0].get_trajectory()[:,1], pe.sample_path.trajectories[0].get_trajectory()[:,2] )
lp.print_stats()
lp_wrapper = lp(pe.parameters_estimation_for_variable_multiple_parents)
lp_wrapper(2, pe.sample_path.trajectories[0].get_times(), pe.sample_path.trajectories[0].transitions[:, 2],
pe.sample_path.trajectories[0].get_trajectory()[:,2], pe.sample_path.trajectories[0].get_trajectory()[:, [0,1]] )
lp.print_stats()"""
"""lp_wrapper = lp(pe.parameters_estimation_for_variable_single_parent_in_place)
lp_wrapper(1, pe.sample_path.trajectories[0].get_times(), pe.sample_path.trajectories[0].transitions[:, 1],
pe.sample_path.trajectories[0].get_trajectory()[:,1], pe.sample_path.trajectories[0].get_trajectory()[:,2], (3,3,3) )
lp.print_stats()"""
"""lp_wrapper = lp(pe.compute_sufficient_statistics_for_trajectory)
lp_wrapper(pe.sample_path.trajectories[0].get_times(), pe.sample_path.trajectories[0].actual_trajectory,
pe.sample_path.trajectories[0].transitions, 3)
lp.print_stats()
lp_wrapper = lp(pe.compute_state_res_time_for_node)
lp_wrapper(0, pe.sample_path.trajectories[0].get_times(),
pe.sample_path.trajectories[0].actual_trajectory, [0], [3], np.zeros([3,3], dtype=np.float))
lp.print_stats()
#pe.compute_state_res_time_for_node(0, pe.sample_path.trajectories[0].get_times(),
#pe.sample_path.trajectories[0].actual_trajectory, [0], [3], np.zeros([3,3], dtype=np.float))"""
"""[[2999.2966 2749.2298 3301.5975]
[3797.1737 3187.8345 2939.2009]
[3432.224 3062.5402 4530.9028]]
[[ 827.6058 838.1515 686.1365]
[1426.384 2225.2093 1999.8528]
[ 745.3068 733.8129 746.2347]
[ 520.8113 690.9502 853.4022]
[1590.8609 1853.0021 1554.1874]
[ 637.5576 643.8822 654.9506]
[ 718.7632 742.2117 998.5844]
[1811.984 1598.0304 2547.988 ]
[ 770.8503 598.9588 984.3304]]
lp_wrapper = lp(pe.compute_state_residence_time_for_all_nodes)
lp_wrapper()
lp.print_stats()
#pe.compute_state_residence_time_for_all_nodes()
print(pe.amalgamated_cims_struct.sets_of_cims[0].state_residence_times)
[[[14472, 3552, 10920],
[12230, 25307, 13077],
[ 9707, 14408, 24115]],
[[22918, 6426, 16492],
[10608, 16072, 5464],
[10746, 11213, 21959]],
[[23305, 6816, 16489],
[ 3792, 19190, 15398],
[13718, 18243, 31961]]])
Raveled [14472 3552 10920 12230 25307 13077 9707 14408 24115 22918 6426 16492
10608 16072 5464 10746 11213 21959 23305 6816 16489 3792 19190 15398
13718 18243 31961]"""
lp_wrapper = lp(pe.compute_parameters)
lp_wrapper()
#for variable in pe.amalgamated_cims_struct.sets_of_cims:
#for cond in variable.get_cims():
#print(cond.cim)
print(pe.amalgamated_cims_struct.get_cims_of_node(1,[2]))
lp.print_stats()