data_type = 'debug-four'
flag = 'four'
#full_data,Z_gen = generate_data(feature_count,data_count + held_out,T,sig,data_type)
full_data, Z_gen = construct_data(data_dim, data_count + held_out, sig,
data_type)
Y = full_data[:data_count, :]
held_out = full_data[data_count:, :]
#we observe half the signal and recover the other half
observe = held_out[:, :T / 2]
trunc = 12 #truncate active features
if algorithm == 'IBP':
iterate = 200
alpha = 2.0
select = 12
Z, W, ll_set, pred_ll = ugibbs_sampler(Y, held_out, alpha, sig_test, sig_w,
iterate, select, trunc, data_dim)
Z_trunc, W_trunc = truncate(Z, W, select)
print_posterior(Z_trunc, W_trunc, data_dim)
recover_ll = recover_IBP(held_out, observe, Z, W, sig_test)
print("Log Recovery")
print(recover_ll)
print("End Uncollapsed IBP")
if algorithm == 'paintbox':
log_res = 10 #log of res
hold = 100 #hold resolution for # iterations
#feature_count = 4 #features
#feature deviation #it was 0.3 for random #0.2 for corr
iterate = 1000
K = 3 #start with K features
ext = 1 #draw one new feature per iteration
trunc = 12 #truncate active features
log_res = 10 #log of res
hold = 100 #hold resolution for # iterations
initialize = True
if initialize:
alpha = 2.0
pre_trunc = 12
#dummy variable
select = 10
Z_init, W_init, _, _, rec_ll, iter_time = ugibbs_sampler(train,
test,
alpha,
sig_test,
sig_w,
init_iter,
select,
pre_trunc,
observe,
obs_indices,
limit,
init=initialize,
display=display)
zip_list = zip(iter_time, rec_ll)
print('IBP')
for z in zip_list:
print(z)
K = Z_init.shape[1]
else:
K = 1 #start with K features
ext = 0 #draw one new feature per iteration
outputs = upaintbox_sample(log_res,
feature = np.zeros((runs,iterate))
lapse_data = np.zeros((runs,iterate))
#pred_data = np.zeros((runs,iterate))
valid = 0
algorithm = 'paintbox'
truncate = 12
#algorithm = 'var'
#algorithm = 'uncollapsed'
#you'll come back to this
while valid < runs:
#print("Trial Number: " + str(valid))
if algorithm == 'paintbox':
#profile.run('upaintbox_sample(log_res,hold,Y,held_out,ext,sig_alg,sig_w,iterate,K,valid)')
ll_list,iter_time,f_count,lapse,Z,W,prob_matrix,pred_ll,tree = upaintbox_sample(data_dim,log_res,hold,Y,held_out,ext,sig,sig_w,iterate,K,truncate)
if algorithm == 'uncollapsed':
Z,W,ll_set = ugibbs_sampler(Y,alpha,sig,sig_w,iterate,Z_gen)
if algorithm == 'var':
nu_set,phi_set,Phi_set,tau_set,pred_ll = run_vi(Y,held_out,alpha,sig_w,sig,iterate,feature_count)
if len(ll_list) == 0:
continue
ll_data[valid,:] = ll_list
ll_time[valid,:] = iter_time
feature[valid,:] = f_count
lapse_data[valid,:] = lapse
#pred_data[valid,:] = pred_ll
valid += 1
# plot_feature = [num for num in range(1,int(np.max(feature[:,iterate-1])+1))]
# plot_lapse = [0 for i in range(len(plot_feature))]
# all_sums = [0 for i in range(len(plot_feature))]
# for i in range(len(plot_feature)):
obs_indices = [i for i in range(int(dimension*obs))]
#print(obs_indices)
observe = test[:,obs_indices]
#print(observe.shape)
if algorithm == 'IBP':
trunc = 12
iterate = 800
alpha = 2.0
#dummy variable
#kept around because we may have to truncate at future date
select = 10
#print("test shape")
#print(test.shape)
Z,W,ll_set,pred_ll,rec_ll,iter_time = ugibbs_sampler(train,test,alpha,sig_test,
sig_w,iterate,select,trunc,
observe,obs_indices,limit)
zip_list = zip(iter_time,rec_ll)
for z in zip_list:
print(z)
#Z_trunc,W_trunc = truncate(Z,W,select)
#print_posterior(Z_trunc,W_trunc,data_dim)
#recover_ll = recover_IBP(test,observe,Z,W,sig_test,obs_indices)
#print("Log Recovery")
#print(recover_ll)
#print("End Uncollapsed IBP")
#np.savetxt("rec" + process + ".txt", rec_ll)
#np.savetxt("time" + process + ".txt", iter_time)
if algorithm == 'paintbox':
trunc = 12 #truncate active features