def model_choice(models, obs):
k = [i for i in xrange(2, 9)]
Statistics = []
for ki in k:
print 'K = ', ki
num_M = models[ki - 2].shape[0]
print 'Num Models: ', num_M
numNbins = len(obs[ki - 2])
numHbins = len(obs[ki - 2][0])
M = theano.shared(
np.asarray(models[ki - 2], dtype=theano.config.floatX))
ObSym = T.matrix(
) # Symbolic tensor for observation batches - indexed elements of Obs shared variable are passed through this
Pred = theano.function(
[], predictiveness_profiles(M, ki, len(
models[ki - 2])))() # This should be dealt with better too...
Pred_n = Pred
Pred = theano.shared(np.asarray(Pred, dtype=theano.config.floatX))
# setup inference schemas and theano symbolic tensors
if INFERENCE == 'underfit':
profiles = make_agression_profiles(num_profiles, num_alpha)
#alpha = theano.shared(np.asmatrix(np.linspace(0.0,1.0, num = num_alpha, endpoint = False), dtype=theano.config.floatX))
Alpha = T.arange(0., 1.0, 1. / num_alpha)
Agression_profiles = T.matrix('Agr')
nAlpha, nM, nO = T.iscalars('', '', '')
Choice_Maker = Underfit_Choice(M, ObSym, nM, nO, ki, nAlpha, Alpha,
Agression_profiles, Pred,
pValue_alg) #only works for 0...
elif INFERENCE == 'bayes':
profiles = make_priors_profiles(num_priors, num_M)
Priors_profiles = T.matrix('Priors')
Loss_funcs = T.arange(
1, 5
) # Loss functions are choices in bayesian_choice numbered [1,4]
nM, nO = T.iscalars('', '')
Choice_Maker = Bayesian_Choice(M, ObSym, nM, nO, ki,
Priors_profiles, Loss_funcs)
else:
print 'unknown inference algorithm...'
quit()
# all data for this K
k_Data = kData(numNbins, numHbins, num_profiles)
for i in xrange(numNbins):
for j in xrange(numHbins):
print 'bin ', i, j
t0 = time.time()
if obs[ki - 2][i][j] == [] or obs[ki -
2][i][j][0].shape[1] == 0:
#there are no observtions in this N*H bin...
continue
else:
num_obs = obs[ki - 2][i][j][0].shape[0]
# allocate for predictiveness of model choice vs universe for each obs for each profile
k_pred = kPred(num_obs, num_profiles)
num_batches = int(np.ceil(num_obs / np.float(BATCH_SIZE)))
for batch_index in xrange(num_batches):
top = BATCH_SIZE * (batch_index + 1) if batch_index < (
num_batches - 1) else num_obs
n_obs = top - BATCH_SIZE * (batch_index)
print 'batch index ', batch_index, '\t num obs: ', top - BATCH_SIZE * batch_index
if INFERENCE == 'underfit':
batch_choice = Choice_Maker.Choice_Profile_F(
profiles, num_alpha, num_M, n_obs,
obs[ki - 2][i][j][0][BATCH_SIZE * batch_index:top])
print batch_choice
for prof in xrange(num_profiles):
k_pred[prof][BATCH_SIZE * (
batch_index):top] = get_predictiveness_array(
batch_choice[prof], obs[ki - 2][i][j][1],
Pred_n, n_obs)
elif INFERENCE == 'bayes':
batch_choice = Choice_Maker.Choice_Profile_F(
profiles, num_M, n_obs,
obs[ki - 2][i][j][0][BATCH_SIZE * batch_index:top])
print batch_choice
for pr in xrange(num_priors):
for lf in xrange(num_loss_funcs):
k_pred[pr * num_loss_funcs +
lf][BATCH_SIZE *
(batch_index
):top] = get_predictiveness_array(
batch_choice[pr][lf],
obs[ki - 2][i][j][1], Pred_n,
n_obs)
for prof in xrange(num_profiles):
pred_moments = get_moments(k_pred[prof], num_obs)
for m in xrange(len(pred_moments)):
k_Data[prof][m][i, j] = pred_moments[m]
t1 = time.time()
print 'single bin takes: ', (t1 - t0) / 60., ' minutes'
Statistics.append(k_Data)
f = open('%s_k%d.pkl' % (name, ki), 'wb')
pickle.dump(k_Data, f)
f.close()
return Statistics
def model_choice(models, obs):
k = [i for i in xrange(2, 9)]
Statistics = []
for ki in k:
print 'K = ', ki
num_M = models[ki-2].shape[0]
print 'Num Models: ', num_M
numNbins = len(obs[ki-2])
numHbins = len(obs[ki-2][0])
M = theano.shared(np.asarray(models[ki-2], dtype = theano.config.floatX))
ObSym = T.matrix() # Symbolic tensor for observation batches - indexed elements of Obs shared variable are passed through this
Pred = theano.function([], predictiveness_profiles(M, ki, len(models[ki-2])))() # This should be dealt with better too...
Pred_n = Pred
Pred = theano.shared(np.asarray(Pred, dtype = theano.config.floatX))
# setup inference schemas and theano symbolic tensors
if INFERENCE == 'underfit':
profiles = make_agression_profiles(num_profiles, num_alpha)
#alpha = theano.shared(np.asmatrix(np.linspace(0.0,1.0, num = num_alpha, endpoint = False), dtype=theano.config.floatX))
Alpha = T.arange(0., 1.0, 1./num_alpha)
Agression_profiles = T.matrix('Agr')
nAlpha, nM, nO = T.iscalars('','','')
Choice_Maker = Underfit_Choice(M, ObSym, nM, nO, ki, nAlpha, Alpha, Agression_profiles, Pred, pValue_alg) #only works for 0...
elif INFERENCE == 'bayes':
profiles = make_priors_profiles(num_priors, num_M)
Priors_profiles = T.matrix('Priors')
Loss_funcs = T.arange(1,5) # Loss functions are choices in bayesian_choice numbered [1,4]
nM, nO = T.iscalars('','')
Choice_Maker = Bayesian_Choice(M, ObSym, nM, nO, ki, Priors_profiles, Loss_funcs)
else:
print 'unknown inference algorithm...'
quit()
# all data for this K
k_Data = kData(numNbins, numHbins, num_profiles)
for i in xrange(numNbins):
for j in xrange(numHbins):
print 'bin ', i, j
t0 = time.time()
if obs[ki-2][i][j] == [] or obs[ki-2][i][j][0].shape[1] == 0:
#there are no observtions in this N*H bin...
continue
else:
num_obs = obs[ki-2][i][j][0].shape[0]
# allocate for predictiveness of model choice vs universe for each obs for each profile
k_pred = kPred(num_obs, num_profiles)
num_batches = int(np.ceil(num_obs/np.float(BATCH_SIZE)))
for batch_index in xrange(num_batches):
top = BATCH_SIZE*(batch_index+1) if batch_index < (num_batches-1) else num_obs
n_obs = top - BATCH_SIZE*(batch_index)
print 'batch index ', batch_index, '\t num obs: ', top - BATCH_SIZE*batch_index
if INFERENCE == 'underfit':
batch_choice = Choice_Maker.Choice_Profile_F(profiles, num_alpha, num_M, n_obs, obs[ki-2][i][j][0][BATCH_SIZE*batch_index:top])
print batch_choice
for prof in xrange(num_profiles):
k_pred[prof][BATCH_SIZE*(batch_index):top] = get_predictiveness_array(batch_choice[prof], obs[ki-2][i][j][1], Pred_n, n_obs)
elif INFERENCE == 'bayes':
batch_choice = Choice_Maker.Choice_Profile_F(profiles, num_M, n_obs, obs[ki-2][i][j][0][BATCH_SIZE*batch_index:top])
print batch_choice
for pr in xrange(num_priors):
for lf in xrange(num_loss_funcs):
k_pred[pr*num_loss_funcs + lf][BATCH_SIZE*(batch_index):top] = get_predictiveness_array(batch_choice[pr][lf], obs[ki-2][i][j][1], Pred_n, n_obs)
for prof in xrange(num_profiles):
pred_moments = get_moments(k_pred[prof], num_obs)
for m in xrange(len(pred_moments)):
k_Data[prof][m][i,j] = pred_moments[m]
t1 = time.time()
print 'single bin takes: ',(t1-t0)/60., ' minutes'
Statistics.append(k_Data)
f = open('%s_k%d.pkl'%(name, ki), 'wb')
pickle.dump(k_Data, f)
f.close()
return Statistics
def model_choice(models, obs):
k = [i for i in xrange(2, 9)]
num_alpha = 10
Choices = []
Statistics = []
Agression_profiles = []
Agression_profiles.append(theano.shared(np.ones((num_alpha, 1), dtype = theano.config.floatX)/num_alpha)) #uniform
Agression_profiles.append(theano.shared(np.array([[i*2./num_alpha] for i in xrange(num_alpha)], dtype = theano.config.floatX))) #agressive
Agression_profiles.append(theano.shared(np.array([[2-i*2./num_alpha] for i in xrange(num_alpha)], dtype = theano.config.floatX))) #cautious
for ki in k:
Choices.append([])
Statistics.append([])
print 'K = ', ki
#Agression = theano.shared(np.ones((num_alpha, 1), dtype = theano.config.floatX)/num_alpha)
num_M = models[ki-2].shape[0]
M = theano.shared(np.asarray(models[ki-2], dtype = theano.config.floatX))
result = predictiveness_profiles(M, ki, len(models[ki-2]))
t0=time.time()
Pred = theano.function([], result)()
t1=time.time()
print 'pred took ', t1-t0, ' s'
numNbins = len(obs[ki-2])
numHbins = len(obs[ki-2][0])
for i in xrange(numNbins):
#Choices[ki-2].append([])
Statistics[ki-2].append([])
for j in xrange(numHbins):
print 'bin ', i, j
t0 = time.time()
if obs[ki-2][i][j] == [] or obs[ki-2][i][j][0].shape[1] == 0:
#there are no observtions in this N*H bin...
#Choices[ki-2][i].append([])
Statistics[ki-2][i].append([])
continue
else:
num_obs = obs[ki-2][i][j][0].shape[0]
print 'num obs ', num_obs
if num_obs < 1000:
Obs = theano.shared(np.asarray(obs[ki-2][i][j][0], dtype = theano.config.floatX))
choice_bin = call_underfit_choice_theano(M, Obs, num_M, num_obs, ki, num_alpha, Agression_profiles, Pred)
else:
Obs = theano.shared(np.asarray(obs[ki-2][i][j][0][0:1000], dtype = theano.config.floatX))
n_obs = len(Obs.get_value())
print n_obs
choice_bin = call_underfit_choice_theano(M, Obs, num_M, n_obs, ki, num_alpha, Agression_profiles, Pred)
for batch_index in xrange(int(np.floor(num_obs/1000))):
top = 1000*(batch_index+2) if batch_index < (int(np.floor(num_obs/1000))-1) else len(obs[ki-2][i][j][0])
Obs = theano.shared(np.asarray(obs[ki-2][i][j][0][1000*(batch_index+1):top]))
n_obs = len(Obs.get_value())
print n_obs
batch_choice = call_underfit_choice_theano(M, Obs, num_M, n_obs, ki, num_alpha, Agression_profiles, Pred)
for ag in xrange(len(Agression_profiles)):
choice_bin[ag] = np.asarray(choice_bin[ag].tolist() + batch_choice[ag].tolist()) #appends batches for each agression function
t1=time.time()
print choice_bin#, obs[ki-2][i][j][1]
print 'choices took ', t1-t0, ' s'
#Choices[ki-2][i].append([choice_bin, obs[ki-2][i][j][1]])
agg_bin = []
for a in xrange(len(Agression_profiles)):
choice_pred = get_predictiveness_array(choices = choice_bin[a], verses = obs[ki-2][i][j][1], Preds = Pred, num_Obs = num_obs)
pred_moments = get_moments(choice_pred, num_obs)
agg_bin.append(pred_moments)
Statistics[ki-2][i].append(agg_bin)
t1 = time.time()
print 'single bin takes: ',(t1-t0)/60., ' minutes'
#print choice_bin
f = open('k%d_%d_chi_statistics.pkl'%(ki, i), 'wb')
pickle.dump(Statistics[ki-2][i], f)
f.close()
f = open('k%d_chi_statistics.pkl'%ki, 'wb')
pickle.dump(Statistics[ki-2],f)
f.close()
return Choices
def model_choice(models, obs):
k = [i for i in xrange(2, 9)]
num_alpha = 10
Choices = []
Statistics = []
Agression_profiles = []
Agression_profiles.append(
theano.shared(
np.ones((num_alpha, 1), dtype=theano.config.floatX) /
num_alpha)) #uniform
Agression_profiles.append(
theano.shared(
np.array([[i * 2. / num_alpha] for i in xrange(num_alpha)],
dtype=theano.config.floatX))) #agressive
Agression_profiles.append(
theano.shared(
np.array([[2 - i * 2. / num_alpha] for i in xrange(num_alpha)],
dtype=theano.config.floatX))) #cautious
for ki in k:
Choices.append([])
Statistics.append([])
print 'K = ', ki
#Agression = theano.shared(np.ones((num_alpha, 1), dtype = theano.config.floatX)/num_alpha)
num_M = models[ki - 2].shape[0]
M = theano.shared(
np.asarray(models[ki - 2], dtype=theano.config.floatX))
result = predictiveness_profiles(M, ki, len(models[ki - 2]))
t0 = time.time()
Pred = theano.function([], result)()
t1 = time.time()
print 'pred took ', t1 - t0, ' s'
numNbins = len(obs[ki - 2])
numHbins = len(obs[ki - 2][0])
for i in xrange(numNbins):
#Choices[ki-2].append([])
Statistics[ki - 2].append([])
for j in xrange(numHbins):
print 'bin ', i, j
t0 = time.time()
if obs[ki - 2][i][j] == [] or obs[ki -
2][i][j][0].shape[1] == 0:
#there are no observtions in this N*H bin...
#Choices[ki-2][i].append([])
Statistics[ki - 2][i].append([])
continue
else:
num_obs = obs[ki - 2][i][j][0].shape[0]
print 'num obs ', num_obs
if num_obs < 1000:
Obs = theano.shared(
np.asarray(obs[ki - 2][i][j][0],
dtype=theano.config.floatX))
choice_bin = call_underfit_choice_theano(
M, Obs, num_M, num_obs, ki, num_alpha,
Agression_profiles, Pred)
else:
Obs = theano.shared(
np.asarray(obs[ki - 2][i][j][0][0:1000],
dtype=theano.config.floatX))
n_obs = len(Obs.get_value())
print n_obs
choice_bin = call_underfit_choice_theano(
M, Obs, num_M, n_obs, ki, num_alpha,
Agression_profiles, Pred)
for batch_index in xrange(int(np.floor(num_obs / 1000))):
top = 1000 * (batch_index + 2) if batch_index < (
int(np.floor(num_obs / 1000)) - 1) else len(
obs[ki - 2][i][j][0])
Obs = theano.shared(
np.asarray(obs[ki -
2][i][j][0][1000 *
(batch_index + 1):top]))
n_obs = len(Obs.get_value())
print n_obs
batch_choice = call_underfit_choice_theano(
M, Obs, num_M, n_obs, ki, num_alpha,
Agression_profiles, Pred)
for ag in xrange(len(Agression_profiles)):
choice_bin[ag] = np.asarray(
choice_bin[ag].tolist() +
batch_choice[ag].tolist()
) #appends batches for each agression function
t1 = time.time()
print choice_bin #, obs[ki-2][i][j][1]
print 'choices took ', t1 - t0, ' s'
#Choices[ki-2][i].append([choice_bin, obs[ki-2][i][j][1]])
agg_bin = []
for a in xrange(len(Agression_profiles)):
choice_pred = get_predictiveness_array(
choices=choice_bin[a],
verses=obs[ki - 2][i][j][1],
Preds=Pred,
num_Obs=num_obs)
pred_moments = get_moments(choice_pred, num_obs)
agg_bin.append(pred_moments)
Statistics[ki - 2][i].append(agg_bin)
t1 = time.time()
print 'single bin takes: ', (t1 - t0) / 60., ' minutes'
#print choice_bin
f = open('k%d_%d_chi_statistics.pkl' % (ki, i), 'wb')
pickle.dump(Statistics[ki - 2][i], f)
f.close()
f = open('k%d_chi_statistics.pkl' % ki, 'wb')
pickle.dump(Statistics[ki - 2], f)
f.close()
return Choices
