def run(model, data, parallel=True):
num_subparts = data["data"].shape[
0] # mmm the first dimension of data represents each subpart?? interesting.
num_resources = len(model["learns"])
result = E_step.run(data, model, 1, int(parallel))
for j in range(num_resources):
result['all_trans_softcounts'][j] = result['all_trans_softcounts'][
j].transpose()
for j in range(num_subparts):
result['all_emission_softcounts'][j] = result[
'all_emission_softcounts'][j].transpose()
state_predictions = predict_onestep_states.run(data, model,
result['alpha'],
int(parallel))
p = state_predictions.shape
state_predictions = state_predictions.flatten(order='C').reshape(p,
order='F')
# multiguess solution, should work
correct_emission_predictions = np.expand_dims(
model["guesses"], axis=1) @ np.expand_dims(
state_predictions[0, :], axis=0) + np.expand_dims(
1 - model["slips"], axis=1) @ np.expand_dims(
state_predictions[1, :], axis=0)
#correct_emission_predictions = model['guesses'] * np.asarray([state_predictions[0,:]]).T + (1 - model['slips']) * np.asarray([state_predictions[1,:]]).T
flattened_predictions = np.take_along_axis(
correct_emission_predictions,
(data['data'] != 0).argmax(axis=0)[:, None].T,
axis=0)
return (flattened_predictions.ravel(), state_predictions)
def EM_fit(model, data, tol=None, maxiter=None, parallel=True):
if tol is None:
tol = 1e-3
if maxiter is None:
maxiter = 100
num_subparts = data["data"].shape[
0] #mmm the first dimension of data represents each subpart?? interesting.
num_resources = len(model["learns"])
log_likelihoods = np.zeros((maxiter, 1))
for i in range(maxiter):
result = E_step.run(data, model, 1, int(parallel))
for j in range(num_resources):
result['all_trans_softcounts'][j] = result['all_trans_softcounts'][
j].transpose()
for j in range(num_subparts):
result['all_emission_softcounts'][j] = result[
'all_emission_softcounts'][j].transpose()
log_likelihoods[i][0] = result['total_loglike']
if (i > 1 and
abs(log_likelihoods[i][0] - log_likelihoods[i - 1][0]) < tol):
break
model = M_step.run(model, result['all_trans_softcounts'],
result['all_emission_softcounts'],
result['all_initial_softcounts'])
return (model, log_likelihoods[:i + 1])
def run(model, data):
num_subparts = data["data"].shape[
0] # mmm the first dimension of data represents each subpart?? interesting.
num_resources = len(model["learns"])
trans_softcounts = np.zeros((num_resources, 2, 2))
emission_softcounts = np.zeros((num_subparts, 2, 2))
init_softcounts = np.zeros((2, 1))
result = {}
result['all_trans_softcounts'] = trans_softcounts
result['all_emission_softcounts'] = emission_softcounts
result['all_initial_softcounts'] = init_softcounts
result = E_step.run(data, model, result['all_trans_softcounts'],
result['all_emission_softcounts'],
result['all_initial_softcounts'], 1)
for j in range(num_resources):
result['all_trans_softcounts'][j] = result['all_trans_softcounts'][
j].transpose()
for j in range(num_subparts):
result['all_emission_softcounts'][j] = result[
'all_emission_softcounts'][j].transpose()
state_predictions = predict_onestep_states.run(data, model,
result['alpha'])
correct_emission_predictions = []
for i in range(len(model["guesses"])):
correct_emission_predictions.append(
model["guesses"][i] * state_predictions[0, :] +
(1 - model["slips"][i]) * state_predictions[1, :])
return (correct_emission_predictions, state_predictions)
def run(model, data):
num_subparts = data["data"].shape[0] # mmm the first dimension of data represents each subpart?? interesting.
num_resources = len(model["learns"])
trans_softcounts = np.zeros((num_resources, 2, 2))
emission_softcounts = np.zeros((num_subparts, 2, 2))
init_softcounts = np.zeros((2, 1))
result = {}
result['all_trans_softcounts'] = trans_softcounts
result['all_emission_softcounts'] = emission_softcounts
result['all_initial_softcounts'] = init_softcounts
result = E_step.run(data, model, result['all_trans_softcounts'], result['all_emission_softcounts'], result['all_initial_softcounts'], 1)
for j in range(num_resources):
result['all_trans_softcounts'][j] = result['all_trans_softcounts'][j].transpose()
for j in range(num_subparts):
result['all_emission_softcounts'][j] = result['all_emission_softcounts'][j].transpose()
state_predictions = predict_onestep_states.run(data, model, result['alpha'])
p = state_predictions.shape
state_predictions = state_predictions.flatten(order = 'C').reshape(p, order = 'F')
# multiguess solution, should work
correct_emission_predictions = np.expand_dims(model["guesses"], axis = 1) @ np.expand_dims(state_predictions[0,:], axis = 0) + np.expand_dims(1-model["slips"], axis = 1) @ np.expand_dims(state_predictions[1,:], axis = 0)
#correct_emission_predictions = model['guesses'] * np.asarray([state_predictions[0,:]]).T + (1 - model['slips']) * np.asarray([state_predictions[1,:]]).T
flattened_predictions = np.zeros((len(correct_emission_predictions[0]),))
for i in range(len(correct_emission_predictions)):
for j in range(len(correct_emission_predictions[0])):
if data["data"][i][j] != 0:
flattened_predictions[j] = correct_emission_predictions[i][j]
return (flattened_predictions, state_predictions)
def EM_fit(model, data, tol = None, maxiter = None):
if tol is None: tol = 1e-3
if maxiter is None: maxiter = 100
check_data.check_data(data)
#print(data["data"])
#print(data["data"].shape)
num_subparts = data["data"].shape[0] #mmm the first dimension of data represents each subpart?? interesting.
num_resources = len(model["learns"])
trans_softcounts = np.zeros((num_resources,2,2))
emission_softcounts = np.zeros((num_subparts,2,2))
init_softcounts = np.zeros((2, 1))
log_likelihoods = np.zeros((maxiter, 1))
result = {}
result['all_trans_softcounts'] = trans_softcounts
result['all_emission_softcounts'] = emission_softcounts
result['all_initial_softcounts'] = init_softcounts
#print(result)
#data["data"] = np.asarray(data["data"], dtype='int32')
#print(data)
for i in range(maxiter):
result = E_step.run(data, model, result['all_trans_softcounts'], result['all_emission_softcounts'], result['all_initial_softcounts'], 1)
for j in range(num_resources):
result['all_trans_softcounts'][j] = result['all_trans_softcounts'][j].transpose()
for j in range(num_subparts):
result['all_emission_softcounts'][j] = result['all_emission_softcounts'][j].transpose()
log_likelihoods[i][0] = result['total_loglike']
if(i > 1 and abs(log_likelihoods[i][0] - log_likelihoods[i-1][0]) < tol):
break
model = M_step.run(model, result['all_trans_softcounts'], result['all_emission_softcounts'], result['all_initial_softcounts'])
return(model, log_likelihoods[:i+1])