def _fit(self, data, skill, forgets, preload=False):
""" Helper function for fitting data. """
num_learns = len(data["resource_names"])
num_gs = len(data["gs_names"])
self._check_manual_param_init(num_learns, num_gs, skill)
num_fit_initializations = self.num_fits
best_likelihood = float("-inf")
best_model = None
for i in range(num_fit_initializations):
fitmodel = random_model_uni.random_model_uni(num_learns, num_gs)
if forgets:
fitmodel["forgets"] = np.random.uniform(
size=fitmodel["forgets"].shape)
if self.manual_param_init and skill in self.fit_model:
for var in self.fit_model[skill]:
if var in fitmodel:
fitmodel[var] = self.fit_model[skill][var]
if not preload:
fitmodel, log_likelihoods = EM_fit.EM_fit(
fitmodel, data, parallel=self.parallel)
if log_likelihoods[-1] > best_likelihood:
best_likelihood = log_likelihoods[-1]
best_model = fitmodel
else:
best_model = fitmodel
fit_model = best_model
fit_model["learns"] = fit_model["As"][:, 1, 0]
fit_model["forgets"] = fit_model["As"][:, 0, 1]
fit_model["prior"] = fit_model["pi_0"][1][0]
fit_model["resource_names"] = data["resource_names"]
fit_model["gs_names"] = data["gs_names"]
return fit_model
truemodel["prior"] = truemodel["pi_0"][1][0]
truemodel["guesses"] = np.full(num_subparts, 0.1, dtype=np.float_)
truemodel["slips"] = np.full(num_subparts, 0.03, dtype=np.float_)
#data!
print("generating data...")
data = synthetic_data.synthetic_data(truemodel, observation_sequence_lengths)
#fit models, starting with random initializations
print('fitting! each dot is a new EM initialization')
best_likelihood = float("-inf")
fitmodel = deepcopy(truemodel) # NOTE: include this line to initialize at the truth
(fitmodel, log_likelihoods) = EM_fit.EM_fit(fitmodel, data)
if(log_likelihoods[-1] > best_likelihood):
best_likelihood = log_likelihoods[-1]
best_model = fitmodel
# compare the fit model to the true model
print('')
print('\ttruth\tlearned')
for r in range(num_resources):
print('learn%d\t%.4f\t%.4f' % (r+1, truemodel['As'][r, 1, 0].squeeze(), best_model['As'][r, 1, 0].squeeze()))
for r in range(num_resources):
print('forget%d\t%.4f\t%.4f' % (r+1, truemodel['As'][r, 0, 1].squeeze(), best_model['As'][r, 0, 1].squeeze()))
for s in range(num_subparts):
print('guess%d\t%.4f\t%.4f' % (s+1, truemodel['guesses'][s], best_model['guesses'][s]))
print("All data okay")
total_auc = 0
total_trials = 0
all_true = []
all_pred = []
for skill in range(skill_count):
num_fit_initializations = 5
best_likelihood = float("-inf")
if len(Data[skill]["resources"]) < 1:
print("No data for skill %s" % skill)
continue
else:
for i in range(num_fit_initializations):
fitmodel = random_model_uni.random_model_uni(1, 1)
(fitmodel, log_likelihoods) = EM_fit.EM_fit(fitmodel, Data[skill])
if (log_likelihoods[-1] > best_likelihood):
best_likelihood = log_likelihoods[-1]
best_model = fitmodel
#print(" ")
#print('\tlearned')
#print('prior\t%.4f' % (best_model["pi_0"][1][0]))
#for r in range(1):
# print('learn%d\t%.4f' % (r+1, best_model['As'][r, 1, 0].squeeze()))
#for r in range(1):
# print('forget%d\t%.4f' % (r+1, best_model['As'][r, 0, 1].squeeze()))
#for s in range(1):
# print('guess%d\t%.4f' % (s+1, best_model['guesses'][s]))
#for s in range(1):
def crossvalidate(data, folds=5, verbose=False, seed=0, return_arrays=False):
if "resource_names" in data:
num_learns = len(data["resource_names"])
else:
num_learns = 1
if "gs_names" in data:
num_gs = len(data["gs_names"])
else:
num_gs = 1
total = 0
acc = 0
area_under_curve = 0
num_fit_initializations = 20
split_size = (len(data["starts"]) // folds)
#create random permutation to act as indices for folds for crossvalidation
shuffle = np.random.RandomState(seed=seed).permutation(len(data["starts"]))
all_true, all_pred = [], []
# crossvalidation on students which are identified by the starts array
for iteration in range(folds):
#create training/test data based on random permutation from earlier
train = np.concatenate(
(shuffle[0:iteration * split_size],
shuffle[(iteration + 1) * split_size:len(data["starts"])]))
test = shuffle[iteration * split_size:(iteration + 1) * split_size]
training_data = fix_data(data, train)
num_fit_initializations = 5
best_likelihood = float("-inf")
for i in range(num_fit_initializations):
fitmodel = random_model_uni.random_model_uni(
num_learns, num_gs
) # include this line to randomly set initial param values
(fitmodel,
log_likelihoods) = EM_fit.EM_fit(fitmodel, training_data)
if (log_likelihoods[-1] > best_likelihood):
best_likelihood = log_likelihoods[-1]
best_model = fitmodel
if verbose:
print(" ")
print('Iteration %d' % (iteration))
print('\tlearned')
print('prior\t%.4f' % (best_model["pi_0"][1][0]))
for r in range(num_learns):
print('learn%d\t%.4f' %
(r + 1, best_model['As'][r, 1, 0].squeeze()))
for r in range(num_learns):
print('forget%d\t%.4f' %
(r + 1, best_model['As'][r, 0, 1].squeeze()))
for s in range(num_gs):
print('guess%d\t%.4f' % (s + 1, best_model['guesses'][s]))
for s in range(num_gs):
print('slip%d\t%.4f' % (s + 1, best_model['slips'][s]))
test_data = fix_data(data, test)
# run model predictions from training data on test data
(correct_predictions,
state_predictions) = predict_onestep.run(best_model, test_data)
flat_true_values = np.zeros((len(test_data["data"][0]), ),
dtype=np.intc)
for i in range(len(test_data["data"])):
for j in range(len(test_data["data"][0])):
if test_data["data"][i][j] != 0:
flat_true_values[j] = test_data["data"][i][j]
flat_true_values = flat_true_values.tolist()
# print(len(flat_true_values))
# print(len(correct_predictions))
# print(auc.compute_auc(flat_true_values, correct_predictions))
all_true.extend(flat_true_values)
all_pred.extend(correct_predictions)
if return_arrays:
return (all_true, all_pred)
# print(len(all_true))
print(len(all_pred))
total += rmse.compute_rmse(all_true, all_pred)
acc += accuracy.compute_acc(all_true, all_pred)
area_under_curve += auc.compute_auc(all_true, all_pred)
if verbose:
print("Average RMSE: ", total)
print("Average Accuracy: ", acc)
print("Average AUC: ", area_under_curve)
return (acc, total, area_under_curve)