def sample_abilities_diffusion(args):
"""Sample the ability vector for this user, from the posterior over user
ability conditioned on the observed exercise performance.
use Metropolis-Hastings with Gaussian proposal distribution.
This is just a wrapper around the corresponding function in mirt_util.
"""
# make sure each student gets a different random sequence
id = multiprocessing.current_process()._identity
if len(id) > 0:
np.random.seed([id[0], time.time() * 1e9])
else:
np.random.seed([time.time() * 1e9])
couplings, state, options, user_index = args
abilities = state['abilities']
correct = state['correct']
exercises_ind = state['exercises_ind']
num_steps = options.sampling_num_steps
abilities, Eabilities, _, _ = mirt_util.sample_abilities_diffusion(
couplings, exercises_ind, correct, abilities, num_steps)
return abilities, Eabilities, user_index
def sample_abilities_diffusion(args):
"""Sample the ability vector for this user, from the posterior over user
ability conditioned on the observed exercise performance.
use Metropolis-Hastings with Gaussian proposal distribution.
This is just a wrapper around the corresponding function in mirt_util.
"""
# TODO(jascha) make this a better sampler (eg, use the HMC sampler from
# TMIRT)
# make sure each student gets a different random sequence
id = multiprocessing.current_process()._identity
if len(id) > 0:
np.random.seed([id[0], time.time() * 1e9])
else:
np.random.seed([time.time() * 1e9])
theta, state, options, user_index = args
abilities = state['abilities']
correct = state['correct']
log_time_taken = state['log_time_taken']
exercises_ind = state['exercises_ind']
num_steps = options.sampling_num_steps
abilities, Eabilities, _, _ = mirt_util.sample_abilities_diffusion(
theta, exercises_ind, correct, log_time_taken,
abilities, num_steps)
return abilities, Eabilities, user_index
def emit_features(user_states, couplings, options, split_desc):
"""Emit a CSV data file of correctness, prediction, and abilities."""
f = open("%s_split=%s.csv" % (options.output, split_desc), 'w+')
for user_state in user_states:
# initialize
abilities = np.zeros((options.num_abilities, 1))
correct = user_state['correct']
exercises_ind = user_state['exercises_ind']
# NOTE: I currently do not output features for the first problem
for i in xrange(1, correct.size):
# TODO(jace) this should probably be the marginal estimation
_, _, abilities, _ = mirt_util.sample_abilities_diffusion(
couplings, exercises_ind[:i], correct[:i],
abilities_init=abilities, num_steps=200)
prediction = mirt_util.conditional_probability_correct(
abilities, couplings, exercises_ind[i:(i + 1)])
print >>f, "%d," % correct[i],
print >>f, "%.4f," % prediction[-1],
print >>f, ",".join(["%.4f" % a for a in abilities])
f.close()
def emit_features(user_states, theta, options, split_desc):
"""Emit a CSV data file of correctness, prediction, and abilities."""
f = open("%s_split=%s.csv" % (options.output, split_desc), 'w+')
for user_state in user_states:
# initialize
abilities = np.zeros((options.num_abilities, 1))
correct = user_state['correct']
log_time_taken = user_state['log_time_taken']
exercises_ind = user_state['exercises_ind']
# NOTE: I currently do not output features for the first problem
for i in xrange(1, correct.size):
# TODO(jace) this should probably be the marginal estimation
_, _, abilities, _ = mirt_util.sample_abilities_diffusion(
theta, exercises_ind[:i], correct[:i], log_time_taken[:i],
abilities_init=abilities, num_steps=200)
prediction = mirt_util.conditional_probability_correct(
abilities, theta, exercises_ind[i:(i + 1)])
print >>f, "%d," % correct[i],
print >>f, "%.4f," % prediction[-1],
print >>f, ",".join(["%.4f" % a for a in abilities])
f.close()
