respRT = helpers.load_subject_data(sub, behavdatadir)
# data trials need to be in the order corresponding to the one in the
# model, remember that indeces in respRT start at 1, not 0
data = respRT.loc[fitind+1, ['response', 'RT']].values
#%% infer with EP-ABC and collect results
print('subject: %02d, model: %s' % (sub, mname))
# the for-loop is a fail-safe procedure: try to run EP-ABC 3 times when
# it unexpectedly breaks, if it didn't run through the 3rd time, just
# continue with the next subject
for run in range(3):
try:
ep_mean, ep_cov, ep_logml, nacc, ntotal, runtime = pyEPABC.run_EPABC(
data, simfun, None, pars.mu, pars.cov, epsilon=epsilon,
minacc=minacc, samplestep=10000, samplemax=samplemax, npass=npass,
alpha=alpha, veps=veps)
except KeyboardInterrupt:
raise KeyboardInterrupt
except:
errors[sub] = 'run %d:\n' % run + traceback.format_exc()
else:
ep_means[si, :] = ep_mean
ep_covs[si, :, :] = ep_cov
ep_summary.loc[sub, 'logml'] = ep_logml
ep_summary.loc[sub, 'nsamples'] = ntotal.sum()
ep_summary.loc[sub, 'runtime'] = runtime.total_seconds()
ep_summary.loc[sub, pars.names] = pars.get_transformed_mode(
ep_mean, ep_cov)
#%% preliminary extra analysis
pars.plot_param_dist()
simfun = lambda data, dind, parsamples: model.gen_distances_with_params(
data[0], data[1], dind, pars.transform(parsamples), pars.names)
epsilon = 0.05
veps = 2 * epsilon
# calling EPABC:
ep_mean, ep_cov, ep_logml, nacc, ntotal, runtime = pyEPABC.run_EPABC(
data[['response', 'RT']].values,
simfun,
None,
pars.mu,
pars.cov,
epsilon=epsilon,
minacc=2000,
samplestep=10000,
samplemax=6000000,
npass=2,
alpha=0.5,
veps=veps)
#%% show posterior distribution
pg = pars.plot_param_dist(ep_mean, ep_cov)
pg.fig.tight_layout()
pg.fig.savefig('%02d_posterior.svg' % sub)
#%% posterior predictive check
choices_post, rts_post = model.gen_response_from_Gauss_posterior(
# epsilon; for response_dist this is: (2*epsilon for the Euclidean distance
# between true and sampled RT and another factor of 2 for the two possible
# responses - timed out responses don't occur here, because they are only a
# point mass, i.e., one particular value of response and RT)
veps = 2 * 2 * epsilon
data = respRT.loc[fitind + 1, ['response', 'RT']].values
# run EPABC
ep_mean, ep_cov, ep_logml, nacc, ntotal, runtime = pyEPABC.run_EPABC(
data,
simfun,
distfun,
pars.mu,
pars.cov,
epsilon=epsilon,
minacc=2000,
samplestep=10000,
samplemax=6000000,
npass=2,
alpha=0.5,
veps=veps)
print('logml = %8.2f' % ep_logml)
#%% evaluate posterior
# sample from EPABC posterior (for all trials)
NP = 15000
choices, rts, samples_pos = model.gen_response_from_Gauss_posterior(
np.arange(L),
pars.names,
pg.fig.tight_layout()
#%% fit with EP-ABC
simfun = lambda data, dind, parsamples: model.gen_distances_with_params(
data[0], data[1], dind, pars.transform(parsamples), pars.names)
# maximum allowed difference in RT for accepting samples
epsilon = 0.05
ep_mean, ep_cov, ep_logml, nacc, ntotal, runtime = pyEPABC.run_EPABC(
np.c_[data.choice, data.RT],
simfun,
None,
pars.mu,
pars.cov,
epsilon=epsilon,
minacc=2000,
samplestep=10000,
samplemax=6000000,
npass=2,
alpha=0.5,
veps=2 * epsilon)
#%% investigate posterior
modes = pars.get_transformed_mode(ep_mean, ep_cov)
fig, axes = pars.compare_pdfs(ep_mean, ep_cov)
ch_post, rt_post = model.gen_response_from_Gauss_posterior(
np.arange(N), pars.names, ep_mean, ep_cov, 500, pars.transform)
data.values[:, 0].astype(dtype=int))
# maximum distance allowed for accepting samples, note that for
# response_dist this is in units of RT
epsilon = 0.05
# normalising constant of the uniform distribution defined by distfun and
# epsilon; for response_dist this is: (2*epsilon for the Euclidean distance
# between true and sampled RT and another factor of 2 for the two possible
# responses - timed out responses don't occur here, because they are only a
# point mass, i.e., one particular value of response and RT)
veps = 2 * 2 * epsilon
# run EPABC
ep_mean, ep_cov, ep_logml, nacc, ntotal, runtime = run_EPABC(
data.values[:, 1:], simfun, response_dist, prior_mean, prior_cov,
epsilon=epsilon, minacc=500, samplestep=10000, samplemax=2000000,
npass=3, alpha=0.3, veps=veps)
# sample from EPABC posterior
samples_pos = np.random.multivariate_normal(ep_mean, ep_cov, N)
samples_pos = pd.DataFrame(paramtransform(samples_pos), columns=parnames)
samples_pos['distribution'] = 'epabc_pos'
samples = samples.append(samples_pos, ignore_index=True)
# see what happens when you ignore RTs, epco = ep choice only
distfun = lambda data, sims: response_dist(data, sims, useRT=False)
epco_mean, epco_cov, epco_logml, epco_nacc, epco_ntotal, epco_runtime = run_EPABC(
data.values[:, 1:], simfun, distfun, prior_mean, prior_cov,
epsilon=epsilon, minacc=10000, samplestep=10000, samplemax=2000000,
npass=3, alpha=0.3, veps=0.5)