def get_ideal_observer(dotpos=None, bound=0.8):
if dotpos is None:
dotpos = load_dots()
D = dotpos.shape[0]
feature_means = np.c_[[-cond, 0], [cond, 0]]
model = rtmodels.discrete_static_gauss(dt=dotdt,
maxrt=dotdt * D,
toresponse=toresponse,
choices=[-1, 1],
Trials=dotpos,
means=feature_means)
# set ideal observer parameters
model.intstd = dotstd
model.noisestd = 1e-15
model.prior = 0.5
model.lapseprob = 0.0
model.ndtmean = -100
model.bound = bound
return model
cond = helpers.cond
feature_means = np.c_[[-cond, 0], [cond, 0]]
toresponse = helpers.toresponse
# sort trials such that fitind are first, because EP-ABC will only fit the
# first couple of trials that are stored in the model
fitind = np.setdiff1d(np.arange(L), testind, assume_unique=True)
allind = np.r_[fitind, testind]
Trials = dotpos[:, :, allind]
#%% make basic model and parameters
model = rtmodels.discrete_static_gauss(dt=dotdt, maxrt=dotdt*D,
toresponse=toresponse, choices=[-1, 1],
Trials=Trials, means=feature_means,
intstd=dotstd)
#%% setup EP-ABC
# normalising constant of the uniform distribution over the data space in
# which samples are accepted as defined by distfun and epsilon; for
# response_dist this is 2*epsilon for the Euclidean distance between true
# and sampled RT
veps = 2 * epsilon
#%% inference
for mname in models:
# generate time-stamped file name
import rtmodels
import numpy as np
from scipy.stats import multivariate_normal
#%% make model
dotstd = 70.0
stimulus = np.random.randint(1, 3, 200)
means = np.c_[[-25., 0], [25, 0]]
features = means[:, stimulus - 1]
features = np.tile(features, (25, 1, 1))
features += np.random.normal(scale=70, size=features.shape)
model = rtmodels.discrete_static_gauss(Trials=features,
dt=0.1,
choices=[1, 2],
means=means,
maxrt=2.5,
toresponse=[0, 5.])
model.noisestd = 1e-10
model.intstd = dotstd
model.bound = 0.99
model.bstretch = 0.2
model.bshape = 0.4
model.lapseprob = 0.0
model.ndtmean = -20
model.prior = 0.34
#%% sample responses from model with numba
ch, rt = model.gen_response(np.arange(200))
model.plot_response_distribution(ch, rt)
import pandas as pd
import matplotlib.pyplot as plt
import pyEPABC
from pyEPABC.parameters import exponential, gaussprob
import rtmodels
#%% load data
data = pd.read_csv(os.path.join(os.path.dirname(__file__), 'responsedata.csv'))
N = data.shape[0]
#%% create response model with some standard parameter settings
model = rtmodels.discrete_static_gauss(dt=0.05,
maxrt=2.5,
choices=[1, 2],
Trials=data.stimulus.values,
means=np.c_[[-25, 0], [25, 0]],
intstd=70,
noisestd=60)
print(model)
ch_sim, rt_sim = model.gen_response(np.arange(N), 1000)
fig, axes = plt.subplots(2, 1, sharex=True, sharey=True)
model.plot_response_distribution(data.choice, data.RT, ax=axes[0])
model.plot_response_distribution(ch_sim, rt_sim, ax=axes[1])
axes[0].set_ylabel('response data')
axes[1].set_ylabel('model simulations')
axes[1].set_xlabel('RT (s)')
pars.add_param('ndtspread', -1.5, 1, exponential())
pars.add_param('noisestd', 4, 1.5, exponential())
pars.add_param('bound', 0, 1, gaussprob(0.5, 0.5))
pars.add_param('prior', 0, 1, gaussprob())
pars.add_param('lapseprob', -1, 1, gaussprob())
pars.add_param('lapsetoprob', 0, 1, gaussprob())
# collapsing bound parameters
if collapse:
pars.add_param('bshape', np.log(1.4), 1.5, exponential())
pars.add_param('bstretch', 0, 1, gaussprob())
# make model
model = rtmodels.discrete_static_gauss(maxrt=dotdt * D,
choices=[-1, 1],
Trials=Trials,
means=feature_means,
intstd=dotstd)
model.dt = dotdt
model.toresponse = toresponse
# this is ignored when ndtmean is fitted, but makes a difference when
# predicting from the model and not providing a value for ndtmean in the call
model.ndtmean = -10
# plot prior distribution of parameters
if S > 0:
pars.plot_param_dist(S=S)
#%% setup and run EP-ABC
# wrapper for sampling from model and directly computing distances
Created on Wed Jun 1 16:02:43 2016 @author: Sebastian Bitzer ([email protected]) """ import numpy as np import matplotlib.pyplot as plt import rtmodels dt = 0.1 N = 2000 # make model model = rtmodels.discrete_static_gauss(maxrt=2.5, choices=[-1, 1], Trials=(np.random.randint(2, size=N) - 0.5) * 2, means=np.c_[[-25, 0], [25, 0]], intstd=70.) # ignore lapses model.lapseprob = 0 # set nondecision time distribution with small, but realistic values model.ndtmean = -0.9 model.ndtspread = 0.41 # noise relatively small model.noisestd = 70 # simulate from model choice, rt = model.gen_response(np.arange(N))
# -*- coding: utf-8 -*-
import numpy as np
import matplotlib.pyplot as plt
import rtmodels
plt.style.use('ggplot')
choices = np.array([-1, 1])
Trials = choices[np.random.randint(2, size=100)]
model = rtmodels.discrete_static_gauss(maxrt=2.5,
choices=choices,
Trials=Trials)
trind = np.random.randint(100, size=10000)
choices, rts = model.gen_response_with_params(trind, params={'ndtspread': 0.3})
model.plot_response_distribution(choices, rts)
print('fraction correct = %6.4f' % np.mean(choices == Trials[trind]))