def run_trials(p, args):
# Model
m = p['model']
# Number of trials
try:
ntrials = int(args[0])
except:
ntrials = 100
ntrials *= m.nconditions
# RNN
rng = np.random.RandomState(p['seed'])
rnn = RNN(p['savefile'], {'dt': p['dt']}, verbose=False)
w = len(str(ntrials))
trials = []
backspaces = 0
try:
for i in xrange(ntrials):
# Condition
k = tasktools.unravel_index(
i % m.nconditions, (len(m.cohs), len(m.left_rights), len(
m.cohs), len(m.left_rights), len(m.contexts)))
coh_m = m.cohs[k[0]]
left_right_m = m.left_rights[k[1]]
coh_c = m.cohs[k[2]]
left_right_c = m.left_rights[k[3]]
context = m.contexts[k[4]]
# Trial
trial_func = m.generate_trial
trial_args = {
'name': 'test',
'catch': False,
'coh_m': coh_m,
'left_right_m': left_right_m,
'coh_c': coh_c,
'left_right_c': left_right_c,
'context': context
}
info = rnn.run(inputs=(trial_func, trial_args), rng=rng)
# Display trial type
s = ("Trial {:>{}}/{}: ({}) m{:>+3}, c{:>+3}".format(
i + 1, w, ntrials, info['context'],
info['left_right_m'] * info['coh_m'],
info['left_right_c'] * info['coh_c']))
sys.stdout.write(backspaces * '\b' + s)
sys.stdout.flush()
backspaces = len(s)
# Save
dt = rnn.t[1] - rnn.t[0]
step = int(p['dt_save'] / dt)
trial = {
't': rnn.t[::step],
'u': rnn.u[:, ::step],
'r': rnn.r[:, ::step],
'z': rnn.z[:, ::step],
'info': info,
}
trials.append(trial)
except KeyboardInterrupt:
pass
print("")
# Save all
filename = get_trialsfile(p)
with open(filename, 'wb') as f:
pickle.dump(trials, f, pickle.HIGHEST_PROTOCOL)
size = os.path.getsize(filename) * 1e-9
print("[ {}.run_trials ] Trials saved to {} ({:.1f} GB)".format(
THIS, filename, size))
# Compute the psychometric function
psychometric_function(filename)
def run_trials(p, args):
"""
Run trials.
"""
# Model
m = p['model']
# Number of trials
try:
ntrials = int(args[0])
except:
ntrials = 100
ntrials *= m.nconditions
# RNN
rng = np.random.RandomState(p['seed'])
rnn = RNN(p['savefile'], {'dt': p['dt']}, verbose=False)
# Trials
w = len(str(ntrials))
trials = []
backspaces = 0
try:
for i in xrange(ntrials):
b = i % m.nconditions
k1, k2 = tasktools.unravel_index(b, (len(m.modalities), len(m.freqs)))
modality = m.modalities[k1]
freq = m.freqs[k2]
# Trial
trial_func = m.generate_trial
trial_args = {
'name': 'test',
'catch': False,
'modality': modality,
'freq': freq
}
info = rnn.run(inputs=(trial_func, trial_args), rng=rng)
# Display trial type
if info['modality'] == 'v':
s = "Trial {:>{}}/{}: v |{:>2}".format(i+1, w, ntrials, info['freq'])
elif info['modality'] == 'a':
s = "Trial {:>{}}/{}: a|{:>2}".format(i+1, w, ntrials, info['freq'])
else:
s = "Trial {:>{}}/{}: va|{:>2}".format(i+1, w, ntrials, info['freq'])
sys.stdout.write(backspaces*'\b' + s)
sys.stdout.flush()
backspaces = len(s)
# Save
dt = rnn.t[1] - rnn.t[0]
step = int(p['dt_save']/dt)
trial = {
't': rnn.t[::step],
'u': rnn.u[:,::step],
'r': rnn.r[:,::step],
'z': rnn.z[:,::step],
'info': info
}
trials.append(trial)
except KeyboardInterrupt:
pass
print("")
# Save all
filename = get_trialsfile(p)
with open(filename, 'wb') as f:
pickle.dump(trials, f, pickle.HIGHEST_PROTOCOL)
size = os.path.getsize(filename)*1e-9
print("[ {}.run_trials ] Trials saved to {} ({:.1f} GB)".format(THIS, filename, size))
# Compute the psychometric function
psychometric_function(filename)
def run_trials(p, args):
"""
Run trials.
"""
# Model
m = p['model']
# Number of trials
try:
ntrials = int(args[0])
except:
ntrials = 100
ntrials *= m.nconditions
# RNN
rng = np.random.RandomState(p['seed'])
rnn = RNN(p['savefile'], {'dt': p['dt']}, verbose=False)
# Trials
w = len(str(ntrials))
trials = []
backspaces = 0
try:
for i in xrange(ntrials):
# Condition
b = i % m.nconditions
k0, k1 = tasktools.unravel_index(b, (len(m.fpairs), len(m.gt_lts)))
fpair = m.fpairs[k0]
gt_lt = m.gt_lts[k1]
# Trial
trial_func = m.generate_trial
trial_args = {
'name': 'test',
'catch': False,
'fpair': fpair,
'gt_lt': gt_lt
}
info = rnn.run(inputs=(trial_func, trial_args), rng=rng)
# Display trial type
if info['f1'] > info['f2']:
gt_lt = '>'
else:
gt_lt = '<'
s = ("Trial {:>{}}/{}: {:>2} {} {:>2}".format(
i + 1, w, ntrials, info['f1'], gt_lt, info['f2']))
sys.stdout.write(backspaces * '\b' + s)
sys.stdout.flush()
backspaces = len(s)
# Save
dt = rnn.t[1] - rnn.t[0]
step = int(p['dt_save'] / dt)
trial = {
't': rnn.t[::step],
'u': rnn.u[:, ::step],
'r': rnn.r[:, ::step],
'z': rnn.z[:, ::step],
'info': info
}
trials.append(trial)
except KeyboardInterrupt:
pass
print("")
# Save all
filename = get_trialsfile(p)
with open(filename, 'wb') as f:
pickle.dump(trials, f, pickle.HIGHEST_PROTOCOL)
size = os.path.getsize(filename) * 1e-9
print("[ {}.run_trials ] Trials saved to {} ({:.1f} GB)".format(
THIS, filename, size))
# Psychometric function
psychometric_function(filename)
def generate_trial(rng, dt, params):
#-------------------------------------------------------------------------------------
# Select task condition
#-------------------------------------------------------------------------------------
catch_trial = False
if params['name'] in ['gradient', 'test']:
if params.get('catch', rng.rand() < pcatch):
catch_trial = True
else:
# Context
context = params.get('context', rng.choice(contexts))
# Coherences
coh_m = params.get('coh_m', rng.choice(cohs))
coh_c = params.get('coh_c', rng.choice(cohs))
# Left/right
left_right_m = params.get('left_right_m', rng.choice(left_rights))
left_right_c = params.get('left_right_c', rng.choice(left_rights))
elif params['name'] == 'validation':
b = params['minibatch_index'] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k = tasktools.unravel_index(b - 1,
(len(contexts), len(cohs), len(cohs),
len(left_rights), len(left_rights)))
context = contexts[k[0]]
coh_m = cohs[k[1]]
coh_c = cohs[k[2]]
left_right_m = left_rights[k[3]]
left_right_c = left_rights[k[4]]
else:
raise ValueError("Unknown trial type.")
#-------------------------------------------------------------------------------------
# Epochs
#-------------------------------------------------------------------------------------
if catch_trial:
epochs = {'T': 2000}
else:
if params['name'] == 'test':
fixation = 400
else:
fixation = 100
stimulus = 800
decision = 300
T = fixation + stimulus + decision
epochs = {
'fixation': (0, fixation),
'stimulus': (fixation, fixation + stimulus),
'decision': (fixation + stimulus, T)
}
epochs['T'] = T
#-------------------------------------------------------------------------------------
# Trial info
#-------------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(
dt, epochs) # Time, task epochs in discrete time
trial = {'t': t, 'epochs': epochs} # Trial
if catch_trial:
trial['info'] = {}
else:
if context == 'm':
left_right = left_right_m
else:
left_right = left_right_c
# Correct choice
if left_right > 0:
choice = 0
else:
choice = 1
# Trial info
trial['info'] = {
'coh_m': coh_m,
'left_right_m': left_right_m,
'coh_c': coh_c,
'left_right_c': left_right_c,
'context': context,
'choice': choice
}
#-------------------------------------------------------------------------------------
# Inputs
#-------------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
# Context
if context == 'm':
X[e['stimulus'], 0] = 1
else:
X[e['stimulus'], 1] = 1
# Motion stimulus
if left_right_m > 0:
choice_m = 0
else:
choice_m = 1
X[e['stimulus'], 2 + choice_m] = scale(+coh_m)
X[e['stimulus'], 2 + (1 - choice_m)] = scale(-coh_m)
# Colour stimulus
if left_right_c > 0:
choice_c = 0
else:
choice_c = 1
X[e['stimulus'], 4 + choice_c] = scale(+coh_c)
X[e['stimulus'], 4 + (1 - choice_c)] = scale(-coh_c)
trial['inputs'] = X
#-------------------------------------------------------------------------------------
# Target output
#-------------------------------------------------------------------------------------
if params.get('target_output', False):
Y = np.zeros((len(t), Nout)) # Output matrix
M = np.zeros_like(Y) # Mask matrix
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e['fixation'], :] = lo
# Decision
Y[e['decision'], choice] = hi
Y[e['decision'], 1 - choice] = lo
# Only care about fixation and decision periods
M[e['fixation'] + e['decision'], :] = 1
# Outputs and mask
trial['outputs'] = Y
trial['mask'] = M
#-------------------------------------------------------------------------------------
return trial
def generate_trial(rng, dt, params):
#-------------------------------------------------------------------------------------
# Select task condition
#-------------------------------------------------------------------------------------
catch_trial = False
if params['name'] in ['gradient', 'test']:
if params.get('catch', rng.rand() < pcatch):
catch_trial = True
else:
modality = params.get('modality', rng.choice(modalities))
freq = params.get('freq', rng.choice(freqs))
elif params['name'] == 'validation':
b = params['minibatch_index'] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k1, k2 = tasktools.unravel_index(b - 1,
(len(modalities), len(freqs)))
modality = modalities[k1]
freq = freqs[k2]
else:
raise ValueError("Unknown trial type.")
#-------------------------------------------------------------------------------------
# Epochs
#-------------------------------------------------------------------------------------
if catch_trial:
epochs = {'T': 2500}
else:
if params['name'] == 'test':
fixation = 500
else:
fixation = 100
stimulus = 1000
decision = 300
T = fixation + stimulus + decision
epochs = {
'fixation': (0, fixation),
'stimulus': (fixation, fixation + stimulus),
'decision': (fixation + stimulus, T)
}
epochs['T'] = T
#-------------------------------------------------------------------------------------
# Trial info
#-------------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(
dt, epochs) # Time, task epochs in discrete time
trial = {'t': t, 'epochs': epochs} # Trial
if catch_trial:
trial['info'] = {}
else:
# Correct choice
if freq > boundary:
choice = 0
else:
choice = 1
# Trial info
trial['info'] = {'modality': modality, 'freq': freq, 'choice': choice}
#-------------------------------------------------------------------------------------
# Inputs
#-------------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
if 'v' in modality:
X[e['stimulus'], VISUAL_P] = scale_v_p(freq)
X[e['stimulus'], VISUAL_N] = scale_v_n(freq)
if 'a' in modality:
X[e['stimulus'], AUDITORY_P] = scale_a_p(freq)
X[e['stimulus'], AUDITORY_N] = scale_a_n(freq)
X[e['stimulus'] + e['decision'], START] = 1
trial['inputs'] = X
#-------------------------------------------------------------------------------------
# Target output
#-------------------------------------------------------------------------------------
if params.get('target_output', False):
Y = np.zeros((len(t), Nout)) # Output matrix
M = np.zeros_like(Y) # Mask matrix
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e['fixation'], :] = lo
# Decision
Y[e['decision'], choice] = hi
Y[e['decision'], 1 - choice] = lo
# Only care about fixation and decision periods
M[e['fixation'] + e['decision'], :] = 1
# Outputs and mask
trial['outputs'] = Y
trial['mask'] = M
#-------------------------------------------------------------------------------------
return trial
def generate_trial(rng, dt, params):
#-------------------------------------------------------------------------------------
# Select task condition
#-------------------------------------------------------------------------------------
catch_trial = False
if params['name'] in ['gradient', 'test']:
if params.get('catch',
rng.rand() < pcatch): #pcatch chance of a catch trial
catch_trial = True
else:
cond = params.get('conditions',
conditions[rng.choice(len(conditions))])
#Revert back to cohs if this brings up an issue
in_out = params.get('in_out', rng.choice(in_outs))
elif params['name'] == 'validation':
b = params['minibatch_index'] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k0, k1 = tasktools.unravel_index(b - 1,
(len(conditions), len(in_outs)))
cond = conditions[k0]
in_out = in_outs[k1]
else:
raise ValueError("Unknown trial type.")
#-------------------------------------------------------------------------------------
# Epochs
#-------------------------------------------------------------------------------------
if catch_trial:
epochs = {'T': 2500}
else:
if params['name'] == 'test':
fixation = 300
else:
fixation = 100
f1 = 100
delay = 200
f2 = 100
decision = 200
# else:
# if params['name'] == 'test':
# firstInterval = 300
# else:
# firstInterval = 100
# secondInterval = 200
# decision = 200
T = fixation + f1 + delay + f2 + decision
epochs = {
'fixation': (0, fixation),
'f1': (fixation, fixation + f1),
'delay': (fixation + f1, fixation + f1 + delay),
'f2': (fixation + f1 + delay, fixation + f1 + delay + f2),
'decision': (fixation + f1 + delay + f2, T)
}
epochs['T'] = T
#-------------------------------------------------------------------------------------
# Trial info
#-------------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(
dt, epochs) # Time, e is array of epochs w/ points
trial = {'t': t, 'epochs': epochs} # Trial
if catch_trial:
trial['info'] = {}
else:
# Correct choice
if cond in [(yes, yes), (no, no)]:
choice = 0
else:
choice = 1
#if in_out == 'equal':
# choice = 0
#else:
# choice = 1
# Trial info
trial['info'] = {'cond': cond, 'in_out': in_out, 'choice': choice}
#-------------------------------------------------------------------------------------
# Inputs
#-------------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
# Stimulus 1
X[e['f1'], POS] = cond[0]
X[e['f1'],
NEG] = yes + no - cond[0] #flipped frequency 1->0.2, 0.2 -> 1
# Stimulus 2
X[e['f2'], POS] = cond[1]
X[e['f2'],
NEG] = yes + no - cond[1] #flipped frequency 1->0.2, 0.2 -> 1
trial['inputs'] = X
#-------------------------------------------------------------------------------------
# Target output
#-------------------------------------------------------------------------------------
if params.get('target_output', False):
Y = np.zeros((len(t), Nout)) # Output
M = np.zeros_like(Y) # Mask
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e['fixation'], :] = lo
# Decision
Y[e['decision'], choice] = hi # One neuron spikes high
Y[e['decision'], 1 - choice] = lo # The other spikes low
# Only care about fixation and decision periods
M[e['fixation'] + e['decision'], :] = 1
# Outputs and mask
trial['outputs'] = Y
trial['mask'] = M
#-------------------------------------------------------------------------------------
return trial
def run_trials(p, args):
"""
Run trials.
"""
# Model
m = p['model']
# Number of trials
try:
ntrials = int(args[0])
except:
ntrials = 100
ntrials *= m.nconditions + 1
# RNN
rng = np.random.RandomState(p['seed'])
rnn = RNN(p['savefile'], {'dt': p['dt']}, verbose=False)
if 'ant_level' in p:
rnn.Wrec[np.where(rnn.Wrec<0)] *= (1 - p['ant_level'])
# Trials
w = len(str(ntrials))
trials = []
backspaces = 0
try:
for i in xrange(ntrials):
b = i % (m.nconditions + 1)
if b == 0:
# Zero-coherence condition
coh = 0
in_out = rng.choice(m.in_outs)
else:
# All other conditions
k1, k2 = tasktools.unravel_index(b-1, (len(m.cohs), len(m.in_outs)))
coh = m.cohs[k1]
in_out = m.in_outs[k2]
# Trial
trial_func = m.generate_trial
trial_args = {
'name': 'test',
'catch': False,
'coh': coh,
'in_out': in_out
}
info = rnn.run(inputs=(trial_func, trial_args), rng=rng)
# Display trial type
#if coh == 0:
# s = "Trial {:>{}}/{}: {:>3}".format(i+1, w, ntrials, info['coh'])
#else:
# s = ("Trial {:>{}}/{}: {:>+3}"
# .format(i+1, w, ntrials, info['in_out']*info['coh']))
# Update the user with the current progress
if (i/100) % 1 == 0:
print("We are {:.2f}% complete".format(100*i/ntrials))
sys.stdout.flush()
# Save
dt = rnn.t[1] - rnn.t[0]
step = int(p['dt_save']/dt)
trial = {
't': rnn.t[::step],
'u': rnn.u[:,::step],
'r': rnn.r[:,::step],
'z': rnn.z[:,::step],
'info': info
}
trials.append(trial)
except KeyboardInterrupt:
pass
print("")
# Save all
filename = get_trialsfile(p)
dump(filename, trials)
size = os.path.getsize(filename)*1e-9
print("[ {}.run_trials ] Trials saved to {} ({:.1f} GB)".format(THIS, filename, size))
# Compute the psychometric function
psychometric_function(filename)
def run_trials(p, args):
"""
Run trials.
"""
# Model
m = p['model']
# Number of trials
try:
ntrials = int(args[0])
except:
ntrials = 100
ntrials *= m.nconditions + 1
# RNN
rng = np.random.RandomState(p['seed'])
rnn = RNN(p['savefile'], {'dt': p['dt']}, verbose=False)
# Trials
w = len(str(ntrials))
trials = []
backspaces = 0
try:
for i in xrange(ntrials):
b = i % (m.nconditions + 1)
if b == 0:
# Zero-coherence condition
coh = 0
in_out = rng.choice(m.in_outs)
else:
# All other conditions
k1, k2 = tasktools.unravel_index(b-1, (len(m.cohs), len(m.in_outs)))
coh = m.cohs[k1]
in_out = m.in_outs[k2]
# Trial
trial_func = m.generate_trial
trial_args = {
'name': 'test',
'catch': False,
'coh': coh,
'in_out': in_out
}
info = rnn.run(inputs=(trial_func, trial_args), rng=rng)
# Display trial type
if coh == 0:
s = "Trial {:>{}}/{}: {:>3}".format(i+1, w, ntrials, info['coh'])
else:
s = ("Trial {:>{}}/{}: {:>+3}"
.format(i+1, w, ntrials, info['in_out']*info['coh']))
sys.stdout.write(backspaces*'\b' + s)
sys.stdout.flush()
backspaces = len(s)
# Save
dt = rnn.t[1] - rnn.t[0]
step = int(p['dt_save']/dt)
trial = {
't': rnn.t[::step],
'u': rnn.u[:,::step],
'r': rnn.r[:,::step],
'z': rnn.z[:,::step],
'info': info
}
trials.append(trial)
except KeyboardInterrupt:
pass
print("")
# Save all
filename = get_trialsfile(p)
with open(filename, 'wb') as f:
pickle.dump(trials, f, pickle.HIGHEST_PROTOCOL)
size = os.path.getsize(filename)*1e-9
print("[ {}.run_trials ] Trials saved to {} ({:.1f} GB)".format(THIS, filename, size))
# Compute the psychometric function
psychometric_function(filename)
def generate_trial(rng, dt, params):
#---------------------------------------------------------------------------------
# Select task condition
#---------------------------------------------------------------------------------
catch_trial = False
if params['name'] in ['gradient', 'test']:
if params.get('catch', rng.rand() < pcatch):
catch_trial = True
else:
fpair = params.get('fpair', fpairs[rng.choice(len(fpairs))])
gt_lt = params.get('gt_lt', rng.choice(gt_lts))
elif params['name'] == 'validation':
b = params['minibatch_index'] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k0, k1 = tasktools.unravel_index(b-1, (len(fpairs), len(gt_lts)))
fpair = fpairs[k0]
gt_lt = gt_lts[k1]
else:
raise ValueError("Unknown trial type.")
#---------------------------------------------------------------------------------
# Epochs
#---------------------------------------------------------------------------------
if catch_trial:
epochs = {'T': 2500}
else:
if params['name'] == 'test':
fixation = 500
else:
fixation = 100
f1 = 500
if params['name'] == 'test':
delay = 3000
else:
delay = tasktools.uniform(rng, dt, 2500, 3500)
f2 = 500
decision = 300
T = fixation + f1 + delay + f2 + decision
epochs = {
'fixation': (0, fixation),
'f1': (fixation, fixation + f1),
'delay': (fixation + f1, fixation + f1 + delay),
'f2': (fixation + f1 + delay, fixation + f1 + delay + f2),
'decision': (fixation + f1 + delay + f2, T)
}
epochs['T'] = T
#---------------------------------------------------------------------------------
# Trial info
#---------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(dt, epochs) # Task epochs in discrete time
trial = {'t': t, 'epochs': epochs} # Trial
if catch_trial:
trial['info'] = {}
else:
# Correct choice
if gt_lt == '>':
f1, f2 = fpair
choice = 0
else:
f2, f1 = fpair
choice = 1
# Info
trial['info'] = {'f1': f1, 'f2': f2, 'choice': choice}
#---------------------------------------------------------------------------------
# Inputs
#---------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
# Stimulus 1
X[e['f1'],POS] = scale_p(f1)
X[e['f1'],NEG] = scale_n(f1)
# Stimulus 2
X[e['f2'],POS] = scale_p(f2)
X[e['f2'],NEG] = scale_n(f2)
trial['inputs'] = X
#---------------------------------------------------------------------------------
# Target output
#---------------------------------------------------------------------------------
if params.get('target_output', False):
Y = np.zeros((len(t), Nout)) # Output matrix
M = np.zeros_like(Y) # Mask matrix
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e['fixation'],:] = lo
# Decision
Y[e['decision'],choice] = hi
Y[e['decision'],1-choice] = lo
# Mask
M[e['fixation']+e['decision'],:] = 1
trial['outputs'] = Y
trial['mask'] = M
#---------------------------------------------------------------------------------
return trial
def generate_trial(rng, dt, params):
#-------------------------------------------------------------------------------------
# Select task condition
#-------------------------------------------------------------------------------------
catch_trial = False
if params['name'] in ['gradient', 'test']:
if params.get('catch', rng.rand() < pcatch):
catch_trial = True
else:
modality = params.get('modality', rng.choice(modalities))
freq = params.get('freq', rng.choice(freqs))
elif params['name'] == 'validation':
b = params['minibatch_index'] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k1, k2 = tasktools.unravel_index(b-1,
(len(modalities), len(freqs)))
modality = modalities[k1]
freq = freqs[k2]
else:
raise ValueError("Unknown trial type.")
#-------------------------------------------------------------------------------------
# Epochs
#-------------------------------------------------------------------------------------
if catch_trial:
epochs = {'T': 2500}
else:
if params['name'] == 'test':
fixation = 500
else:
fixation = 100
stimulus = 1000
decision = 300
T = fixation + stimulus + decision
epochs = {
'fixation': (0, fixation),
'stimulus': (fixation, fixation + stimulus),
'decision': (fixation + stimulus, T)
}
epochs['T'] = T
#-------------------------------------------------------------------------------------
# Trial info
#-------------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(dt, epochs) # Time, task epochs in discrete time
trial = {'t': t, 'epochs': epochs} # Trial
if catch_trial:
trial['info'] = {}
else:
# Correct choice
if freq > boundary:
choice = 0
else:
choice = 1
# Trial info
trial['info'] = {'modality': modality, 'freq': freq, 'choice': choice}
#-------------------------------------------------------------------------------------
# Inputs
#-------------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
if 'v' in modality:
X[e['stimulus'],VISUAL_P] = scale_v_p(freq)
X[e['stimulus'],VISUAL_N] = scale_v_n(freq)
if 'a' in modality:
X[e['stimulus'],AUDITORY_P] = scale_a_p(freq)
X[e['stimulus'],AUDITORY_N] = scale_a_n(freq)
X[e['stimulus'] + e['decision'],START] = 1
trial['inputs'] = X
#-------------------------------------------------------------------------------------
# Target output
#-------------------------------------------------------------------------------------
if params.get('target_output', False):
Y = np.zeros((len(t), Nout)) # Output matrix
M = np.zeros_like(Y) # Mask matrix
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e['fixation'],:] = lo
# Decision
Y[e['decision'],choice] = hi
Y[e['decision'],1-choice] = lo
# Only care about fixation and decision periods
M[e['fixation']+e['decision'],:] = 1
# Outputs and mask
trial['outputs'] = Y
trial['mask'] = M
#-------------------------------------------------------------------------------------
return trial
def generate_trial(rng, dt, params):
# -------------------------------------------------------------------------------------
# Select task condition
# -------------------------------------------------------------------------------------
catch_trial = False
if params["name"] in ["gradient", "test"]:
if params.get("catch", rng.rand() < pcatch):
catch_trial = True
else:
coh = params.get("coh", rng.choice(cohs))
in_out = params.get("in_out", rng.choice(in_outs))
elif params["name"] == "validation":
b = params["minibatch_index"] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k0, k1 = tasktools.unravel_index(b - 1, (len(cohs), len(in_outs)))
coh = cohs[k0]
in_out = in_outs[k1]
else:
raise ValueError("Unknown trial type.")
# -------------------------------------------------------------------------------------
# Epochs
# -------------------------------------------------------------------------------------
if catch_trial:
epochs = {"T": 2000}
else:
if params["name"] == "test":
fixation = 300
stimulus = 1500
else:
fixation = 100
stimulus = 800
no_reward = 300
T = fixation + stimulus
epochs = {"fixation": (0, fixation), "stimulus": (fixation, T), "decision": (fixation + no_reward, T)}
epochs["T"] = T
# -------------------------------------------------------------------------------------
# Trial info
# -------------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(dt, epochs) # Time, task epochs in discrete time
trial = {"t": t, "epochs": epochs} # Trial
if catch_trial:
trial["info"] = {}
else:
# Correct choice
if in_out > 0:
choice = 0
else:
choice = 1
# Trial info
trial["info"] = {"coh": coh, "in_out": in_out, "choice": choice}
# -------------------------------------------------------------------------------------
# Inputs
# -------------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
# Stimulus
X[e["stimulus"], choice] = scale(+coh)
X[e["stimulus"], 1 - choice] = scale(-coh)
# Start cue
X[e["stimulus"], START] = 1
trial["inputs"] = X
# -------------------------------------------------------------------------------------
# Target output
# -------------------------------------------------------------------------------------
if params.get("target_output", False):
Y = np.zeros((len(t), Nout)) # Output matrix
M = np.zeros_like(Y) # Mask matrix
# Hold values
hi = 1.2
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e["fixation"], :] = lo
# Decision
Y[e["decision"], choice] = hi
Y[e["decision"], 1 - choice] = lo
# Only care about fixation and decision periods
M[e["fixation"] + e["decision"], :] = 1
# Outputs and mask
trial["outputs"] = Y
trial["mask"] = M
# -------------------------------------------------------------------------------------
return trial
def run_trials(p, args):
# Model
m = p['model']
# Number of trials
try:
ntrials = int(args[0])
except:
ntrials = 100
ntrials *= m.nconditions
# RNN
rng = np.random.RandomState(p['seed'])
rnn = RNN(p['savefile'], {'dt': p['dt']}, verbose=False)
w = len(str(ntrials))
trials = []
backspaces = 0
try:
for i in xrange(ntrials):
# Condition
k = tasktools.unravel_index(i % m.nconditions,
(len(m.cohs), len(m.left_rights),
len(m.cohs), len(m.left_rights),
len(m.contexts)))
coh_m = m.cohs[k[0]]
left_right_m = m.left_rights[k[1]]
coh_c = m.cohs[k[2]]
left_right_c = m.left_rights[k[3]]
context = m.contexts[k[4]]
# Trial
trial_func = m.generate_trial
trial_args = {
'name': 'test',
'catch': False,
'coh_m': coh_m,
'left_right_m': left_right_m,
'coh_c': coh_c,
'left_right_c': left_right_c,
'context': context
}
info = rnn.run(inputs=(trial_func, trial_args), rng=rng)
# Display trial type
s = ("Trial {:>{}}/{}: ({}) m{:>+3}, c{:>+3}"
.format(i+1, w, ntrials, info['context'],
info['left_right_m']*info['coh_m'],
info['left_right_c']*info['coh_c']))
sys.stdout.write(backspaces*'\b' + s)
sys.stdout.flush()
backspaces = len(s)
# Save
dt = rnn.t[1] - rnn.t[0]
step = int(p['dt_save']/dt)
trial = {
't': rnn.t[::step],
'u': rnn.u[:,::step],
'r': rnn.r[:,::step],
'z': rnn.z[:,::step],
'info': info,
}
trials.append(trial)
except KeyboardInterrupt:
pass
print("")
# Save all
filename = get_trialsfile(p)
with open(filename, 'wb') as f:
pickle.dump(trials, f, pickle.HIGHEST_PROTOCOL)
size = os.path.getsize(filename)*1e-9
print("[ {}.run_trials ] Trials saved to {} ({:.1f} GB)".format(THIS, filename, size))
# Compute the psychometric function
psychometric_function(filename)
def generate_trial(rng, dt, params):
# -------------------------------------------------------------------------------------
# Select task condition
# -------------------------------------------------------------------------------------
catch_trial = False
if params["name"] in ["gradient", "test"]:
if params.get("catch", rng.rand() < pcatch):
catch_trial = True
else:
# Context
context = params.get("context", rng.choice(contexts))
# Coherences
coh_m = params.get("coh_m", rng.choice(cohs))
coh_c = params.get("coh_c", rng.choice(cohs))
# Left/right
left_right_m = params.get("left_right_m", rng.choice(left_rights))
left_right_c = params.get("left_right_c", rng.choice(left_rights))
elif params["name"] == "validation":
b = params["minibatch_index"] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k = tasktools.unravel_index(
b - 1, (len(contexts), len(cohs), len(cohs), len(left_rights), len(left_rights))
)
context = contexts[k[0]]
coh_m = cohs[k[1]]
coh_c = cohs[k[2]]
left_right_m = left_rights[k[3]]
left_right_c = left_rights[k[4]]
else:
raise ValueError("Unknown trial type.")
# -------------------------------------------------------------------------------------
# Epochs
# -------------------------------------------------------------------------------------
if catch_trial:
epochs = {"T": 2000}
else:
if params["name"] == "test":
fixation = 400
else:
fixation = 100
stimulus = 800
decision = 300
T = fixation + stimulus + decision
epochs = {
"fixation": (0, fixation),
"stimulus": (fixation, fixation + stimulus),
"decision": (fixation + stimulus, T),
}
epochs["T"] = T
# -------------------------------------------------------------------------------------
# Trial info
# -------------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(dt, epochs) # Time, task epochs in discrete time
trial = {"t": t, "epochs": epochs} # Trial
if catch_trial:
trial["info"] = {}
else:
if context == "m":
left_right = left_right_m
else:
left_right = left_right_c
# Correct choice
if left_right > 0:
choice = 0
else:
choice = 1
# Trial info
trial["info"] = {
"coh_m": coh_m,
"left_right_m": left_right_m,
"coh_c": coh_c,
"left_right_c": left_right_c,
"context": context,
"choice": choice,
}
# -------------------------------------------------------------------------------------
# Inputs
# -------------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
# Context
if context == "m":
X[e["stimulus"], 0] = 1
else:
X[e["stimulus"], 1] = 1
# Motion stimulus
if left_right_m > 0:
choice_m = 0
else:
choice_m = 1
X[e["stimulus"], 2 + choice_m] = scale(+coh_m)
X[e["stimulus"], 2 + (1 - choice_m)] = scale(-coh_m)
# Colour stimulus
if left_right_c > 0:
choice_c = 0
else:
choice_c = 1
X[e["stimulus"], 4 + choice_c] = scale(+coh_c)
X[e["stimulus"], 4 + (1 - choice_c)] = scale(-coh_c)
trial["inputs"] = X
# -------------------------------------------------------------------------------------
# Target output
# -------------------------------------------------------------------------------------
if params.get("target_output", False):
Y = np.zeros((len(t), Nout)) # Output matrix
M = np.zeros_like(Y) # Mask matrix
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e["fixation"], :] = lo
# Decision
Y[e["decision"], choice] = hi
Y[e["decision"], 1 - choice] = lo
# Only care about fixation and decision periods
M[e["fixation"] + e["decision"], :] = 1
# Outputs and mask
trial["outputs"] = Y
trial["mask"] = M
# -------------------------------------------------------------------------------------
return trial
def run_trials(p, args): #args are the number of trials
"""
Run trials.
"""
# Model
m = p['model']
# Number of trials
try:
ntrials = int(args[0])
except:
ntrials = 100
ntrials *= m.nconditions
# RNN
rng = np.random.RandomState(p['seed'])
rnn = RNN(p['savefile'], {'dt': p['dt']}, verbose=False)
# Trials
w = len(str(ntrials))
trials = []
backspaces = 0
try:
for i in xrange(ntrials):
# Condition
b = i % m.nconditions #iterate through all conditions
k0, k1 = tasktools.unravel_index(b, (len(m.pairs), 1))
pair = m.pairs[k0]
# Trial
trial_func = m.generate_trial
trial_args = {
'name': 'test',
'catch': False,
'pair': pair,
}
info = rnn.run(inputs=(trial_func, trial_args), rng=rng)
# Display trial type
s = ("Trial {:>{}}/{}: {:>2} {:>2}"
.format(i+1, w, ntrials, info['f1'], info['f2']))
sys.stdout.write(backspaces*'\b' + s)
sys.stdout.flush()
backspaces = len(s)
# Save
dt = rnn.t[1] - rnn.t[0]
step = int(p['dt_save']/dt)
trial = {
't': rnn.t[::step],
'u': rnn.u[:,::step],
'r': rnn.r[:,::step],
'z': rnn.z[:,::step],
'info': info
}
trials.append(trial)
except KeyboardInterrupt:
pass
print("")
# Save all
filename = get_trialsfile(p)
with open(filename, 'wb') as f:
pickle.dump(trials, f, pickle.HIGHEST_PROTOCOL)
size = os.path.getsize(filename)*1e-9
print("[ {}.run_trials ] Trials saved to {} ({:.1f} GB)".format(THIS, filename, size))
# Psychometric function
psychometric_function(filename)
def generate_trial(rng, dt, params):
#-------------------------------------------------------------------------------------
# Select task condition
#-------------------------------------------------------------------------------------
catch_trial = False
if params['name'] in ['gradient', 'test']:
if params.get('catch', rng.rand() < pcatch):
catch_trial = True
else:
coh = params.get('coh', rng.choice(cohs))
in_out = params.get('in_out', rng.choice(in_outs))
elif params['name'] == 'validation':
b = params['minibatch_index'] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k0, k1 = tasktools.unravel_index(b - 1, (len(cohs), len(in_outs)))
coh = cohs[k0]
in_out = in_outs[k1]
else:
raise ValueError("Unknown trial type.")
#-------------------------------------------------------------------------------------
# Epochs
#-------------------------------------------------------------------------------------
if catch_trial:
epochs = {'T': 2000}
else:
if params['name'] == 'test':
fixation = 300
else:
fixation = 100
stimulus = 800
decision = 300
T = fixation + stimulus + decision
epochs = {
'fixation': (0, fixation),
'stimulus': (fixation, fixation + stimulus),
'decision': (fixation + stimulus, T)
}
epochs['T'] = T
#-------------------------------------------------------------------------------------
# Trial info
#-------------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(
dt, epochs) # Time, task epochs in discrete time
trial = {'t': t, 'epochs': epochs} # Trial
if catch_trial:
trial['info'] = {}
else:
# Correct choice
if in_out > 0:
choice = 0
else:
choice = 1
# Trial info
trial['info'] = {'coh': coh, 'in_out': in_out, 'choice': choice}
#-------------------------------------------------------------------------------------
# Inputs
#-------------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
X[e['stimulus'], choice] = scale(+coh)
X[e['stimulus'], 1 - choice] = scale(-coh)
trial['inputs'] = X
#-------------------------------------------------------------------------------------
# Target output
#-------------------------------------------------------------------------------------
if params.get('target_output', False):
Y = np.zeros((len(t), Nout)) # Output
M = np.zeros_like(Y) # Mask
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e['fixation'], :] = lo
# Decision
Y[e['decision'], choice] = hi
Y[e['decision'], 1 - choice] = lo
# Only care about fixation and decision periods
M[e['fixation'] + e['decision'], :] = 1
# Outputs and mask
trial['outputs'] = Y
trial['mask'] = M
#-------------------------------------------------------------------------------------
return trial
def generate_trial(rng, dt, params):
#---------------------------------------------------------------------------------
# Select task condition
#---------------------------------------------------------------------------------
catch_trial = False
if params['name'] in ['gradient', 'test']:
if params.get('catch', rng.rand() < pcatch): #pcatch probability of catch trial
catch_trial = True
else:
pair = params.get('pair', pairs[rng.choice(len(pairs))]) #random pair
#gt_lt = params.get('gt_lt', rng.choice(gt_lts)) #random order
elif params['name'] == 'validation':
b = params['minibatch_index'] % (nconditions + 1) #0 to 4
if b == 0:
catch_trial = True
else: #b-1 is 0 to 3, len(pairs) = 0 to 1
k0, k1 = tasktools.unravel_index(b-1, (len(pairs),1))#len(gt_lts)))
pair = pairs[b-1]
#gt_lt = gt_lts[k1]
else:
raise ValueError("Unknown trial type.")
#---------------------------------------------------------------------------------
# Epochs
#---------------------------------------------------------------------------------
if catch_trial:
epochs = {'T': 2500}
else:
if params['name'] == 'test':
fixation = 500
else:
fixation = 100
f1 = 500
if params['name'] == 'test':
delay = 3000
else:
delay = 3000
f2 = 500
decision = 300
T = fixation + f1 + delay + f2 + decision
epochs = {
'fixation': (0, fixation),
'f1': (fixation, fixation + f1),
'delay': (fixation + f1, fixation + f1 + delay),
'f2': (fixation + f1 + delay, fixation + f1 + delay + f2),
'decision': (fixation + f1 + delay + f2, T)
}
epochs['T'] = T
#---------------------------------------------------------------------------------
# Trial info
#---------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(dt, epochs) # Task epochs in discrete time
trial = {'t': t, 'epochs': epochs} # Trial
if catch_trial:
trial['info'] = {}
else:
a1, a2 = pair
# Correct choice
if a1 == a2:
choice = 0
else:
choice = 1
# Info
trial['info'] = {'f1': a1, 'f2': a2, 'choice': choice}
#---------------------------------------------------------------------------------
# Inputs
#---------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
# Stimulus 1
X[e['f1'],POS] = scale_p(a1)
#X[e['f1'],NEG] = scale_n(f1)
# Stimulus 2
X[e['f2'],POS] = scale_p(a2)
#X[e['f2'],NEG] = scale_n(f2)
trial['inputs'] = X
#---------------------------------------------------------------------------------
# Target output
#---------------------------------------------------------------------------------
if params.get('target_output', False):
Y = np.zeros((len(t), Nout)) # Output matrix by time
M = np.zeros_like(Y) # Mask matrix by time
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e['fixation'],:] = lo #while fixating it should not be spiking
# Decision
Y[e['decision'],choice] = hi #output corresponding to choice = hi
Y[e['decision'],1-choice] = lo #other one goes to lo
# Mask
M[e['fixation']+e['decision'],:] = 1
trial['outputs'] = Y
trial['mask'] = M
#---------------------------------------------------------------------------------
return trial
def run_trials(p, args):
"""
Run trials.
"""
# Model
m = p['model']
# Number of trials
try:
ntrials = int(args[0])
except:
ntrials = 100
ntrials *= m.nconditions + 1
# RNN
rng = np.random.RandomState(p['seed'])
rnn = RNN(p['savefile'], {'dt': p['dt']}, verbose=False)
# Trials
w = len(str(ntrials))
trials = []
backspaces = 0
try:
for i in xrange(ntrials):
b = i % (m.nconditions + 1)
if b == 0:
# Zero-coherence condition
coh = 0
in_out = rng.choice(m.in_outs)
else:
# All other conditions
k1, k2 = tasktools.unravel_index(b - 1,
(len(m.cohs), len(m.in_outs)))
coh = m.cohs[k1]
in_out = m.in_outs[k2]
# Trial
trial_func = m.generate_trial
trial_args = {
'name': 'test',
'catch': False,
'coh': coh,
'in_out': in_out
}
info = rnn.run(inputs=(trial_func, trial_args), rng=rng)
# Display trial type
if coh == 0:
s = "Trial {:>{}}/{}: {:>3}".format(i + 1, w, ntrials,
info['coh'])
else:
s = ("Trial {:>{}}/{}: {:>+3}".format(
i + 1, w, ntrials, info['in_out'] * info['coh']))
sys.stdout.write(backspaces * '\b' + s)
sys.stdout.flush()
backspaces = len(s)
# Save
dt = rnn.t[1] - rnn.t[0]
step = int(p['dt_save'] / dt)
trial = {
't': rnn.t[::step],
'u': rnn.u[:, ::step],
'r': rnn.r[:, ::step],
'z': rnn.z[:, ::step],
'info': info
}
trials.append(trial)
except KeyboardInterrupt:
pass
print("")
# Save all
filename = get_trialsfile(p)
with open(filename, 'wb') as f:
pickle.dump(trials, f, pickle.HIGHEST_PROTOCOL)
size = os.path.getsize(filename) * 1e-9
print("[ {}.run_trials ] Trials saved to {} ({:.1f} GB)".format(
THIS, filename, size))
# Compute the psychometric function
psychometric_function(filename)
def generate_trial(rng, dt, params):
#---------------------------------------------------------------------------------
# Select task condition
#---------------------------------------------------------------------------------
catch_trial = False
if params['name'] in ['gradient', 'test']:
if params.get('catch', rng.rand() < pcatch):
catch_trial = True
else:
fpair = params.get('fpair', fpairs[rng.choice(len(fpairs))])
gt_lt = params.get('gt_lt', rng.choice(gt_lts))
elif params['name'] == 'validation':
b = params['minibatch_index'] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k0, k1 = tasktools.unravel_index(b - 1, (len(fpairs), len(gt_lts)))
fpair = fpairs[k0]
gt_lt = gt_lts[k1]
else:
raise ValueError("Unknown trial type.")
#---------------------------------------------------------------------------------
# Epochs
#---------------------------------------------------------------------------------
if catch_trial:
epochs = {'T': 2500}
else:
if params['name'] == 'test':
fixation = 500
else:
fixation = 100
f1 = 500
if params['name'] == 'test':
delay = 3000
else:
delay = tasktools.uniform(rng, dt, 2500, 3500)
f2 = 500
decision = 300
T = fixation + f1 + delay + f2 + decision
epochs = {
'fixation': (0, fixation),
'f1': (fixation, fixation + f1),
'delay': (fixation + f1, fixation + f1 + delay),
'f2': (fixation + f1 + delay, fixation + f1 + delay + f2),
'decision': (fixation + f1 + delay + f2, T)
}
epochs['T'] = T
#---------------------------------------------------------------------------------
# Trial info
#---------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(dt, epochs) # Task epochs in discrete time
trial = {'t': t, 'epochs': epochs} # Trial
if catch_trial:
trial['info'] = {}
else:
# Correct choice
if gt_lt == '>':
f1, f2 = fpair
choice = 0
else:
f2, f1 = fpair
choice = 1
# Info
trial['info'] = {'f1': f1, 'f2': f2, 'choice': choice}
#---------------------------------------------------------------------------------
# Inputs
#---------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
# Stimulus 1
X[e['f1'], POS] = scale_p(f1)
X[e['f1'], NEG] = scale_n(f1)
# Stimulus 2
X[e['f2'], POS] = scale_p(f2)
X[e['f2'], NEG] = scale_n(f2)
trial['inputs'] = X
#---------------------------------------------------------------------------------
# Target output
#---------------------------------------------------------------------------------
if params.get('target_output', False):
Y = np.zeros((len(t), Nout)) # Output matrix
M = np.zeros_like(Y) # Mask matrix
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e['fixation'], :] = lo
# Decision
Y[e['decision'], choice] = hi
Y[e['decision'], 1 - choice] = lo
# Mask
M[e['fixation'] + e['decision'], :] = 1
trial['outputs'] = Y
trial['mask'] = M
#---------------------------------------------------------------------------------
return trial
def generate_trial(rng, dt, params):
#-------------------------------------------------------------------------------------
# Select task condition
#-------------------------------------------------------------------------------------
catch_trial = False
if params['name'] in ['gradient', 'test']:
if params.get('catch', rng.rand() < pcatch):
catch_trial = True
else:
coh = params.get('coh', rng.choice(cohs))
in_out = params.get('in_out', rng.choice(in_outs))
elif params['name'] == 'validation':
b = params['minibatch_index'] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k0, k1 = tasktools.unravel_index(b-1, (len(cohs), len(in_outs)))
coh = cohs[k0]
in_out = in_outs[k1]
else:
raise ValueError("Unknown trial type.")
#-------------------------------------------------------------------------------------
# Epochs
#-------------------------------------------------------------------------------------
if catch_trial:
epochs = {'T': 2000}
else:
if params['name'] == 'test':
fixation = 300
else:
fixation = 100
stimulus = 800
decision = 300
T = fixation + stimulus + decision
epochs = {
'fixation': (0, fixation),
'stimulus': (fixation, fixation + stimulus),
'decision': (fixation + stimulus, T)
}
epochs['T'] = T
#-------------------------------------------------------------------------------------
# Trial info
#-------------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(dt, epochs) # Time, task epochs in discrete time
trial = {'t': t, 'epochs': epochs} # Trial
if catch_trial:
trial['info'] = {}
else:
# Correct choice
if in_out > 0:
choice = 0
else:
choice = 1
# Trial info
trial['info'] = {'coh': coh, 'in_out': in_out, 'choice': choice}
#-------------------------------------------------------------------------------------
# Inputs
#-------------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
X[e['stimulus'],choice] = scale(+coh)
X[e['stimulus'],1-choice] = scale(-coh)
trial['inputs'] = X
#-------------------------------------------------------------------------------------
# Target output
#-------------------------------------------------------------------------------------
if params.get('target_output', False):
Y = np.zeros((len(t), Nout)) # Output
M = np.zeros_like(Y) # Mask
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e['fixation'],:] = lo
# Decision
Y[e['decision'],choice] = hi
Y[e['decision'],1-choice] = lo
# Only care about fixation and decision periods
M[e['fixation']+e['decision'],:] = 1
# Outputs and mask
trial['outputs'] = Y
trial['mask'] = M
#-------------------------------------------------------------------------------------
return trial
def generate_trial(rng, dt, params):
#-------------------------------------------------------------------------------------
# Select task condition
#-------------------------------------------------------------------------------------
catch_trial = False
if params['name'] in ['gradient', 'test']:
if params.get('catch', rng.rand() < pcatch):
catch_trial = True
else:
# Context
context = params.get('context', rng.choice(contexts))
# Coherences
coh_m = params.get('coh_m', rng.choice(cohs))
coh_c = params.get('coh_c', rng.choice(cohs))
# Left/right
left_right_m = params.get('left_right_m', rng.choice(left_rights))
left_right_c = params.get('left_right_c', rng.choice(left_rights))
elif params['name'] == 'validation':
b = params['minibatch_index'] % (nconditions + 1)
if b == 0:
catch_trial = True
else:
k = tasktools.unravel_index(b-1, (len(contexts),
len(cohs), len(cohs),
len(left_rights), len(left_rights)))
context = contexts[k[0]]
coh_m = cohs[k[1]]
coh_c = cohs[k[2]]
left_right_m = left_rights[k[3]]
left_right_c = left_rights[k[4]]
else:
raise ValueError("Unknown trial type.")
#-------------------------------------------------------------------------------------
# Epochs
#-------------------------------------------------------------------------------------
if catch_trial:
epochs = {'T': 2000}
else:
if params['name'] == 'test':
fixation = 400
else:
fixation = 100
stimulus = 800
decision = 300
T = fixation + stimulus + decision
epochs = {
'fixation': (0, fixation),
'stimulus': (fixation, fixation + stimulus),
'decision': (fixation + stimulus, T)
}
epochs['T'] = T
#-------------------------------------------------------------------------------------
# Trial info
#-------------------------------------------------------------------------------------
t, e = tasktools.get_epochs_idx(dt, epochs) # Time, task epochs in discrete time
trial = {'t': t, 'epochs': epochs} # Trial
if catch_trial:
trial['info'] = {}
else:
if context == 'm':
left_right = left_right_m
else:
left_right = left_right_c
# Correct choice
if left_right > 0:
choice = 0
else:
choice = 1
# Trial info
trial['info'] = {
'coh_m': coh_m,
'left_right_m': left_right_m,
'coh_c': coh_c,
'left_right_c': left_right_c,
'context': context,
'choice': choice
}
#-------------------------------------------------------------------------------------
# Inputs
#-------------------------------------------------------------------------------------
X = np.zeros((len(t), Nin))
if not catch_trial:
# Context
if context == 'm':
X[e['stimulus'],0] = 1
else:
X[e['stimulus'],1] = 1
# Motion stimulus
if left_right_m > 0:
choice_m = 0
else:
choice_m = 1
X[e['stimulus'],2+choice_m] = scale(+coh_m)
X[e['stimulus'],2+(1-choice_m)] = scale(-coh_m)
# Colour stimulus
if left_right_c > 0:
choice_c = 0
else:
choice_c = 1
X[e['stimulus'],4+choice_c] = scale(+coh_c)
X[e['stimulus'],4+(1-choice_c)] = scale(-coh_c)
trial['inputs'] = X
#-------------------------------------------------------------------------------------
# Target output
#-------------------------------------------------------------------------------------
if params.get('target_output', False):
Y = np.zeros((len(t), Nout)) # Output matrix
M = np.zeros_like(Y) # Mask matrix
# Hold values
hi = 1
lo = 0.2
if catch_trial:
Y[:] = lo
M[:] = 1
else:
# Fixation
Y[e['fixation'],:] = lo
# Decision
Y[e['decision'],choice] = hi
Y[e['decision'],1-choice] = lo
# Only care about fixation and decision periods
M[e['fixation']+e['decision'],:] = 1
# Outputs and mask
trial['outputs'] = Y
trial['mask'] = M
#-------------------------------------------------------------------------------------
return trial
