def run():
defaults = PaperDefaults()
#David's globals
size = 51
csize = 9
npoints = 32
scale = 2.0
cval = 0.5
csvfiles = [[ defaults._DATADIR + \
'/TB2015_%i_%s.csv' % (i, s) \
for i in range(-90, 90, 30)] for s in ('PS', 'PO')
]
# experiment parameters
ppop = {
'kind': 'circular',
'npoints': npoints,
'scale': scale,
'fdomain': (0, 1),
}
vals_ang = sp.array([-90., -60., -30., 0., 30., 60.])
vals = (vals_ang + 90.) / 180.
imc1 = stim.get_center_surround(size=size,
csize=csize,
cval=cval,
sval=sp.nan)
x1 = model_utils.get_population(imc1, **ppop)
x = sp.zeros((2, len(vals), npoints, size, size))
for vdx, v in enumerate(vals):
imc2 = stim.get_center_surround(size=size,
csize=csize,
cval=v,
sval=sp.nan)
ims = stim.get_center_surround(size=size,
csize=csize,
cval=sp.nan,
sval=v)
x2 = model_utils.get_population(imc2, **ppop)
xs = model_utils.get_population(ims, **ppop)
x[0, vdx] = (x1 + x2) / 2.
x[1, vdx] = (x1 + x2) / 2. + xs
x.shape = (2 * len(vals), npoints, size, size)
# trott and born 2015 data
gt = get_gt(npoints, csvfiles)
extra_vars = {}
extra_vars['scale'] = scale
extra_vars['npoints'] = npoints
extra_vars['cval'] = cval
extra_vars['size'] = size
extra_vars['csize'] = csize
extra_vars['vals'] = vals
extra_vars['figure_name'] = 'tbp'
extra_vars['return_var'] = 'O'
optimize_model(x, gt, extra_vars, defaults)
def run(hps=None):
defaults = PaperDefaults()
#David's globals
size=51
csize=9
npoints=37
scale=1.
_DEFAULT_BWC_CSV_CTS = sp.array([0.0, .06, .12, .25, .50]) * 100
csvfiles=sp.array([[os.path.join(defaults._DATADIR, 'BWC2009_%i_%i.csv' \
% (i, j)) for i in _DEFAULT_BWC_CSV_CTS] for j in _DEFAULT_BWC_CSV_CTS]).T
# experiment parameters
im = sp.array([
stim.get_center_surround(
size=size, csize=csize, cval=.25, sval=sp.nan),
stim.get_center_surround(
size=size, csize=csize, cval=.75, sval=sp.nan)])
# populations for vertical (masking) and horizontal (driving) stimuli
#####################################################################
xv = model_utils.get_population(im[0],
kind='circular', npoints=npoints, scale=scale)
xh = model_utils.get_population(im[1],
kind='circular', npoints=npoints, scale=scale)
# superimposed populations
##########################
v_contrasts = [0.0, .06, .12, .25, .50]
h_contrasts = [0.0, .06, .12, .25, .50]
nv, nh = len(v_contrasts), len(h_contrasts)
x = sp.array([[h*xh + v*xv for h in h_contrasts] for v in v_contrasts])
x.shape = (nv * nh,) + x.shape[2:]
# busse and wade data
t_paper = sp.zeros((nv, nh, 13))
y_paper = sp.zeros((nv, nh, 13))
for idx in range(nv):
for jdx in range(nh):
t_paper[idx, jdx], y_paper[idx, jdx] = \
sp.genfromtxt(csvfiles[idx, jdx], delimiter=',').T
res_y_paper = sp.zeros((y_paper.shape[0],y_paper.shape[1],npoints))
for r in range(y_paper.shape[0]):
for c in range(y_paper.shape[1]):
res_y_paper[r,c,:] = sp.signal.resample(y_paper[r,c,:],npoints)
gt = [t_paper,res_y_paper]
extra_vars = {}
extra_vars['scale'] = scale
extra_vars['npoints'] = npoints
extra_vars['size'] = size
extra_vars['csize'] = csize
extra_vars['nv'] = nv
extra_vars['nh'] = nh
extra_vars['figure_name'] = 'bw'
extra_vars['return_var'] = 'O'
optimize_model(x,gt,extra_vars,defaults)
def run():
defaults = PaperDefaults()
#David's globals
_DEFAULT_TILTEFFECT_DEGPERPIX = .25 # <OToole77>
_DEFAULT_TILTEFFECT_SIZE = 51 #101
_DEFAULT_TILTEFFECT_CSIZE = iround(2. / _DEFAULT_TILTEFFECT_DEGPERPIX)
_DEFAULT_TILTEFFECT_SSIZE = iround(8. / _DEFAULT_TILTEFFECT_DEGPERPIX)
_DEFAULT_TILTEFFECT_CVAL = .5
_DEFAULT_TILTEFFECT_SVALS = np.linspace(0.0, 0.5, 10)
_DEFAULT_TILTEFFECT_SCALES = {'ow77': 0.40, 'ms79': 0.60} #0.45
_DEFAULT_TILTEFFECT_NPOINTS = 25 #100
_DEFAULT_TILTEFFECT_DECODER_TYPE = 'circular_vote'
_DEFAULT_TILTEFFECT_CSV = {
'ow77': os.path.join(defaults._DATADIR, 'OW_fig4_Black.csv'),
'ms79': os.path.join(defaults._DATADIR, 'MS1979.csv'),
}
# experiment parameters
cpt = (_DEFAULT_TILTEFFECT_SIZE // 2, _DEFAULT_TILTEFFECT_SIZE // 2)
spt = (_DEFAULT_TILTEFFECT_SIZE // 2,
_DEFAULT_TILTEFFECT_SIZE // 2 + _DEFAULT_TILTEFFECT_CSIZE)
dt_in = _DEFAULT_TILTEFFECT_CVAL - _DEFAULT_TILTEFFECT_SVALS
# simulate populations
im = sp.array([[
stim.get_center_nfsurrounds(size=_DEFAULT_TILTEFFECT_SIZE,
csize=_DEFAULT_TILTEFFECT_CSIZE,
nsize=_DEFAULT_TILTEFFECT_CSIZE,
fsize=_DEFAULT_TILTEFFECT_SSIZE,
cval=_DEFAULT_TILTEFFECT_CVAL,
nval=_DEFAULT_TILTEFFECT_CVAL,
fval=sval,
bgval=sp.nan)
] for sval in _DEFAULT_TILTEFFECT_SVALS])
# get shifts for model for both papers, and from digitized data
sortidx = sp.argsort(dt_in) # re-order in increasing angular differences
# O'Toole and Wenderoth (1977)
_, ds_ow77_paper_y = sp.genfromtxt(_DEFAULT_TILTEFFECT_CSV['ow77'],
delimiter=',').T
extra_vars = {}
extra_vars['scale'] = _DEFAULT_TILTEFFECT_SCALES['ow77']
extra_vars['decoder'] = _DEFAULT_TILTEFFECT_DECODER_TYPE
extra_vars['npoints'] = _DEFAULT_TILTEFFECT_NPOINTS
extra_vars['npoints'] = _DEFAULT_TILTEFFECT_NPOINTS
extra_vars['cval'] = _DEFAULT_TILTEFFECT_CVAL
extra_vars['sortidx'] = sortidx
extra_vars['cpt'] = cpt
extra_vars['spt'] = spt
extra_vars['sval'] = sval
extra_vars['kind'] = 'circular'
extra_vars['figure_name'] = 'f3a'
extra_vars['return_var'] = 'O'
extra_vars['hp_file'] = os.path.join(defaults._FIGURES, 'best_hps.npz')
optimize_model(im, ds_ow77_paper_y, extra_vars, defaults)
def run():
defaults = PaperDefaults()
#David's globals
size = 51
csize = 5
npoints = 64
scale = 2.0
neuron_theta = 0.50
cval = 0.5
csvfiles = [ defaults._DATADIR + \
'/TB2015_Fig1B_%s.csv' % (s,) \
for s in range(-90, 90, 30) + ['CO']]
# experiment parameters
cvals = (sp.arange(-90, 90, 30) + 90.) / 180.
svals = sp.linspace(0.0, 1.0, 6).tolist() + [sp.nan]
neuron_thetas = sp.linspace(0.0, 1.0, npoints)
neuron_idx = sp.argmin(sp.absolute(neuron_thetas - neuron_theta))
stims = [
stim.get_center_surround(size=size, csize=csize, cval=cv, sval=sv)
for cv in cvals for sv in svals
]
x = sp.array([
model_utils.get_population(im,
npoints=npoints,
kind='circular',
scale=scale,
fdomain=(0, 1)) for im in stims
])
# [Array shapes]
# trott and born 2015 data
gt = get_gt(csvfiles)
extra_vars = {}
extra_vars['scale'] = scale
extra_vars['npoints'] = npoints
extra_vars['cval'] = cval
extra_vars['cvals'] = cvals
extra_vars['svals'] = svals
extra_vars['size'] = size
extra_vars['csize'] = csize
extra_vars['neuron_idx'] = neuron_idx
extra_vars['figure_name'] = 'tbtcso'
extra_vars['return_var'] = 'O'
extra_vars['hp_file'] = os.path.join(defaults._FIGURES, 'best_hps.npz')
optimize_model(x, gt, extra_vars, defaults)
def run():
defaults = PaperDefaults()
#David's globals
size = 51
npoints = 64
cval1 = 0.25
cval2 = 0.75
sval = 0.75
test_contrasts = sp.array([0., 8., 32.])
mask_contrasts = sp.array([0., 8., 32.])
# experiment parameters
idx1 = int(cval1 * npoints)
idx2 = int(cval2 * npoints)
# simulate populations
imc = stim.get_center_surround(size=size, csize=9, cval=cval1, sval=sp.nan)
ims = stim.get_center_surround(size=size, csize=9, cval=sp.nan, sval=sval)
x1 = utils.get_population(imc,
npoints=npoints,
kind='gaussian',
scale=0.1,
fdomain=(0, 1))
x2 = sp.roll(x1, int((cval2 - cval1) * npoints), axis=-3)
xs = utils.get_population(ims,
npoints=npoints,
kind='gaussian',
scale=0.1,
fdomain=(0, 1))
x = []
for k1 in test_contrasts:
for k2 in mask_contrasts:
x.append(k1 / 100. * x1 + k2 / 100. * x2)
x = sp.array(x) + sp.array([xs])
# Experimental data
extra_vars = {}
extra_vars['size'] = size
extra_vars['npoints'] = npoints
extra_vars['sval'] = sval
extra_vars['figure_name'] = 'cross_orientation_suppression'
extra_vars['return_var'] = 'O'
extra_vars['idx1'] = idx1
extra_vars['idx2'] = idx2
extra_vars['test_contrasts'] = test_contrasts
extra_vars['mask_contrasts'] = mask_contrasts
optimize_model(x, [], extra_vars, defaults)
def run():
defaults = PaperDefaults()
# David's globals
size = 51
nr = 17
npoints = 32 // 2
ncontrasts = 5
# experiment parameters
# generate stimuli
##################
im = []
for k in range(nr):
im_ = sp.zeros((size, size)) + sp.nan
im_[size // 2 - k:size // 2 + k + 1,
size // 2 - k:size // 2 + k + 1] = 0.5
im.append(im_)
im = sp.array(im)
# generate populations
######################
contrasts = sp.linspace(1., 0., ncontrasts, endpoint=False)[::-1]
# contrasts = sp.logspace(-2, 0., ncontrasts)
x = sp.array([
utils.get_population(xim_, 'gaussian', npoints=npoints) for xim_ in im
])
ax = [c * x for c in contrasts]
cx = np.concatenate(ax[:], axis=0)
# Experimental data
extra_vars = {}
extra_vars['size'] = size
extra_vars['npoints'] = npoints
extra_vars['nr'] = nr
extra_vars['stimsizes'] = 2 * sp.arange(nr) + 1
extra_vars['ssn'] = defaults._DEFAULT_PARAMETERS['ssn']
extra_vars['ssf'] = defaults._DEFAULT_PARAMETERS['ssf']
extra_vars['hp_file'] = os.path.join(defaults._FIGURES, 'best_hps.npz')
extra_vars['figure_name'] = 'size_tuning'
extra_vars['return_var'] = 'O'
extra_vars['contrasts'] = contrasts
extra_vars['curvecols'] = sns.cubehelix_palette(ncontrasts)
extra_vars['curvelabs'] = [
'Single-cell response at contrast %g' % (cst, ) for cst in contrasts
]
optimize_model(cx, None, extra_vars, defaults)
def run(create_stim=True):
defaults = PaperDefaults()
# David's globals
_DEFAULT_TILTEFFECT_DEGPERPIX = .25 # <OToole77>
_DEFAULT_TILTEFFECT_SIZE = 51 # 101
_DEFAULT_TILTEFFECT_CSIZE = iround(2. / _DEFAULT_TILTEFFECT_DEGPERPIX)
_DEFAULT_TILTEFFECT_CVAL = .5
_DEFAULT_TILTEFFECT_SCALES = {'ow77': 0.40, 'ms79': 0.60} # 0.45
_DEFAULT_TILTEFFECT_NPOINTS = 25 # 100
_DEFAULT_TILTEFFECT_DECODER_TYPE = 'circular_vote'
# experiment parameters
cpt = (_DEFAULT_TILTEFFECT_SIZE // 2, _DEFAULT_TILTEFFECT_SIZE // 2)
spt = (_DEFAULT_TILTEFFECT_SIZE // 2,
_DEFAULT_TILTEFFECT_SIZE // 2 + _DEFAULT_TILTEFFECT_CSIZE)
# simulate populations
fl = 'conv2_2'
if create_stim:
sys.path.append('../../')
from MIRC_tests import features_vgg16
im, im_names = features_vgg16.baseline_vgg16(
images='/home/drew/Desktop/nrsa_png',
num_images=25,
feature_layer='content_vgg/' + fl + '/Relu:0',
im_ext='.png',
batch_size=1)
target_size = [
400, 400
] # [_DEFAULT_TILTEFFECT_SIZE, _DEFAULT_TILTEFFECT_SIZE]
im = normalize(resize(im, target_size=target_size).astype(float))
np.save('raw_' + fl, im)
extra_vars = {}
extra_vars['scale'] = _DEFAULT_TILTEFFECT_SCALES['ow77']
extra_vars['decoder'] = _DEFAULT_TILTEFFECT_DECODER_TYPE
extra_vars['npoints'] = _DEFAULT_TILTEFFECT_NPOINTS
extra_vars['npoints'] = _DEFAULT_TILTEFFECT_NPOINTS
extra_vars['cval'] = _DEFAULT_TILTEFFECT_CVAL
extra_vars['cpt'] = cpt
extra_vars['spt'] = spt
extra_vars['kind'] = 'circular'
extra_vars['figure_name'] = 'cnn_features'
extra_vars['return_var'] = 'O'
extra_vars['save_file'] = 'proc_' + fl
optimize_model(im, None, extra_vars, defaults)
from ops.db_utils import init_db, generate_combos, create_and_execute_daemons, prepare_settings from ops.parameter_defaults import PaperDefaults import sys defaults = PaperDefaults() print 'Initializing database' init_db(sys.argv[1]) print 'Generating initial ' + sys.argv[2] + ' hyperparameter combos' generate_combos(int(sys.argv[2]))
def run():
defaults = PaperDefaults()
#David's globals
_DEFAULT_KW97_TILTEFFECT_DEGPERPIX = .45 # <OToole77>
_DEFAULT_TILTEFFECT_SIZE = 101 #101
_DEFAULT_KW97_TILTEFFECT_CSIZE = iround(3.6 /
_DEFAULT_KW97_TILTEFFECT_DEGPERPIX)
_DEFAULT_KW97_TILTEFFECT_NSIZE = iround(5.4 /
_DEFAULT_KW97_TILTEFFECT_DEGPERPIX)
_DEFAULT_KW97_TILTEFFECT_FSIZE = iround(10.7 /
_DEFAULT_KW97_TILTEFFECT_DEGPERPIX)
_DEFAULT_TILTEFFECT_CVAL = .5
_DEFAULT_TILTEFFECT_SVALS = np.linspace(0.0, 0.5, 10)
_DEFAULT_KW97_TILTEFFECT_SCALE = 1.25
_DEFAULT_TILTEFFECT_NPOINTS = 25 #100
_DEFAULT_TILTEFFECT_CIRCULAR = True
_DEFAULT_TILTEFFECT_DECODER_TYPE = 'circular_vote'
csvfiles = [
os.path.join(defaults._DATADIR, 'KW97_GH.csv'),
os.path.join(defaults._DATADIR, 'KW97_JHK.csv'),
os.path.join(defaults._DATADIR, 'KW97_LL.csv'),
os.path.join(defaults._DATADIR, 'KW97_SJL.csv'),
]
# experiment parameters
cpt = (_DEFAULT_TILTEFFECT_SIZE // 2, _DEFAULT_TILTEFFECT_SIZE // 2)
spt = (_DEFAULT_TILTEFFECT_SIZE // 2,
_DEFAULT_TILTEFFECT_SIZE // 2 + _DEFAULT_KW97_TILTEFFECT_CSIZE)
dt_in = _DEFAULT_TILTEFFECT_CVAL - _DEFAULT_TILTEFFECT_SVALS
# simulate populations
im = sp.array([[
stim.get_center_nfsurrounds(size=_DEFAULT_TILTEFFECT_SIZE,
csize=_DEFAULT_KW97_TILTEFFECT_CSIZE,
nsize=_DEFAULT_KW97_TILTEFFECT_NSIZE,
fsize=_DEFAULT_KW97_TILTEFFECT_FSIZE,
cval=_DEFAULT_TILTEFFECT_CVAL,
nval=sp.nan,
fval=sval,
bgval=sp.nan)
] for sval in _DEFAULT_TILTEFFECT_SVALS])
# get shifts for model for both papers, and from digitized data
sortidx = sp.argsort(dt_in) # re-order in increasing angular differences
# O'Toole and Wenderoth (1977)
n_subjects = len(csvfiles)
ds_kw97_paper_x = sp.zeros((n_subjects, 9))
ds_kw97_paper_y = sp.zeros((n_subjects, 9))
for sidx, csv in enumerate(csvfiles):
ds_kw97_paper_x[sidx], ds_kw97_paper_y[sidx] = \
sp.genfromtxt(csv, delimiter=',').T
ds_kw97_paper_x = (ds_kw97_paper_x + 360.) % 360. - 45.
ds_kw97_paper_y = 45. - ds_kw97_paper_y
for sidx in range(n_subjects):
ds_kw97_paper_x[sidx] = ds_kw97_paper_x[sidx][sp.argsort(
ds_kw97_paper_x[sidx])]
extra_vars = {}
extra_vars['scale'] = _DEFAULT_KW97_TILTEFFECT_SCALE
extra_vars['decoder'] = _DEFAULT_TILTEFFECT_DECODER_TYPE
extra_vars['npoints'] = _DEFAULT_TILTEFFECT_NPOINTS
extra_vars['npoints'] = _DEFAULT_TILTEFFECT_NPOINTS
extra_vars['cval'] = _DEFAULT_TILTEFFECT_CVAL
extra_vars['sortidx'] = sortidx
extra_vars['cpt'] = cpt
extra_vars['spt'] = spt
extra_vars['sval'] = sval
extra_vars['kind'] = 'circular'
extra_vars['figure_name'] = 'f3b'
extra_vars['return_var'] = 'O'
extra_vars['hp_file'] = os.path.join(defaults._FIGURES, 'best_hps.npz')
adjusted_gt = signal.resample(np.mean(ds_kw97_paper_y, axis=0), 10)
optimize_model(im, adjusted_gt, extra_vars, defaults)
import sys
import numpy as np
import tensorflow as tf
import model_utils
from copy import deepcopy
from timeit import default_timer as timer
from model_defs.model_cpu_port_scan_optim import ContextualCircuit, _sgw, _sdw
import scipy as sp
from ops.db_utils import update_data, get_lesion_rows_from_db, count_sets
from model_utils import iceil
from ops.parameter_defaults import PaperDefaults
defaults = PaperDefaults().__dict__
def adjust_parameters(defaults, hps):
hps_keys = hps.keys()
for k in defaults._DEFAULT_PARAMETERS.iteritems():
if k in hps_keys:
defaults._DEFAULT_PARAMETERS[k] = hps[k]
return defaults
def printProgress(iteration,
total,
prefix='',
suffix='',
decimals=1,
bar_length=100):
str_format = "{0:." + str(decimals) + "f}"
percents = str_format.format(100 * (iteration / float(total)))
def run(make_stims=False):
defaults = PaperDefaults()
#David's globals
test_colors = ['orange', 'turquoise']
size = 77
csize = 3
ssize = None
_MINPERPIX = 2.
_DEFAULT_SM2003_CSIZE_MIN = 6.0
_DEFAULT_SM2003_SIZE_MIN = 153.0
_DEFAULT_SM2003_CPDS = sp.array([0., 1., 2., 3.3, 5., 10.]) # <from paper>
_DEFAULT_SM2003_CPMS = _DEFAULT_SM2003_CPDS / 60.
_DEFAULT_SM2003_CPSS = iround(1 + _DEFAULT_SM2003_CPMS * \
(_DEFAULT_SM2003_SIZE_MIN - _DEFAULT_SM2003_CSIZE_MIN)/2.0) # <realistic>
_DEFAULT_SM2003_CSIZE = iround(_DEFAULT_SM2003_CSIZE_MIN / _MINPERPIX)
cpss = _DEFAULT_SM2003_CPSS
#Use color parameters
defaults._DEFAULT_PARAMETERS['continuous'] = False
defaults._DEFAULT_PARAMETERS['srf'] = _DEFAULT_SM2003_CSIZE
csvfiles = {
'ObsML': {
'PL': defaults._DATADIR + '/SM2003_Fig5_ObsML_PL.csv',
'LP': defaults._DATADIR + '/SM2003_Fig5_ObsML_LP.csv',
},
'ObsPM': {
'PL': defaults._DATADIR + '/SM2003_Fig5_ObsPM_PL.csv',
'LP': defaults._DATADIR + '/SM2003_Fig5_ObsPM_LP.csv',
},
}
_DEFAULT_SM2003_COLORS_RGB = {
'orange': sp.array([.99, .65, .47]),
'purple': sp.array([.65, .54, .77]),
# adjusted to have Y=.30 #sp.array([.60, .49, .71]),
'purple_alt': sp.array([.60, .48, .70]),
'lime': sp.array([.56, .60, .49]),
# adjusted to have Y=.30 #sp.array([.61, .65, .53]),
'lime_alt': sp.array([.51, .61, .53]),
'turquoise': sp.array([.44, .78, .73]),
'neutral': sp.array([.66, .66, .66]),
'lilac': sp.array([.89, .62, .80]),
'citrus': sp.array([.67, 1.0, .50]),
'EEW': sp.array([.58, .58, .58])
}
# O'Toole and Wenderoth (1977)
# load digitzed data from original paper
########################################
shift_phase_paper = sp.array([
sp.genfromtxt(csvfiles['ObsML']['PL'], delimiter=',').T[1],
sp.genfromtxt(csvfiles['ObsPM']['PL'], delimiter=',').T[1]
])
shift_anti_paper = sp.array([
sp.genfromtxt(csvfiles['ObsML']['LP'], delimiter=',').T[1],
sp.genfromtxt(csvfiles['ObsPM']['LP'], delimiter=',').T[1]
])
shift_phase_paper = np.mean(shift_phase_paper, axis=0) #double check this
shift_anti_paper = np.mean(shift_anti_paper, axis=0)
gt = np.vstack((shift_phase_paper, shift_anti_paper))
#Also preload regs for postprocessing
regpath = os.path.join(defaults._WORKDIR,\
'ShevellMonnier2003.reg.pkl.std.srf%issn%issf%i' \
% (defaults._DEFAULT_PARAMETERS['srf'], defaults._DEFAULT_PARAMETERS['ssn'], defaults._DEFAULT_PARAMETERS['ssf']))
reg_X_SO = joblib.load(regpath)['reg_X_SO']
reg_Y_SO = joblib.load(regpath)['reg_Y_SO']
reg_Z_SO = joblib.load(regpath)['reg_Z_SO']
scaler_SO = joblib.load(regpath)['scaler_SO']
reg_X_SX = joblib.load(regpath)['reg_X_SX']
reg_Y_SX = joblib.load(regpath)['reg_Y_SX']
reg_Z_SX = joblib.load(regpath)['reg_Z_SX']
scaler_SX = joblib.load(regpath)['scaler_SX']
so2image = cutils.get_XYZ2RGB_predictor(reg_X_SO, reg_Y_SO, reg_Z_SO,
scaler_SO)
sx2image = cutils.get_XYZ2RGB_predictor(reg_X_SX, reg_Y_SX, reg_Z_SX,
scaler_SX)
#Add to a dictionary
extra_vars = {}
extra_vars['n_cps'] = len(cpss)
extra_vars['cpss'] = cpss
extra_vars['test_colors'] = test_colors
extra_vars['_DEFAULT_SM2003_COLORS_RGB'] = _DEFAULT_SM2003_COLORS_RGB
extra_vars['size'] = size
extra_vars['csize'] = csize
extra_vars['ssize'] = ssize
extra_vars['kind'] = 'circular'
extra_vars['so2image'] = so2image
extra_vars['sx2image'] = sx2image
extra_vars['n_col'] = len(test_colors)
extra_vars['figure_name'] = 'f7'
extra_vars['return_var'] = 'O'
extra_vars['hp_file'] = os.path.join(defaults._FIGURES, 'best_hps.npz')
# measure shift for phase & antiphase inducers, for each test
if make_stims:
create_stimuli(gt, extra_vars, defaults)
else:
optimize_model(gt, extra_vars, defaults)
def run():
defaults = PaperDefaults()
#David's globals
size = 51
mpp = 0.76 # 0.76 # 1.11
scale = 0.23 # 0.23 # 0.22
csv_file_x = os.path.join(defaults._DATADIR, 'WL1987_corrected_X.csv')
csv_file_y = os.path.join(defaults._DATADIR, 'WL1987_corrected_Y.csv')
# experiment parameters
dd = (-150., 150.) # in seconds of arc
sec2u = lambda s: (s - dd[0]) / (dd[1] - dd[0])
u2sec = lambda u: u * (dd[1] - dd[0]) + dd[0]
min2pix = lambda m: iround(m / float(mpp))
npoints = 50
ndists = 10
dists = sp.linspace(0.0, 12., ndists)
lh, lw = 1, 4.
ph, pw = 2., 2.
center_disp = 0.0
flanker_disp = -33.3
mp0 = size // 2
# Need to scale up the ecrfs
defaults._DEFAULT_PARAMETERS[
'srf'] = defaults._DEFAULT_PARAMETERS['srf'] * 2 - 1
defaults._DEFAULT_PARAMETERS[
'ssn'] = defaults._DEFAULT_PARAMETERS['ssn'] * 2 - 1
defaults._DEFAULT_PARAMETERS[
'ssf'] = defaults._DEFAULT_PARAMETERS['ssf'] * 2 - 1
# simulate populations
im = get_wl87_stim(size=size,
dists=min2pix(dists),
cval=sec2u(center_disp),
sval=sec2u(flanker_disp),
ch=min2pix(lh),
cw=min2pix(lw),
sh=min2pix(ph),
sw=min2pix(pw))
# Get ground truth data
paper_data_x = sp.genfromtxt(csv_file_x, delimiter=',')
paper_data_y = sp.genfromtxt(csv_file_y, delimiter=',') * -1
paper_fit_y = sfit(sp.linspace(dists.min(), dists.max(), 100),
paper_data_x,
sp.nanmean(paper_data_y, axis=0),
k=2,
t=[5.])
paper_fit_y = paper_fit_y[np.round(
np.linspace(0, paper_fit_y.shape[0] - 1, ndists)).astype(int)]
extra_vars = {}
extra_vars['scale'] = scale
extra_vars['kind'] = 'gaussian'
extra_vars['decoder'] = 'circular_vote'
extra_vars['npoints'] = npoints
extra_vars['cval'] = sec2u(center_disp)
extra_vars['sval'] = sec2u(flanker_disp)
extra_vars['figure_name'] = 'f5'
extra_vars['u2sec'] = u2sec
extra_vars['min2pix'] = min2pix
extra_vars['dists'] = dists
extra_vars['flanker_disp'] = flanker_disp
extra_vars['mp0'] = mp0
extra_vars['lh'] = lh
extra_vars['pw'] = pw
extra_vars['size'] = size
extra_vars['gt_x'] = paper_data_x
extra_vars['return_var'] = 'O'
optimize_model(im, paper_fit_y, extra_vars, defaults)
def run():
defaults = PaperDefaults()
#David's globals
_DEFAULT_MURAKAMISHIMOJO96_DPP = 1.0/10. # _PARAMETER_SET_VERSION == 'paper'
_DEFAULT_MURAKAMISHIMOJO96_DPP = .125 # _PARAMETER_SET_VERSION == 'v2'
_DEFAULT_MURAKAMISHIMOJO96_SIZE = 201 # 51 <works>
_DEFAULT_MURAKAMISHIMOJO96_NTRIALS = 25 # 20 <works>
_DEFAULT_MURAKAMISHIMOJO96_WS = sp.unique(sp.around(
1.0 / _DEFAULT_MURAKAMISHIMOJO96_DPP * \
sp.array([0.5, 0.75, 1., 1.5, 2., 3., 4., 5., 8.]))) # <paper> <FULL3_LARGE>
_DEFAULT_MURAKAMISHIMOJO96_RDD = 0.50 # <works> | sp.pi**2/72 # <paper>
_DEFAULT_MURAKAMISHIMOJO96_NUNITS = 32
_DEFAULT_MURAKAMISHIMOJO96_NDIRS = 10
_DEFAULT_MURAKAMISHIMOJO96_NCOH = 21
_DEFAULT_MURAKAMISHIMOJO96_NCOH4FIT = 101
_DEFAULT_MURAKAMISHIMOJO96_SCALE = 1.0
_DEFAULT_MURAKAMISHIMOJO96_VALUE_UP = .80
_DEFAULT_MURAKAMISHIMOJO96_VALUE_DOWN = .30
_DEFAULT_MURAKAMISHIMOJO96_CSVS = {
'%s' % (subj,): [os.path.join(defaults._DATADIR, 'MS96_%s_%g.csv') % (subj, deg) \
for deg in [0, 2, 3, 4.5, 6, 9]] for subj in ['IM', 'SM']}
# experiment parameters
w_paper, PSE_paper = [], [],
w_paper_subs, PSE_paper_subs = [], []
for idx, subject in enumerate(_DEFAULT_MURAKAMISHIMOJO96_CSVS.values()):
for jdx, csvfile in enumerate(subject):
w_, pse_ = sp.genfromtxt(csvfile, delimiter=',').T
w_paper += w_.tolist()
PSE_paper += pse_.tolist()
w_paper_subs.append(w_)
PSE_paper_subs.append(pse_)
gt_data = np.zeros((len(PSE_paper_subs),len(PSE_paper_subs[0])));
max_size = np.max(np.asarray([len(x) for x in PSE_paper_subs]))
#Assign to 9 bins np.concatenate(w_paper_subs)
#Take the average within each bin as the GT
for idx,g in enumerate(PSE_paper_subs):
if len(g) < max_size:
g = signal.resample(g,max_size)
gt_data[idx,:] = g
gt = np.mean(gt_data,axis=0)
#
extra_vars = {}
extra_vars['ws'] = _DEFAULT_MURAKAMISHIMOJO96_WS
extra_vars['rdd'] = _DEFAULT_MURAKAMISHIMOJO96_RDD
extra_vars['size'] = _DEFAULT_MURAKAMISHIMOJO96_SIZE
extra_vars['ntrials'] = _DEFAULT_MURAKAMISHIMOJO96_NTRIALS
extra_vars['nunits'] = _DEFAULT_MURAKAMISHIMOJO96_NUNITS
extra_vars['ncoh'] = _DEFAULT_MURAKAMISHIMOJO96_NCOH
extra_vars['scale'] = _DEFAULT_MURAKAMISHIMOJO96_SCALE
extra_vars['ndirs'] = _DEFAULT_MURAKAMISHIMOJO96_NDIRS
extra_vars['value_up'] = _DEFAULT_MURAKAMISHIMOJO96_VALUE_UP
extra_vars['value_down'] = _DEFAULT_MURAKAMISHIMOJO96_VALUE_DOWN
extra_vars['ncoh4fit'] = _DEFAULT_MURAKAMISHIMOJO96_NCOH4FIT
extra_vars['kind'] = 'circular'
extra_vars['figure_name'] = 'f6'
extra_vars['return_var'] = 'O'
extra_vars['hp_file'] = os.path.join(defaults._FIGURES, 'best_hps.npz')
optimize_model(gt,extra_vars,defaults)
import sys
import numpy as np
import scipy as sp
from scipy import stats
import tensorflow as tf
from copy import deepcopy
sys.path.append('../')
from ops.parameter_defaults import PaperDefaults
from tf_helper_functions import *
from collections import namedtuple
defaults = PaperDefaults().__dict__
module = sys.modules[__name__] # add to the global namespace
for name, value in defaults.iteritems():
setattr(module, name, value)
CircuitParameters = namedtuple('CircuitParameters',
defaults['_DEFAULT_PARAMETERS'].keys(),
verbose=False,
rename=False)
_DEFAULT_PARAMETERS_TEMPLATE = deepcopy(defaults['_DEFAULT_PARAMETERS'])
def sampler(x):
return abs(np.random.uniform(low=x - 1, high=x + 1) +
x)**np.random.uniform(low=-2., high=2.) # previously did [0, 2]
def makeGaussian(size, fwhm=3, center=None):
""" Make a square gaussian kernel.
size is the length of a side of the square
fwhm is full-width-half-maximum, which
def run(initialize_model=False):
defaults = PaperDefaults()
#David's globals
_DEFAULT_KW2015_SO_PARAMETERS = {
'filters': {
'name': 'gabors',
'aspect_ratio': .6,
'sizes': sp.array([9]),
'spatial_frequencies': sp.array([[9.0]]),
'orientations': sp.arange(2) * sp.pi / 2,
'phases': sp.array([0]),
'with_center_surround': False,
'padding': 'reflect',
'corr': False,
'ndp': False
},
'model': {
'channels_so': ('R+G-', 'B+Y-', 'R+C-', 'Wh+Bl-', 'G+R-', 'Y+B-',
'C+R-', 'Bl+Wh-'),
'normalize':
False
},
'dnp_so': None,
'selected_channels': [0, 1, 3, 4, 5, 7],
'norm_channels': [0, 1, 3, 4, 5, 7]
}
size = 51
csize = 9
n_train = 32
n_t_hues = 16
n_s_hues = 16
csvfiles = [
defaults._DATADIR + '/KW2015_%i.csv' % (i, )
for i in range(0, 360, 45)
]
#Load data from experiments
kw2015_fig2_x = sp.zeros((len(csvfiles), 16))
kw2015_fig2_y = sp.zeros((len(csvfiles), 16))
for idx, csv in enumerate(csvfiles):
kw2015_fig2_x[idx], kw2015_fig2_y[idx] = \
sp.genfromtxt(csv, delimiter=',')[1:].T
# experiment stimuli
extra_vars = {}
extra_vars['_DEFAULT_KW2015_SO_PARAMETERS'] = _DEFAULT_KW2015_SO_PARAMETERS
extra_vars['_DEFAULT_FLOATX_NP'] = defaults._DEFAULT_FLOATX_NP
extra_vars['size'] = size
extra_vars['csize'] = csize
extra_vars['n_train'] = n_train
extra_vars['n_t_hues'] = n_t_hues
extra_vars['n_s_hues'] = n_s_hues
extra_vars['figure_name'] = 'f4'
extra_vars['gt_x'] = kw2015_fig2_x
extra_vars['f4_stimuli_file'] = defaults.f4_stimuli_file
extra_vars['return_var'] = 'I'
extra_vars['precalculated_x'] = True
extra_vars['aux_y'] = []
extra_vars['percent_reg_train'] = 80.
if initialize_model:
create_stims(extra_vars)
stim_files = np.load(extra_vars['f4_stimuli_file'])
extra_vars['stims_all_lms'] = stim_files['stims_all_lms']
#Run model
#cx.run(so_all, from_gpu=False)
#sx_all[:] = cx.Y.get()[:, :, size//2, size//2]
adj_gt = np.mean(kw2015_fig2_y, axis=0)
im = stim_files['so_ind'].reshape(
n_t_hues * n_s_hues,
len(_DEFAULT_KW2015_SO_PARAMETERS['norm_channels']), size, size)
extra_vars['aux_data'] = stim_files['so_all'].transpose(0, 2, 3, 1)
extra_vars['cs_hue_diff'] = stim_files['cs_hue_diff']
optimize_model(im, adj_gt, extra_vars, defaults)
zs = np.zeros((len(lesion_data)))
mus = np.zeros((len(lesion_data)))
for idx, row in enumerate(lesion_data):
figure_fits = [row[k] for k in defaults.db_problem_columns]
# if None in figure_fits:
# raise TypeError
# summarize with a z score
zs[idx] = np.nanmean(figure_fits) / np.nanstd(figure_fits)
mus[idx] = np.nanmax(figure_fits)
if len(zs) > 0:
return lesion_data[np.argmax(zs)], lesion_data, zs, mus
else:
return None, None, None, None
defaults = PaperDefaults()
max_row = {}
lesion_dict = {}
score_dict = {}
mus = {}
for lesion in defaults.lesions:
lesion_data = get_all_lesion_data(lesion)
it_max, lesion_dict[lesion], score_dict[lesion], mus[
lesion] = find_best_fit(lesion_data, defaults)
if it_max is not None:
it_max = dict(it_max)
max_row[lesion] = it_max
# bootstrap here
max_zs = {k: v.max() for k, v in score_dict.iteritems()}
# Reorder the dimensions
t_stats = {}
