def __rec_map( map_as_dict, fits, model ): import numpy as N if model=='dog': from dog import reconstruct else: from edog import reconstruct rec_W = N.zeros(map_as_dict['W'].shape) for key in fits: fit = fits[key] p = fit['p'] rec_W[key] = reconstruct(p, map_as_dict['mode'], map_as_dict['vis']**2, map_as_dict['vis'], map_as_dict['W'][key].shape) return N.copy(rec_W)
def make_proto_filter(
args_file,
args_odir,
args_model,
synthetic_pos=True,
mean_rec=True,
max_dist_to_center=2,
meanlen=None,
min_err_metric=True,
abs_weight_threshold=0.7,
indices=None,
debug=False,
odir_nosubdir=False,
):
hand_automatic = indices == None
'''loading precomputed fits'''
from util import depickle_fits
res = depickle_fits(args_file, suffix='cfits')
'''retrieving weightmatrix from fits dic'''
W = res['map']['W']
vis = res['map']['vis']
map_as_dict = res['map']
assert map_as_dict['mode'] == 'rgb'
'''get fit and cluster data'''
fits = res['fits']
fit_keys = fits.keys()
model = res['model']
cl = res['dic_cluster']
num_channel = cl['num_clusters']
'''move RF ids in an appropiate data struc, ignore zeros'''
clusters_by_index = [[] for c in [None]*num_channel]
keys = sorted(fit_keys)
for i, key in enumerate(keys):
cl = fits[key]['cl']
clusters_by_index[cl].append(key)
from ..base.receptivefield import convert_rfvector_to_rgbmatrix
from ..base.plots import colormap_for_mode
cmap = colormap_for_mode(map_as_dict['mode'])
from ..base import make_filename
from cluster import __transpose_zero_to_one
if mean_rec:
'''of every channel mean the fitted parameters'''
import numpy as N
num_param = len(fits[fit_keys[-1]]['p'])
clusters_p_mean = [N.zeros(num_param) for c in [None]*num_channel]
for c in xrange(0,len(clusters_by_index)):
cl_ids = clusters_by_index[c]
if meanlen is None:
len_cl_id = len(cl_ids)
else:
len_cl_id = meanlen
p_mean_tmp = []
for i in xrange(0, len_cl_id):
p = fits[cl_ids[i]]['p']
p_mean_tmp.append(p)
p_mean = N.mean(p_mean_tmp, axis=0)
clusters_p_mean[c] = N.copy(p_mean)
if model=='dog': from dog import reconstruct
else: from edog import reconstruct
else:
import numpy as N
hand_by_index = [[] for c in [None]*num_channel]
zpx = vis/2.
zpy = vis/2.
max_absmax = 0
if not hand_automatic:
'''use provided indices ... handchosen'''
for i, index in enumerate(indices):
if i == num_channel:
break
hand_by_index[i] = index
else:
'''chose automatically:
for each channel find the RF with smallest error (and max value > .7),
in distance close to the center vis field.'''
for c in xrange(0,len(clusters_by_index)):
cl_ids = clusters_by_index[c]
len_cl_id = len(cl_ids)
min_err, min_err_id, min_dist, min_id = N.inf, -1, N.inf, -1
for i in xrange(0, len_cl_id):
p = fits[cl_ids[i]]['real_pix_center_coords']
n = fits[cl_ids[i]]['n']
err = fits[cl_ids[i]]['err']
dist = N.sqrt((zpy - p[0])**2 + (zpx - p[1])**2)
absmax = N.max(N.abs(W[n]))
if err < min_err and absmax > abs_weight_threshold and dist < max_dist_to_center:
min_err = err
min_err_id = n
if dist < min_dist and absmax > abs_weight_threshold:
min_dist = dist
min_id = n
'''statistic'''
if absmax > max_absmax:
max_absmax = absmax
if min_err_metric:
hand_by_index[c] = min_err_id
else:
hand_by_index[c] = min_id
'''store distance of fitted center point to center of visual field'''
trans_rf = []
for n in hand_by_index:
if type(n) == list:
print 'Not enough RF indices for all channels given.\nnum channels:', num_channel, 'num indices:', len(indices), '\n'
exit()
if n == -1:
print 'No prototype RF found. \nTry lowering wheight treshold\nparameter -thr is:', abs_weight_threshold, 'RF max abs weight:', absmax, '\n'
exit()
p = fits[n]['real_pix_center_coords']
dist_y = int(N.floor(zpy - p[0]))
dist_x = int(N.floor(zpx - p[1]))
trans_rf.append( (dist_y, dist_x) )
filters = []
plots = []
for c in xrange(0,num_channel):
if mean_rec:
p = clusters_p_mean[c]
if synthetic_pos:
p = N.concatenate([[map_as_dict['vis']/2., map_as_dict['vis']/2.], p[2:]])
proto_filter = reconstruct(p,
map_as_dict['mode'],
map_as_dict['vis']**2,
map_as_dict['vis'],
map_as_dict['W'][-1].shape)
else:
proto_filter = W[hand_by_index[c]]
'''convert RF vector to matrix and normalize values'''
proto_filter_matr = convert_rfvector_to_rgbmatrix(
proto_filter,
map_as_dict['vis'],
map_as_dict['vis'],
map_as_dict['mode'])
proto_filter_matr = __transpose_zero_to_one(proto_filter_matr)
'''move RF to visual fields center'''
if not mean_rec:
proto_filter_matr = N.roll(proto_filter_matr, trans_rf[c][0], axis=0)
proto_filter_matr = N.roll(proto_filter_matr, trans_rf[c][1], axis=1)
plots.append({
# 'name':('RF: '+str(hand_by_index[c])+' ' if not hand_automatic else '') + 'ch: '+str(c),
'name':'RF: '+str(hand_by_index[c])+' ' + ' ch: '+str(c),
'value':proto_filter_matr,
'maxmin':True,
'cmap':cmap,
'balance':False,
'invert':False,
})
filters.append(N.copy(proto_filter_matr))
if not mean_rec:
if hand_automatic: misc_str = 'auto_'
else: misc_str = 'hand_'
if min_err_metric: metr_str = 'err_'
else: metr_str = 'dist_'
else:
misc_str = 'mean_'
metr_str = ''
if not odir_nosubdir:
import os
mod_odir = os.path.join(args_odir, '../')
from ..base import make_working_dir_sub
work_dir = make_working_dir_sub(mod_odir, 'proto')
else:
work_dir = args_odir
fname = make_filename(args_file,misc_str+metr_str+'conv_proto','.png', work_dir)
def numplots_to_rowscols(num):
sq = int(num**.5)+1
return sq, sq
from ..base.plots import write_row_col_fig
row, col = numplots_to_rowscols(num_channel)
write_row_col_fig(plots, row, col, fname+'.png', dpi=144, alpha=1.0, fontsize=6)
dic = {
'mode': res['map']['mode'],
'num_chn': num_channel,
'filters': filters
}
writefilename = make_filename(args_file,misc_str+metr_str+'conv_proto','.kern', work_dir)
from util import pickle_fits
pickle_fits(writefilename+'.kern', dic)
if debug:
from ..base.plots import write_rf_fit_debug_fig
for i, fmat in enumerate(filters):
fmat = N.swapaxes(fmat, 0, 1)
fvec = fmat.reshape(fmat.shape[0]*fmat.shape[1], fmat.shape[2])
fvecflat = N.copy(N.concatenate([fvec.T[0].T, fvec.T[1].T, fvec.T[2].T]))
fvecflat -= .5
fvecflat *= 2.
p = fits[hand_by_index[i]]['p']
p[0] = p[0] + trans_rf[i][0]
p[1] = p[1] + trans_rf[i][1]
if res['map']['mode'] == 'dog': from dog import reconstruct
else: from edog import reconstruct
rec = reconstruct(p, res['map']['mode'], vis**2, vis, fvecflat.shape)
err = (fvecflat - rec)**2
# err = None
# fname = make_filename(args_file,str(i)+'_debug','.png', work_dir)
fname = work_dir + '/' + str(i)+'_debug'+'.png'
write_rf_fit_debug_fig(fname, fvecflat, vis, 'rgb', p, rec, err, model,
scale=2.8, s_scale=3.8, alpha=.5, dpi=300, draw_ellipsoid=True,
no_title=True, ellipsoid_line_width=1.2)