def main():
if clean_dir:
clean_outdir(outdir)
x0 = generate_vf_landmarks(magnification,
ecc0,
ecc1,
n_prototypes,
vf='full',
filename=full_filename('vf', outdir))
b0 = generate_V2_boundary_retinotopy_for_DM(boundary_len,
map_h,
filename=full_filename(
'boundary', outdir))
y = initialize_map(map_h,
map_w,
x0,
filename=full_filename('init', outdir),
random_seed=random_seed)
optimize(x0,
b0,
b1,
b2,
map_h,
map_w,
y,
k0=k0,
kr=kr,
eta0=eta0,
outdir=outdir,
save_interval=save_interval,
n=n_iterations)
def save_map(sess,
y,
b1,
b2,
i,
outdir="",
save_interval=0,
quiet=False,
suffix=''):
if (outdir != "") and save_interval > 0:
if i % save_interval == 0:
filename = "y-{b1:4.3f}-{b2:4.3f}-{i}".format(b1=b1,
b2=b2,
i=str(i).zfill(5))
zz = sess.run(y)
if (not quiet):
print("saving {filename}".format(filename=filename))
np.savetxt(full_filename(filename,
outdir,
ext=".data",
suffix=suffix),
np.array(zz),
fmt="%f",
delimiter=",")
def main(rseed, outputdir):
if clean_dir:
clean_outdir(outputdir)
x0 = generate_vf_landmarks(magnification,
ecc0,
ecc1,
n_prototypes,
vf='full',
filename=full_filename('vf', outputdir))
b0 = generate_V2_boundary_retinotopy(boundary_len,
map_h,
branch='vd',
filename=full_filename(
'boundary', outputdir),
tilt=10.0)
y = initialize_map(map_h,
map_w,
x0,
filename=full_filename('init', outdir),
random_seed=rseed)
# loop through b1 and b2
b1s = [0.03 * 1.6**i for i in range(0, 8)]
b2s = [0.03 * 1.6**i for i in range(0, 8)]
print("b1:", b1s)
print("b2:", b2s)
for b1 in b1s:
for b2 in b2s:
print("b1 = %10.4f \t b2 = %10.4f" % (b1, b2))
optimize(
x0,
b0,
b1,
b2,
map_h,
map_w,
y,
k0=k0,
kr=kr,
eta0=eta0,
outdir=outputdir,
save_interval=0, # disable saving intermediate maps
n=n_iterations,
report_iterations=False,
log_iterations=False)
def train(opt,
total_cost,
partial_costs,
tf_placeholders,
y,
b1,
b2,
outdir="",
save_interval=0,
n=1000,
k0=30.0,
kr=0.005,
quiet_save=False,
report_iterations=True,
log_iterations=True,
logfile='log',
suffix='',
yx_gauss=None):
"""
b1: weights for the smoothness term
b2: weights for the congruency term
n: number of iterations
k0: initial annealing parameter
kr: reduction of the annealing parameter at each iteration
outdir: where to save the results.
save_intervals: save map every several iterations. if set to 0, saving is disabled
"""
yx_cost = partial_costs[0]
reg1 = partial_costs[1]
reg2 = partial_costs[2]
beta1 = tf_placeholders[0]
beta2 = tf_placeholders[1]
kappa = tf_placeholders[2]
sess = tf.Session()
sess.run(tf.global_variables_initializer())
# save initial map
# always save the initial map (by setting save_interval to 1)
save_map(sess,
y,
b1,
b2,
0,
outdir,
save_interval=1,
quiet=quiet_save,
suffix=suffix)
k = k0
for i in range(1, n + 1):
# itertaively optimize the total cost function
sess.run(opt, {kappa: k, beta1: b1, beta2: b2})
report_and_log(partial_costs,
tf_placeholders,
i,
k,
b1,
b2,
sess,
reportP=report_iterations,
logP=log_iterations,
logfile=full_filename(logfile, outdir, suffix=suffix))
if i > 0:
save_map(sess,
y,
b1,
b2,
i,
outdir,
save_interval,
quiet=quiet_save,
suffix=suffix)
k = k - k * kr
# save the final result
# set save_interval to 1, so that the final result is saved even if intermediate saving is disabled by save_interval=0
save_map(sess,
y,
b1,
b2,
i,
outdir,
save_interval=1,
quiet=quiet_save,
suffix=suffix)
report_and_log(partial_costs,
tf_placeholders,
i,
k,
b1,
b2,
sess,
reportP=False,
logP=True,
logfile=full_filename(logfile, outdir, suffix=suffix))