def run_network_on_image():
import make_seqs
ims1, ims1a, ims2=make_seqs.make_seqs(slength=6,num_seqs=1000)
input_var1 = T.tensor4('inputs')
input_var2 = T.tensor4('inputs_comp')
network = compare_net.build_cnn_new_conv(input_var1, input_var2)
if (os.path.isfile('net.npy')):
spars=np.load('net.npy')
lasagne.layers.set_all_param_values(network,spars)
test_corr = lasagne.layers.get_output(network, deterministic=True)
test_fn = theano.function([input_var1, input_var2], [test_corr])
tcorr_same=test_fn(ims1,ims1a)
tcorr_diff=test_fn(ims1,ims2)
tt_same_std=np_standardize(tcorr_same[0])
tt_diff_std=np_standardize(tcorr_diff[0])
corrs_same=get_shifted_correlations(tt_same_std)
corrs_diff=get_shifted_correlations(tt_diff_std)
dps=optimize_dp(corrs_same)
dpd=optimize_dp(corrs_diff)
print(np.min(dps),np.max(dps),np.min(dpd),np.max(dpd))
import pylab as py
py.figure(1)
py.hist(dps,alpha=.5)
py.hist(dpd,alpha=.5)
py.show()
print('done ')
def run_network_on_all_pairs(num_seqs=40):
import make_seqs
ims1, ims1a, ims2=make_seqs.make_seqs(slength=6,num_seqs=num_seqs, from_font=False)
input_var1 = T.tensor4('inputs')
input_var2 = T.tensor4('inputs_comp')
network = compare_net.build_cnn_new_conv(input_var1, input_var2)
if (os.path.isfile('net.npy')):
spars=np.load('net.npy')
lasagne.layers.set_all_param_values(network,spars)
test_corr = lasagne.layers.get_output(network, deterministic=True)
test_fn = theano.function([input_var1, input_var2], [test_corr])
tcorr_same=test_fn(ims1,ims1a)
tt_same_std=np_standardize(tcorr_same[0])
temp=np.copy(tt_same_std)
ii=np.arange(0,num_seqs)
np.random.shuffle(ii)
iii=np.copy(ii)
temp[1,]=tt_same_std[1,iii]
dps=np.zeros((num_seqs,num_seqs))
for n in range(num_seqs):
temp[1,]=tt_same_std[1,np.roll(ii,-n),]
corrs_same=get_shifted_correlations(temp)
dps[n,]=optimize_dp(corrs_same)
dps=dps.transpose()
dpss=dps.copy()
for n in range(num_seqs):
dpss[n,]=np.roll(dps[n,],n)
print(dpss)
print(iii)
print("done")
dpss=np.max(dpss)-dpss
#dps=dps.transpose()
match_them(dpss,iii)
print('done ')
