# create a folder to save the results
dir_save = model_rbm.make_dir_save(parent_dir_save, prefix, learn_rate_a,
learn_rate_b, learn_rate_W, reg_lambda_W,
reg_alpha_W, visible_type_fixed_param,
hidden_type_fixed_param, maxiter,
normalization_method)
# make a copy of this script in dir_save
shutil.copy(workdir + "main_test_ExpRBM_FASHIONMNIST.py", dir_save)
shutil.copy(workdir + "restricted_boltzmann_machine.py", dir_save)
# plot some some training samples
train_set_x_3way = numpy.reshape(train_set_x[:, 0:100], newshape=(28, 28, 100))
cl.plot_image_subplots(dir_save + "fig_" + prefix + "_train_samples.pdf",
data=train_set_x_3way[:, :, 0:100],
figwidth=6,
figheight=6,
colormap="gray",
num_col=10,
wspace=0.01,
hspace=0.001)
# train RBM
model_rbm.train(X=train_set_x,
X_validate=test_set_x,
batch_size=batch_size,
pcdk=pcdk,
NS=NS,
maxiter=maxiter,
learn_rate_a=learn_rate_a,
learn_rate_b=learn_rate_b,
learn_rate_W=learn_rate_W,
change_rate=change_rate,
Xg, XMg = model_dbn.generate_x(pcdk=10 * sampling_pcdk,
NS=sampling_NS,
X0=None,
persistent=True,
rand_init=True,
init=True)
for s in range(sampling_time):
Xg, XMg = model_dbn.generate_x(pcdk=sampling_pcdk, NS=sampling_NS)
# plot sampled data
sample_set_x_3way = numpy.reshape(XMg, newshape=(28, 28, 100))
print(s)
cl.plot_image_subplots(dir_save + "/fig_" + prefix +
"_pretrain_generated_samples_randinit_" + str(s) +
".pdf",
data=sample_set_x_3way,
figwidth=6,
figheight=6,
colormap=None,
num_col=10,
wspace=0.01,
hspace=0.001)
# estimate the lower bound of the log-likelihood
T = 10000
S = 10
model_dbn.estimate_log_likelihood(X=test_set_x,
Hr=None,
HMr=None,
a_hat_gen=None,
b_hat_gen=None,
estimate_logZ=True,
base_rate_type="prior",
NS=num_sampled_points,
X0=None,
init_sampling=True)
for s in range(sampling_time):
chainX, chainXM = model_mdbn.generate_x(pcdk=pcdk_sampling,
init_sampling=False)
# plot sampled data
sample_set_x_3way = numpy.reshape(chainXM[0],
newshape=(28, 28, num_sampled_points))
filename = dir_save + prefix + "_generated_samples_randinit_" + str(
rand_init) + "_" + str(s) + "_V0.pdf"
cl.plot_image_subplots(filename,
data=sample_set_x_3way,
figwidth=6,
figheight=6,
colormap=None,
num_col=num_col,
wspace=0.01,
hspace=0.001)
#filename=dir_save+"mdbn_MNIST_generated_samples_randinit_"+str(rand_init)+"_"+str(s)+"_V1_1D.txt"
#sample_set_y_1way=numpy.argmax(chainXM[1],axis=0)
#numpy.savetxt(filename,sample_set_y_1way,fmt="%s",delimiter="\t")
filename = dir_save + prefix + "_generated_samples_randinit_" + str(
rand_init) + "_" + str(s) + "_V1_2D.txt"
sample_set_y_2way = numpy.reshape(
numpy.argmax(chainXM[1], axis=0),
newshape=(num_col, int(num_sampled_points / num_col)))
numpy.savetxt(filename, sample_set_y_2way, fmt="%s", delimiter="\t")
filename = dir_save + prefix + "_generated_samples_randinit_" + str(
rand_init) + "_" + str(s) + "_V1_prob.txt"
numpy.savetxt(filename,
clamp_visible=False,
persistent=True,
rand_init_X=rand_init,
rand_init_H=False,
init_sampling=True)
for s in range(sampling_time):
chainX, chainH, chainXM, chainXP, chainHM, chain_length_ = model_dbm.pcd_sampling(
pcdk=1000, init_sampling=False)
# plot sampled data
sample_set_x_3way = numpy.reshape(chainXM, newshape=(28, 28, 100))
print(s)
cl.plot_image_subplots(dir_save +
"/fig_DBM_MNIST_generated_samples_randinit_" +
str(rand_init) + "_" + str(s) + ".pdf",
data=sample_set_x_3way,
figwidth=6,
figheight=6,
colormap=None,
num_col=10,
wspace=0.01,
hspace=0.001)
# estimate log-likelihood
loglh_test, logZ_test, Z_std_test, energy_test, entropy_test = model_dbm.estimate_log_likelihood(
X=test_set_x,
NMF=100,
base_type="prior",
beta=None,
step_base=0.999,
T=10000,
S=10)
loglh_train, logZ_train, Z_std_train, energy_train, entropy_train = model_dbm.estimate_log_likelihood(