def load_rbm1(args):
if os.path.isdir(args.rbm1_dirpath):
print("\nLoading RBM #1 ...\n\n")
rbm1 = BernoulliRBM.load_model(args.rbm1_dirpath)
return rbm1
else:
print("model could not be found!")
def make_rbm2(Q, args):
if os.path.isdir(args.rbm2_dirpath):
print("\nLoading RBM #2 ...\n\n")
rbm2 = BernoulliRBM.load_model(args.rbm2_dirpath)
else:
print("\nTraining RBM #2 ...\n\n")
epochs = args.epochs[1]
n_every = args.increase_n_gibbs_steps_every
n_gibbs_steps = np.arange(args.n_gibbs_steps[1],
args.n_gibbs_steps[1] + epochs / n_every)
learning_rate = args.lr[1] / np.arange(1, 1 + epochs / n_every)
n_gibbs_steps = np.repeat(n_gibbs_steps, n_every)
learning_rate = np.repeat(learning_rate, n_every)
rbm2 = BernoulliRBM(n_visible=args.n_hiddens[0],
n_hidden=args.n_hiddens[1],
W_init=0.005,
vb_init=0.,
hb_init=0.,
n_gibbs_steps=n_gibbs_steps,
learning_rate=learning_rate,
momentum=[0.5] * 5 + [0.9],
max_epoch=max(args.epochs[1], n_every),
batch_size=args.batch_size[1],
l2=args.l2[1],
sample_h_states=True,
sample_v_states=True,
sparsity_cost=0.,
dbm_last=True, # !!!
metrics_config=dict(
msre=True,
pll=True,
train_metrics_every_iter=500,
),
verbose=True,
display_filters=0,
display_hidden_activations=24,
random_seed=args.random_seed[1],
dtype='float32',
tf_saver_params=dict(max_to_keep=1),
model_path=args.rbm2_dirpath)
rbm2.fit(Q)
return rbm2
def init_rbm(args): # initializes standard RBM
# if os.path.isdir(args.model_dirpath):
# print ("\nLoading model ...\n\n")
# rbm = BernoulliRBM.load_model(args.model_dirpath)
# else:
# print ("\nTraining model ...\n\n")
rbm = BernoulliRBM(
n_visible=args.n_visible,
n_hidden=args.n_hidden,
W_init=args.w_init,
vb_init=args.vb_init,
hb_init=args.hb_init,
prune=args.prune,
freeze_weights=args.freeze_weights,
filter_shape=args.filter_shape,
n_gibbs_steps=args.n_gibbs_steps,
learning_rate=args.lr,
# momentum=np.geomspace(0.5, 0.9, 8),
momentum=args.momentum,
max_epoch=args.epochs,
batch_size=args.batch_size,
l2=args.l2,
sample_v_states=args.sample_v_states,
sample_h_states=True,
dropout=args.dropout,
sparsity_target=args.sparsity_target,
sparsity_cost=args.sparsity_cost,
sparsity_damping=args.sparsity_damping,
metrics_config=dict(
msre=False,
pll=False,
feg=False,
train_metrics_every_iter=10,
val_metrics_every_epoch=1,
feg_every_epoch=4,
n_batches_for_feg=10,
),
verbose=True,
random_seed=args.random_seed,
dtype=args.dtype,
tf_saver_params=dict(max_to_keep=1),
model_path=args.model_dirpath)
rbm.init()
return rbm
def make_rbm1(X, args):
if os.path.isdir(args.rbm1_dirpath):
print("\nLoading RBM #1 ...\n\n")
rbm1 = BernoulliRBM.load_model(args.rbm1_dirpath)
else:
print("\nTraining RBM #1 ...\n\n")
rbm1 = BernoulliRBM(n_visible=784,
n_hidden=args.n_hiddens[0],
W_init=0.001,
vb_init=0.,
hb_init=0.,
n_gibbs_steps=args.n_gibbs_steps[0],
learning_rate=args.lr[0],
momentum=[0.5] * 5 + [0.9],
max_epoch=args.epochs[0],
batch_size=args.batch_size[0],
l2=args.l2[0],
sample_h_states=True,
sample_v_states=True,
sparsity_cost=0.,
dbm_first=True, # !!!
metrics_config=dict(
msre=True,
pll=True,
train_metrics_every_iter=500,
),
verbose=True,
display_filters=30,
display_hidden_activations=24,
v_shape=(28, 28),
random_seed=args.random_seed[0],
dtype='float32',
tf_saver_params=dict(max_to_keep=1),
model_path=args.rbm1_dirpath)
rbm1.fit(X)
return rbm1
def init_rbm2(args):
if os.path.isdir(args.rbm2_dirpath):
print("\nLoading RBM #2 ...\n\n")
rbm2 = BernoulliRBM.load_model(args.rbm2_dirpath)
else:
print("\nInitializing RBM #2 ...\n\n")
epochs = args.epochs[1]
n_every = args.increase_n_gibbs_steps_every
n_gibbs_steps = np.arange(args.n_gibbs_steps[1],
args.n_gibbs_steps[1] + epochs / n_every)
learning_rate = args.lr[1] / np.arange(1, 1 + epochs / n_every)
n_gibbs_steps = np.repeat(n_gibbs_steps, n_every)
learning_rate = np.repeat(learning_rate, n_every)
rbm2 = BernoulliRBM(
n_visible=args.n_hidden[0],
n_hidden=args.n_hidden[1],
W_init=args.w_init[1],
vb_init=args.vb_init[1],
hb_init=args.hb_init[1],
n_gibbs_steps=n_gibbs_steps,
learning_rate=learning_rate,
momentum=args.momentum,
max_epoch=max(args.epochs[1], n_every),
batch_size=args.batch_size[1],
l2=args.l2[1],
sample_h_states=True,
sample_v_states=True,
sparsity_cost=0.,
dbm_last=True, # !!!
metrics_config=dict(
msre=True,
pll=True,
train_metrics_every_iter=500,
),
verbose=True,
#display_filters=0,
#display_hidden_activations=24,
random_seed=args.random_seed[1],
v_shape=(20, 20), # what's the v_shape in this case?
freeze_weights=args.freeze_weights,
prune=args.prune,
#filter_shape = (10,10), #,args.filter_shape[1],
dtype='float32',
tf_saver_params=dict(max_to_keep=1),
model_path=args.rbm2_dirpath)
#run on cpu
config = tf.ConfigProto(device_count={'GPU': 0})
rbm2._tf_session_config = config
rbm2.init()
return rbm2
def make_rbm1(X, args):
if os.path.isdir(args.rbm1_dirpath):
print("\nLoading RBM #1 ...\n\n")
rbm1 = BernoulliRBM.load_model(args.rbm1_dirpath)
else:
print("\nTraining RBM #1 ...\n\n")
if not hasattr(args, 'double_rf'):
args.double_rf = False
rbm1 = BernoulliRBM(
n_visible=args.n_vis,
n_hidden=args.n_hidden[0],
W_init=args.w_init[0],
vb_init=args.vb_init[0],
hb_init=args.hb_init[0],
n_gibbs_steps=args.n_gibbs_steps[0],
learning_rate=args.lr[0],
momentum=args.momentum,
max_epoch=args.epochs[0],
batch_size=args.batch_size[0],
l2=args.l2[0],
sample_h_states=True,
sample_v_states=True,
sparsity_cost=0.,
dbm_first=True, # !!!
metrics_config=dict(
msre=True,
pll=True,
train_metrics_every_iter=500,
),
verbose=True,
#display_filters=30,
#display_hidden_activations=24,
v_shape=args.v_shape,
freeze_weights=args.freeze_weights,
prune=args.prune,
filter_shape=args.filter_shape[0],
double_rf=args.double_rf,
random_seed=args.random_seed[0],
dtype='float32',
tf_saver_params=dict(max_to_keep=1),
model_path=args.rbm1_dirpath)
#run on cpu
config = tf.ConfigProto(device_count={'GPU': 0})
rbm1._tf_session_config = config
rbm1.fit(X)
return rbm1