def test_vae():
minibatch_size = 100
random_state = np.random.RandomState(1999)
graph = OrderedDict()
X_sym, y_sym = add_datasets_to_graph([X, y], ["X", "y"], graph)
l1_enc = relu_layer([X_sym, y_sym], graph, 'l1_enc', proj_dim=20,
random_state=random_state)
mu = linear_layer([l1_enc], graph, 'mu', proj_dim=10,
random_state=random_state)
log_sigma = linear_layer([l1_enc], graph, 'log_sigma', proj_dim=10,
random_state=random_state)
samp = gaussian_log_sample_layer([mu], [log_sigma], graph,
'gaussian_log_sample',
random_state=random_state)
l1_dec = relu_layer([samp], graph, 'l1_dec', proj_dim=20,
random_state=random_state)
out = sigmoid_layer([l1_dec], graph, 'out', proj_dim=X.shape[1],
random_state=random_state)
kl = gaussian_log_kl([mu], [log_sigma], graph, 'gaussian_kl').mean()
cost = binary_crossentropy(out, X_sym).mean() + kl
params, grads = get_params_and_grads(graph, cost)
learning_rate = 0.001
opt = sgd(params)
updates = opt.updates(params, grads, learning_rate)
train_function = theano.function([X_sym, y_sym], [cost], updates=updates,
mode="FAST_COMPILE")
iterate_function(train_function, [X, y], minibatch_size,
list_of_output_names=["cost"], n_epochs=1)
def test_vae():
minibatch_size = 10
random_state = np.random.RandomState(1999)
graph = OrderedDict()
X_sym = add_datasets_to_graph([X], ["X"], graph)
l1_enc = softplus_layer([X_sym], graph, 'l1_enc', proj_dim=100,
random_state=random_state)
mu = linear_layer([l1_enc], graph, 'mu', proj_dim=50,
random_state=random_state)
log_sigma = linear_layer([l1_enc], graph, 'log_sigma', proj_dim=50,
random_state=random_state)
samp = gaussian_log_sample_layer([mu], [log_sigma], graph,
'gaussian_log_sample',
random_state=random_state)
l1_dec = softplus_layer([samp], graph, 'l1_dec', proj_dim=100,
random_state=random_state)
out = sigmoid_layer([l1_dec], graph, 'out', proj_dim=X.shape[1],
random_state=random_state)
kl = gaussian_log_kl([mu], [log_sigma], graph, 'gaussian_kl').mean()
cost = binary_crossentropy(out, X_sym).mean() + kl
params, grads = get_params_and_grads(graph, cost)
learning_rate = 0.00000
opt = sgd(params, learning_rate)
updates = opt.updates(params, grads)
fit_function = theano.function([X_sym], [cost], updates=updates,
mode="FAST_COMPILE")
cost_function = theano.function([X_sym], [cost],
mode="FAST_COMPILE")
checkpoint_dict = {}
train_indices = np.arange(len(X))
valid_indices = np.arange(len(X))
early_stopping_trainer(fit_function, cost_function,
train_indices, valid_indices,
checkpoint_dict, [X],
minibatch_size,
list_of_train_output_names=["cost"],
valid_output_name="valid_cost",
n_epochs=1)
minibatch_size = 100
n_code = 400
n_enc_layer = [600, 600]
n_dec_layer = [600, 600]
width = 48
height = 48
n_input = width * height
# encode path aka q
l1_enc = softplus_layer([X_sym], graph, 'l1_enc', n_enc_layer[0], random_state)
l2_enc = softplus_layer([l1_enc], graph, 'l2_enc', n_enc_layer[1],
random_state)
code_mu = linear_layer([l2_enc], graph, 'code_mu', n_code, random_state)
code_log_sigma = linear_layer([l2_enc], graph, 'code_log_sigma', n_code,
random_state)
kl = gaussian_log_kl([code_mu], [code_log_sigma], graph, 'kl').mean()
samp = gaussian_log_sample_layer([code_mu], [code_log_sigma], graph, 'samp',
random_state)
# decode path aka p
l1_dec = softplus_layer([samp], graph, 'l1_dec', n_dec_layer[0], random_state)
l2_dec = softplus_layer([l1_dec], graph, 'l2_dec', n_dec_layer[1], random_state)
out = linear_layer([l2_dec], graph, 'out', n_input, random_state)
nll = squared_error(out, X_sym).mean()
# log p(x) = -nll so swap sign
# want to minimize cost in optimization so multiply by -1
cost = -1 * (-nll - kl)
params, grads = get_params_and_grads(graph, cost)
learning_rate = 0.0003
def test_gaussian_log_kl():
graph = OrderedDict()
X_sym = add_datasets_to_graph([X], ["X"], graph)
kl = gaussian_log_kl([X_sym, X_sym], [X_sym, X_sym], graph,
'gaussian_log_kl')
theano.function([X_sym], [kl], mode="FAST_COMPILE")
n_code = 100
n_hid = 200
width = 28
height = 28
n_input = width * height
# encode path aka q
l1_enc = softplus([X_sym], [X.shape[1]], proj_dim=n_hid, name='l1_enc',
random_state=random_state)
l2_enc = softplus([l1_enc], [n_hid], proj_dim=n_hid, name='l2_enc',
random_state=random_state)
code_mu = linear([l2_enc], [n_hid], proj_dim=n_code, name='code_mu',
random_state=random_state)
code_log_sigma = linear([l2_enc], [n_hid], proj_dim=n_code,
name='code_log_sigma', random_state=random_state)
kl = gaussian_log_kl([code_mu], [code_log_sigma]).mean()
sample_state = np.random.RandomState(2177)
samp = gaussian_log_sample([code_mu], [code_log_sigma], name='samp',
random_state=sample_state)
# decode path aka p
l1_dec = softplus([samp], [n_code], proj_dim=n_hid, name='l1_dec',
random_state=random_state)
l2_dec = softplus([l1_dec], [n_hid], proj_dim=n_hid, name='l2_dec',
random_state=random_state)
out = sigmoid([l2_dec], [n_hid], proj_dim=X.shape[1], name='out',
random_state=random_state)
nll = binary_crossentropy(out, X_sym).mean()
# See https://arxiv.org/pdf/1406.5298v2.pdf, eq 5
# log p(x | z) = -nll so swap sign
proj_dim=n_hid,
name='l1_enc',
random_state=random_state)
l2_enc = softplus([l1_enc], [n_hid],
proj_dim=n_hid,
name='l2_enc',
random_state=random_state)
code_mu = linear([l2_enc], [n_hid],
proj_dim=n_code,
name='code_mu',
random_state=random_state)
code_log_sigma = linear([l2_enc], [n_hid],
proj_dim=n_code,
name='code_log_sigma',
random_state=random_state)
kl = gaussian_log_kl([code_mu], [code_log_sigma]).mean()
sample_state = np.random.RandomState(2177)
samp = gaussian_log_sample([code_mu], [code_log_sigma],
name='samp',
random_state=sample_state)
# decode path aka p
l1_dec = softplus([samp], [n_code],
proj_dim=n_hid,
name='l1_dec',
random_state=random_state)
l2_dec = softplus([l1_dec], [n_hid],
proj_dim=n_hid,
name='l2_dec',
random_state=random_state)
out = sigmoid([l2_dec], [n_hid],
def test_gaussian_log_kl():
kl = gaussian_log_kl([X_sym, X_sym], [X_sym, X_sym])
theano.function([X_sym], [kl], mode="FAST_COMPILE")
def test_gaussian_log_kl():
kl = gaussian_log_kl([X_sym, X_sym], [X_sym, X_sym])
theano.function([X_sym], [kl], mode="FAST_COMPILE")
def test_vae():
minibatch_size = 10
random_state = np.random.RandomState(1999)
graph = OrderedDict()
X_sym = add_datasets_to_graph([X], ["X"], graph)
l1_enc = softplus_layer([X_sym],
graph,
'l1_enc',
proj_dim=100,
random_state=random_state)
mu = linear_layer([l1_enc],
graph,
'mu',
proj_dim=50,
random_state=random_state)
log_sigma = linear_layer([l1_enc],
graph,
'log_sigma',
proj_dim=50,
random_state=random_state)
samp = gaussian_log_sample_layer([mu], [log_sigma],
graph,
'gaussian_log_sample',
random_state=random_state)
l1_dec = softplus_layer([samp],
graph,
'l1_dec',
proj_dim=100,
random_state=random_state)
out = sigmoid_layer([l1_dec],
graph,
'out',
proj_dim=X.shape[1],
random_state=random_state)
kl = gaussian_log_kl([mu], [log_sigma], graph, 'gaussian_kl').mean()
cost = binary_crossentropy(out, X_sym).mean() + kl
params, grads = get_params_and_grads(graph, cost)
learning_rate = 0.00000
opt = sgd(params)
updates = opt.updates(params, grads, learning_rate)
fit_function = theano.function([X_sym], [cost],
updates=updates,
mode="FAST_COMPILE")
cost_function = theano.function([X_sym], [cost], mode="FAST_COMPILE")
checkpoint_dict = {}
train_indices = np.arange(len(X))
valid_indices = np.arange(len(X))
early_stopping_trainer(fit_function,
cost_function,
checkpoint_dict, [X],
minibatch_size,
train_indices,
valid_indices,
fit_function_output_names=["cost"],
cost_function_output_name="valid_cost",
n_epochs=1)