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)
l1_dec = softplus_layer([samp, y_sym],
graph,
'l1_dec',
n_dec_layer,
random_state=random_state)
l2_dec = softplus_layer([l1_dec],
graph,
'l2_dec',
n_dec_layer,
random_state=random_state)
l3_dec = softplus_layer([l2_dec],
graph,
'l3_dec',
n_dec_layer,
random_state=random_state)
out = sigmoid_layer([l3_dec], graph, 'out', n_input, random_state=random_state)
nll = binary_crossentropy(out, X_sym).mean()
# log p(x) = -nll so swap sign
# want to minimize cost in optimization so multiply by -1
base_cost = -1 * (-nll - kl)
# -log q(y | x) is negative log likelihood already
alpha = 0.1
err = categorical_crossentropy(y_pred, y_sym).mean()
cost = base_cost + alpha * err
params, grads = get_params_and_grads(graph, cost)
learning_rate = 0.0001
opt = adam(params, learning_rate)
X_mb, X_mb_mask, y_mb, y_mb_mask = next(train_itr)
train_itr.reset()
valid_itr = minibatch_iterator([X, y], minibatch_size, make_mask=True, axis=1)
datasets_list = [X_mb, X_mb_mask, y_mb, y_mb_mask]
names_list = ["X", "X_mask", "y", "y_mask"]
X_sym, X_mask_sym, y_sym, y_mask_sym = add_datasets_to_graph(
datasets_list, names_list, graph, list_of_test_values=datasets_list)
n_hid = 256
n_out = 8
h = location_attention_tanh_recurrent_layer(
[X_sym], [y_sym], X_mask_sym, y_mask_sym, n_hid, graph, 'l1_att_rec',
random_state=random_state)
X_hat = sigmoid_layer([h], graph, 'output', proj_dim=n_out,
random_state=random_state)
cost = binary_crossentropy(X_hat, X_sym).mean()
cost = masked_cost(cost, X_mask_sym).mean()
params, grads = get_params_and_grads(graph, cost)
opt = adadelta(params)
updates = opt.updates(params, grads)
fit_function = theano.function([X_sym, X_mask_sym, y_sym, y_mask_sym],
[cost], updates=updates)
valid_function = theano.function([X_sym, X_mask_sym, y_sym, y_mask_sym], [cost])
checkpoint_dict = {}
checkpoint_dict["fit_function"] = fit_function
checkpoint_dict["valid_function"] = valid_function
TL = TrainingLoop(fit_function, valid_function, train_itr, valid_itr,
checkpoint_dict=checkpoint_dict,
list_of_train_output_names=["train_cost"],
# 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 = sigmoid_layer([l2_dec], graph, 'out', n_input, random_state)
nll = binary_crossentropy(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
opt = adam(params)
updates = opt.updates(params, grads, learning_rate)
# Checkpointing
try:
checkpoint_dict = load_last_checkpoint()
fit_function = checkpoint_dict["fit_function"]
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)