def test_lstm_rnn():
# random state so script is deterministic
random_state = np.random.RandomState(1999)
# home of the computational graph
graph = OrderedDict()
# number of hidden features
n_hid = 10
# number of output_features = input_features
n_out = X.shape[-1]
# input (where first dimension is time)
datasets_list = [X, X_mask, y, y_mask]
names_list = ["X", "X_mask", "y", "y_mask"]
test_values_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=test_values_list)
# Setup weights
l1 = linear_layer([X_sym], graph, 'l1_proj', n_hid, random_state)
h = lstm_recurrent_layer([l1], X_mask_sym, n_hid, graph, 'l1_rec',
random_state)
# linear output activation
y_hat = linear_layer([h], graph, 'l2_proj', n_out, random_state)
# error between output and target
cost = squared_error(y_hat, y_sym)
cost = masked_cost(cost, y_mask_sym).mean()
# Parameters of the model
params, grads = get_params_and_grads(graph, cost)
# Use stochastic gradient descent to optimize
opt = sgd(params)
learning_rate = 0.01
updates = opt.updates(params, grads, learning_rate)
fit_function = theano.function([X_sym, X_mask_sym, y_sym, y_mask_sym],
[cost],
updates=updates,
mode="FAST_COMPILE")
cost_function = theano.function([X_sym, X_mask_sym, y_sym, y_mask_sym],
[cost],
mode="FAST_COMPILE")
checkpoint_dict = {}
train_indices = np.arange(X.shape[1])
valid_indices = np.arange(X.shape[1])
early_stopping_trainer(fit_function,
cost_function,
checkpoint_dict, [X, y],
minibatch_size,
train_indices,
valid_indices,
fit_function_output_names=["cost"],
cost_function_output_name="valid_cost",
n_epochs=1)
def test_lstm_rnn():
# random state so script is deterministic
random_state = np.random.RandomState(1999)
# home of the computational graph
graph = OrderedDict()
# number of hidden features
n_hid = 10
# number of output_features = input_features
n_out = X.shape[-1]
# input (where first dimension is time)
datasets_list = [X, X_mask, y, y_mask]
names_list = ["X", "X_mask", "y", "y_mask"]
test_values_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=test_values_list)
# Setup weights
l1 = linear_layer([X_sym], graph, 'l1_proj', n_hid,
random_state=random_state)
h = lstm_recurrent_layer([l1], X_mask_sym, n_hid, graph, 'l1_rec',
random_state)
# linear output activation
y_hat = linear_layer([h], graph, 'l2_proj', n_out,
random_state=random_state)
# error between output and target
cost = squared_error(y_hat, y_sym)
cost = masked_cost(cost, y_mask_sym).mean()
# Parameters of the model
params, grads = get_params_and_grads(graph, cost)
# Use stochastic gradient descent to optimize
learning_rate = 0.01
opt = sgd(params, learning_rate)
updates = opt.updates(params, grads)
fit_function = theano.function([X_sym, X_mask_sym, y_sym, y_mask_sym],
[cost], updates=updates, mode="FAST_COMPILE")
cost_function = theano.function([X_sym, X_mask_sym, y_sym, y_mask_sym],
[cost], mode="FAST_COMPILE")
checkpoint_dict = {}
train_indices = np.arange(X.shape[1])
valid_indices = np.arange(X.shape[1])
early_stopping_trainer(fit_function, cost_function,
train_indices, valid_indices,
checkpoint_dict,
[X, y], minibatch_size,
list_of_train_output_names=["cost"],
valid_output_name="valid_cost",
n_epochs=1)
n_hid = 300
rnn_dim = 1200
X_mb, X_mb_mask = make_masked_minibatch(X, slice(0, minibatch_size))
y_mb, y_mb_mask = make_masked_minibatch(y, slice(0, minibatch_size))
datasets_list = [X_mb, X_mb_mask, y_mb, y_mb_mask]
names_list = ["X", "X_mask", "y", "y_mask"]
graph = OrderedDict()
X_sym, X_mask_sym, y_sym, y_mask_sym = add_datasets_to_graph(
datasets_list, names_list, graph)
l1 = relu_layer([X_sym], graph, 'l1', proj_dim=n_hid,
random_state=random_state)
h = lstm_recurrent_layer([l1], X_mask_sym, rnn_dim, graph, 'l1_rec',
random_state=random_state)
l2 = relu_layer([h], graph, 'l2', proj_dim=n_hid,
random_state=random_state)
rval = bernoulli_and_correlated_log_gaussian_mixture_layer(
[l2], graph, 'hw', proj_dim=2, n_components=20, random_state=random_state)
binary, coeffs, mus, sigmas, corr = rval
cost = bernoulli_and_correlated_log_gaussian_mixture_cost(
binary, coeffs, mus, sigmas, corr, y_sym)
cost = masked_cost(cost, y_mask_sym).sum(axis=0).mean()
params, grads = get_params_and_grads(graph, cost)
learning_rate = 0.0003
opt = adam(params, learning_rate)
clipped_grads = gradient_clipping(grads)
updates = opt.updates(params, clipped_grads)
train_itr = list_iterator([X, y], minibatch_size, axis=1, make_mask=True,
stop_index=train_end)
X_mb, X_mb_mask, y_mb, y_mb_mask = next(train_itr)
train_itr.reset()
datasets_list = [X_mb, X_mb_mask, y_mb, y_mb_mask]
names_list = ["X", "X_mask", "y", "y_mask"]
graph = OrderedDict()
X_sym, X_mask_sym, y_sym, y_mask_sym = add_datasets_to_graph(
datasets_list, names_list, graph)
l1 = relu_layer([X_sym], graph, 'l1', proj_dim=n_hid,
random_state=random_state)
h = lstm_recurrent_layer([l1], X_mask_sym, rnn_dim, graph, 'l1_rec',
random_state=random_state)
l2 = relu_layer([h], graph, 'l2', proj_dim=n_hid,
random_state=random_state)
rval = bernoulli_and_correlated_log_gaussian_mixture_layer(
[l2], graph, 'hw', proj_dim=2, n_components=20, random_state=random_state)
binary, coeffs, mus, sigmas, corr = rval
cost = bernoulli_and_correlated_log_gaussian_mixture_cost(
binary, coeffs, mus, sigmas, corr, y_sym)
cost = masked_cost(cost, y_mask_sym).sum(axis=0).mean()
params, grads = get_params_and_grads(graph, cost)
learning_rate = 0.0003
opt = adam(params, learning_rate)
clipped_grads = gradient_clipping(grads)
updates = opt.updates(params, clipped_grads)
