def test_rnn_correlated_mixture_density():
# graph holds information necessary to build layers from parents
random_state = np.random.RandomState(1999)
graph = OrderedDict()
minibatch_size = 5
X_seq = np.array([bernoulli_X for i in range(minibatch_size)])
y_seq = np.array([bernoulli_y for i in range(minibatch_size)])
X_mb, X_mb_mask = make_masked_minibatch(X_seq, slice(0, minibatch_size))
y_mb, y_mb_mask = make_masked_minibatch(y_seq, slice(0, minibatch_size))
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)
n_hid = 5
train_indices = np.arange(len(X_seq))
valid_indices = np.arange(len(X_seq))
l1 = tanh_layer([X_sym], graph, 'l1', proj_dim=n_hid,
random_state=random_state)
h = gru_recurrent_layer([l1], X_mask_sym, n_hid, graph, 'l1_rec',
random_state=random_state)
rval = bernoulli_and_correlated_log_gaussian_mixture_layer(
[h], graph, 'hw', proj_dim=2, n_components=3,
random_state=random_state)
binary, coeffs, mus, log_sigmas, corr = rval
cost = bernoulli_and_correlated_log_gaussian_mixture_cost(
binary, coeffs, mus, log_sigmas, corr, y_sym)
cost = masked_cost(cost, y_mask_sym).mean()
cost_function = theano.function([X_sym, X_mask_sym, y_sym, y_mask_sym],
[cost],
mode="FAST_COMPILE")
checkpoint_dict = create_checkpoint_dict(locals())
epoch_results = fixed_n_epochs_trainer(
cost_function, cost_function, train_indices, valid_indices,
checkpoint_dict, [X_seq, y_seq],
minibatch_size,
list_of_minibatch_functions=[make_masked_minibatch,
make_masked_minibatch],
list_of_train_output_names=["train_cost"],
valid_output_name="valid_cost",
n_epochs=1)
def normalize(x):
return np.hstack((x[:, 0][:, None], (x[:, 1:] - X_mean) / (X_std)))
def unnormalize(x):
return np.hstack((x[:, 0][:, None], (x[:, 1:] * X_std) + X_mean))
X = np.array([normalize(x) for x in X])
y = np.array([x[1:] for x in X])
X = np.array([x[:-1] for x in X])
minibatch_size = 20 # Size must match size in training, same for above preproc
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))
running_mb = X_mb[:1] * 0
running_mask = X_mb_mask[:1] * 0
def gen_sample(rval, random_state, idx=-1):
# binary
# coeffs
# mus
# log_sigmas
# corr
binary, coeffs, mus, sigmas, corr = rval
binary = binary[idx, :, 0]
coeffs = coeffs[idx, :, :]
mu_x = mus[idx, :, 0, :]
import numpy as np import theano from collections import OrderedDict random_state = np.random.RandomState(1999) graph = OrderedDict() data = make_ocr(["Hello", "humans"]) X = data["data"] y = data["target"] vocab_size = data["vocabulary_size"] vocab = data["vocabulary"] train_indices = data["train_indices"] valid_indices = train_indices X_mb, X_mb_mask = make_masked_minibatch(X, slice(0, len(X))) y_mb, y_mb_mask = make_masked_minibatch(y, slice(0, len(y))) n_hid = 256 n_out = vocab_size + 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) h = gru_recurrent_layer([X_sym], X_mask_sym, n_hid, graph, "l1_rec", random_state=random_state) y_pred = softmax_layer([h], graph, "l2_proj", n_out, random_state=random_state) cost = log_ctc_cost(y_sym, y_mask_sym, y_pred, X_mask_sym).mean() params, grads = get_params_and_grads(graph, cost)
import numpy as np
import theano
from collections import OrderedDict
random_state = np.random.RandomState(1999)
graph = OrderedDict()
data = make_ocr(["Hello", "humans"])
X = data["data"]
y = data["target"]
vocab_size = data["vocabulary_size"]
vocab = data["vocabulary"]
train_indices = data["train_indices"]
valid_indices = train_indices
X_mb, X_mb_mask = make_masked_minibatch(X, slice(0, len(X)))
y_mb, y_mb_mask = make_masked_minibatch(y, slice(0, len(y)))
n_hid = 256
n_out = vocab_size + 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)
h = gru_recurrent_layer([X_sym],
X_mask_sym,
n_hid,
graph,
'l1_rec',
from collections import OrderedDict
random_state = np.random.RandomState(1999)
graph = OrderedDict()
data = fetch_fruitspeech()
X = data["specgrams"]
y = data["target"]
vocab_size = data["vocabulary_size"]
vocab = data["vocabulary"]
train_indices = data["train_indices"]
valid_indices = data["valid_indices"]
minibatch_size = 10
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))
n_hid = 500
n_out = vocab_size + 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)
l1 = maxout_layer([X_sym], graph, 'l1', n_hid, random_state=random_state)
h = bidirectional_gru_recurrent_layer([l1],
X_mask_sym,
n_hid,
graph,
