from net import RecurrentNetwork, load_librispeech, labels_to_chars
import numpy as np
import matplotlib.pyplot as plt
(train_x, train_y), (valid_x, valid_y), (test_x, test_y) = load_librispeech()
clf = RecurrentNetwork(learning_alg="rmsprop",
hidden_layer_sizes=[
500,
],
max_iter=100,
cost="ctc",
bidirectional=True,
learning_rate=0.0001,
momentum=0.9,
recurrent_activation="lstm",
random_seed=1999)
print(labels_to_chars(train_y[2]))
means = np.mean(train_x[2], axis=0)
std = np.std(train_x[2], axis=0)
tx = (train_x[2] - means) / std
clf.fit(tx, train_y[2])
from IPython import embed
embed()
from net import RecurrentNetwork, load_librispeech, labels_to_chars
import numpy as np
import matplotlib.pyplot as plt
(train_x, train_y), (valid_x, valid_y), (test_x, test_y) = load_librispeech()
clf = RecurrentNetwork(learning_alg="rmsprop", hidden_layer_sizes=[500,],
max_iter=100, cost="ctc", bidirectional=True,
learning_rate=0.0001, momentum=0.9,
recurrent_activation="lstm",
random_seed=1999)
print(labels_to_chars(train_y[2]))
means = np.mean(train_x[2], axis=0)
std = np.std(train_x[2], axis=0)
tx = (train_x[2] - means) / std
clf.fit(tx, train_y[2])
from IPython import embed; embed()
time_steps = 10
n_seq = 100
# n_y is equal to the number of calsses
random_state = np.random.RandomState(1999)
seq = random_state.randn(n_seq, time_steps, n_u)
targets = np.zeros((n_seq, time_steps), dtype=np.int32)
thresh = 0.5
targets[:, 2:][seq[:, 1:-1, 1] > seq[:, :-2, 0] + thresh] = 1
targets[:, 2:][seq[:, 1:-1, 1] < seq[:, :-2, 0] - thresh] = 2
clf = RecurrentNetwork(learning_alg="sgd",
hidden_layer_sizes=[n_h],
max_iter=10,
cost="softmax",
learning_rate=0.1,
momentum=0.99,
recurrent_activation="lstm",
random_seed=1999)
clf.fit(seq, targets)
clf.predict(seq)
plt.close('all')
fig = plt.figure()
plt.grid()
ax1 = plt.subplot(211)
plt.scatter(np.arange(time_steps), targets[1], marker='o', c='b')
plt.grid()
from net import RecurrentNetwork, load_fruitspeech, labels_to_chars
import matplotlib.pyplot as plt
import numpy as np
(train_x, train_y), (valid_x, valid_y), (test_x, test_y) = load_fruitspeech()
clf = RecurrentNetwork(learning_alg="rmsprop",
hidden_layer_sizes=[500],
max_iter=100,
cost="encdec",
bidirectional=True,
learning_rate=0.00002,
momentum=0.9,
recurrent_activation="lstm",
random_seed=1999)
all_frames = np.vstack(train_x)
means = np.mean(all_frames, axis=0)
std = np.std(all_frames, axis=0)
for n, t in enumerate(train_x):
train_x[n] = (t - means) / std
for n, v in enumerate(valid_x):
valid_x[n] = (v - means) / std
from IPython import embed
embed()
clf.fit(train_x, train_y, valid_x, valid_y)
y_hat = labels_to_chars(clf.predict(valid_x[0])[0])
y = labels_to_chars(valid_y[0])
print(y_hat)
n_h = 6
n_y = 3
time_steps = 10
n_seq = 100
# n_y is equal to the number of calsses
random_state = np.random.RandomState(1999)
seq = random_state.randn(n_seq, time_steps, n_u)
targets = np.zeros((n_seq, time_steps), dtype=np.int32)
thresh = 0.5
targets[:, 2:][seq[:, 1:-1, 1] > seq[:, :-2, 0] + thresh] = 1
targets[:, 2:][seq[:, 1:-1, 1] < seq[:, :-2, 0] - thresh] = 2
clf = RecurrentNetwork(learning_alg="sgd", hidden_layer_sizes=[n_h, n_h],
max_iter=1E3, cost="softmax", learning_rate=0.1,
momentum=0.99, recurrent_activation="lstm",
bidirectional=True, random_seed=1999)
clf.fit(seq, targets)
plt.close('all')
fig = plt.figure()
plt.grid()
ax1 = plt.subplot(211)
plt.scatter(np.arange(time_steps), targets[1], marker='o', c='b')
plt.grid()
guess = clf.predict_proba(seq[1])
guessed_probs = plt.imshow(guess[0].T, interpolation='nearest', cmap='gray')
ax1.set_title('blue points: true class, grayscale: model output (white mean class)')
from net import RecurrentNetwork, load_cmuarctic, labels_to_chars
import matplotlib.pyplot as plt
import numpy as np
(train_x, train_y), (valid_x, valid_y), (test_x, test_y) = load_cmuarctic()
clf = RecurrentNetwork(learning_alg="sfg",
hidden_layer_sizes=[500],
max_col_norm=1.9635,
max_iter=1000, cost="ctc", bidirectional=True,
learning_rate=0.0001, momentum=0.9,
recurrent_activation="lstm",
random_seed=1999)
tx = train_x[2]
tx = (tx - tx.mean()) / tx.std()
clf.fit(train_x[2], train_y[2])
y = labels_to_chars(train_y[2])
print(y)
from net import RecurrentNetwork, load_fruitspeech, labels_to_chars
import matplotlib.pyplot as plt
import numpy as np
(train_x, train_y), (valid_x, valid_y), (test_x, test_y) = load_fruitspeech()
clf = RecurrentNetwork(learning_alg="rmsprop",
hidden_layer_sizes=[500],
max_iter=100, cost="encdec", bidirectional=True,
learning_rate=0.00002, momentum=0.9,
recurrent_activation="lstm",
random_seed=1999)
all_frames = np.vstack(train_x)
means = np.mean(all_frames, axis=0)
std = np.std(all_frames, axis=0)
for n, t in enumerate(train_x):
train_x[n] = (t - means) / std
for n, v in enumerate(valid_x):
valid_x[n] = (v - means) / std
from IPython import embed; embed()
clf.fit(train_x, train_y, valid_x, valid_y)
y_hat = labels_to_chars(clf.predict(valid_x[0])[0])
y = labels_to_chars(valid_y[0])
print(y_hat)
print(y)