train_cost, cv_cost = \
nn.train_backprop(train_input, train_output,
d_f_list=[d_sigmoid, d_sigmoid],
goal=log_Bernoulli_likelihood,
d_goal=d_log_Bernoulli_likelihood,
batch_size=None,
max_iter=2500,
learning_rate=0.1,
momentum_rate=0.9,
neural_local_gain=(0.005, 0.995, 0.001, 1000),
stop_threshold=0.05,
cv_input_data=cv_input,
cv_output_data=cv_output,
#regularization_rate=0.1,
#regularization_norm=l2,
#d_regularization_norm=d_l2
verbose=True
)
t = np.argmax(train_output, axis=1)
y = np.argmax(nn.compute_output(train_input), axis=1)
print('%s / %s' % (sum(t == y), train_output.shape[0]))
t = np.argmax(test_output, axis=1)
y = np.argmax(nn.compute_output(test_input), axis=1)
print('%s / %s' % (sum(t == y), test_output.shape[0]))
activation_functions=[tanh, softmax],
rng=(lambda n: np.random.normal(0, 0.01, n)))
train_cost, cv_cost = \
nn.train_backprop(train_input, train_output,
d_f_list=[d_tanh, d_softmax],
goal=cross_entropy,
d_goal=d_cross_entropy,
batch_size=1,
max_iter=100,
learning_rate=0.01,
momentum_rate=0.9,
#neural_local_gain=(0.0005, 0.9995, 0.001, 1000),
stop_threshold=0.05,
cv_input_data=cv_input,
cv_output_data=cv_output,
#regularization_rate=0.1,
#regularization_norm=l2,
#d_regularization_norm=d_l2,
verbose=True
)
t = np.argmax(train_output, axis=1)
y = np.argmax(nn.compute_output(train_input), axis=1)
print('%s / %s' % (sum(t == y), train_output.shape[0]))
t = np.argmax(test_output, axis=1)
y = np.argmax(nn.compute_output(test_input), axis=1)
print('%s / %s' % (sum(t == y), test_output.shape[0]))
