def train(hidden, output, sampling_points):
"""Train the given layers."""
print('Training...')
last_training_run = np.zeros(NUM_SAMPLES)
last_training_errors = np.zeros(NUM_SAMPLES)
for epoch in range(NUM_EPOCHS):
current_index = epoch % NUM_SAMPLES
next_index = (current_index + 1) % NUM_SAMPLES
if epoch % (20 * NUM_SAMPLES) == 0:
# change training frequency
frequency_factor = float(np.random.random_integers(NUM_TRAINING_FREQUENCIES))
sine_input = np.sin(frequency_factor * sampling_points)
frequency_input = [frequency_factor] * NUM_SAMPLES
input_data = np.insert(
frequency_input,
range(NUM_SAMPLES),
sine_input
).reshape(NUM_SAMPLES, 2)
hidden.clear()
# process inputs
outputs = output.process(hidden.process(input_data[current_index]))
# backpropagate errors
output.errors = sine_input[next_index] - outputs
hidden.errors.appendleft(expit_prime(hidden.h[0]) * np.dot(output.errors, output.weights))
for index in range(1, len(hidden.errors)):
hidden.errors[index] = expit_prime(hidden.h[index]) * np.dot(hidden.errors[index-1], hidden.recurrent_weights)
# learn
output.update()
hidden.update()
# log last run
if epoch >= (NUM_EPOCHS - NUM_SAMPLES):
last_training_run[next_index] = outputs
last_training_errors[next_index] = output.errors
return sine_input, last_training_run, last_training_errors
hidden = Layer(NUM_HIDDEN_NODES, input_data.shape[0], expit)
output = Layer(len(input_files), NUM_HIDDEN_NODES, softmax)
average_errors = np.ones(len(input_files))
accepted_errors = np.full(len(input_files), ACCEPTED_ERROR)
while not np.all(np.less(average_errors, accepted_errors)):
# get a random date
input_data, expected_value = get_input_data(input_files)
# process inputs
outputs = output.process(hidden.process(input_data))
# calculate errors
output.errors = expected_value - outputs
hidden.errors = expit_prime(hidden.h) * np.dot(output.errors, output.weights)
# update weights and biases
output.update()
hidden.update()
average_errors = update_ema(output.errors, average_errors)
print('last errors:\n', output.errors)
print('last outputs:\n', outputs)
print('last outputs rounded:\n', outputs.round())
#index = expected_value.nonzero()[0][0]
#print('index of class:', index)
#print('output of class:', outputs[index])
#print('error for class:', output.errors[index])
