def plot_estimates():
X, y = gen_data()
y_predictions = y_pred(X)
ax = plt.gca()
ax.plot(y_predictions[0, :, 0], label="estimate")
ax.plot(y[0, :, 0], label="ground truth")
# ax.plot(X[0,:,0], label='aggregate')
ax.legend()
plt.show()
def plot_estimates():
X, y = gen_data()
y_predictions = y_pred(X)
ax = plt.gca()
ax.plot(y_predictions[0, :, 0], label='estimate')
ax.plot(y[0, :, 0], label='ground truth')
# ax.plot(X[0,:,0], label='aggregate')
ax.legend()
plt.show()
def run_training():
costs = np.zeros(N_ITERATIONS)
for n in range(N_ITERATIONS):
X, y = gen_data()
# you should use your own training data mask instead of mask_val
costs[n] = train(X, y)
if not n % 10:
cost_val = compute_cost(X_val, y_val)
print "Iteration {} validation cost = {}".format(n, cost_val)
plt.plot(costs)
plt.xlabel("Iteration")
plt.ylabel("Cost")
plt.show()
def run_training():
costs = np.zeros(N_ITERATIONS)
for n in range(N_ITERATIONS):
X, y = gen_data()
# you should use your own training data mask instead of mask_val
costs[n] = train(X, y)
if not n % 10:
cost_val = compute_cost(X_val, y_val)
print "Iteration {} validation cost = {}".format(n, cost_val)
plt.plot(costs)
plt.xlabel('Iteration')
plt.ylabel('Cost')
plt.show()
import theano.tensor as T import lasagne from gen_data_029 import gen_data, N_BATCH, LENGTH theano.config.compute_test_value = "raise" # Number of units in the hidden (recurrent) layer N_HIDDEN = 5 # SGD learning rate LEARNING_RATE = 1e-1 # Number of iterations to train the net N_ITERATIONS = 200 # Generate a "validation" sequence whose cost we will periodically compute X_val, y_val = gen_data() n_features = X_val.shape[-1] n_output = y_val.shape[-1] assert X_val.shape == (N_BATCH, LENGTH, n_features) assert y_val.shape == (N_BATCH, LENGTH, n_output) # Construct LSTM RNN: One LSTM layer and one dense output layer l_in = lasagne.layers.InputLayer(shape=(N_BATCH, LENGTH, n_features)) # setup fwd and bck LSTM layer. l_fwd = lasagne.layers.LSTMLayer(l_in, N_HIDDEN, backwards=False, learn_init=True, peepholes=True) l_bck = lasagne.layers.LSTMLayer(l_in, N_HIDDEN, backwards=True, learn_init=True, peepholes=True) # concatenate forward and backward LSTM layers
"""
tanh output
lower learning rate
* does just about learn something sensible, but not especially convincing,
even after 2000 iterations.
"""
# Number of units in the hidden (recurrent) layer
N_HIDDEN = 5
# SGD learning rate
LEARNING_RATE = 1e-2
# Number of iterations to train the net
N_ITERATIONS = 2000
# Generate a "validation" sequence whose cost we will periodically compute
X_val, y_val = gen_data()
n_features = X_val.shape[-1]
n_output = y_val.shape[-1]
assert X_val.shape == (N_BATCH, LENGTH, n_features)
assert y_val.shape == (N_BATCH, LENGTH, n_output)
# Construct LSTM RNN: One LSTM layer and one dense output layer
l_in = lasagne.layers.InputLayer(shape=(N_BATCH, LENGTH, n_features))
# setup fwd and bck LSTM layer.
l_fwd = lasagne.layers.LSTMLayer(l_in,
N_HIDDEN,
backwards=False,
learn_init=True,
peepholes=True)