def test_fprop_double_layer_one_weights_positive_output_values_relu(self):
n_vis = 8
n_hid = 2
hidden_layer_1 = HiddenLayer(n_vis=n_vis,
n_hid=n_vis / 2,
layer_name='h1',
activation='relu',
param_init_range=0,
alpha=0)
hidden_layer_2 = HiddenLayer(n_vis=n_vis / 2,
n_hid=n_hid,
layer_name='h2',
activation='relu',
param_init_range=0,
alpha=0)
W = theano.shared(value=np.ones((n_vis, n_vis / 2)),
name='h1_W',
borrow=True)
hidden_layer_1.W = W
W = theano.shared(value=np.ones((n_vis / 2, n_hid)),
name='h2_W',
borrow=True)
hidden_layer_2.W = W
mlp = QNetwork([hidden_layer_1, hidden_layer_2],
discount=1,
learning_rate=1)
features = np.ones(n_vis)
actual = list(mlp.fprop(features).eval())
expected = [32., 32.]
self.assertSequenceEqual(actual, expected)
def test_fprop_single_layer_zero_weights_positive_input_values_relu(self):
hidden_layer = HiddenLayer(n_vis=4, n_hid=2, layer_name='h', activation='relu', param_init_range=0, alpha=0)
mlp = QNetwork([hidden_layer], discount=1, learning_rate=1)
features = [1, 2, 3, 4]
actual = list(mlp.fprop(features).eval())
expected = [0., 0.]
self.assertSequenceEqual(actual, expected)
def test_fprop_single_layer_one_weights_negative_output_values_relu(self):
n_vis = 4
n_hid = 2
hidden_layer = HiddenLayer(n_vis=n_vis, n_hid=n_hid, layer_name='h', activation='relu', param_init_range=0, alpha=0)
W = theano.shared(value=np.ones((n_vis, n_hid)), name='h_W', borrow=True)
hidden_layer.W = W
mlp = QNetwork([hidden_layer], discount=1, learning_rate=1)
features = [-1, -2, -3, 4]
actual = list(mlp.fprop(features).eval())
expected = [0., 0.]
self.assertSequenceEqual(actual, expected)
def test_fprop_single_layer_zero_weights_positive_input_values_relu(self):
hidden_layer = HiddenLayer(n_vis=4,
n_hid=2,
layer_name='h',
activation='relu',
param_init_range=0,
alpha=0)
mlp = QNetwork([hidden_layer], discount=1, learning_rate=1)
features = [1, 2, 3, 4]
actual = list(mlp.fprop(features).eval())
expected = [0., 0.]
self.assertSequenceEqual(actual, expected)
def test_fprop_double_layer_one_weights_negative_output_values_relu(self):
n_vis = 8
n_hid = 2
hidden_layer_1 = HiddenLayer(n_vis=n_vis, n_hid=n_vis / 2, layer_name='h1', activation='relu', param_init_range=0, alpha=0)
hidden_layer_2 = HiddenLayer(n_vis=n_vis / 2, n_hid=n_hid, layer_name='h2', activation='relu', param_init_range=0, alpha=0)
W = theano.shared(value=np.ones((n_vis, n_vis / 2)), name='h1_W', borrow=True)
hidden_layer_1.W = W
W = theano.shared(value=np.ones((n_vis / 2, n_hid)), name='h2_W', borrow=True)
hidden_layer_2.W = W
mlp = QNetwork([hidden_layer_1, hidden_layer_2], discount=1, learning_rate=1)
features = [-5, -4, -3, -2, -1, 0, 1, 2]
actual = list(mlp.fprop(features).eval())
expected = [0., 0.]
self.assertSequenceEqual(actual, expected)
def test_fprop_single_layer_one_weights_negative_output_values_relu(self):
n_vis = 4
n_hid = 2
hidden_layer = HiddenLayer(n_vis=n_vis,
n_hid=n_hid,
layer_name='h',
activation='relu',
param_init_range=0,
alpha=0)
W = theano.shared(value=np.ones((n_vis, n_hid)),
name='h_W',
borrow=True)
hidden_layer.W = W
mlp = QNetwork([hidden_layer], discount=1, learning_rate=1)
features = [-1, -2, -3, 4]
actual = list(mlp.fprop(features).eval())
expected = [0., 0.]
self.assertSequenceEqual(actual, expected)
