def test_conv2d_layer():
random_state = np.random.RandomState(42)
graph = OrderedDict()
# 3 channel mnist
X_r = np.random.randn(10, 3, 28, 28).astype(theano.config.floatX)
X_sym = add_datasets_to_graph([X_r], ["X"], graph)
l1 = conv2d_layer([X_sym], graph, 'l1', 5, random_state=random_state)
# test that they can stack as well
l2 = conv2d_layer([l1], graph, 'l2', 6, random_state=random_state)
f = theano.function([X_sym], [l1, l2], mode="FAST_COMPILE")
l1, l2 = f(X_r)
train_indices = mnist["train_indices"]
valid_indices = mnist["valid_indices"]
X = mnist["images"]
y = mnist["target"]
n_targets = 10
y = convert_to_one_hot(y, n_targets)
minibatch_size = 128
# graph holds information necessary to build layers from parents
graph = OrderedDict()
X_sym, y_sym = add_datasets_to_graph([X[:minibatch_size], y[:minibatch_size]],
["X", "y"], graph)
# random state so script is deterministic
random_state = np.random.RandomState(1999)
l1 = conv2d_layer([X_sym], graph, 'conv1', 8, random_state=random_state)
l2 = pool2d_layer([l1], graph, 'pool1')
l3 = conv2d_layer([l2], graph, 'conv2', 16, random_state=random_state)
l4 = pool2d_layer([l3], graph, 'pool2')
l5 = l4.reshape((l4.shape[0], -1))
y_pred = softmax_layer([l5],
graph,
'y_pred',
n_targets,
random_state=random_state)
nll = categorical_crossentropy(y_pred, y_sym).mean()
cost = nll
params, grads = get_params_and_grads(graph, cost)
learning_rate = 0.001