def get_accuracy(structure):
structure = [25] + list(structure) + [10]
instances_by_label = util.read_training("train.csv")
for label in instances_by_label:
for instance in instances_by_label[label]:
instance /= max(instance.max(), -instance.min())
np.random.seed(0)
test_instances = []
test_labels = []
for label in instances_by_label:
np.random.shuffle(instances_by_label[label])
test_instances.extend(instances_by_label[label][:1000])
test_labels.extend([label]*1000)
instances_by_label[label] = instances_by_label[label][1000:]
max_accuracy, paras = train_nn(structure, instances_by_label, test_instances, test_labels)
return max_accuracy[1]
np.random.shuffle(instances_by_label[label])
test_instances.extend(instances_by_label[label][:1000])
test_labels.extend([label]*1000)
instances_by_label[label] = instances_by_label[label][1000:]
max_accuracy, paras = train_nn(structure, instances_by_label, test_instances, test_labels)
return max_accuracy[1]
# definitons of vector-version tanh and exp
vec_tanh = np.vectorize(math.tanh)
vec_exp = np.vectorize(math.exp)
if __name__ == "__main__":
structure = [25, 100, 10]
instances_by_label = util.read_training("train.csv")
# normalize so that tha absolute value in an input doesn't exceed 1
for label in instances_by_label:
for instance in instances_by_label[label]:
instance /= max(instance.max(), -instance.min())
np.random.seed(0)
test_instances = []
test_labels = []
for label in instances_by_label:
np.random.shuffle(instances_by_label[label])
test_instances.extend(instances_by_label[label][:1000])
test_labels.extend([label]*1000)
instances_by_label[label] = instances_by_label[label][1000:]
