def run():
## inputs = csv.open_card_csv('train.csv')
## #numbers hardcoded for now, and no functions specified
## out_node = Node(f)
## mid_layer1 = Layer(f_list1)
## mid_layer2 = Layer(f_list2)
## mid_layers = [mid_layer1,mid_layer2]
## network = Network(5,mid_layers,out_node)
## network = network.add_edges
## alg.train(inputs,network,10000)
## network.print_weights
## return
t, v, r = icsv.csv_to_binary('train.csv')
network = neural_net.Net([52, 30, 10])
network.SGD(t, 30, 10, 3.0, test_data=r)
def train_ann_classifier(step_num, eta, active_func_type):
#preprocess() # run at the first time to split the data
train_data = load_data("./data/iris_train.data")
labels = load_label("./data/iris_label.data")
feature_num = len(train_data[0]) - 1
hidden_node_num = feature_num + 8
label_num = len(labels)
'''
print "feature_num: %d" % feature_num
print "hidden_node_num: %d" % hidden_node_num
print "label_num: %d" % label_num
'''
# initialize simple ann
ann_net = neural_net.Net(feature_num, hidden_node_num, label_num,
active_func_type)
# train
for i in xrange(step_num):
for ele in train_data:
ann_net.train(ele[:-1], labels.index(ele[-1]), eta)
# print accuracy of every 100 iterion
if i % 100 == 0:
m_count = 0
for ele in train_data:
m_count += ann_net.test(ele[:-1], labels.index(ele[-1]))
print "accuracy of %d steps: %f" % (i, 1.0 * m_count /
len(train_data))
# test
test_data = load_data("./data/iris_train.data")
m_count = 0
for ele in test_data:
m_count += ann_net.test(ele[:-1], labels.index(ele[-1]))
print "final accuracy: %f" % (1.0 * m_count / len(test_data))
import game
import neural_net
import time
import random
net = neural_net.Net([198, 40, 1])
gammon = game.Game()
print "Starting game..."
# 2 players roll dice, the largest goes first using the roll
# r1 is white, r2 is black
r1, r2 = (0, 0)
turn = 0
# keep rolling if get same
while r1 == r2:
r1, r2 = gammon.roll_dice()
if r1 > r2:
turn = 1
elif r2 > r1:
turn = 0
start = 1
moves = 0
zero_actions = []
one_actions = []
while not gammon.game_over():
actions = []
if start:
actions = gammon.get_actions((r1, r2), turn)
start = 0
else:
roll = gammon.roll_dice()
actions = gammon.get_actions(roll, turn)
import mnist_loader import neural_net # load dataset training_data, validation_data, test_data = mnist_loader.load_data_wrapper() # create subset of trainng data training_data_subset = [] for i in range(0,10000): training_data_subset.append(training_data[i]) # neural_net.Net will only accept [784, 30, 10] -- will not work otherwise net = neural_net.Net([784, 30, 10]) # subset used to save time because it takes forever O(n^2) to train # pass in subset of training data and learning rate #net.train(training_data_subset, .3) # load saved network net.load() # solve for inputs net.imagine(3) # test the trained network #net.test(test_data) # save weights net.save()
def train_ann_classifier(step_num, eta):
raw_train_data = load_one_hot_data("./data/train.csv")
feature_num = raw_train_data.shape[1] - 1
hidden_node_num = 13
label_num = 10
batch_size = 50
print "feature_num: %d" % feature_num
print "hidden_node_num: %d" % hidden_node_num
print "label_num: %d" % label_num
print "batch_size: %d" % batch_size
# load pre-trained weight and bias
input_hidden_w = load_pre_train_weight("./model/input_hidden_w.bin",
(feature_num, hidden_node_num))
input_hidden_b = load_pre_train_weight("./model/input_hidden_b.bin",
(1, hidden_node_num))
hidden_output_w = load_pre_train_weight("./model/hidden_output_w.bin",
(hidden_node_num, label_num))
hidden_output_b = load_pre_train_weight("./model/hidden_output_b.bin",
(1, label_num))
# initialize simple ann
ann_net = neural_net.Net(batch_size, feature_num, hidden_node_num,
label_num, input_hidden_w, input_hidden_b,
hidden_output_w, hidden_output_b)
# train
np.random.shuffle(raw_train_data)
train_len = int(raw_train_data.shape[0] * 0.8)
train_data, valid_data = raw_train_data[:train_len], raw_train_data[
train_len:]
valid_features, valid_labels = valid_data[:,
1:] / 255.0, labels_to_one_hot(
valid_data[:, 0])
for i in xrange(step_num):
features_data, labels_data = gen_batch_data(train_data, batch_size)
#start_t = time.time()
ann_net.train(features_data, labels_data, eta)
#print "%d batch cost time: %d ms" % (i, int((time.time() - start_t)*1000))
# print accuracy of every one epoch
if (i + 1) == 1 or (i + 1) % 840 == 0 or (i + 1) == step_num:
print "accuracy of %d batch (%d epoch): %f" % (
(i + 1), (i + 1) / 840,
ann_net.get_accuracy(valid_features, valid_labels))
ann_net.input_hidden_w.tofile("./model/input_hidden_w.bin")
ann_net.input_hidden_b.tofile("./model/input_hidden_b.bin")
ann_net.hidden_output_w.tofile("./model/hidden_output_w.bin")
ann_net.hidden_output_b.tofile("./model/hidden_output_b.bin")
# test
test_data = load_one_hot_data("./data/test.csv")
predict_label = ann_net.predict(test_data / 255.0)
predict_label = [np.arange(1, 1 + len(predict_label)), predict_label]
predict_label = np.transpose(predict_label)
np.savetxt('submission.csv',
predict_label,
fmt='%i,%i',
header='ImageId,Label',
comments='')
import neural_net
import mnist_loader
import time
import os.path
if os.path.isfile("../data/weights_file1"):
network = neural_net.Net([784, 20, 30, 10])
network.load()
print("Network loaded.")
else:
training_data, validation_data, test_data = mnist_loader.load_data_wrapper()
#network = neural_net.Net([784, 30, 10])
network = neural_net.Net([784, 20, 30, 10])
network.load()
print("Training network...")
# Training takes O(n^2) time now (a really long time for the whole data set) because I wanted to be very explicit about how errors and weights were being calculated.
start = time.time()
network.train(training_data, 3.0)
end = time.time()
print("Training time:")
print(end - start)
print("Test network...")
start = time.time()
network.test(test_data)
end = time.time()
print("Test time:")
print(end - start)
network.save()
print("Network successfully saved.")
import numpy as np import neural_net as nn import matplotlib.pyplot as plt # Build fake data for y = 1 if (x > 5) else 0 # X in range [0,10] X_train = np.random.random((100, 64, 1, 1)) * 10 X_test = np.random.random((1000, 1, 1)) * 10 # Calculate y values for each X y_train = (X_train > 5).astype(int) y_test = (X_test > 5).astype(int) # Build and train the network L = [1, 16, 16, 1] A = [nn.sigmoid, nn.sigmoid, nn.sigmoid] net = nn.Net(L, A) net.train(X_train, y_train, X_test, y_test, 1000, nn.squared_error) _, _, y_pred = net.forward(X_test) plt.scatter(X_test.reshape(-1), y_pred.reshape(-1)) plt.show()
