class Application(tornado.web.Application):
def __init__(self):
self.net = ConvNet()
self.net.setup_layers([
InputLayer(InputLayerSettings(in_shape=(28, 28, 1))),
ConvolutionalLayer(ConvolutionalLayerSettings(filters_count=16, filter_size=5, stride=1, zero_padding=0)),
ReluLayer(ReluLayerSettings(activation='max')),
PoolingLayer(PoolingLayerSettings(filter_size=2, stride=2)),
ConvolutionalLayer(ConvolutionalLayerSettings(filters_count=8, filter_size=5, stride=1, zero_padding=0)),
ReluLayer(ReluLayerSettings(activation='max')),
PoolingLayer(PoolingLayerSettings(filter_size=3, stride=3)),
FullConnectedLayer(FullConnectedLayerSettings(neurons_count=10, activation='sigmoid')),
])
s = {
'debug': convsettings.DEBUG,
'template_path': convsettings.TEMPLATES_DIR,
'static_path': convsettings.STATIC_DIR,
'static_url_prefix': convsettings.STATIC_PREFIX
}
handlers = [
(r"/", convnet_web.handlers.VisualHandler),
(r"/config", convnet_web.handlers.ConfigHandler),
(r"/api/config", convnet_web.handlers.ApiConfig, dict(net=self.net)),
(r"/api/config_mnist", convnet_web.handlers.ApiConfigMnist, dict(net=self.net)),
(r"/api/predict", convnet_web.handlers.ApiPredict, dict(net=self.net)),
]
super(Application, self).__init__(handlers, **s)
def _test_basic(self):
X = [np.zeros((3, 3, 2)) for _ in xrange(2)]
y = [np.zeros((1, 1, 2)) for _ in xrange(2)]
net = ConvNet()
net.setup_layers([
InputLayer(InputLayerSettings(in_shape=X[0].shape)),
ConvolutionalLayer(ConvolutionalLayerSettings(filter_size=2, filters_count=2, stride=1)),
PoolingLayer(PoolingLayerSettings(filter_size=2, stride=1)),
ReluLayer(ReluLayerSettings(activation='max')),
FullConnectedLayer(FullConnectedLayerSettings(neurons_count=y[0].shape[-1])),
ReluLayer(ReluLayerSettings(activation='sigmoid')),
])
net.fit(X, y)
def test_backward(self):
arr1 = np.empty((3, 3, 1))
arr1[:, :, 0] = np.array([
[1, 0, 0],
[0, 1, 0],
[0, 0, 1]])
arr2 = np.empty((3, 3, 1))
arr2[:, :, 0] = np.array([
[0, 0, 1],
[0, 1, 0],
[1, 0, 0]])
samples = [arr1, arr2]
expected_res1 = np.empty((1, 1, 2))
expected_res1[0, 0, 0] = 1
expected_res1[0, 0, 1] = -1
expected_res2 = np.empty((1, 1, 2))
expected_res2[0, 0, 0] = -1
expected_res2[0, 0, 1] = 1
expected_res = [expected_res1, expected_res2]
net = ConvNet()
net.setup_layers([
InputLayer(InputLayerSettings(in_shape=arr1.shape)),
ConvolutionalLayer(ConvolutionalLayerSettings(filter_size=3,
stride=1,
filters_count=2,
zero_padding=0))
])
# before learn
res_before = []
for i in xrange(len(samples)):
res = net.predict(samples[i])
res_before.append(res)
# calc distance before
dist_before = get_dist(res_before, expected_res)
# learning
net.fit(samples, expected_res)
# after learn
res_after = []
for i in xrange(len(samples)):
res = net.predict(samples[i])
res_after.append(res)
# calc distance after
dist_after = get_dist(res_after, expected_res)
print 'Before: %s' % dist_before
print 'After: %s' % dist_after
self.assertLess(dist_after, dist_before)
def _test_crossings(self):
X = to_3d([
np.asarray([
[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
]),
np.asarray([
[0, 0, 0],
[0, 1, 0],
[0, 0, 1],
]),
np.asarray([
[0, 0, 1],
[0, 1, 0],
[1, 0, 0],
]),
np.asarray([
[0, 0, 1],
[0, 1, 0],
[0, 0, 0],
]),
])
Y = [
np.asarray([1, 0]),
np.asarray([1, 0]),
np.asarray([0, 1]),
np.asarray([0, 1]),
]
net = ConvNet(iterations_count=1000, learning_rate=0.01)
net.setup_layers([
InputLayer(InputLayerSettings(in_shape=X[0].shape)),
ConvolutionalLayer(ConvolutionalLayerSettings(filter_size=3, filters_count=2, stride=1)),
ReluLayer(ReluLayerSettings(activation='max')),
# PoolingLayer(PoolingLayerSettings(filter_size=1, stride=1)),
# FullConnectedLayer(FullConnectedLayerSettings(neurons_count=Y[0].shape[-1])),
# ReluLayer(ReluLayerSettings(activation='sigmoid')),
])
net.fit(X, y_to_3d(Y))
# net = ConvNet.load_net('/home/igor/Desktop/net.pkl')
for x, y in zip(X, Y):
h = net.predict(x)
print("predicted = {}; \nreal = {}\n\n".format(h, y))
pass
def test_mnist(self):
script_path = os.path.dirname(os.path.abspath(__file__))
f = gzip.open(os.path.join(script_path, '../mnist/mnist.pkl.gz'), 'rb')
train_set, valid_set, test_set = cPickle.load(f)
f.close()
X_train, Y_train = train_set
X_train, Y_train = self.get_examples(X_train, Y_train, labels=np.arange(0, 4), count=10)
X_train, Y_train = self.shuffle_in_unison_inplace(X_train, Y_train)
X_train = self.transform_X(X_train)
Y_train = self.transform_Y(Y_train)
net = ConvNet(iterations_count=10, batch_size=1, learning_rate=0.001, momentum=0.8, weight_decay=0.001)
net.setup_layers([
InputLayer(InputLayerSettings(in_shape=X_train[0].shape)),
ConvolutionalLayer(ConvolutionalLayerSettings(filters_count=8, filter_size=5, stride=1, zero_padding=0)),
ReluLayer(ReluLayerSettings(activation='max')),
PoolingLayer(PoolingLayerSettings(filter_size=2, stride=2)),
ConvolutionalLayer(ConvolutionalLayerSettings(filters_count=16, filter_size=5, stride=1, zero_padding=0)),
ReluLayer(ReluLayerSettings(activation='max')),
PoolingLayer(PoolingLayerSettings(filter_size=3, stride=3)),
FullConnectedLayer(FullConnectedLayerSettings(neurons_count=Y_train[0].shape[-1], activation='sigmoid')),
# ReluLayer(ReluLayerSettings(activation='sigmoid')),
])
examples_count = 100000
net.fit(X_train[:examples_count], Y_train[:examples_count])
matched = 0
for x, y in zip(X_train[:examples_count], Y_train[:examples_count]):
h = net.predict(x)
h_res = h.argmax()
y_res = y.argmax()
print("predicted = {}; max = {}".format(h, h.argmax()))
print("real = {}; max = {}".format(y, y.argmax()))
print("\n")
matched += int(h_res == y_res)
print("Accuracy {}/{}".format(matched, len(X_train[:examples_count])))
def mnist():
script_path = os.path.dirname(os.path.abspath(__file__))
f = gzip.open(os.path.join(script_path, './mnist.pkl.gz'), 'rb')
train_set, valid_set, test_set = cPickle.load(f)
f.close()
X_train, Y_train = train_set
#X_train, Y_train = get_examples(X_train, Y_train, labels=np.arange(0, 10), count=200)
#X_train, Y_train = shuffle_in_unison_inplace(X_train, Y_train)
X_train = transform_X(X_train)
Y_train = transform_Y(Y_train)
print("Train set size: {}".format(len(X_train)))
X_cv, Y_cv = valid_set
#X_cv, Y_cv = get_examples(X_cv, Y_cv, labels=np.arange(0, 10), count=15)
#X_cv, Y_cv = shuffle_in_unison_inplace(X_cv, Y_cv)
X_cv = transform_X(X_cv)
Y_cv = transform_Y(Y_cv)
print("Cross validation set size: {}".format(len(X_cv)))
X_test, Y_test = test_set
#X_test, Y_test = get_examples(X_test, Y_test, labels=np.arange(0, 10), count=100)
#X_test, Y_test = shuffle_in_unison_inplace(X_test, Y_test)
X_test = transform_X(X_test)
Y_test = transform_Y(Y_test)
print("Test set size: {}".format(len(X_test)))
net = ConvNet(iterations_count=1000, batch_size=200, learning_rate=0.001, momentum=0.9, weight_decay=0.001)
net.setup_layers([
InputLayer(InputLayerSettings(in_shape=X_train[0].shape)),
ConvolutionalLayer(ConvolutionalLayerSettings(filters_count=8, filter_size=5, stride=1, zero_padding=0)),
ReluLayer(ReluLayerSettings(activation='max')),
PoolingLayer(PoolingLayerSettings(filter_size=2, stride=2)),
ConvolutionalLayer(ConvolutionalLayerSettings(filters_count=16, filter_size=5, stride=1, zero_padding=0)),
ReluLayer(ReluLayerSettings(activation='max')),
PoolingLayer(PoolingLayerSettings(filter_size=3, stride=3)),
FullConnectedLayer(FullConnectedLayerSettings(neurons_count=Y_train[0].shape[-1], activation='sigmoid')),
])
# net = ConvNet.load_net(os.path.join(script_path, './convnet1.pkl'))
try:
net.fit(X_train, Y_train)
except KeyboardInterrupt:
print("Training stopped")
train_matched = 0
for x, y in zip(X_train, Y_train):
h = net.predict(x)
h_res = h.argmax()
y_res = y.argmax()
# print("predicted = {}; max = {}".format(h, h.argmax()))
# print("real = {}; max = {}".format(y, y.argmax()))
# print("\n")
train_matched += int(h_res == y_res)
test_matched = 0
for x, y in zip(X_test, Y_test):
h = net.predict(x)
h_res = h.argmax()
y_res = y.argmax()
# print("predicted = {}; max = {}".format(h, h.argmax()))
# print("real = {}; max = {}".format(y, y.argmax()))
# print("\n")
test_matched += int(h_res == y_res)
print("Accuracy train {}/{}".format(train_matched, len(X_train)))
print("Accuracy test {}/{}".format(test_matched, len(X_test)))
path = os.path.join(script_path, "./convnet1.pkl")
net.dump_net(path)
print("Dumped to {}".format(path))
