def Model_Linear_Sigmoid_2_HingeLoss():
name = '2_Sigmoid_HingeLoss'
model = Network()
model.add(Linear('fc1', 784, 441, 0.01))
model.add(Sigmoid('a1'))
model.add(Linear('fc2', 441, 196, 0.01))
model.add(Sigmoid('a2'))
model.add(Linear('fc3', 196, 10, 0.01))
loss = HingeLoss(name='loss')
return name, model, loss
def test_backward(self):
layer = Sigmoid()
x = np.random.rand(2)
y = layer.forward(x)
deriv_grad = layer.backward(np.ones(1))
numerical_grad_matrix = numerical_gradient.calc(layer.forward, x)
# the numerical grad in this case is a matrix made of zeros with
# dJ/dx_i only in the diagonal
num_grad = np.diagonal(numerical_grad_matrix)
numerical_gradient.assert_are_similar(deriv_grad, num_grad)
def network_setup(model_file_path=None):
freq_count = 4000
count_bins = 88 * 20
dataset = MapsDB('../db',
freq_count=freq_count,
count_bins=count_bins,
batch_size=128,
start_time=0.5,
duration=0.5)
model = Network()
model.add(Linear('fc1', dataset.get_vec_input_width(), 2048, 0.001))
model.add(Sigmoid('sigmoid1'))
model.add(Linear('fc2', 2048, dataset.get_label_width(), 0.001))
model.add(Softmax('softmax2'))
loss = CrossEntropyLoss(name='xent')
# loss = EuclideanLoss(name='r2')
optim = SGDOptimizer(learning_rate=0.00001, weight_decay=0.005, momentum=0.9)
# optim = AdagradOptimizer(learning_rate=0.001, eps=1e-6)
input_placeholder = T.fmatrix('input')
label_placeholder = T.fmatrix('label')
label_active_size_placeholder = T.ivector('label_active_size')
if model_file_path:
model.loads(model_file_path)
else:
dataset.load_cache()
model.compile(input_placeholder, label_placeholder, label_active_size_placeholder, loss, optim)
return model, dataset, freq_count, count_bins
def two_layer_sigmoid():
model = Network()
model.add(Linear('fc1', 784, 256, 0.001))
model.add(Sigmoid('sg1'))
model.add(Linear('fc2', 256, 10, 0.001))
model.add(Sigmoid('sg2'))
config = {
'learning_rate': 0.01,
'weight_decay': 0.005,
'momentum': 0.9,
'batch_size': 100,
'max_epoch': 20,
'disp_freq': 50,
'test_epoch': 5
}
return model, config
def test_LinearSigmoid(self):
model = Seq()
model.add(Linear(2, 1, initialize='ones'))
model.add(Sigmoid())
data = np.array([2., 3.])
out = model.forward(data)
self.assertEqual(round(out, 2), 1.)
def build_model(config):
model = Network()
layer_num = 0
for layer in config['use_layer']:
if layer['type'] == "Linear":
in_num = layer['in_num']
out_num = layer['out_num']
if "init_std" in layer.keys():
model.add(
Linear(layer['type'] + str(layer_num),
in_num,
out_num,
init_std=layer['init_std']))
else:
model.add(
Linear(layer['type'] + str(layer_num), in_num, out_num))
layer_num += 1
elif layer['type'] == 'Relu':
model.add(Relu(layer['type'] + str(layer_num)))
layer_num += 1
elif layer['type'] == 'Sigmoid':
model.add(Sigmoid(layer['type'] + str(layer_num)))
layer_num += 1
else:
assert 0
loss_name = config['use_loss']
if loss_name == 'EuclideanLoss':
loss = EuclideanLoss(loss_name)
elif loss_name == 'SoftmaxCrossEntropyLoss':
loss = SoftmaxCrossEntropyLoss(loss_name)
else:
assert 0
return model, loss
def getNetwork():
'''
to obtain network structure from specified file
'''
file_name = "models/structure.json"
if len(sys.argv)>1:
file_name = sys.argv[1]
f = file(file_name, "r")
s = f.read()
f.close()
networks = json.loads(s)
for network in networks:
config = network['config']
dis_model = network['model']
model = Network()
for layer in dis_model:
if layer['type'] == 'Linear':
model.add(Linear(layer['name'], layer['in_num'], layer['out_num'], layer['std']))
if layer['type'] == 'Relu':
model.add(Relu(layer['name']))
if layer['type'] == 'Sigmoid':
model.add(Sigmoid(layer['name']))
if layer['type'] == 'Softmax':
model.add(Softmax(layer['name']))
loss = EuclideanLoss('loss')
if 'loss' in config:
if config['loss'] == 'CrossEntropyLoss':
loss = CrossEntropyLoss('loss')
yield network['name'], model, config, loss
def Model_Linear_Sigmoid_1_SoftmaxCrossEntropyLoss():
name = '1_Sigmoid_SoftmaxCrossEntropyLoss'
model = Network()
model.add(Linear('fc1', 784, 256, 0.01))
model.add(Sigmoid('a1'))
model.add(Linear('fc2', 256, 10, 0.01))
loss = SoftmaxCrossEntropyLoss(name='loss')
return name, model, loss
def Model_Linear_Sigmoid_1_HingeLoss():
name = '1_Sigmoid_HingeLoss'
model = Network()
model.add(Linear('fc1', 784, 256, 0.01))
model.add(Sigmoid('a1'))
model.add(Linear('fc2', 256, 10, 0.01))
loss = HingeLoss(name='loss')
return name, model, loss
def testNetwork(): # noqa D103
net = Network([Linear(10, 64), ReLU(), Linear(64, 2), Sigmoid()])
x = np.random.randn(32, 10)
y = np.random.randn(32, 2)
mse = MSE()
optim = SGD(0.001, 0.001)
pred = net(x)
_ = mse(pred, y)
_ = net.backward(mse.grad)
optim.step(net)
def test_Perceptron(self):
train_set, test_set = gen_data()
model = Seq([
Linear(2, 5, initialize='random'),
Sigmoid(),
Linear(5, 1, initialize='random'),
Sigmoid(),
])
OnlineTrainer().train(
model,
train_set=train_set,
loss=SquaredLoss(),
# optimizer=SGD(learning_rate=0.1),
optimizer=MomentumSGD(learning_rate=0.1, momentum=0.9),
# optimizer=AdaGrad(learning_rate=0.9),
# optimizer=RMSProp(learning_rate=0.1, decay_rate=0.9),
epochs=200,
save_progress=False)
# model.learn_minibatch(
# input_data=train_data,
# target_data=train_targets,
# loss=SquaredLoss(),
# batch_size=5,
# # optimizer=SGD(learning_rate=0.1),
# # optimizer=MomentumSGD(learning_rate=0.1, momentum=0.9),
# optimizer=AdaGrad(learning_rate=0.9),
# # optimizer=RMSProp(learning_rate=0.1, decay_rate=0.9),
#
# epochs=100,
# save_progress=True)
model.save_to_file('perceptron.pkl')
scatter_test_data(test_set, model)
# model.plot_errors_history()
# model.plot_loss_gradient_history()
plt.show()
def addSigmoidLayer(self, **kwargs):
"""
Add sigmoid classification layer.
"""
input_layer = self.input_layer if not self.all_layers \
else self.all_layers[-1]
new_layer = Sigmoid(input_layer, **kwargs)
self.all_layers += (new_layer, )
self.n_layers = len(self.all_layers)
def test_TwoLinearSigmoidLayers(self):
x = np.random.rand(5)
real_model = Seq([
Linear(5, 3, initialize='ones'),
Sigmoid(),
Linear(3, 5, initialize='ones'),
Sigmoid()
])
y = real_model.forward(x)
real_grad = real_model.backward(np.ones(5))
num_model = Seq([
Linear(5, 3, initialize='ones'),
Sigmoid(),
Linear(3, 5, initialize='ones'),
Sigmoid()
])
num_grad = numerical_gradient.calc(num_model.forward, x)
num_grad = np.sum(num_grad, axis=1)
numerical_gradient.assert_are_similar(real_grad, num_grad)
def __init__(self, input_size: int, hidden_size: int, output_size: int):
W1 = 0.01 * np.random.randn(input_size, hidden_size)
b1 = np.zeros(hidden_size)
W2 = 0.01 * np.random.randn(hidden_size, output_size)
b2 = np.zeros(output_size)
self.layers = [Affine(W1, b1), Sigmoid(), Affine(W2, b2)]
self.loss_layer = SoftmaxWithLoss()
self.params, self.grads = [], []
for layer in self.layers:
self.params += layer.params
self.grads += layer.grads
def __init__(self, input_size, hidden_size, output_size):
I, H, O = input_size, hidden_size, output_size
w1 = np.random.randn(I, H) * 0.01
b1 = np.zeros(H) #np.random.randn(H)
w2 = np.random.randn(H, O) * 0.01
b2 = np.zeros(O) #np.random.randn(O)
self.layers = [Affine(w1, b1), Sigmoid(), Affine(w2, b2)]
self.loss_layer = SoftmaxWithLoss()
self.params, self.grads = [], []
for l in self.layers:
self.params += l.params
self.grads += l.grads # 勾配まとめはこのときだけ。-> 各layerの勾配更新は参照場所を動かさないようにする。
def __init__(self, input_size, hidden_size, output_size):
I, H, O = input_size, hidden_size, output_size
W1 = 0.01 * np.random.randn(I, H)
b1 = np.zeros(H)
W2 = 0.01 * np.random.randn(H, O)
b2 = np.zeros(O)
self.layers = [Affine(W1, b1), Sigmoid(), Affine(W2, b2)]
self.loss_layer = SoftmaxWithLoss()
self.params, self.grads = [], []
for layer in self.layer:
self.params += layer.params
self.grads += layer.grads
def create_model(hidden_layer, nb_nodes):
if hidden_layer <= 0:
raise ValueError("Model needs at least 1 hidden layer")
elif hidden_layer > 10:
raise ValueError("Maximum number of hidden layers is 10")
model = Model()
input_length = 5
for i in range(hidden_layer):
model.append_layer(
Dense(node_count=nb_nodes[i], input_length=input_length))
model.append_layer(Sigmoid())
input_length = nb_nodes[i]
return model
def __init__(self, input_size, hidden_size, output_size):
I, H, O = input_size, hidden_size, output_size
# 重みとバイアスの初期化
W1 = 0.01 * np.random.randn(I, H)
b1 = np.random.randn(H)
W2 = 0.01 * np.random.randn(H, O)
b2 = np.random.randn(O)
# レイヤの生成
self.layers = [Affine(W1, b1), Sigmoid(), Affine(W2, b2)]
self.loss_layer = SoftmaxWithLoss()
# すべての重みをリストにまとめる
self.params, self.grads = [], []
for layer in self.layers:
self.params += layer.params
self.grads += layer.grads
def __init__(self, input_size, hidden_size, output_size):
I, H, O = input_size, hidden_size, output_size
# initialize weight and bias
W1 = 0.01 * cp.random.randn(I, H)
b1 = cp.zeros(H)
W2 = 0.01 * cp.random.randn(H, O)
b2 = cp.zeros(O)
# create layer
self.layers = [Affine(W1, b1), Sigmoid(), Affine(W2, b2)]
self.loss_layer = SoftmaxWithLoss()
# combine all weight and grads into list
self.params, self.grads = [], []
for layer in self.layers:
self.params += layer.params
self.grads += layer.grads
def build_model_from_string(def_str):
model = Network()
def_str.strip()
layer_strs = def_str.split(';')
for layer_str in layer_strs:
tokens = layer_str.split(',')
if (len(tokens) <= 1):
raise Exception(
"Invalid token: {} in layer definition".format(layer_str))
type = tokens[0].strip()
name = tokens[1].strip()
if (type == "linear"):
model.add(
Linear(name, int(tokens[2]), int(tokens[3]), float(tokens[4])))
elif (type == "sigmoid"):
model.add(Sigmoid(name))
elif (type == "relu"):
model.add(Relu(name))
else:
raise NotImplementedError("Unsupported layer type {}".format(type))
print("=" * 50 + "\nModel Summary:\n{}\n".format(model) + "=" * 50 + "\n")
return model
def __init__(self,
sizes,
batch_size,
epoch_num,
learning_rate,
use_trained_params=False,
filename=None):
self.num_layers = len(sizes)
self.sizes = sizes
self.batch_size = batch_size
self.epoch_num = epoch_num
self.learning_rate = learning_rate
if use_trained_params:
path = os.path.dirname(os.path.abspath(__file__))
loaded_params = np.load(os.path.join(path, filename))
self.W1 = loaded_params['W1']
self.b1 = loaded_params['b1']
self.W2 = loaded_params['W2']
self.b2 = loaded_params['b2']
else:
np.random.seed(12)
self.W1 = np.sqrt(1 / sizes[0]) * np.random.randn(
sizes[0], sizes[1]) #（784,50）
self.b1 = np.sqrt(1 / sizes[0]) * np.random.randn(sizes[1])
self.W2 = np.sqrt(1 / sizes[1]) * np.random.randn(
sizes[1], sizes[2]) #（50,10)
self.b2 = np.sqrt(1 / sizes[1]) * np.random.randn(sizes[2])
# layers of network
self.layers = {}
self.layers['FullyConnected1'] = FullyConnected(self.W1, self.b1)
self.layers['Activation'] = Sigmoid()
self.layers['FullyConnected2'] = FullyConnected(self.W2, self.b2)
self.lastLayer = SoftmaxLoss()
from network import Network
from utils import LOG_INFO
from layers import Relu, Sigmoid, Linear
from loss import EuclideanLoss
from solve_net import train_net, test_net
from load_data import load_mnist_2d
import numpy as np
train_data, test_data, train_label, test_label = load_mnist_2d('data')
# Your model defintion here
# You should explore different model architecture
model = Network()
model.add(Linear('fc1', 784, 100, 0.01))
model.add(Sigmoid('Sigmoid1'))
model.add(Linear('fc2', 100, 10, 0.01))
loss = EuclideanLoss(name='loss')
# Training configuration
# You should adjust these hyperparameters
# NOTE: one iteration means model forward-backwards one batch of samples.
# one epoch means model has gone through all the training samples.
# 'disp_freq' denotes number of iterations in one epoch to display information.
config = {
'learning_rate': 0.08,
'weight_decay': 0.001,
'momentum': 0.9,
'batch_size': 80,
'max_epoch': 100,
kernel_weights_0 = np.array([[5, -5],
[5, -5]])
bias_weights_1 = np.array([-5])
kernel_weights_1 = np.array([[5],
[5]])
saved_bias_0 = saved_weights(bias_weights_0)
saved_kernel_0 = saved_weights(kernel_weights_0)
saved_bias_1 = saved_weights(bias_weights_1)
saved_kernel_1 = saved_weights(kernel_weights_1)
model = Sequential()
model.add(Dense(2, 2, kernel_initializer=saved_kernel_0, bias_initializer=saved_bias_0, alpha=50.0))
model.add(Sigmoid())
model.add(Dense(1, 2, kernel_initializer=saved_kernel_1, bias_initializer=saved_bias_1, alpha=50.0))
model.add(Sigmoid())
X = np.array([[0, 0],
[0, 1],
[1, 0],
[1, 1]])
y = np.array([[1],
[0],
[0],
[1]])
print("Prediction")
p = model.predict(X)
print(p)
from network import Network
from utils import LOG_INFO
from layers import Relu, Sigmoid, Linear
from loss import EuclideanLoss
from solve_net import train_net, test_net
from load_data import load_mnist_2d
train_data, test_data, train_label, test_label = load_mnist_2d('data')
# Your model defintion here
# You should explore different model architecture
model = Network()
model.add(Linear('fc1', 784, 10, 0.01))
model.add(Sigmoid('fc2'))
#model.add(Linear('fc3', 10, 10, 0.01))
#model.add(Sigmoid('fc4'))
loss = EuclideanLoss(name='loss')
# Training configuration
# You should adjust these hyperparameters
# NOTE: one iteration means model forward-backwards one batch of samples.
# one epoch means model has gone through all the training samples.
# 'disp_freq' denotes number of iterations in one epoch to display information.
config = {
'learning_rate': 0.01,
'weight_decay': 0.0,
'momentum': 0.9,
'batch_size': 100,
def test_forward(self):
layer = Sigmoid()
x = np.array([2., 3., 4.])
y = layer.forward(x)
import numpy as np
### GLOBALS ###
BATCH_SIZE = 50
EPOCHS = 1000
LEARNING_RATE = 0.001
if __name__ == '__main__':
rawData, rawLabels = extract_raw()
trainingData, labels = raw_2_numpy(rawData, rawLabels)
total = trainingData.shape[0]
trainingData, labels = data_batches(trainingData, labels, BATCH_SIZE)
layer1 = Dense(trainingData.shape[2], 16)
activation1 = Sigmoid()
layer2 = Dense(16, 10)
activation2 = SoftMax()
cost = CostMeanSquared()
for epoch in range(EPOCHS):
print('Epoch: ' + str(epoch + 1) + '/' + str(EPOCHS))
print('')
correct = 0
for batch in range(total // BATCH_SIZE):
### SOCHASIC GRADIENT DESCENT ###
layer1.forward(trainingData[batch])
activation1.forward(layer1.outputs)
layer2.forward(activation1.outputs)
def testSigmoid(): # noqa D103
s = Sigmoid()
x = np.random.randn(32, 10) # batch size by in_features
output = s(x)
assert output.shape == (32, 10)
#width1 = 2048
#width2 = 256
width2 = 1024
#width2 = 2048
init_std_dev = 1 / sqrt(784)
init_std_dev_l1 = 1 / sqrt(width1)
init_std_dev_l2 = 1 / sqrt(width2)
model = Network()
if usedNN == '1 layer':
model.add(Linear('fc1', 784, width1, init_std_dev))
if usedActivation == 'sigmoid':
model.add(Sigmoid('act1'))
else:
model.add(Relu('act1'))
model.add(Linear('fc2', width1, 10, init_std_dev))
if usedNN == '2 layers':
model.add(Linear('fc1', 784, width1, init_std_dev))
if usedActivation == 'sigmoid':
model.add(Sigmoid('act1'))
else:
model.add(Relu('act1'))
model.add(Linear('fc2', width1, width2, init_std_dev))
if usedActivation == 'sigmoid':
model.add(Sigmoid('act2'))
else:
model.add(Relu('act2'))
model.add(Linear('fc3', width2, 10, init_std_dev))
# Sigmoid Layer Test
import os
import sys
from pathlib import Path
import numpy as np
try:
sys.path.append(os.path.join(Path(os.getcwd()).parent, 'lib'))
from layers import Sigmoid
except ImportError:
print('Library Module Can Not Found')
# Test1(Vector)
layer = Sigmoid()
x = np.array([0.1, -0.2, 0.3, -0.4, 0.5])
print(x)
y = layer.forward(x)
print(y)
print(layer.out)
dout = np.array([-0.1, -0.2, -0.3, 0.4, -0.5])
dout = layer.backward(dout)
print(dout)
print('=========================================')
# Test2(Matrix)
x = np.array([
from utils import LOG_INFO
from layers import Relu, Sigmoid, Linear, Leaky_Relu, ELU
from loss import EuclideanLoss
from solve_net import train_net, test_net
from load_data import load_mnist_2d
train_data, test_data, train_label, test_label = load_mnist_2d('data')
# Your model defintion here
# You should explore different model architecture
model = Network()
model.add(Linear('fc1', 784, 1000, 0.01))
#model.add(ELU('elu1', 0.5))
#model.add(Leaky_Relu('lrelu1', 0.1))
#model.add(Relu('relu1'))
model.add(Sigmoid('sigmoid1'))
model.add(Linear('fc2', 1000, 100, 0.01))
#model.add(Relu('relu2'))
model.add(Sigmoid('sigmoid2'))
#model.add(Leaky_Relu('lrelu2', 0.1))
model.add(Linear('fc3', 100, 10, 0.01))
#model.add(Sigmoid('sigmoid2'))
loss = EuclideanLoss(name='loss')
# Training configuration
# You should adjust these hyperparameters
# NOTE: one iteration means model forward-backwards one batch of samples.
# one epoch means model has gone through all the training samples.
# 'disp_freq' denotes number of iterations in one epoch to display information.
