def main():
n = np.random.randint(2, 20)
# Generate an undirected graph
graph = np.random.randint(2, size=[n, n])
graph = np.tril(graph, -1) + np.tril(graph, -1).T
label = np.random.randint(2)
print(graph)
print(graph.shape)
print('label', label)
dim_feature = 10
x = np.zeros([n, dim_feature])
x[:, 0] = 1
W = np.random.normal(0, 0.4, [dim_feature, dim_feature])
A = np.random.normal(0, 0.4, dim_feature)
b = np.array([0.])
model = GraphNeuralNetwork(W, A, b, 2)
optimizer = SGD(model, lr=0.001)
for i in range(500):
grads_flat = calc_grads(model, graph, x, label, lossfunc=bce_with_logit, eps=1e-4)
outputs = model(graph, x)
train_loss = bce_with_logit(outputs, label)
optimizer.update(grads_flat)
print('step: %d, train_loss: %.15f' % (i, train_loss))
def __train(weight_init_std):
#初始化网络结构
bn_network=multi_net_extend(input_size=784, hidden_size_list=[100, 100, 100, 100, 100], output_size=10,
weight_init_std=weight_init_std, use_batchnorm=True)
network=multi_net_extend(input_size=784, hidden_size_list=[100, 100, 100, 100, 100], output_size=10,
weight_init_std=weight_init_std)
train_acc_list=[]
bn_train_acc_list=[]
optimizer=SGD(lr=0.01)
per_epoch=50
epoch_cnt = 0
for i in range(1000000000):
##minibatch
batch_mask=np.random.choice(train_size,batch_size)
x_batch=x_train[batch_mask]
t_batch=t_train[batch_mask]
##梯度下降
for _network in (bn_network, network):
grads = _network.gradient(x_batch, t_batch)
optimizer.update(_network.params, grads)
if i %10 == 0:
train_acc = network.accuracy(x_train, t_train)
bn_train_acc = bn_network.accuracy(x_train, t_train)
train_acc_list.append(train_acc)
bn_train_acc_list.append(bn_train_acc)
print("epoch:" + str(epoch_cnt) + " | " + str(train_acc) + " - " + str(bn_train_acc))
epoch_cnt += 1
if epoch_cnt >= max_epochs:
break
return train_acc_list, bn_train_acc_list
for key, weight in weight_type.items():
networks[key] = multi_layer_net(input_size=784,
hidden_layer_list=[100, 100, 100, 100],
output_size=10,
weight_init_std=weight)
train_loss[key] = []
for i in range(iter_num):
##取batch数据
batch_mask = np.random.choice(train_size, batch_size)
x_batch = x_train[batch_mask]
t_batch = t_train[batch_mask]
#计算梯度
for key in weight_type.keys():
grads = networks[key].gradient(x_batch, t_batch)
optimizer.update(networks[key].params, grads)
#计算损失
batch_loss = networks[key].loss(x_batch, t_batch)
train_loss[key].append(batch_loss)
if i % 100 == 0:
print('+++++++++++')
for key in weight_type.keys():
loss = networks[key].loss(x_batch, t_batch)
print(key + ":" + str(loss))
#绘图
makers = {'std=0.01': 'o', 'Xavier': 's', 'He': 'D'}
x = np.arange(iter_num)
for key in weight_type.keys():