def main(): (x_train, t_train), (x_test, t_test) = load_mnist(normalize=True, one_hot_label=True) x_train = x_train[:300] t_train = t_train[:300] max_epochs = 201 train_size = x_train.shape[0] batch_size = 100 learning_rate = 0.01 use_dropout = True dropout_ratio = 0.2 train_acc_list = [] test_acc_list = [] network = MultiLayerNetExtend( input_size=784, hidden_size_list=[100, 100, 100, 100, 100, 100], output_size=10, use_dropout=use_dropout, dropout_ration=dropout_ratio) optimizer = SGD(lr=learning_rate) iter_per_epoch = max(train_size / batch_size, 1) epoch_cnt = 0 for i in range(1000000000): batch_mask = np.random.choice(train_size, batch_size) x_batch = x_train[batch_mask] t_batch = t_train[batch_mask] grads = network.gradient(x_batch, t_batch) optimizer.update(network.params, grads) if i % iter_per_epoch == 0: train_acc = network.accuracy(x_train, t_train) test_acc = network.accuracy(x_test, t_test) train_acc_list.append(train_acc) test_acc_list.append(test_acc) print("epoch:" + str(epoch_cnt) + ", train acc:" + str(train_acc) + \ ", test acc:" + str(test_acc)) epoch_cnt += 1 if epoch_cnt >= max_epochs: break markers = {'train': 'o', 'test': 's'} x = np.arange(max_epochs) plt.plot(x, train_acc_list, marker='o', label='train', markevery=10) plt.plot(x, test_acc_list, marker='s', label='test', markevery=10) plt.xlabel("epochs") plt.ylabel("accuracy") plt.ylim(0, 1.0) plt.legend(loc='lower right') plt.show()
def main(): # 读入数据 (train_x, train_label), _ = load_mnist(one_hot_label=True) # 构造神经网络 network = MultiLayerNetExtend(input_size=784, hidden_size_list=[100, 100], output_size=10, use_batchnorm=True) # 仅用一个训练样本来测试 batch_x = train_x[:1] batch_label = train_label[:1] # 用反向传播和数值方法分别计算梯度 grad_backprop = network.gradient(batch_x, batch_label) grad_numerical = network.numerical_gradient(batch_x, batch_label) # 比较两种方法的计算结果 for key in grad_numerical.keys(): diff = np.average(np.abs(grad_backprop[key] - grad_numerical[key])) print(key + ":" + str(diff))
# coding: utf-8 import sys, os sys.path.append(os.pardir) # 为了导入父目录的文件而进行的设定 import numpy as np from dataset.mnist import load_mnist from common.multi_layer_net_extend import MultiLayerNetExtend # 读入数据 (x_train, t_train), (x_test, t_test) = load_mnist(normalize=True, one_hot_label=True) network = MultiLayerNetExtend(input_size=784, hidden_size_list=[100, 100], output_size=10, use_batchnorm=True) x_batch = x_train[:1] t_batch = t_train[:1] grad_backprop = network.gradient(x_batch, t_batch) grad_numerical = network.numerical_gradient(x_batch, t_batch) for key in grad_numerical.keys(): diff = np.average( np.abs(grad_backprop[key] - grad_numerical[key]) ) print(key + ":" + str(diff))
bn_train_accuracies = [] # 배치 정규화를 사용하는 신경망의 정확도를 기록 # optimizer = Sgd(learning_rate) # 파라미터 최적화 알고리즘이 SGD가 아닌 경우에는 신경망 개수만큼 optimizer를 생성. optimizer = Sgd(learning_rate) bn_optimizer = Sgd(learning_rate) # 학습하면서 정확도의 변화를 기록 for i in range(iterations): # 미니 배치를 랜덤하게 선택(0~999 숫자들 중 128개를 랜덤하게 선택) mask = np.random.choice(train_size, batch_size) x_batch = X_train[mask] y_batch = Y_train[mask] # 배치 정규화를 사용하지 않는 신경망에서 gradient를 계산. gradients = neural_net.gradient(x_batch, y_batch) # 파라미터 업데이트(갱신) - W(가중치), b(편향)을 업데이트 optimizer.update(neural_net.params, gradients) # 업데이트된 파라미터들을 사용해서 배치 데이터의 정확도 계산 acc = neural_net.accuracy(x_batch, y_batch) # 정확도를 기록 train_accuracies.append(acc) # 배치 정규화를 사용하는 신경망에서 같은 작업을 수행. bn_gradients = bn_neural_net.gradient(x_batch, y_batch) # gradient 계산 bn_optimizer.update(bn_neural_net.params, bn_gradients) # W, b 업데이트 bn_acc = bn_neural_net.accuracy(x_batch, y_batch) # 정확도 계산 bn_train_accuracies.append(bn_acc) # 정확도 기록 print(f'iteration #{i}: without={acc}, with={bn_acc}')
epochs = 200 # 1 에포크: 모든 학습 데이터가 1번씩 학습된 경우 mini_batch_size = 100 # 1번 forward에 보낼 데이터 샘플 개수 train_size = X_train.shape[0] iter_per_epoch = int(max(train_size / mini_batch_size, 1)) # 학습하면서 학습/테스트 데이터의 정확도를 각 에포크마다 기록 train_accuracies = [] test_accuracies = [] optimizer = Sgd(learning_rate=0.01) # optimizer for epoch in range(epochs): for i in range(iter_per_epoch): x_batch = X_train[(i * mini_batch_size):((i + 1) * mini_batch_size)] y_batch = Y_train[(i * mini_batch_size):((i + 1) * mini_batch_size)] gradients = neural_net.gradient(x_batch, y_batch) optimizer.update(neural_net.params, gradients) train_acc = neural_net.accuracy(X_train, Y_train) train_accuracies.append(train_acc) test_acc = neural_net.accuracy(X_test, Y_test) test_accuracies.append(test_acc) print(f'epoch #{epoch}: train={train_acc}, test={test_acc}') x = np.arange(epochs) plt.plot(x, train_accuracies, label='Train') plt.plot(x, test_accuracies, label='Test') plt.legend() plt.title(f'Weight Decay (lambda={wd_rate})') plt.xlabel('epoch') plt.ylabel('accuracy')
# coding: utf-8 import sys, os sys.path.append(os.pardir) # 親ディレクトリのファイルをインポートするための設定 import numpy as np from dataset.mnist import load_mnist from common.multi_layer_net_extend import MultiLayerNetExtend # データの読み込み (x_train, t_train), (x_test, t_test) = load_mnist(normalize=True, one_hot_label=True) network = MultiLayerNetExtend(input_size=784, hidden_size_list=[100, 100], output_size=10, use_batchnorm=True) x_batch = x_train[:1] t_batch = t_train[:1] grad_backprop = network.gradient(x_batch, t_batch) grad_numerical = network.numerical_gradient(x_batch, t_batch) for key in grad_numerical.keys(): diff = np.average( np.abs(grad_backprop[key] - grad_numerical[key]) ) print(key + ":" + str(diff))
class Agent: """ 各エージェントの動作を main からこっちに書き写す形で. """ n = int() AdjG_init = np.zeros((n, n)) #require to be {0 or 1} to all arguments WeiG_init = np.zeros((n, n)) maxdeg = int() train_size = 0 batch_size = 100 # weight graph 作成規則 ===== # wtype = "maximum-degree" wtype = "local-degree" # =========================== def __init__(self, idx, x_train, t_train, x_test, t_test, optimizer, weight_decay_lambda=0.0): self.idx = idx # self.layer = MultiLayerNet(input_size=784, hidden_size_list=[100], output_size=10, # weight_decay_lambda=weight_decay_lambda) self.layer = MultiLayerNetExtend( input_size=784, hidden_size_list=[50, 50, 50], output_size=10, weight_decay_lambda=weight_decay_lambda, use_dropout=False, dropout_ration=0.0, use_batchnorm=False) self.optimizer = optimizer self.rec_param = np.array([{} for i in range(self.n)]) self.x_train = x_train self.t_train = t_train self.x_test = x_test self.t_test = t_test self.z_vec = np.zeros(self.n) self.z_vec[idx] = 1 self.rec_z = np.zeros((self.n, self.n)) self.AdjG = np.zeros(self.n) #require to be {0 or 1} to all arguments self.WeiG = np.zeros(self.n) if np.all(self.WeiG_init == 0): self.makeWeiGraph(self.AdjG_init) else: self.WeiG = self.WeiG_init[self.idx] self.makeAdjGraph() self.train_loss = 0 self.train_acc = 0 self.test_acc = 0 def send(self, k, agent): """sending params to other nodes (return "self.layer.params"): send(agent)""" return (self.layer.params.copy(), self.z_vec.copy()) def receive(self, agent, getparams, getz): """receiving other node's params: receive(agent, new_params)""" self.rec_param[agent] = getparams.copy() self.rec_z[agent] = getz.copy() def selectData(self, train_size, batch_size): batch_mask = np.random.choice(train_size, batch_size) x_batch = self.x_train[batch_mask] t_batch = self.t_train[batch_mask] return x_batch, t_batch def consensus(self): self.weightConsensus() self.subvalConsensus() def weightConsensus(self): for key in self.layer.params.keys(): self.layer.params[key] *= self.WeiG[self.idx] for idn in np.nonzero(self.AdjG)[0]: self.layer.params[ key] += self.WeiG[idn] * self.rec_param[idn][key] def subvalConsensus(self): self.rec_z[self.idx] = self.z_vec self.z_vec = np.dot(self.WeiG, self.rec_z) def update(self, k=1): x_batch, t_batch = self.selectData(self.train_size, self.batch_size) grads = self.layer.gradient(x_batch, t_batch) self.optimizer.update(self.layer.params, grads, self.z_vec[self.idx], k) # self.optimizer.update(self.layer.params, grads) def calcLoss(self): self.train_acc = self.layer.accuracy(self.x_train, self.t_train) self.test_acc = self.layer.accuracy(self.x_test, self.t_test) self.train_loss = self.layer.loss(self.x_train, self.t_train) def makeAdjGraph(self): """make Adjecency Graph""" self.AdjG = self.AdjG_init[self.idx] def makeWeiGraph(self, lAdjG): """make Weight matrix""" if self.n is 1: tmpWeiG = np.ones([1]) else: if self.wtype == "maximum-degree": tmpWeiG = (1 / (self.maxdeg + 1)) * lAdjG[self.idx] tmpWeiG[self.idx] = 1 - np.sum(tmpWeiG) elif self.wtype == "local-degree": ### count degrees ### #degMat = np.kron(np.dot(lAdjG,np.ones([self.n,1])), np.ones([1,self.n])) degMat = np.kron( np.dot(lAdjG, np.ones([self.n, 1])) + 1, np.ones([1, self.n])) ### take max() for each elements ### degMat = np.maximum(degMat, degMat.T) ### divide for each elememts ### tmpAllWeiG = lAdjG / degMat selfDegMat = np.eye(self.n) - np.diag( np.sum(tmpAllWeiG, axis=1)) tmpAllWeiG = tmpAllWeiG + selfDegMat tmpWeiG = tmpAllWeiG[self.idx, :] else: try: raise ValueError("Error: invalid weight-type") except ValueError as e: print(e) self.WeiG = tmpWeiG ######## # debugging functions ######## def degub_numericalGrad(self): return self.layer.numerical_gradient(self.x_train[:3], self.t_train[:3]) def debug_backpropGrad(self): return self.layer.gradient(self.x_train[:3], self.t_train[:3]) def debug_consensus(self): params = self.layer.params.copy() self.weightConsensus() self.subvalConsensus() if self.idx == 0: ano_params = self.layer.params.copy() for key in params.keys(): diff = np.average(np.abs(params[key] - ano_params[key])) print(key + ":" + str(diff))
train_loss_list = [] test_loss_list = [] train_acc_list = [] test_acc_list = [] iter_per_epoch = max(train_size / batch_size, 1) epoch_cnt = 0 for i in range(10000): batch_mask = np.random.choice(train_size, batch_size) x_batch = x_train[batch_mask] t_batch = t_train[batch_mask] # 学習フェーズ grads = network.gradient(x_train, t_train) optimizer.update(network.params, grads) if i % iter_per_epoch == 0: #loss_train = network.loss(x_train, t_train) #loss_test = network.loss(x_test, t_test) #train_loss_list.append(loss_train) #test_loss_list.append(loss_test) train_acc = network.accuracy(x_train, t_train) test_acc = network.accuracy(x_test, t_test) train_acc_list.append(train_acc) test_acc_list.append(test_acc) epoch_cnt += 1 print("---{}/{}---".format(epoch_cnt, max_epochs))
#平均計算をするために使用されるfloat型 count = 0.0 train_sum = 0.0 test_sum = 0.0 #処理前の時刻 t1 = time.time() for i in range(iters_num): #イテレーションの回数分for文を回している batch_mask = np.random.choice(train_size, batch_size) #batch_size分のbatch_size行列の乱数を作成 x_batch = x_train[batch_mask] #x_subsetの作成 t_batch = t_train[batch_mask] #t_subsetの作成 # 勾配 #grad = network.numerical_gradient(x_batch, t_batch) grad = network.gradient(x_batch, t_batch) optimizers.update(network.params, grad) # 更新 #for key in ('W1', 'b1', 'W2', 'b2'): #network.params[key] -= learning_rate * grad[key] loss = network.loss(x_batch, t_batch) #損失を計算 train_loss_list.append(loss) #損失リストに格納する if i % iter_per_epoch == 0: #イテレーションの回数分回している中で、iがiter_per_epochの倍数の時に以下を実行 #それぞれの認識精度を計算 train_acc = network.accuracy(x_train, t_train) test_acc = network.accuracy(x_test, t_test) #認識精度のリストに格納する
class Agent: """アルゴリズム構築に必要なAgentの機能 Function: send : 隣接するagentの以下の状態変数を送る Args: layer.param (np.array) : 各層のパラメータ receive : 隣接するagentから状態変数を受け取る Args: layer.param (np.arrar) : 各層のパラメータ optimizer(SGD) : 確率勾配をバックプロパゲーションとランダムシードを用いて実装 For example: self.optimizer = optimizer(SGD(lr)) x_batch, t_batch = self.selectData(self.train_size, self.batch_size) grads = self.layer.gradient(x_batch, t_batch) self.optimizer.update(self.layer.params, grads, k) """ """ 各エージェントの動作を main からこっちに書き写す形で. """ n = int() AdjG_init = np.zeros((n, n)) #require to be {0 or 1} to all arguments WeiG_init = np.zeros((n, n)) maxdeg = int() train_size = 0 batch_size = 100 # weight graph 作成規則 ===== # wtype = "maximum-degree" wtype = "local-degree" # =========================== def __init__(self, idx, x_train, t_train, x_test, t_test, optimizer, weight_decay_lambda=0.0): """各Agentの初期状態変数 Args: idx : Agentのインデックス layer : Agent内のニューラルネットワークの層 optimizer : 最適化を行うアルゴリズムの選択 rec_param : 隣接するエージェントから受け取るパラメータ z_vec : 左の固有ベクトル rec_z : 隣接するエージェントから受け取る左固有ベクトル AdjG : 隣接行列?? WeiG : 重み行列?? """ self.idx = idx # self.layer = MultiLayerNet(input_size=784, hidden_size_list=[100], output_size=10, # weight_decay_lambda=weight_decay_lambda) self.layer = MultiLayerNetExtend( input_size=784, hidden_size_list=[ 500, 400, 300, 300, 200, 200, 100, 100, 100, 100, 100, 50, 50 ], output_size=10, weight_decay_lambda=weight_decay_lambda, use_dropout=True, dropout_ration=0.3, use_batchnorm=True) #dropout_ratio=0.03, use_batchnorm=True, hidden_size_list=[500,400,300,300,200,200,100,100,100,50,50,50] weightdecay=0.01 → 0.9428 #hidden_size_list=[500,400,300,300,200,200,100,100,100,50,50,50], output_size=10,weight_decay_lambda=weight_decay_lambda,use_dropout=True, dropout_ration=0.05, use_batchnorm=True 一番いい #hidden_size_list=[100,100,100,100,100] weightdecay=0.3, dropout_ration=0.3 self.optimizer = optimizer self.rec_param = np.array([{} for i in range(self.n)]) self.send_param = np.array([{} for i in range(self.n)]) #Initialize self.rec_param[self.idx] = self.layer.params.copy() self.send_param[self.idx] = self.layer.params.copy() self.x_train = x_train self.t_train = t_train self.x_test = x_test self.t_test = t_test self.AdjG = np.zeros(self.n) #require to be {0 or 1} to all arguments self.WeiG = np.zeros(self.n) if np.all(self.WeiG_init == 0): self.makeWeiGraph(self.AdjG_init) else: self.WeiG = self.WeiG_init[self.idx] self.makeAdjGraph() self.train_loss = 0 self.train_acc = 0 self.test_acc = 0 #山下さんのアルゴリズム構築に必要な関数 # def send(self, k, agent): # """sending params to other nodes (return "self.layer.params"): send(agent)""" # return (self.layer.params.copy(), self.z_vec.copy()) # def receive(self, agent, getparams, getz): # """receiving other node's params: receive(agent, new_params)""" # self.rec_param[agent] = getparams.copy() # self.rec_z[agent] = getz.copy() def send(self, k, agent): """sending params to other nodes (return "self.layer.params"): send(agent)""" self.send_param[self.idx] = self.layer.params.copy() return self.layer.params.copy() def receive(self, agent, getparams): """receiving other node's params: receive(agent, new_params)""" for key in getparams.keys(): self.rec_param[agent][key] = getparams[key].copy() def selectData(self, train_size, batch_size): batch_mask = np.random.choice(train_size, batch_size) x_batch = self.x_train[batch_mask] t_batch = self.t_train[batch_mask] return x_batch, t_batch def consensus(self): self.weightConsensus() # self.subvalConsensus() def weightConsensus(self): for key in self.layer.params.keys(): self.layer.params[key] *= self.WeiG[self.idx] for idn in np.nonzero(self.AdjG)[0]: self.layer.params[ key] += self.WeiG[idn] * self.rec_param[idn][key] # def subvalConsensus(self): # self.rec_z[self.idx] = self.z_vec # self.z_vec = np.dot(self.WeiG, self.rec_z) def update(self, k=1): x_batch, t_batch = self.selectData(self.train_size, self.batch_size) grads = self.layer.gradient(x_batch, t_batch) # self.optimizer.update(self.layer.params, grads, self.z_vec[self.idx], k) self.send_param[self.idx], self.rec_param[ self.idx] = self.optimizer.update(self.layer.params, grads, self.rec_param[self.idx], self.send_param[self.idx], self.WeiG[self.idx], k) def calcLoss(self): self.train_acc = self.layer.accuracy(self.x_train, self.t_train) self.test_acc = self.layer.accuracy(self.x_test, self.t_test) self.train_loss = self.layer.loss(self.x_train, self.t_train) def makeAdjGraph(self): """make Adjecency Graph""" self.AdjG = self.AdjG_init[self.idx] def makeWeiGraph(self, lAdjG): """make Weight matrix 2020/01/28 山下さんは有効グラフを作成している.無向グラフに変更("maximum-degree"の方のみ) Args: tmpWeiG (np.array) : 一次的な重み行列 """ if self.n is 1: tmpWeiG = np.ones([1]) else: if self.wtype == "maximum-degree": tmpWeiG = (1 / (self.maxdeg + 1)) * lAdjG[self.idx] tmpWeiG[self.idx] = 1 - np.sum(tmpWeiG) elif self.wtype == "local-degree": ### count degrees ### #degMat = np.kron(np.dot(lAdjG,np.ones([self.n,1])), np.ones([1,self.n])) degMat = np.kron( np.dot(lAdjG, np.ones([self.n, 1])) + 1, np.ones([1, self.n])) ### take max() for each elements ### degMat = np.maximum(degMat, degMat.T) ### divide for each elememts ### tmpAllWeiG = lAdjG / degMat selfDegMat = np.eye(self.n) - np.diag( np.sum(tmpAllWeiG, axis=1)) tmpAllWeiG = tmpAllWeiG + selfDegMat tmpWeiG = tmpAllWeiG[self.idx, :] else: try: raise ValueError("Error: invalid weight-type") except ValueError as e: print(e) self.WeiG = tmpWeiG ######## # debugging functions ######## def degub_numericalGrad(self): return self.layer.numerical_gradient(self.x_train[:3], self.t_train[:3]) def debug_backpropGrad(self): return self.layer.gradient(self.x_train[:3], self.t_train[:3]) def debug_consensus(self): params = self.layer.params.copy() self.weightConsensus() self.subvalConsensus() if self.idx == 0: ano_params = self.layer.params.copy() for key in params.keys(): diff = np.average(np.abs(params[key] - ano_params[key])) print(key + ":" + str(diff))