TUKRの実装@田中

Lecture_TUKR/tanaka/animal.py

@@ -0,0 +1,28 @@
+import numpy as np
+import os
+
+
+def load_data(retlabel_animal=True, retlabel_feature=False):
+    datastore_name = 'datastore/animal'
+    file_name = 'features.txt'
+
+    directory_path = os.path.join(os.path.dirname(__file__), datastore_name)
+    file_path = os.path.join(directory_path, file_name)
+
+    x = np.loadtxt(file_path)
+
+    return_objects = [x]
+
+    if retlabel_animal:
+        label_name = 'labels_animal.txt'
+        label_path = os.path.join(directory_path, label_name)
+        label_animal = np.genfromtxt(label_path, dtype=str)
+        return_objects.append(label_animal)
+
+    if retlabel_feature:
+        label_name = 'labels_feature.txt'
+        label_path = os.path.join(directory_path, label_name)
+        label_feature = np.genfromtxt(label_path, dtype=str)
+        return_objects.append(label_feature)
+
+    return return_objects

Lecture_TUKR/tanaka/data_scratch_tanaka.py

@@ -0,0 +1,42 @@
+import numpy as np
+import matplotlib.pyplot as plt
+from sklearn.preprocessing import MinMaxScaler
+
+def load_kura_tsom(xsamples, ysamples, missing_rate=None,retz=False):
+    # z1 = np.random.rand(xsamples)/2
+    z1 = np.linspace(-1,1,xsamples)
+    # z2 = np.random.rand(ysamples)/2
+    z2 = np.linspace(-1,1,ysamples)
+
+    z1_repeated, z2_repeated = np.meshgrid(z1,z2)
+    x1 = z1_repeated
+    x2 = z2_repeated
+    x3 = (x1**2-x2**2)
+    #ノイズを加えたい時はここをいじる,locがガウス分布の平均、scaleが分散,size何個ノイズを作るか
+    #このノイズを加えることによって三次元空間のデータ点は上下に動く
+
+    x = np.concatenate((x1[:, :, np.newaxis], x2[:, :, np.newaxis], x3[:, :, np.newaxis]), axis=2)
+    truez = np.concatenate((z1_repeated[:, :, np.newaxis], z2_repeated[:, :, np.newaxis]), axis=2)
+    # print(x.shape)
+
+    if missing_rate == 0 or missing_rate == None:
+        if retz:
+            return x, truez, z1, z2
+        else:
+            return x
+
+if __name__ == '__main__':
+    import matplotlib.pyplot as plt
+    from mpl_toolkits.mplot3d import Axes3D
+
+    xsamples = 10
+    ysamples = 10
+
+    x, truez = load_kura_tsom(xsamples,ysamples,retz=True)
+
+    fig = plt.figure(figsize=[5, 5])
+    ax_x = fig.add_subplot(projection='3d')
+    ax_x.scatter(x[:, :, 0].flatten(), x[:, :, 1].flatten(), x[:, :, 2].flatten(), c=x[:, :, 0].flatten())
+    ax_x.set_title('Generated three-dimensional data')
+    plt.show()
+

Lecture_TUKR/tanaka/datastore/animal/features.txt

@@ -0,0 +1,17 @@
+1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.5 0.5 0.0 0.0 0.0
+0.0 1.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.5 0.5 0.0 0.0 0.0
+0.0 1.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 1.0 0.5 0.5 0.0 0.0 0.0
+0.0 1.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 1.0 1.0 1.0 0.5 0.5 0.0 0.0 0.0
+0.0 1.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 0.3 0.0 0.0 1.0 1.0 0.0 0.5 0.5 0.0 0.0 0.0
+0.0 1.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 1.0 0.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0
+0.0 0.0 1.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 1.0 0.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0
+0.0 0.0 1.0 1.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 1.0 0.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0
+0.0 1.0 0.0 0.5 0.0 1.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.3 0.7 1.0 0.0 0.0
+0.0 1.0 0.0 0.0 0.0 1.0 1.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 1.0 0.0 1.0
+0.0 1.0 0.0 1.0 0.0 1.0 1.0 0.0 0.0 0.0 0.0 1.0 1.0 0.0 0.0 0.0 0.0 1.0 1.0 0.0 0.0
+1.0 0.0 0.0 0.5 0.0 1.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 1.0 1.0
+0.0 0.0 1.0 0.5 0.0 1.0 1.0 0.0 0.0 0.0 1.0 1.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 1.0 0.0
+0.0 0.0 1.0 0.0 0.0 1.0 1.0 0.0 1.0 0.0 0.0 1.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 1.0 0.0
+0.0 0.0 1.0 0.0 0.0 1.0 1.0 1.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 1.0 1.0 0.0 0.0 0.0 0.0
+0.0 0.0 1.0 0.0 0.0 1.0 1.0 1.0 1.0 0.0 1.0 0.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0
+0.0 0.0 1.0 0.0 0.0 1.0 1.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 1.0 0.0 0.0 0.0 0.0

Lecture_TUKR/tanaka/datastore/animal/labels_animal.txt

@@ -0,0 +1,17 @@
+dove
+cock
+duck
+w_duck
+owl
+hawk
+eagle
+crow
+fox
+dog
+wolf
+cat
+tiger
+lion
+horse
+zebra
+cattle

Lecture_TUKR/tanaka/datastore/animal/labels_feature.txt

@@ -0,0 +1,21 @@
+small
+medium
+large
+nocturnality
+two_legs
+four_legs
+hair
+hoof
+mane
+wing
+stripe
+hunt
+run
+fly
+swim
+domestic
+herbivorous
+carnivore
+canidae
+felidae
+pet

Lecture_TUKR/tanaka/tukr.py

@@ -0,0 +1,175 @@
+import numpy as np
+import jax,jaxlib
+import jax.numpy as jnp
+import tensorflow as tf
+from tqdm import tqdm #プログレスバーを表示させてくれる
+from sklearn.datasets import load_iris
+
+
+
+class TUKR:
+    def __init__(self, X, nb_samples1, nb_samples2, latent_dim1, latent_dim2, sigma1, sigma2, prior='random', Uinit=None, Vinit=None):
+        #--------初期値を設定する．---------
+        self.X = X
+        #ここから下は書き換えてね
+
+        if X.ndim == 3:
+            self.nb_samples1, self.nb_samples2, self.ob_dim = self.X.shape
+        else:
+            self.nb_samples1, self.nb_samples2 = self.X.shape
+            self.ob_dim = 1
+            self.X = X[:,:,None]
+
+        self.sigma1 = sigma1
+        self.sigma2 = sigma2
+        self.latent_dim1 = latent_dim1
+        self.latent_dim2 = latent_dim2
+        self.alpha = alpha
+        self.norm = norm
+
+        if Uinit is None:
+            if prior == 'random': #一様事前分布のとき
+                self.U = np.random.uniform(-0.1, 0.1, size=(self.nb_samples1, self.latent_dim1))
+                #(平均,標準偏差,配列のサイズ)
+            else: #ガウス事前分布のとき
+                self.U = np.random.uniform(0, 0.1 * self.sigma1, (self.nb_samples1, self.latent_dim1))
+        else: #Zの初期値が与えられた時
+            self.U = Uinit
+
+        self.history = {}
+
+        if Vinit is None:
+            if prior == 'random':  # 一様事前分布のとき
+                self.V = np.random.normal(-0.1, 0.1 , size=(self.nb_samples2, self.latent_dim2))
+                # (平均,標準偏差,配列のサイズ)
+            else: #ガウス事前分布のとき
+                self.V = np.random.normal(0, 0.1 * self.sigma2, (self.nb_samples2, self.latent_dim2))
+        else:  # Zの初期値が与えられた時
+            self.V = Vinit
+
+        self.history = {}
+
+    def f(self, U, V): #写像の計算
+        DistU = jnp.sum((U[:, None, :] - U[None, :, :]) ** 2, axis=2)
+        DistV = jnp.sum((V[:, None, :] - V[None, :, :]) ** 2, axis=2)
+        HU = jnp.exp((-1 * DistU) / (2 * (self.sigma1) ** 2))
+        HV = jnp.exp((-1 * DistV) / (2 * (self.sigma2) ** 2))
+        # GU = jnp.sum(HU, axis=1)[:, None]
+        # GV = jnp.sum(HV, axis=1)[:, None]
+        # RU = HU / GU
+        # RV = HV / GV
+        f = jnp.einsum('li,kj,ijd->lkd', HU, HV, self.X)
+        f1 = jnp.einsum('li,kj->lk', HU, HV)
+        f2 = f1[:, :, None]
+        return f / f2
+
+    def ff(self, U, V, epoch): #写像の計算
+        DistU = jnp.sum((U[:, None, :] - self.history['u'][epoch][None, :, :]) ** 2, axis=2)
+        DistV = jnp.sum((V[:, None, :] - self.history['v'][epoch][None, :, :]) ** 2, axis=2)
+        HU = jnp.exp((-1 * DistU) / (2 * (self.sigma1) ** 2))
+        HV = jnp.exp((-1 * DistV) / (2 * (self.sigma2) ** 2))
+        # GU = jnp.sum(HU, axis=1)[:, None]
+        # GV = jnp.sum(HV, axis=1)[:, None]
+        # RU = HU / GU
+        # RV = HV / GV
+        f = jnp.einsum('li,kj,ijd->lkd', HU, HV, self.X)
+        f1 = jnp.einsum('li,kj->lk', HU, HV)
+        f2 = f1[:, :, None]
+        return f / f2
+
+    def E(self,U,V,X,alpha,norm):#目的関数の計算
+        Y = self.f(U,V)
+        e = jnp.sum((X - Y) ** 2)
+        r = alpha*(jnp.sum(U**norm)+jnp.sum(V**norm))
+        e = e/(self.nb_samples1*self.nb_samples2)
+        r = r/(self.nb_samples1*self.nb_samples2)
+        return e + r
+
+    def fit(self, nb_epoch: int, eta: float,alpha,norm):
+        # 学習過程記録用
+        self.history['u'] = np.zeros((nb_epoch, self.nb_samples1, self.latent_dim2))
+        self.history['v'] = np.zeros((nb_epoch, self.nb_samples2, self.latent_dim1))
+        self.history['f'] = np.zeros((nb_epoch, self.nb_samples1, self.nb_samples2, self.ob_dim))
+        self.history['error'] = np.zeros(nb_epoch)
+
+        for epoch in tqdm(range(nb_epoch)):
+
+            # U,Vの更新
+            dEdu = jax.grad(self.E, argnums=0)(self.U, self.V, self.X, alpha, norm) / self.nb_samples1
+            self.U = self.U - eta * dEdu
+
+            dEdv = jax.grad(self.E, argnums=1)(self.U, self.V, self.X, alpha, norm) / self.nb_samples2
+            self.V = self.V - eta * dEdv
+
+            # 学習過程記録用
+            self.history['u'][epoch] = self.U
+            self.history['v'][epoch] = self.V
+            self.history['f'][epoch] = self.f(self.U, self.V)
+            self.history['error'][epoch] = self.E(self.U, self.V, self.X, alpha, norm)
+
+    #--------------以下描画用(上の部分が実装できたら実装してね)---------------------
+    def calc_approximate_f(self, resolution): #fのメッシュ描画用，resolution:一辺の代表点の数
+        nb_epoch = self.history['u'].shape[0]
+        self.history['y'] = np.zeros((nb_epoch,self.nb_samples1,self.nb_samples2, self.ob_dim))
+        for epoch in tqdm(range(nb_epoch)):
+            Uzeta = self.create_Uzeta(self.history['u'][epoch],resolution)
+            Vzeta = self.create_Vzeta(self.history['v'][epoch],resolution)
+
+            y = self.ff(Uzeta,Vzeta,epoch)
+            self.history['y'][epoch] = y
+
+    def create_Uzeta(self, U, resolution): #fのメッシュの描画用に潜在空間に代表点zetaを作る．
+        Uzeta = np.linspace(np.min(U), np.max(U),self.nb_samples1).reshape(-1,1)
+
+        return Uzeta
+
+    def create_Vzeta(self, V, resolution):  # fのメッシュの描画用に潜在空間に代表点zetaを作る．
+        Vzeta = np.linspace(np.min(V), np.max(V),self.nb_samples2).reshape(-1,1)
+
+        return Vzeta
+
+
+if __name__ == '__main__':
+    # from Lecture_TUKR.tanaka.animal import load_data
+    from Lecture_TUKR.tanaka.data_scratch_tanaka import load_kura_tsom
+    # from Lecture_TUKR.tanaka.data_scratch_tanaka import create_rasen
+    # from Lecture_TUKR.tanaka.data_scratch_tanaka import create_2d_sin_curve
+    from visualizer import visualize_history
+    # from visualizer_animal import visualize_history
+
+    #各種パラメータ変えて遊んでみてね．
+    epoch = 200 #学習回数
+    sigma1 = 0.2
+    sigma2 = 0.1 #カーネルの幅
+    eta = 50  #学習率
+    latent_dim1 = 2 #潜在空間の次元
+    latent_dim2 = 2 #潜在空間の次元
+    alpha = 0.1
+    norm = 2
+    seed = 4
+    np.random.seed(seed)
+
+
+
+    #入力データ（詳しくはdata.pyを除いてみると良い）
+    nb_samples1 = 10 #データ数
+    nb_samples2 = 20 #データ数
+    # data = load_data(retlabel_animal=True, retlabel_feature=True)
+    # X = load_iris()
+    X = load_kura_tsom(nb_samples1,nb_samples2) #鞍型データ　ob_dim=3, 真のL=2
+    # X = create_rasen(nb_samples) #らせん型データ　ob_dim=3, 真のL=1
+    # X = create_2d_sin_curve(nb_samples) #sin型データ　ob_dim=2, 真のL=1
+
+    # X = data[0]
+    # animal_label = data[1]
+    # feature_label = data[2]
+    X, truez, z1, z2 = load_kura_tsom(nb_samples1, nb_samples2, retz=True)
+    allZ =[truez,z1,z2]
+    tukr = TUKR(X, nb_samples1, nb_samples2, latent_dim1, latent_dim2, sigma1, sigma2, prior='random')
+    tukr.fit(epoch, eta, alpha, norm)
+    # visualize_history(X, tukr.history['f'], tukr.history['u'],tukr.history['v'], tukr.history['error'], save_gif=False, filename="tmp")
+
+    #----------描画部分が実装されたらコメントアウト外す----------
+    tukr.calc_approximate_f(resolution=10)
+    # visualize_history(X, tukr.history['y'], tukr.history['u'],tukr.history['v'], tukr.history['error'],animal_label, feature_label, save_gif=False, filename="tmp")
+    visualize_history(X, tukr.history['y'], tukr.history['u'],tukr.history['v'], tukr.history['error'], save_gif=False, filename="tmp")