Create GAN.py

neural_network/gan.py

@@ -0,0 +1,391 @@
+import matplotlib.gridspec as gridspec
+import matplotlib.pyplot as plt
+import numpy as np
+from sklearn.utils import shuffle
+import input_data
+
+random_numer = 42
+
+np.random.seed(random_numer)
+def ReLu(x):
+    mask = (x>0) * 1.0
+    return mask *x
+def d_ReLu(x):
+    mask = (x>0) * 1.0
+    return mask
+
+def arctan(x):
+    return np.arctan(x)
+def d_arctan(x):
+    return 1 / (1 + x ** 2)
+
+def log(x):
+    return 1 / ( 1+ np.exp(-1*x))
+def d_log(x):
+    return log(x) * (1 - log(x))
+
+def tanh(x):
+    return np.tanh(x)
+def d_tanh(x):
+    return 1 - np.tanh(x) ** 2
+
+def plot(samples):
+    fig = plt.figure(figsize=(4, 4))
+    gs = gridspec.GridSpec(4, 4)
+    gs.update(wspace=0.05, hspace=0.05)
+
+    for i, sample in enumerate(samples):
+        ax = plt.subplot(gs[i])
+        plt.axis('off')
+        ax.set_xticklabels([])
+        ax.set_yticklabels([])
+        ax.set_aspect('equal')
+        plt.imshow(sample.reshape(28, 28), cmap='Greys_r')
+
+    return fig
+
+
+
+# 1. Load Data and declare hyper
+print('--------- Load Data ----------')
+mnist = input_data.read_data_sets('MNIST_data', one_hot=False)
+temp = mnist.test
+images, labels = temp.images, temp.labels
+images, labels = shuffle(np.asarray(images),np.asarray(labels))
+num_epoch = 10
+learing_rate = 0.00009
+G_input = 100
+hidden_input,hidden_input2,hidden_input3 = 128,256,346
+hidden_input4,hidden_input5,hidden_input6 = 480,560,686
+
+
+
+print('--------- Declare Hyper Parameters ----------')
+# 2. Declare Weights
+D_W1 = np.random.normal(size=(784,hidden_input),scale=(1. / np.sqrt(784 / 2.)))   *0.002
+# D_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.)))       *0.002
+D_b1 = np.zeros(hidden_input)
+
+D_W2 = np.random.normal(size=(hidden_input,1),scale=(1. / np.sqrt(hidden_input / 2.)))     *0.002
+# D_b2 = np.random.normal(size=(1),scale=(1. / np.sqrt(1 / 2.)))           *0.002
+D_b2 = np.zeros(1)
+
+
+G_W1 = np.random.normal(size=(G_input,hidden_input),scale=(1. / np.sqrt(G_input / 2.)))   *0.002
+# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.)))      *0.002
+G_b1 = np.zeros(hidden_input)
+
+G_W2 = np.random.normal(size=(hidden_input,hidden_input2),scale=(1. / np.sqrt(hidden_input / 2.)))   *0.002
+# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.)))      *0.002
+G_b2 = np.zeros(hidden_input2)
+
+G_W3 = np.random.normal(size=(hidden_input2,hidden_input3),scale=(1. / np.sqrt(hidden_input2 / 2.)))   *0.002
+# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.)))      *0.002
+G_b3 = np.zeros(hidden_input3)
+
+G_W4 = np.random.normal(size=(hidden_input3,hidden_input4),scale=(1. / np.sqrt(hidden_input3 / 2.)))   *0.002
+# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.)))      *0.002
+G_b4 = np.zeros(hidden_input4)
+
+G_W5 = np.random.normal(size=(hidden_input4,hidden_input5),scale=(1. / np.sqrt(hidden_input4 / 2.)))   *0.002
+# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.)))      *0.002
+G_b5 = np.zeros(hidden_input5)
+
+G_W6 = np.random.normal(size=(hidden_input5,hidden_input6),scale=(1. / np.sqrt(hidden_input5 / 2.)))   *0.002
+# G_b1 = np.random.normal(size=(128),scale=(1. / np.sqrt(128 / 2.)))      *0.002
+G_b6 = np.zeros(hidden_input6)
+
+G_W7 = np.random.normal(size=(hidden_input6,784),scale=(1. / np.sqrt(hidden_input6 / 2.)))  *0.002
+# G_b2 = np.random.normal(size=(784),scale=(1. / np.sqrt(784 / 2.)))      *0.002
+G_b7 = np.zeros(784)
+
+# 3. For Adam Optimzier
+v1,m1 = 0,0
+v2,m2 = 0,0
+v3,m3 = 0,0
+v4,m4 = 0,0
+
+v5,m5 = 0,0
+v6,m6 = 0,0
+v7,m7 = 0,0
+v8,m8 = 0,0
+v9,m9 = 0,0
+v10,m10 = 0,0
+v11,m11 = 0,0
+v12,m12 = 0,0
+
+v13,m13 = 0,0
+v14,m14 = 0,0
+
+v15,m15 = 0,0
+v16,m16 = 0,0
+
+v17,m17 = 0,0
+v18,m18 = 0,0
+
+
+beta_1,beta_2,eps = 0.9,0.999,0.00000001
+
+print('--------- Started Training ----------')
+for iter in range(num_epoch):
+
+    random_int = np.random.randint(len(images) - 5)
+    current_image = np.expand_dims(images[random_int],axis=0)
+
+    # Func: Generate The first Fake Data
+    Z = np.random.uniform(-1., 1., size=[1, G_input])
+    Gl1 = Z.dot(G_W1) + G_b1
+    Gl1A = arctan(Gl1)
+    Gl2 = Gl1A.dot(G_W2) + G_b2
+    Gl2A = ReLu(Gl2)
+    Gl3 = Gl2A.dot(G_W3) + G_b3
+    Gl3A = arctan(Gl3)
+
+    Gl4 = Gl3A.dot(G_W4) + G_b4
+    Gl4A = ReLu(Gl4)
+    Gl5 = Gl4A.dot(G_W5) + G_b5
+    Gl5A = tanh(Gl5)
+    Gl6 = Gl5A.dot(G_W6) + G_b6
+    Gl6A = ReLu(Gl6)
+    Gl7 = Gl6A.dot(G_W7) + G_b7
+
+    current_fake_data = log(Gl7)
+
+    # Func: Forward Feed for Real data
+    Dl1_r = current_image.dot(D_W1) + D_b1
+    Dl1_rA = ReLu(Dl1_r)
+    Dl2_r = Dl1_rA.dot(D_W2) + D_b2
+    Dl2_rA = log(Dl2_r)
+
+    # Func: Forward Feed for Fake Data
+    Dl1_f = current_fake_data.dot(D_W1) + D_b1
+    Dl1_fA = ReLu(Dl1_f)
+    Dl2_f = Dl1_fA.dot(D_W2) + D_b2
+    Dl2_fA = log(Dl2_f)
+
+    # Func: Cost D
+    D_cost = -np.log(Dl2_rA) + np.log(1.0- Dl2_fA)
+
+    # Func: Gradient
+    grad_f_w2_part_1 =  1/(1.0- Dl2_fA)
+    grad_f_w2_part_2 =  d_log(Dl2_f)
+    grad_f_w2_part_3 =   Dl1_fA
+    grad_f_w2 =       grad_f_w2_part_3.T.dot(grad_f_w2_part_1 * grad_f_w2_part_2)
+    grad_f_b2 = grad_f_w2_part_1 * grad_f_w2_part_2
+
+    grad_f_w1_part_1 =  (grad_f_w2_part_1 * grad_f_w2_part_2).dot(D_W2.T)
+    grad_f_w1_part_2 =  d_ReLu(Dl1_f)
+    grad_f_w1_part_3 =   current_fake_data
+    grad_f_w1 =       grad_f_w1_part_3.T.dot(grad_f_w1_part_1 * grad_f_w1_part_2)
+    grad_f_b1 =      grad_f_w1_part_1 * grad_f_w1_part_2
+
+    grad_r_w2_part_1 =  - 1/Dl2_rA
+    grad_r_w2_part_2 =  d_log(Dl2_r)
+    grad_r_w2_part_3 =   Dl1_rA
+    grad_r_w2 =       grad_r_w2_part_3.T.dot(grad_r_w2_part_1 * grad_r_w2_part_2)
+    grad_r_b2 =       grad_r_w2_part_1 * grad_r_w2_part_2
+
+    grad_r_w1_part_1 =  (grad_r_w2_part_1 * grad_r_w2_part_2).dot(D_W2.T)
+    grad_r_w1_part_2 =  d_ReLu(Dl1_r)
+    grad_r_w1_part_3 =   current_image
+    grad_r_w1 =       grad_r_w1_part_3.T.dot(grad_r_w1_part_1 * grad_r_w1_part_2)
+    grad_r_b1 =       grad_r_w1_part_1 * grad_r_w1_part_2
+
+    grad_w1 =grad_f_w1 + grad_r_w1
+    grad_b1 =grad_f_b1 + grad_r_b1
+
+    grad_w2 =grad_f_w2 + grad_r_w2
+    grad_b2 =grad_f_b2 + grad_r_b2
+
+    # ---- Update Gradient ----
+    m1 = beta_1 * m1 + (1 - beta_1) * grad_w1
+    v1 = beta_2 * v1 + (1 - beta_2) * grad_w1 ** 2
+
+    m2 = beta_1 * m2 + (1 - beta_1) * grad_b1
+    v2 = beta_2 * v2 + (1 - beta_2) * grad_b1 ** 2
+
+    m3 = beta_1 * m3 + (1 - beta_1) * grad_w2
+    v3 = beta_2 * v3 + (1 - beta_2) * grad_w2 ** 2
+
+    m4 = beta_1 * m4 + (1 - beta_1) * grad_b2
+    v4 = beta_2 * v4 + (1 - beta_2) * grad_b2 ** 2
+
+    D_W1 = D_W1 - (learing_rate / (np.sqrt(v1 /(1-beta_2) ) + eps)) * (m1/(1-beta_1))
+    D_b1 = D_b1 - (learing_rate / (np.sqrt(v2 /(1-beta_2) ) + eps)) * (m2/(1-beta_1))
+
+    D_W2 = D_W2 - (learing_rate / (np.sqrt(v3 /(1-beta_2) ) + eps)) * (m3/(1-beta_1))
+    D_b2 = D_b2 - (learing_rate / (np.sqrt(v4 /(1-beta_2) ) + eps)) * (m4/(1-beta_1))
+
+    # Func: Forward Feed for G
+    Z = np.random.uniform(-1., 1., size=[1, G_input])
+    Gl1 = Z.dot(G_W1) + G_b1
+    Gl1A = arctan(Gl1)
+    Gl2 = Gl1A.dot(G_W2) + G_b2
+    Gl2A = ReLu(Gl2)
+    Gl3 = Gl2A.dot(G_W3) + G_b3
+    Gl3A = arctan(Gl3)
+
+    Gl4 = Gl3A.dot(G_W4) + G_b4
+    Gl4A = ReLu(Gl4)
+    Gl5 = Gl4A.dot(G_W5) + G_b5
+    Gl5A = tanh(Gl5)
+    Gl6 = Gl5A.dot(G_W6) + G_b6
+    Gl6A = ReLu(Gl6)
+    Gl7 = Gl6A.dot(G_W7) + G_b7
+
+    current_fake_data = log(Gl7)
+
+    Dl1 = current_fake_data.dot(D_W1) + D_b1
+    Dl1_A = ReLu(Dl1)
+    Dl2 = Dl1_A.dot(D_W2) + D_b2
+    Dl2_A = log(Dl2)
+
+    # Func: Cost G
+    G_cost = -np.log(Dl2_A)
+
+    # Func: Gradient
+    grad_G_w7_part_1 = ((-1/Dl2_A) * d_log(Dl2).dot(D_W2.T) * (d_ReLu(Dl1))).dot(D_W1.T)
+    grad_G_w7_part_2 = d_log(Gl7)
+    grad_G_w7_part_3 = Gl6A
+    grad_G_w7 = grad_G_w7_part_3.T.dot(grad_G_w7_part_1 * grad_G_w7_part_1)
+    grad_G_b7 = grad_G_w7_part_1 * grad_G_w7_part_2
+
+    grad_G_w6_part_1 = (grad_G_w7_part_1 * grad_G_w7_part_2).dot(G_W7.T)
+    grad_G_w6_part_2 = d_ReLu(Gl6)
+    grad_G_w6_part_3 = Gl5A
+    grad_G_w6 = grad_G_w6_part_3.T.dot(grad_G_w6_part_1 * grad_G_w6_part_2)
+    grad_G_b6 = (grad_G_w6_part_1 * grad_G_w6_part_2)
+
+    grad_G_w5_part_1 = (grad_G_w6_part_1 * grad_G_w6_part_2).dot(G_W6.T)
+    grad_G_w5_part_2 = d_tanh(Gl5)
+    grad_G_w5_part_3 = Gl4A
+    grad_G_w5 = grad_G_w5_part_3.T.dot(grad_G_w5_part_1 * grad_G_w5_part_2)
+    grad_G_b5 = (grad_G_w5_part_1 * grad_G_w5_part_2)
+
+    grad_G_w4_part_1 = (grad_G_w5_part_1 * grad_G_w5_part_2).dot(G_W5.T)
+    grad_G_w4_part_2 = d_ReLu(Gl4)
+    grad_G_w4_part_3 = Gl3A
+    grad_G_w4 = grad_G_w4_part_3.T.dot(grad_G_w4_part_1 * grad_G_w4_part_2)
+    grad_G_b4 = (grad_G_w4_part_1 * grad_G_w4_part_2)
+
+    grad_G_w3_part_1 = (grad_G_w4_part_1 * grad_G_w4_part_2).dot(G_W4.T)
+    grad_G_w3_part_2 = d_arctan(Gl3)
+    grad_G_w3_part_3 = Gl2A
+    grad_G_w3 = grad_G_w3_part_3.T.dot(grad_G_w3_part_1 * grad_G_w3_part_2)
+    grad_G_b3 = (grad_G_w3_part_1 * grad_G_w3_part_2)
+
+    grad_G_w2_part_1 = (grad_G_w3_part_1 * grad_G_w3_part_2).dot(G_W3.T)
+    grad_G_w2_part_2 = d_ReLu(Gl2)
+    grad_G_w2_part_3 = Gl1A
+    grad_G_w2 = grad_G_w2_part_3.T.dot(grad_G_w2_part_1 * grad_G_w2_part_2)
+    grad_G_b2 = (grad_G_w2_part_1 * grad_G_w2_part_2)
+
+    grad_G_w1_part_1 = (grad_G_w2_part_1 * grad_G_w2_part_2).dot(G_W2.T)
+    grad_G_w1_part_2 = d_arctan(Gl1)
+    grad_G_w1_part_3 = Z
+    grad_G_w1 = grad_G_w1_part_3.T.dot(grad_G_w1_part_1 * grad_G_w1_part_2)
+    grad_G_b1 = grad_G_w1_part_1 * grad_G_w1_part_2
+
+    # ---- Update Gradient ----
+    m5 = beta_1 * m5 + (1 - beta_1) * grad_G_w1
+    v5 = beta_2 * v5 + (1 - beta_2) * grad_G_w1 ** 2
+
+    m6 = beta_1 * m6 + (1 - beta_1) * grad_G_b1
+    v6 = beta_2 * v6 + (1 - beta_2) * grad_G_b1 ** 2
+
+    m7 = beta_1 * m7 + (1 - beta_1) * grad_G_w2
+    v7 = beta_2 * v7 + (1 - beta_2) * grad_G_w2 ** 2
+
+    m8 = beta_1 * m8 + (1 - beta_1) * grad_G_b2
+    v8 = beta_2 * v8 + (1 - beta_2) * grad_G_b2 ** 2
+
+    m9 = beta_1 * m9 + (1 - beta_1) * grad_G_w3
+    v9 = beta_2 * v9 + (1 - beta_2) * grad_G_w3 ** 2
+
+    m10 = beta_1 * m10 + (1 - beta_1) * grad_G_b3
+    v10 = beta_2 * v10 + (1 - beta_2) * grad_G_b3 ** 2
+
+    m11 = beta_1 * m11 + (1 - beta_1) * grad_G_w4
+    v11 = beta_2 * v11 + (1 - beta_2) * grad_G_w4 ** 2
+
+    m12 = beta_1 * m12 + (1 - beta_1) * grad_G_b4
+    v12 = beta_2 * v12 + (1 - beta_2) * grad_G_b4 ** 2
+
+    m13 = beta_1 * m13 + (1 - beta_1) * grad_G_w5
+    v13 = beta_2 * v13 + (1 - beta_2) * grad_G_w5 ** 2
+
+    m14 = beta_1 * m14 + (1 - beta_1) * grad_G_b5
+    v14 = beta_2 * v14 + (1 - beta_2) * grad_G_b5 ** 2
+
+    m15 = beta_1 * m15 + (1 - beta_1) * grad_G_w6
+    v15 = beta_2 * v15 + (1 - beta_2) * grad_G_w6 ** 2
+
+    m16 = beta_1 * m16 + (1 - beta_1) * grad_G_b6
+    v16 = beta_2 * v16 + (1 - beta_2) * grad_G_b6 ** 2
+
+    m17 = beta_1 * m17 + (1 - beta_1) * grad_G_w7
+    v17 = beta_2 * v17 + (1 - beta_2) * grad_G_w7 ** 2
+
+    m18 = beta_1 * m18 + (1 - beta_1) * grad_G_b7
+    v18 = beta_2 * v18 + (1 - beta_2) * grad_G_b7 ** 2
+
+    G_W1 = G_W1 - (learing_rate / (np.sqrt(v5 /(1-beta_2) ) + eps)) * (m5/(1-beta_1))
+    G_b1 = G_b1 - (learing_rate / (np.sqrt(v6 /(1-beta_2) ) + eps)) * (m6/(1-beta_1))
+
+    G_W2 = G_W2 - (learing_rate / (np.sqrt(v7 /(1-beta_2) ) + eps)) * (m7/(1-beta_1))
+    G_b2 = G_b2 - (learing_rate / (np.sqrt(v8 /(1-beta_2) ) + eps)) * (m8/(1-beta_1))
+
+    G_W3 = G_W3 - (learing_rate / (np.sqrt(v9 /(1-beta_2) ) + eps)) * (m9/(1-beta_1))
+    G_b3 = G_b3 - (learing_rate / (np.sqrt(v10 /(1-beta_2) ) + eps)) * (m10/(1-beta_1))
+
+    G_W4 = G_W4 - (learing_rate / (np.sqrt(v11 /(1-beta_2) ) + eps)) * (m11/(1-beta_1))
+    G_b4 = G_b4 - (learing_rate / (np.sqrt(v12 /(1-beta_2) ) + eps)) * (m12/(1-beta_1))
+
+    G_W5 = G_W5 - (learing_rate / (np.sqrt(v13 /(1-beta_2) ) + eps)) * (m13/(1-beta_1))
+    G_b5 = G_b5 - (learing_rate / (np.sqrt(v14 /(1-beta_2) ) + eps)) * (m14/(1-beta_1))
+
+    G_W6 = G_W6 - (learing_rate / (np.sqrt(v15 /(1-beta_2) ) + eps)) * (m15/(1-beta_1))
+    G_b6 = G_b6 - (learing_rate / (np.sqrt(v16 /(1-beta_2) ) + eps)) * (m16/(1-beta_1))
+
+    G_W7 = G_W7 - (learing_rate / (np.sqrt(v17 /(1-beta_2) ) + eps)) * (m17/(1-beta_1))
+    G_b7 = G_b7 - (learing_rate / (np.sqrt(v18 /(1-beta_2) ) + eps)) * (m18/(1-beta_1))
+
+    # --- Print Error ----
+    #print("Current Iter: ",iter, " Current D cost:",D_cost, " Current G cost: ", G_cost,end='\r')
+
+    if iter == 0:
+        learing_rate = learing_rate * 0.01
+    if iter == 40:
+        learing_rate = learing_rate * 0.01
+
+    # ---- Print to Out put ----
+    if iter%10 == 0:
+
+        print("Current Iter: ",iter, " Current D cost:",D_cost, " Current G cost: ", G_cost,end='\r')
+        print('--------- Show Example Result See Tab Above ----------')
+        print('--------- Wait for the image to load ---------')
+        Z = np.random.uniform(-1., 1., size=[16, G_input])
+
+        Gl1 = Z.dot(G_W1) + G_b1
+        Gl1A = arctan(Gl1)
+        Gl2 = Gl1A.dot(G_W2) + G_b2
+        Gl2A = ReLu(Gl2)
+        Gl3 = Gl2A.dot(G_W3) + G_b3
+        Gl3A = arctan(Gl3)
+
+        Gl4 = Gl3A.dot(G_W4) + G_b4
+        Gl4A = ReLu(Gl4)
+        Gl5 = Gl4A.dot(G_W5) + G_b5
+        Gl5A = tanh(Gl5)
+        Gl6 = Gl5A.dot(G_W6) + G_b6
+        Gl6A = ReLu(Gl6)
+        Gl7 = Gl6A.dot(G_W7) + G_b7
+
+        current_fake_data = log(Gl7)
+
+        fig = plot(current_fake_data)
+        fig.savefig('Click_Me_{}.png'.format(str(iter).zfill(3)+"_Ginput_"+str(G_input)+ \
+        "_hiddenone"+str(hidden_input) + "_hiddentwo"+str(hidden_input2) + "_LR_" + str(learing_rate)
+        ), bbox_inches='tight')
+#for complete explanation visit https://towardsdatascience.com/only-numpy-implementing-gan-general-adversarial-networks-and-adam-optimizer-using-numpy-with-2a7e4e032021
+# -- end code --