3/14/22, 6:14 PM NumberGAN.

ipynb - Colaboratory

1) Importing Python Packages for GAN

from keras.datasets import mnist

from keras.models import Sequential
from keras.layers import BatchNormalization
from keras.layers import Dense, Reshape, Flatten
from keras.layers.advanced_activations import LeakyReLU
from tensorflow.keras.optimizers import Adam

import numpy as np
!mkdir generated_images

2) Variables for Neural Networks & Data

img_width = 28
img_height = 28
channels = 1
img_shape = (img_width, img_height, channels)
latent_dim = 100
adam = Adam(lr=0.0001)

/usr/local/lib/python3.7/dist-packages/keras/optimizer_v2/adam.py:105: UserWarning: T
super(Adam, self).__init__(name, **kwargs)

3) Building Generator

def build_generator():

  model = Sequential()

  model.add(Dense(256, input_dim=latent_dim))

  model.add(LeakyReLU(alpha=0.2))

  model.add(BatchNormalization(momentum=0.8))

  model.add(Dense(256))

  model.add(LeakyReLU(alpha=0.2))

  model.add(BatchNormalization(momentum=0.8))

  model.add(Dense(256))

  model.add(LeakyReLU(alpha=0.2))

  model.add(BatchNormalization(momentum=0.8))

  model.add(Dense(np.prod(img_shape), activation='tanh'))

  model.add(Reshape(img_shape))

https://colab.research.google.com/drive/1IXODAITOAo5BGLfViCY_ExK-8UHBQFNY#scrollTo=XPShgQpg1EMy 1/7
3/14/22, 6:14 PM NumberGAN.ipynb - Colaboratory

  model.summary()

  return model

generator = build_generator()

Model: "sequential"

_________________________________________________________________

Layer (type) Output Shape Param #

=================================================================

dense (Dense) (None, 256) 25856

leaky_re_lu (LeakyReLU) (None, 256) 0

batch_normalization (BatchN (None, 256) 1024

ormalization)

dense_1 (Dense) (None, 256) 65792

leaky_re_lu_1 (LeakyReLU) (None, 256) 0

batch_normalization_1 (Batc (None, 256) 1024

hNormalization)

dense_2 (Dense) (None, 256) 65792

leaky_re_lu_2 (LeakyReLU) (None, 256) 0

batch_normalization_2 (Batc (None, 256) 1024

hNormalization)

dense_3 (Dense) (None, 784) 201488

reshape (Reshape) (None, 28, 28, 1) 0

=================================================================

Total params: 362,000

Trainable params: 360,464

Non-trainable params: 1,536

_________________________________________________________________

4) Building Discriminator

def build_discriminator():

  model = Sequential()

  model.add(Flatten(input_shape=img_shape))

  model.add(Dense(512))

  model.add(LeakyReLU(alpha=0.2))

  model.add(Dense(256))

  model.add(Dense(1, activation='sigmoid'))

  model.summary()

  return model

https://colab.research.google.com/drive/1IXODAITOAo5BGLfViCY_ExK-8UHBQFNY#scrollTo=XPShgQpg1EMy 2/7
3/14/22, 6:14 PM NumberGAN.ipynb - Colaboratory

discriminator = build_discriminator()

discriminator.compile(loss='binary_crossentropy', optimizer=adam, metrics=['accuracy'])

Model: "sequential_1"

_________________________________________________________________

Layer (type) Output Shape Param #

=================================================================

flatten (Flatten) (None, 784) 0

dense_4 (Dense) (None, 512) 401920

leaky_re_lu_3 (LeakyReLU) (None, 512) 0

dense_5 (Dense) (None, 256) 131328

dense_6 (Dense) (None, 1) 257

=================================================================

Total params: 533,505

Trainable params: 533,505

Non-trainable params: 0

_________________________________________________________________

5) Connecting Neural Networks to build GAN

GAN = Sequential()

discriminator.trainable = False

GAN.add(generator)

GAN.add(discriminator)

GAN.compile(loss='binary_crossentropy', optimizer=adam)

GAN.summary()

Model: "sequential_2"

_________________________________________________________________

Layer (type) Output Shape Param #

=================================================================

sequential (Sequential) (None, 28, 28, 1) 362000

sequential_1 (Sequential) (None, 1) 533505

=================================================================

Total params: 895,505

Trainable params: 360,464

Non-trainable params: 535,041

_________________________________________________________________

6) Outputting Images

#@title

## **7) Outputting Images**

import matplotlib.pyplot as plt

https://colab.research.google.com/drive/1IXODAITOAo5BGLfViCY_ExK-8UHBQFNY#scrollTo=XPShgQpg1EMy 3/7
3/14/22, 6:14 PM NumberGAN.ipynb - Colaboratory

import glob

import imageio

import PIL

save_name = 0.00000000

def save_imgs(epoch):

    r, c = 5, 5

    noise = np.random.normal(0, 1, (r * c, latent_dim))

    gen_imgs = generator.predict(noise)

    global save_name

    save_name += 0.00000001

    print("%.8f" % save_name)

    # Rescale images 0 - 1

    gen_imgs = 0.5 * gen_imgs + 0.5

    fig, axs = plt.subplots(r, c)
    cnt = 0

    for i in range(r):

        for j in range(c):

            axs[i,j].imshow(gen_imgs[cnt, :,:,0], cmap='gray')

            # axs[i,j].imshow(gen_imgs[cnt])

            axs[i,j].axis('off')

            cnt += 1

    fig.savefig("generated_images/%.8f.png" % save_name)

    print('saved')

    plt.close()

7) Training GAN

def train(epochs, batch_size=64, save_interval=200):

  (X_train, _), (_, _) = mnist.load_data()

  # print(X_train.shape)

  #Rescale data between -1 and 1

  X_train = X_train / 127.5 -1.

  # X_train = np.expand_dims(X_train, axis=3)

  # print(X_train.shape)

  #Create our Y for our Neural Networks

  valid = np.ones((batch_size, 1))

  fakes = np.zeros((batch_size, 1))

  for epoch in range(epochs):

    #Get Random Batch

    idx = np.random.randint(0, X_train.shape[0], batch_size)

    imgs = X_train[idx]

    #Generate Fake Images

    noise = np.random.normal(0, 1, (batch_size, latent_dim))

https://colab.research.google.com/drive/1IXODAITOAo5BGLfViCY_ExK-8UHBQFNY#scrollTo=XPShgQpg1EMy 4/7
3/14/22, 6:14 PM NumberGAN.ipynb - Colaboratory

    gen_imgs = generator.predict(noise)

    #Train discriminator

    d_loss_real = discriminator.train_on_batch(imgs, valid)

    d_loss_fake = discriminator.train_on_batch(gen_imgs, fakes)

    d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)

    noise = np.random.normal(0, 1, (batch_size, latent_dim))

    #inverse y label

    g_loss = GAN.train_on_batch(noise, valid)

    print("******* %d [D loss: %f, acc: %.2f%%] [G loss: %f]" % (epoch, d_loss[0], 100* d_

    if(epoch % save_interval) == 0:

      save_imgs(epoch)

  # print(valid)

train(30000, batch_size=64, save_interval=200)

******* 29942 [D loss: 0.559609, acc: 70.31%] [G loss: 1.274176]

******* 29943 [D loss: 0.562384, acc: 71.09%] [G loss: 1.345835]

******* 29944 [D loss: 0.468171, acc: 78.91%] [G loss: 1.376247]

******* 29945 [D loss: 0.539233, acc: 71.09%] [G loss: 1.593389]

******* 29946 [D loss: 0.609176, acc: 68.75%] [G loss: 1.719936]

******* 29947 [D loss: 0.447626, acc: 80.47%] [G loss: 1.609250]

******* 29948 [D loss: 0.570068, acc: 71.88%] [G loss: 1.363200]

******* 29949 [D loss: 0.519664, acc: 71.88%] [G loss: 1.314612]

******* 29950 [D loss: 0.536247, acc: 71.88%] [G loss: 1.293153]

******* 29951 [D loss: 0.601880, acc: 66.41%] [G loss: 1.475299]

******* 29952 [D loss: 0.522597, acc: 72.66%] [G loss: 1.409748]

******* 29953 [D loss: 0.495711, acc: 79.69%] [G loss: 1.676774]

******* 29954 [D loss: 0.462763, acc: 82.03%] [G loss: 1.689979]

******* 29955 [D loss: 0.574412, acc: 70.31%] [G loss: 1.281035]

******* 29956 [D loss: 0.477904, acc: 78.91%] [G loss: 1.460865]

******* 29957 [D loss: 0.460951, acc: 79.69%] [G loss: 1.607989]

******* 29958 [D loss: 0.573022, acc: 69.53%] [G loss: 1.745356]

******* 29959 [D loss: 0.518633, acc: 69.53%] [G loss: 1.448777]

******* 29960 [D loss: 0.628339, acc: 71.09%] [G loss: 1.390859]

******* 29961 [D loss: 0.569294, acc: 72.66%] [G loss: 1.289187]

******* 29962 [D loss: 0.524241, acc: 75.00%] [G loss: 1.568775]

******* 29963 [D loss: 0.568505, acc: 70.31%] [G loss: 1.456203]

******* 29964 [D loss: 0.561265, acc: 75.00%] [G loss: 1.549096]

******* 29965 [D loss: 0.595656, acc: 67.97%] [G loss: 1.352225]

******* 29966 [D loss: 0.598945, acc: 68.75%] [G loss: 1.322804]

******* 29967 [D loss: 0.545177, acc: 68.75%] [G loss: 1.389703]

******* 29968 [D loss: 0.560595, acc: 74.22%] [G loss: 1.494714]

******* 29969 [D loss: 0.613210, acc: 68.75%] [G loss: 1.534424]

******* 29970 [D loss: 0.508624, acc: 78.12%] [G loss: 1.334824]

******* 29971 [D loss: 0.522550, acc: 71.09%] [G loss: 1.595869]

******* 29972 [D loss: 0.496946, acc: 75.78%] [G loss: 1.426164]

******* 29973 [D loss: 0.666016, acc: 69.53%] [G loss: 1.390148]

******* 29974 [D loss: 0.567081, acc: 70.31%] [G loss: 1.159011]

******* 29975 [D loss: 0.492483, acc: 74.22%] [G loss: 1.361443]

******* 29976 [D loss: 0.514459, acc: 76.56%] [G loss: 1.619536]

******* 29977 [D loss: 0.522412, acc: 74.22%] [G loss: 1.687408]

******* 29978 [D loss: 0.558854, acc: 73.44%] [G loss: 1.327003]

******* 29979 [D loss: 0.562206, acc: 70.31%] [G loss: 1.694671]

******* 29980 [D loss: 0.525805, acc: 70.31%] [G loss: 1.349991]

******* 29981 [D loss: 0.527241, acc: 73.44%] [G loss: 1.485349]

******* 29982 [D loss: 0.484577, acc: 77.34%] [G loss: 1.573783]

******* 29983 [D loss: 0.487552, acc: 79.69%] [G loss: 1.571594]

******* 29984 [D loss: 0.624213, acc: 69.53%] [G loss: 1.419726]

******* 29985 [D loss: 0.632475, acc: 66.41%] [G loss: 1.326593]

******* 29986 [D loss: 0.488182, acc: 73.44%] [G loss: 1.428414]

******* 29987 [D loss: 0.591298, acc: 68.75%] [G loss: 1.369640]

******* 29988 [D loss: 0.459753, acc: 78.12%] [G loss: 1.578751]

******* 29989 [D loss: 0.509766, acc: 78.91%] [G loss: 1.669813]

******* 29990 [D loss: 0.491158, acc: 77.34%] [G loss: 1.348950]

******* 29991 [D loss: 0.625076, acc: 66.41%] [G loss: 1.167850]

******* 29992 [D loss: 0.632827, acc: 67.19%] [G loss: 1.197102]

******* 29993 [D loss: 0.500343, acc: 78.12%] [G loss: 1.425504]

******* 29994 [D loss: 0.520639, acc: 78.91%] [G loss: 1.719311]

******* 29995 [D loss: 0.424208, acc: 81.25%] [G loss: 1.839022]

******* 29996 [D loss: 0.471873, acc: 76.56%] [G loss: 1.578932]

******* 29997 [D loss: 0.498148, acc: 73.44%] [G loss: 1.425658]

******* 29998 [D loss: 0.583471, acc: 72.66%] [G loss: 1.398373]

******* 29999 [D loss: 0.541282, acc: 71.09%] [G loss: 1.406367]

8) Making GIF

# Display a single image using the epoch number
# def display_image(epoch_no):
#   return PIL.Image.open('generated_images/%.8f.png'.format(epoch_no))

anim_file = 'dcgan.gif'

with imageio.get_writer(anim_file, mode='I') as writer:
  filenames = glob.glob('generated_images/*.png')
  filenames = sorted(filenames)
  for filename in filenames:
    image = imageio.imread(filename)
    writer.append_data(image)
  image = imageio.imread(filename)
  writer.append_data(image)

https://colab.research.google.com/drive/1IXODAITOAo5BGLfViCY_ExK-8UHBQFNY#scrollTo=XPShgQpg1EMy 6/7
3/14/22, 6:14 PM NumberGAN.ipynb - Colaboratory

check 2s completed at 6:13 PM

https://colab.research.google.com/drive/1IXODAITOAo5BGLfViCY_ExK-8UHBQFNY#scrollTo=XPShgQpg1EMy 7/7