Denoising autoencoders

import numpy as np
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import matplotlib.pyplot as plt
 
data = input_data.read_data_sets("./mnist/", one_hot=True)
 
# Print shapes of data
print("Training X: ", data.train.images.shape)
print("Training Y: ", data.train.labels.shape)
print("Test X: ", data.test.images.shape)
print("Test Y: ", data.test.labels.shape)
 
def gaussian_additive_noise(x, std):
    return x + tf.random_normal(shape=tf.shape(x), dtype=tf.float32, mean=0.0, stddev=std)
 
imgs = tf.placeholder(tf.float32, shape=[None, 28*28], name="Input")
 
noise = gaussian_additive_noise(imgs, 0.1)
corrupted_imgs_test = noise.eval(session=tf.Session(), feed_dict={imgs: data.test.images})
 
def plot_mnist(imgs, lbls):
    classes = np.argmax(lbls, 1)
    for i in range(10):
        ids = (classes == i)
        images = imgs[ids][0:10]
        for j in range(3):   
            plt.subplot(5, 10, i + j*10 + 1)
            plt.imshow(images[j].reshape(28, 28), cmap='gray')
            if j == 0:
                plt.title(i)
            plt.axis('off')
    plt.show()
 
def autoencoder(dims=[28*28, 512, 256, 128, 64, 32], std=0.01):
    x = tf.placeholder(tf.float32, shape=[None, dims[0]], name="Input")
    cur = gaussian_additive_noise(x, 0.1)
    Ws = []
    bs = []
    # encoder
    for i, n_out in enumerate(dims[1:]):
        n_inp = int(cur.get_shape()[1])
        W = tf.Variable(tf.random_normal(shape=[n_inp, n_out], mean=0.0, stddev=std, dtype=tf.float32))
        b = tf.Variable(tf.random_normal(shape=[n_out], mean=0.0, stddev=std, dtype=tf.float32))
        Ws.append(W)
        bs.append(b)
        out = tf.nn.tanh(cur @ W + b)
        cur = out
    z = cur
    Ws.reverse()
    bs.reverse()
    # decoder
    for i, n_out in enumerate(dims[:-1][::-1]):
        W = tf.transpose(Ws[i])
        b = tf.Variable(tf.random_normal(shape=[n_out], mean=0.0, stddev=std, dtype=tf.float32))
        out = tf.nn.tanh(cur @ W + b)
        cur = out
    y = cur
    loss = tf.reduce_mean(tf.square(y - x))
    return (x, z, y, loss)
 
lr = 0.001
batch_size = 64
n_epochs = 50
n_batchs = data.train.num_examples // batch_size
 
x, z, y, loss = autoencoder(dims=[28*28, 512, 256, 64], std=0.01)
optimizer = tf.train.AdamOptimizer(lr).minimize(loss)
 
S = tf.Session()
S.run(tf.global_variables_initializer())
 
for i_epoch in range(1, n_epochs+1):
    loss_avg = 0.0
    for i_batch in range(1, n_batchs+1):
        b, _ = data.train.next_batch(batch_size)
        _, loss_val = S.run([optimizer, loss], feed_dict={x: b})
        loss_avg = (loss_val / batch_size)
    print(i_epoch, loss_avg)
    loss_avg = 0.0
 
n_samples = 10
reconstructed = S.run([y], feed_dict={x: corrupted_imgs_test})
 
reconstructed = reconstructed[0]
 
print("\t\t Original Images")
plot_mnist(data.test.images, data.test.labels)
print("\t\t Corrupted Images")
plot_mnist(corrupted_imgs_test, data.test.labels)
print("\t\t Reconstructed Images")
plot_mnist(reconstructed, data.test.labels)
 

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
 
from tensorflow.keras import layers
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Model
 
 
def preprocess(array):
    """
    Normalizes the supplied array and reshapes it into the appropriate format.
    """
 
    array = array.astype("float32") / 255.0
    array = np.reshape(array, (len(array), 28, 28, 1))
    return array
 
 
def noise(array):
    """
    Adds random noise to each image in the supplied array.
    """
 
    noise_factor = 0.4
    noisy_array = array + noise_factor * np.random.normal(
        loc=0.0, scale=1.0, size=array.shape
    )
 
    return np.clip(noisy_array, 0.0, 1.0)
 
 
def display(array1, array2):
    """
    Displays ten random images from each one of the supplied arrays.
    """
 
    n = 10
 
    indices = np.random.randint(len(array1), size=n)
    images1 = array1[indices, :]
    images2 = array2[indices, :]
 
    plt.figure(figsize=(20, 4))
    for i, (image1, image2) in enumerate(zip(images1, images2)):
        ax = plt.subplot(2, n, i + 1)
        plt.imshow(image1.reshape(28, 28))
        plt.gray()
        ax.get_xaxis().set_visible(False)
        ax.get_yaxis().set_visible(False)
 
        ax = plt.subplot(2, n, i + 1 + n)
        plt.imshow(image2.reshape(28, 28))
        plt.gray()
        ax.get_xaxis().set_visible(False)
        ax.get_yaxis().set_visible(False)
 
    plt.show()
 
#Prepare the data
 
# Since we only need images from the dataset to encode and decode, we
# won't use the labels.
(train_data, _), (test_data, _) = mnist.load_data()
 
# Normalize and reshape the data
train_data = preprocess(train_data)
test_data = preprocess(test_data)
 
# Create a copy of the data with added noise
noisy_train_data = noise(train_data)
noisy_test_data = noise(test_data)
 
# Display the train data and a version of it with added noise
display(train_data, noisy_train_data)
 
#Build the autoencoder
 
input = layers.Input(shape=(28, 28, 1))
 
# Encoder
x = layers.Conv2D(32, (3, 3), activation="relu", padding="same")(input)
x = layers.MaxPooling2D((2, 2), padding="same")(x)
x = layers.Conv2D(32, (3, 3), activation="relu", padding="same")(x)
x = layers.MaxPooling2D((2, 2), padding="same")(x)
 
# Decoder
x = layers.Conv2DTranspose(32, (3, 3), strides=2, activation="relu", padding="same")(x)
x = layers.Conv2DTranspose(32, (3, 3), strides=2, activation="relu", padding="same")(x)
x = layers.Conv2D(1, (3, 3), activation="sigmoid", padding="same")(x)
 
# Autoencoder
autoencoder = Model(input, x)
autoencoder.compile(optimizer="adam", loss="binary_crossentropy")
autoencoder.summary()
 
autoencoder.fit(
    x=train_data,
    y=train_data,
    epochs=50,
    batch_size=128,
    shuffle=True,
    validation_data=(test_data, test_data),
)
 
predictions = autoencoder.predict(test_data)
display(test_data, predictions)
 
autoencoder.fit(
    x=noisy_train_data,
    y=train_data,
    epochs=100,
    batch_size=128,
    shuffle=True,
    validation_data=(noisy_test_data, test_data),
)
 
predictions = autoencoder.predict(noisy_test_data)
display(noisy_test_data, predictions)
 

import torch
from torchvision import datasets
from torchvision import transforms
import matplotlib.pyplot as plt
import numpy as np
import torch.nn as nn
 
tensor_transform = transforms.ToTensor()
 
dataset = datasets.MNIST(root = "./data",
                         train = True,
                         download = True,
                         transform = tensor_transform)
 
train_loader = torch.utils.data.DataLoader(dataset = dataset,
                                     batch_size = 100,
                                     shuffle = True)
dataset2 = datasets.MNIST(root = "./data",
                         train = False,
                         download = True,
                         transform = tensor_transform)
test_loader = torch.utils.data.DataLoader(dataset = dataset2,
                                     batch_size = 100,
                                     shuffle = True)
 
dataiter = iter(train_loader)
images,labels = dataiter.next()
print(torch.min(images),torch.max(images))
 
class Autoencoder(nn.Module):
  def __init__(self):
    super().__init__()
    self.encoder= nn.Sequential(
        nn.Conv2d(1,16,3,stride=2,padding=1), #[(inputsize+2*padding-filter_size)/stride] + 1
        nn.ReLU(),
        nn.Conv2d(16,32,3,stride=2,padding=1),
        nn.ReLU(),
        nn.Conv2d(32,64,5),
        nn.ReLU()        
    )
    self.decoder=nn.Sequential(
        nn.ConvTranspose2d(64,32,5),
        nn.ReLU(),
        nn.ConvTranspose2d(32,16,3,stride=2,padding=1,output_padding=1),
        nn.ReLU(),
        nn.ConvTranspose2d(16,1,3,stride=2,padding=1,output_padding=1), #(inputsize-1)*stride + kernal_size + output_padding - 2*padding
        nn.Sigmoid()
 
    )
  def forward(self,x):
    encoded = self.encoder(x)
    decoded = self.decoder(encoded)
    return decoded
 
model = Autoencoder()
loss_function = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(),lr=0.001)
 
def add_noise(img):
  noise_factor = 0.5
  noise_img = img + torch.randn_like(img)*noise_factor
  noise_img = torch.clip(noise_img,0.,1.)
  return noise_img
 
losses = []
l = len(train_loader)
running_loss =0 
for epoch in range(5):
  for (img,_) in train_loader:
    noisy_img = add_noise(img)
    reconstruction = model(noisy_img)
    loss = loss_function(reconstruction,img)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
    running_loss += loss.item()
  losses.append(running_loss/l)
  print(f"Epoch : {epoch+1}, loss : {losses[epoch]} ")
  running_loss=0
plt.ylabel("Loss")
plt.xlabel("Epoch")
plt.plot(losses)
 
outputs = {}
img, _ = list(test_loader)[-3] 
out = model(img)
outputs["original_img"] = img
outputs['img'] = add_noise(img)
outputs['out'] = out
 
counter = 1
print("Original Images")
for j in range(6):
  val= outputs['original_img']
  plt.subplot(1,6,counter)
  plt.imshow(val[j].reshape(28,28),cmap='gray')
  counter += 1
plt.show()
print("Noisy Images")
for i in range(6):
  val = outputs['img']
  plt.subplot(1, 6, i+1)
 
 
  plt.imshow(val[i].reshape(28, 28), cmap='gray')
  counter += 1
plt.show()
val = outputs['out'].detach().numpy()
print("Reconstructed Images")
for i in range(6):
  plt.subplot(1, 6, i+1)
  plt.imshow(val[i].reshape(28, 28), cmap='gray')
  counter += 1
plt.show()
 

import torch
from torchvision import datasets
from torchvision import transforms
import matplotlib.pyplot as plt
import numpy as np
import torch.nn as nn
import torchvision
 
transform = transforms.ToTensor()
 
dataset = datasets.FashionMNIST(root = "./data",
                         train = True,
                         download = True,
                         transform = transform)
 
train_loader = torch.utils.data.DataLoader(dataset = dataset,
                                     batch_size = 100,
                                     shuffle = True)
dataset2 = datasets.FashionMNIST(root = "./data",
                         train = False,
                         download = True,
                         transform = transform)
test_loader = torch.utils.data.DataLoader(dataset = dataset2,
                                     batch_size = 100,
                                     shuffle = True)
 
def imshow(img):
  img = img/2 + 0.5
  npimg = img.numpy()
  plt.imshow(np.transpose(npimg,(1,2,0)))
  plt.show()
dataiter = iter(train_loader)
images,labels = dataiter.next()
 
imshow(torchvision.utils.make_grid(images))
 
dataiter = iter(train_loader)
images,labels = dataiter.next()
print(torch.min(images),torch.max(images))
 
class Autoencoder(nn.Module):
  def __init__(self):
    super().__init__()
    self.encoder= nn.Sequential(
        nn.Conv2d(1,16,3,stride=2,padding=1),
        nn.ReLU(),
        nn.Conv2d(16,32,3,stride=2,padding=1),
        nn.ReLU(),
        nn.Conv2d(32,64,5),
        nn.ReLU()        
    )
    self.decoder=nn.Sequential(
        nn.ConvTranspose2d(64,32,5),
        nn.ReLU(),
        nn.ConvTranspose2d(32,16,3,stride=2,padding=1,output_padding=1),
        nn.ReLU(),
        nn.ConvTranspose2d(16,1,3,stride=2,padding=1,output_padding=1),
        nn.Sigmoid()
 
    )
  def forward(self,x):
    encoded = self.encoder(x)
    decoded = self.decoder(encoded)
    return decoded
 
model = Autoencoder()
loss_function = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(),lr=0.001,weight_decay=1e-5)
 
def add_noise(img):
  noise_factor = 0.3
  noise_img = img + torch.randn_like(img)*noise_factor
  noise_img = torch.clip(noise_img,0.,1.)
  return noise_img
 
running_loss = 0
losses = []
l = len(train_loader)
for epoch in range(5):
  for (img,_) in train_loader:
    noisy_img = add_noise(img)
    reconstruction = model(noisy_img)
    loss = loss_function(reconstruction,img)
    optimizer.zero_grad()
    loss.backward()
    running_loss += loss.item()
    optimizer.step()
  losses.append(running_loss/l)
  print(f"Epoch : {epoch+1}, loss : {losses[epoch]:.5f} ")
  running_loss = 0
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.plot(losses)
 
outputs = {}
img, _ = list(test_loader)[-3] 
out = model(img)
plt.figure(figsize=(14, 4))
outputs['img'] = add_noise(img)
outputs['out'] = out
outputs['Original_img'] = img
for i in range(6):
	val = outputs['Original_img']
	plt.subplot(1,6,i+1)
	plt.title("Original")
	plt.imshow(val[i].reshape(28,28),cmap='gray')
plt.show()
plt.figure(figsize=(14, 4))
counter = 1
for i in range(6):
	val = outputs['img']
	plt.subplot(1, 6, i+1)
	plt.title("Noisy")
	plt.imshow(val[i].reshape(28, 28), cmap='gray')
	counter += 1
plt.show()
plt.figure(figsize=(14, 4))
val = outputs['out'].detach().numpy()
for i in range(6):
	plt.subplot(1, 6, i+1)
	plt.title("Reconstructed")
	plt.imshow(val[i].reshape(28, 28), cmap='gray')
	# plt.figure(figsize=(18, 5))
	counter += 1
 
plt.show()