我一直在尝试在高光谱数据上训练 GAN。数据 (Indian Pines) 的大小为 145x145x200，在执行 PCA (K=3) 后已分解为 64x64x3 的块。生成器在接收到 100x1x1 潜在暗淡后生成此大小的图像，而鉴别器接收真实类标签并生成图像或真实图像块以提供两个单独的输出-
注意-数据集中有 16 个类别，我为生成的图像分配了第 17 个类别。
使用的数据预处理实用程序：
def applyPCA(X, numComponents=75):
newX = np.reshape(X, (-1, X.shape[2]))
pca = PCA(n_components=numComponents, whiten=True)
newX = pca.fit_transform(newX)
newX = np.reshape(newX, (X.shape[0],X.shape[1], numComponents))
return newX, pca
def padWithZeros(X, margin=2):
newX = np.zeros((X.shape[0] + 2 * margin, X.shape[1] + 2* margin, X.shape[2]))
x_offset = margin
y_offset = margin
newX[x_offset:X.shape[0] + x_offset, y_offset:X.shape[1] + y_offset, :] = X
return newX
def createImageCubes(X, y, windowSize=5, removeZeroLabels = True):
margin = int((windowSize) / 2)
zeroPaddedX = padWithZeros(X, margin=margin)
# split patches
patchesData = np.zeros((X.shape[0] * X.shape[1], windowSize, windowSize, X.shape[2]))
patchesLabels = np.zeros((X.shape[0] * X.shape[1]))
patchIndex = 0
for r in range(margin, zeroPaddedX.shape[0] - margin):
for c in range(margin, zeroPaddedX.shape[1] - margin):
patch = zeroPaddedX[r - margin:r + margin, c - margin:c + margin]
patchesData[patchIndex, :, :, :] = patch
patchesLabels[patchIndex] = y[r-margin, c-margin]
patchIndex = patchIndex + 1
if removeZeroLabels:
patchesData = patchesData[patchesLabels>0,:,:,:]
patchesLabels = patchesLabels[patchesLabels>0]
patchesLabels -= 1
return patchesData, patchesLabels
def splitTrainTestSet(X, y, testRatio, randomState=345):
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=testRatio, random_state=randomState, stratify=y)
return X_train, X_test, y_train, y_test
模型架构和训练循环：
def build_generator(latent_dim=(100, 1, 1)):
model = Sequential()
model.add(Dense(2 * 13 * 13, input_shape=latent_dim))
model.add(Reshape((13, 13, 200)))
# model.add(Dropout(0.2))
model.add(Conv2DTranspose(512, kernel_size=4, strides=1, padding='valid'))
model.add(BatchNormalization())
model.add(ReLU())
model.add(Dropout(0.2))
model.add(Conv2DTranspose(256, kernel_size=4, strides=2, padding='same'))
model.add(BatchNormalization())
model.add(ReLU())
model.add(Dropout(0.2))
model.add(Conv2DTranspose(128, kernel_size=4, strides=2, padding='same'))
model.add(BatchNormalization())
model.add(ReLU())
model.add(Dropout(0.2))
model.add(Conv2DTranspose(64, kernel_size=4, strides=2, padding='same'))
model.add(BatchNormalization())
model.add(ReLU())
model.add(Dropout(0.2))
model.add(Conv2D(3, kernel_size=4, strides=2, padding='same', activation='tanh'))
return model
generator = build_generator()
generator.compile(
loss='binary_crossentropy', optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5)
)
generator.summary()
from keras import backend as be
def build_discriminator_net(img_shape=(64,64,3), num_classes=17):
img_input = Input(shape=img_shape)
label_input = Input(shape=(1,), dtype='int32')
label_embedding = Flatten()(Embedding(num_classes+1, np.prod(img_shape))(label_input))
label_embedding = Reshape(img_shape)(label_embedding)
x = Concatenate()([img_input, label_embedding])
x = Conv2D(64, kernel_size=4, strides=2, padding='same')(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Dropout(0.5)(x)
x = Conv2D(128, kernel_size=4, strides=2, padding='same')(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Dropout(0.5)(x)
x = Conv2D(256, kernel_size=4, strides=2, padding='same')(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Dropout(0.5)(x)
x = Conv2D(512, kernel_size=4, strides=2, padding='same')(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Dropout(0.5)(x)
x = Conv2D(64, kernel_size=4, strides=1, padding='valid')(x)
x = Flatten()(x)
real_or_fake = Dense(2, activation='sigmoid', name='real_or_fake')(x)
class_label = Dense(num_classes, activation='softmax', name='class_label')(x)
model = Model(inputs=[img_input, label_input], outputs=[real_or_fake, class_label])
return model
discriminator = build_discriminator_net()
discriminator.compile(
loss=['binary_crossentropy', 'categorical_crossentropy'],
optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5),
metrics=["accuracy"],
)
discriminator.summary()
gan_input_noise = keras.Input(shape=latent_dim)
gan_input_label = keras.Input(shape=(1,))
gan_output = discriminator([generator(gan_input_noise), gan_input_label])
gan = keras.Model(inputs=[gan_input_noise, gan_input_label], outputs=gan_output, name="gan")
gan.compile(
loss=['binary_crossentropy', 'categorical_crossentropy'],
optimizer=keras.optimizers.Adam(learning_rate=0.0002, beta_1=0.5),
)
epochs = 500
batch_size = 64
def preprocess_images(images):
images = (images - 127.5) / 127.5
return images
x_train = preprocess_images(Xtrain)
y_train = to_categorical(ytrain, num_classes+1)
def batch(batch_size=batch_size, x_train=x_train, y_train=y_train):
idx = np.random.randint(0, Xtrain.shape[0], batch_size)
imgs = x_train[idx]
labels = y_train[idx]
return imgs, labels
for epoch in range(epochs):
real_images, real_labels = batch()
noise = tf.random.normal([batch_size, 100, 1, 1])
fake_images = generator.predict(noise)
fake_labels = np.ones((batch_size, 17)) * 17
y_real = np.ones((batch_size, 1), dtype=np.int32)
y_fake = np.zeros((batch_size, 1), dtype=np.int32)
d_loss_real = discriminator.train_on_batch([real_images, real_labels], [y_real, real_labels])
d_loss_fake = discriminator.train_on_batch([fake_images, fake_labels], [y_fake, fake_labels])
d_loss = d_loss_real + d_loss_fake
noise = tf.random.normal([batch_size, 100, 1, 1])
gen_labels = np.ones((batch_size, 1)) * 17
g_loss = gan.train_on_batch(noise, gen_labels)
if epoch % 100 == 0:
print(f"Epoch {epoch+1}/{epochs}, Discriminator Loss: {d_loss}, Generator Loss: {g_loss}")