尝试训练 GAN 时出现空白图像

Question

我在 python 中有一个完全可用的自定义 GAN 脚本。 它运行没有任何错误，但它给出的图像只是灰色/空白，我不知道错误在哪里，如果我训练它错误或者我以错误的方式显示图像。 我也在不同的应用程序上运行它，但它没有工作。

训练数据为：1463张图片，全是RGB，全是.jpeg，全是256X256p。

来自 0、2、4、6、8 时期的图像

10 个 epoch 后生成的图像

完整代码：

import tensorflow as tf
import os
import pathlib
import numpy as np

tf.random.set_seed(1)

print(len(tf.config.list_physical_devices("GPU")))

config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
session = tf.compat.v1.Session(config=config)

data_dir = "C:/Users/m8/Desktop/test_fro_tensorflow/train/training/"

dataset = tf.io.read_file("C:/Users/m8/Desktop/test_fro_tensorflow/val/validation/img (1).jpg")
tensor = tf.io.decode_image(dataset, channels=3, dtype=tf.dtypes.float32)
tensor = tf.image.resize(tensor, [128, 128])
dataset = tf.expand_dims(tensor, axis=0)

for file in os.listdir(data_dir):
    f = os.path.join(data_dir, file)
    full_path = data_dir + file
    img = tf.io.read_file(full_path)
    tensor = tf.io.decode_image(img, channels=3, dtype=tf.dtypes.float32)
    tensor = tf.image.resize(tensor, [128, 128])
    new_tensor = tf.expand_dims(tensor, axis=0)
    dataset = np.concatenate([dataset, new_tensor])

print(dataset.shape)

dataset = np.expand_dims(dataset, -1).astype("float32") / 255

BATCH_SIZE = 16

dataset = np.reshape(dataset, (-1, 128, 128, 3))
dataset = tf.data.Dataset.from_tensor_slices(dataset)
dataset = dataset.shuffle(buffer_size=1024).batch(BATCH_SIZE)

from tensorflow import keras
from tensorflow.keras import layers

NOISE_DIM = 150


generator = keras.models.Sequential([
    keras.layers.InputLayer(input_shape=(NOISE_DIM,)),
    layers.Dense(32*32*256),
    layers.Reshape(target_shape=(32, 32, 256)),
    layers.Conv2DTranspose(256, 3, activation="LeakyReLU", strides=2, padding="same"),
    layers.Conv2DTranspose(128, 3, activation="LeakyReLU", strides=2, padding="same"),
    layers.Conv2DTranspose(3, 3, activation="sigmoid", padding="same"),
])

generator.summary()

discriminator = keras.models.Sequential([
    keras.layers.InputLayer(input_shape=(128, 128, 3)),
    layers.Conv2D(256, 3, activation="relu", strides=2, padding="same"),
    layers.Conv2D(128, 3, activation="relu", strides=2, padding="same"),
    layers.Dense(64, activation="relu"),
    layers.Flatten(),
    layers.Dropout(0.2),
    layers.Dense(1, activation="sigmoid")
])

discriminator.summary()

optimizerG = keras.optimizers.Adam(learning_rate=0.001, beta_1=0.5)
optimizerD = keras.optimizers.Adam(learning_rate=0.003, beta_1=0.5)

lossFn = keras.losses.BinaryCrossentropy(from_logits=True)

gAccMetric = tf.keras.metrics.BinaryAccuracy()
dAccMetric = tf.keras.metrics.BinaryAccuracy()


@tf.function
def trainDStep(data):
    batchSize = tf.shape(data)[0]

    noise = tf.random.normal(shape=(batchSize, NOISE_DIM))
    y_true = tf.concat(
        [
            tf.ones(batchSize, 1),

            tf.zeros(batchSize, 1)
        ],
        axis=0
    )

    with tf.GradientTape() as tape:
        fake = generator(noise)
        x = tf.concat([data, fake], axis=0)
        y_pred = discriminator(x)
        discriminatorLoss = lossFn(y_true, y_pred)

    grads = tape.gradient(discriminatorLoss, discriminator.trainable_weights)
    optimizerD.apply_gradients(zip(grads, discriminator.trainable_weights))

    dAccMetric.update_state(y_true, y_pred)

    return {
        "discriminator_loss": discriminatorLoss,
        "discriminator_accuracy": dAccMetric.result()
    }

@tf.function
def trainGStep(data):
    batchSize = tf.shape(data)[0]
    noise = tf.random.normal(shape=(batchSize, NOISE_DIM))
    y_true = tf.ones(batchSize, 1)

    with tf.GradientTape() as tape:
        y_pred = discriminator(generator(noise))
        generatorLoss = lossFn(y_true, y_pred)

    grads = tape.gradient(generatorLoss, generator.trainable_weights)
    optimizerG.apply_gradients(zip(grads, generator.trainable_weights))

    gAccMetric.update_state(y_true, y_pred)

    return {
        "generator_loss": generatorLoss,
        "generator_accuracy": gAccMetric.result()
    }

from matplotlib import pyplot as plt

def plotImages(model):
    images = model(np.random.normal(size=(4, NOISE_DIM)))

    plt.figure(figsize=(2, 2))

    for i, image in enumerate(images):
        plt.subplot(2,2,i+1)
        plt.imshow(image)
        plt.axis("off")

    plt.show();

for epoch in range(10):

    dLossSum = 0
    gLossSum = 0
    dAccSum = 0
    gAccSum = 0
    cnt = 0


    for batch in dataset:

        dLoss = trainDStep(batch)
        dLossSum += dLoss["discriminator_loss"]
        dAccSum += dLoss["discriminator_accuracy"]

        gLoss = trainGStep(batch)
        gLossSum += dLoss["discriminator_loss"]
        gAccSum += dLoss["discriminator_accuracy"]


        cnt += 1


    print("E:{}, Loss G:{:0.4f}, Loss D:{:0.4f}, Acc G:%{:0.2f}, Acc D:%{:0.2f}".format(
        epoch,
        gLossSum/cnt,
        dLossSum/cnt,
        100 * gAccSum/cnt,
        100 * dAccSum/cnt
    ))

    if epoch % 2 == 0:
        plotImages(generator)