tensorflow - Keras 报告错误的准确性

Question

我正在 Keras 中训练生成对抗网络 (GAN)。

我的日志报告两个网络(鉴别器和组合模型)都达到了 100% 的准确度。这表明出现了问题。

我尝试运行推理，发现鉴别器确实 100% 准确，但生成器只生成噪声，根本无法欺骗鉴别器。

我的问题:为什么 Keras 报告我的组合模型的准确性为 100%？

代码:

generator = create_generator(input_shape=(374,))
in_vector = Input(shape=(374,))
fake_images = generator(in_vector)

discriminator = create_discriminator()
disc_optimizer = keras.optimizers.SGD(lr=1e-4)
discriminator.compile(optimizer=disc_optimizer, loss='binary_crossentropy', metrics=['accuracy'])

discriminator.trainable = False
for l in discriminator.layers:
    l.trainable = False

gan_output = discriminator(fake_images)
gan = Model(in_vector, gan_output)
gan_optimizer = keras.optimizers.RMSprop(lr=1e-5)
gan.compile(optimizer=gan_optimizer, loss='binary_crossentropy', metrics=['accuracy'])

start_time = datetime.datetime.now()
tensorboard = TensorBoard(log_dir=f'data/logs/gawwn/{start_time}')
tensorboard.set_model(gan)

d_train_logs = ['train_discriminator_loss',
                'train_discriminator_accuracy']
g_train_logs = ['train_generator_loss',
                'train_generator_accuracy']
val_logs = ['val_discriminator_loss',
            'val_discriminator_accuracy',
            'val_generator_loss',
            'val_generator_accuracy']

d_train_step, g_train_step, val_step = 0, 0, 0

valid = np.ones((batch_size, 1))
fake = np.zeros((batch_size, 1))

noise_sigma = 0.00
noise_decay = 0.95

for epoch in range(1, 1 + epochs):
    d_loss = [1]
    while d_loss[0] > d_loss_thres:
        for i, (x_vectors, x_images, y) in enumerate(train_loader.load_batch(batch_size)):
            # ---------------------
            #  Train Discriminator
            # ---------------------

            # Generate a batch of new images
            gen_imgs = generator.predict(x_vectors)

            # Train the discriminator
            data = np.concatenate([y, gen_imgs], axis=0)
            labels = np.concatenate([valid[:len(y)], fake[:len(y)]])
            train_batch = list(zip(data, labels))
            np.random.shuffle(train_batch)
            data, labels = zip(*train_batch)
            data, labels = np.array(data), np.array(labels)
            d_loss = discriminator.train_on_batch(data, labels)

#             d_loss_real = discriminator.train_on_batch(y, valid[:len(y)])
#             d_loss_fake = discriminator.train_on_batch(gen_imgs, fake[:len(y)])
#             d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
            write_log(tensorboard, d_train_logs, d_loss, d_train_step)
            d_train_step += 1
            time_elaped = datetime.datetime.now() - start_time
            print(f'D step {d_train_step}: loss={d_loss[0]}; acc={d_loss[1]}; time={time_elaped}')

    g_loss = [1]
    while g_loss[0] > g_loss_thres:
        for i, (x_vectors, x_images, y) in enumerate(train_loader.load_batch(batch_size)):
            # ---------------------
            #  Train Generator
            # ---------------------

            # Train the generator (to have the discriminator label samples as valid)
            g_loss = gan.train_on_batch(x_vectors, valid[:len(y)])

            # Plot the progress
            write_log(tensorboard, g_train_logs, g_loss, g_train_step)
            g_train_step += 1

            time_elaped = datetime.datetime.now() - start_time
            print(f'G step {g_train_step}: loss={g_loss[0]}; acc={g_loss[1]}; time={time_elaped}')

    # If at save interval => save generated image samples
    if epoch % sample_interval == 0:
        d_losses = []
        g_losses = []
        for x_vectors, x_images, y in val_loader.load_batch(batch_size):
            gen_imgs = generator.predict(x_vectors)
            d_loss_real = discriminator.test_on_batch(y, valid[:len(y)])
            d_loss_fake = discriminator.test_on_batch(gen_imgs, fake[:len(y)])
            d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
            d_losses.append(d_loss)

            g_loss = gan.test_on_batch(x_vectors, valid[:len(y)])
            g_losses.append(g_loss)

        d_loss = np.average(d_losses, axis=0)
        g_loss = np.average(g_losses, axis=0)
        write_log(tensorboard, val_logs, [d_loss[0], d_loss[1], g_loss[0], g_loss[1]], val_step)
        val_step += 1
        sample_images(val_loader, generator, epoch)
        save_model(generator, epoch, 'generator')
        save_model(discriminator, epoch, 'discriminator')

最后几个步骤的结果:

D step 349: loss=0.09932675957679749; acc=1.0; time=0:05:58.468997
D step 350: loss=0.10563915222883224; acc=0.9900000095367432; time=0:05:59.088657
D step 351: loss=0.09658461064100266; acc=1.0; time=0:05:59.533442
G step 214: loss=0.167491614818573; acc=0.9800000190734863; time=0:06:00.196747
G step 215: loss=0.13409791886806488; acc=1.0; time=0:06:00.891946
G step 216: loss=0.1523411124944687; acc=0.9722222089767456; time=0:06:01.402974
D step 352: loss=0.10553492605686188; acc=0.9900000095367432; time=0:06:02.015083
D step 353: loss=0.10318870842456818; acc=0.9900000095367432; time=0:06:02.654599
D step 354: loss=0.07871382683515549; acc=1.0; time=0:06:03.131933
G step 217: loss=0.1493617743253708; acc=0.9800000190734863; time=0:06:03.827815
G step 218: loss=0.12147567421197891; acc=0.9599999785423279; time=0:06:04.537494
G step 219: loss=0.17327196896076202; acc=1.0; time=0:06:05.099841
D step 355: loss=0.10441411286592484; acc=0.9900000095367432; time=0:06:05.768096
D step 356: loss=0.09612423181533813; acc=1.0; time=0:06:06.451947
D step 357: loss=0.1072489321231842; acc=0.9861111044883728; time=0:06:06.937882

推论:

>>> np.reshape(discriminator.predict(ground_truth), (5, 10))

array([[0.5296475 , 0.52787906, 0.5270807 , 0.5260455 , 0.528732  ,
        0.52820367, 0.53157693, 0.52730876, 0.5244186 , 0.52673554],
       [0.5229454 , 0.5239704 , 0.53051734, 0.52862865, 0.52718925,
        0.52680767, 0.52621156, 0.5308223 , 0.52489233, 0.5297055 ],
       [0.53033316, 0.5260847 , 0.5300899 , 0.52788675, 0.529595  ,
        0.52183014, 0.5321261 , 0.5251559 , 0.52876014, 0.52384466],
       [0.528658  , 0.52737784, 0.53003156, 0.52685475, 0.53047454,
        0.52759105, 0.52710444, 0.52546424, 0.52709824, 0.52520245],
       [0.5283209 , 0.52810913, 0.52451426, 0.5196351 , 0.5299184 ,
        0.5274567 , 0.52686375, 0.5269972 , 0.5248108 , 0.5263274 ]],
      dtype=float32)

>>> np.reshape(gan.predict(input_vector), (5, 10))

array([[0.4719111 , 0.47217596, 0.47209665, 0.47233126, 0.4741753 ,
        0.4712048 , 0.4721919 , 0.47193947, 0.47010162, 0.47092766],
       [0.47291884, 0.47334394, 0.4714141 , 0.46976995, 0.47092718,
        0.47233835, 0.47164065, 0.47276756, 0.47107005, 0.47187868],
       [0.47153524, 0.47157907, 0.4706026 , 0.47128928, 0.47320494,
        0.47089615, 0.47108623, 0.47432283, 0.47186196, 0.47404772],
       [0.47164053, 0.47348404, 0.4701542 , 0.4741918 , 0.4702833 ,
        0.47303212, 0.4726331 , 0.47118646, 0.47191456, 0.47318774],
       [0.47043982, 0.47027725, 0.47308347, 0.47376725, 0.4733549 ,
        0.47157207, 0.47205287, 0.47177386, 0.47119975, 0.4707804 ]],
      dtype=float32)

Answer 1

注意gan = Model(in_vector, gan_output)，因此您的模型被定义为从输入向量到判别器输出的所有层，包括中间的生成器。所以当你打电话

gan.compile(optimizer=gan_optimizer, loss='binary_crossentropy',metrics=['accuracy'])，它自动使用鉴别器的输出来确定准确性。因此，为了获得生成器的准确性，您可以使用回调并手动计算“准确性”，但是这可能是为您的生成器定义的(加上生成器没有典型的准确性指标，当您考虑时，您打算做什么比较一下？)。 此外，如果您的生成器产生随机噪声，这并不意味着准确度应该为 0，并且由于您仅使用判别器的准确度并且它成功识别输出不属于基础分布，因此准确度仍为 100百分比(这很容易，因为生成器的输出是随机噪声)。简而言之，判别器的高精度并不意味着生成器成功地欺骗了判别器。事实上，当判别器的准确率接近 50% 时，这意味着生成器确实很好地对输入数据进行了建模，而判别器无法区分两者并进行随机猜测。因此，您所看到的是预期行为