我正在Keras训练一个Generative Adversarial Network（GAN）。
我的日志报告两个网络（鉴别器和组合模型）的准确率达到100％。这表明出现了问题。
我尝试运行推理并看到鉴别器确实100％准确，但是生成器只产生噪声，并且根本没有欺骗鉴别器。
我的问题：为什么Keras将我的组合模型的准确度报告为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)
答案 0 :(得分：0)
注意gan = Model(in_vector, gan_output)，因此您的模型被定义为从输入向量到鉴别器输出的所有层，包括中间的生成器。所以当你打电话时
gan.compile(optimizer=gan_optimizer, loss='binary_crossentropy', metrics=['accuracy'])，它会自动使用鉴别器的输出来确定准确性。因此，为了获得发电机精度，您可以使用回调并手动计算“精度”，但是可以为您的发电机定义（当您考虑它时，发电机没有一个典型的精度指标，您打算如何比较它？）。 此外，如果您的发电机产生随机噪声，这并不意味着精度应该为0，并且由于您只使用鉴别器的精度并且它成功识别输出不属于基础分布，因此精度仍为100百分比（这很容易，因为发电机的输出是随机噪声）。简而言之，鉴别器的高精度并不意味着发生器成功地欺骗了鉴别器。事实上，当鉴别器的准确度接近50％时，这意味着发生器确实对输入数据进行了很好的建模，并且鉴别器无法区分两者并且正在进行随机猜测。因此，您所看到的是预期的行为