【29】遇到不平衡资料(Imbalanced Data) 时 使用 SMOTE 解决实验

Colab连结

今天要介绍处理不平衡资料的方法叫 SMOTE (Synthetic Minority Oversampling Technique)，其原理就是透过将少数样本以合成的方式来增加，让资料集变得相对平衡，而这边的合成方式是什麽？并没有一定的方式，以一般资料数值分析为例，它从样本的邻近区间生成一个近似值来当新的少数样本。

但是我们这次要介绍的都是图片相关的分类任务，图片的话要如何合成样本呢？没错！就是使用 GAN 啦！所以今天的实验分成两阶段。

1. 先用 GAN 透过不平衡训练集训练生成器。
2. 利用上述的生成器产生不平衡的少数样本(数字6,8,9)，再进行一次分类训练。

本次文章因为主要想实验利用 GAN 产生的样本能够对训练任务带来多少效益，所以不会着重在 GAN 的讲解上，我们会使用 Conditional GAN ，这种可以指定标签的生成器作为示范，其 Conditional GAN 的程序码我会从 Keras 官方 Conditional GAN 教学文件修改来使用，比较不同的地方是 Keras 范例是使用完整的训练集和测试及产生的，我则是使用6,8,9个只有100笔的不平衡资料来模拟真实情况。

实验一：用 cGAN (Conditional GAN) 直接跑不平衡资料(6,8,9的样本只有100笔)

dataset = tf.data.Dataset.from_tensor_slices((train_images_imbalanced, train_label_imbalanced))

def normalize_img(image, label):
  return tf.cast(image, tf.float32) / 255., tf.one_hot(label, num_classes)

dataset = dataset.map(
    normalize_img, num_parallel_calls=tf.data.experimental.AUTOTUNE)
dataset = dataset.cache()
dataset = dataset.shuffle(1000)
dataset = dataset.batch(128)
dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)

cond_gan = ConditionalGAN(
    discriminator=discriminator, generator=generator, latent_dim=latent_dim
)
cond_gan.compile(
    d_optimizer=tf.keras.optimizers.Adam(learning_rate=0.0003),
    g_optimizer=tf.keras.optimizers.Adam(learning_rate=0.0003),
    loss_fn=tf.keras.losses.BinaryCrossentropy(from_logits=True),
)

cond_gan.fit(dataset, epochs=20)

产出：

Epoch 30/30
g_loss: 0.7804 - d_loss: 0.6826

和 Keras 官网的 loss 比较，因为我们用不平衡资料，导致生成器(g_loss)偏高。

values = [6, 8, 9]
n_values = np.max(values) + 1
one_hot = np.eye(n_values)[values]

for v in one_hot:
    fake = np.random.rand(128)
    label = np.asarray(v, dtype=np.float32)
    fake = np.concatenate((fake,label))
    fake = cond_gan.generator.predict(np.expand_dims(fake, axis=0))

    fake *= 255.0
    converted_images = fake.astype(np.uint8)
    converted_images = tf.image.resize(converted_images, (28, 28)).numpy().astype(np.uint8).squeeze()
    plt.imshow(converted_images)
    plt.show()

生成器产生的1,2,3:

生成器产生的6,8,9:

将生成器的6,8,9印出来後，长得很歪...看来我故意挑6,8,9这三组数字是真有搞到模型ＸＤ。

虽然这个生成器产生的样本有些奇怪，但我们仍使用这些长歪的6,8,9来做分类任务，我们产生4900个合成样本加上原本100个本来的样本来 Oversampling。

DUP_COUNT=4900

idx_we_want = list(range(sum(counts[:6]))) + list(range(sum(counts[:7]) ,sum(counts[:7])+counts[7])) # [0,5] + [7,7]
train_label_imbalanced = train_labels_sorted[idx_we_want]
train_images_imbalanced = train_images_sorted[idx_we_want]

idx_we_want = list(range(sum(counts[:6]),sum(counts[:6])+100)) + list(range(sum(counts[:8]),sum(counts[:8])+100)) + list(range(sum(counts[:9]),sum(counts[:9])+100))
train_label_689 = train_labels_sorted[idx_we_want]
train_images_689 = train_images_sorted[idx_we_want]

train_label_689_dup = np.asarray([6,8,9]).repeat(DUP_COUNT)

values = [6, 8, 9]
n_values = np.max(values) + 1
one_hot = np.eye(n_values)[values]
train_images_689_dup = np.zeros((DUP_COUNT*3,28,28,1))
for bucket, v in enumerate(one_hot):
    for idx in range(DUP_COUNT):
        fake = np.random.rand(128)
        label = np.asarray(v, dtype=np.float32)
        fake = np.concatenate((fake,label))
        fake = cond_gan.generator.predict(np.expand_dims(fake, axis=0))

        fake *= 255.0
        fake = fake.astype(np.uint8)
        fake = tf.image.resize(fake, (28, 28)).numpy().astype(np.uint8).squeeze()
        train_images_689_dup[bucket*DUP_COUNT+idx,:,:,0] = fake

train_label_imbalanced = np.concatenate((train_label_imbalanced, train_label_689, train_label_689_dup))
train_images_imbalanced = np.concatenate((train_images_imbalanced, train_images_689, train_images_689_dup), axis=0)

train_images_imbalanced, train_label_imbalanced = shuffle(train_images_imbalanced, train_label_imbalanced)

确认一下各个样本数量是否正确

{0: 5923,
 1: 6742,
 2: 5958,
 3: 6131,
 4: 5842,
 5: 5421,
 6: 5000,
 7: 6265,
 8: 5000,
 9: 5000}

训练模型。

model = tf.keras.Sequential()
model.add(tf.keras.layers.Conv2D(32, [3, 3], activation='relu', input_shape=(28,28,1)))
model.add(tf.keras.layers.Conv2D(64, [3, 3], activation='relu'))
model.add(tf.keras.layers.MaxPooling2D(pool_size=(2, 2)))
model.add(tf.keras.layers.Dropout(0.25))
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(128, activation='relu'))
model.add(tf.keras.layers.Dropout(0.5))
model.add(tf.keras.layers.Dense(10))

model.compile(
    optimizer=tf.keras.optimizers.SGD(LR),
    loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.SparseCategoricalAccuracy()],
)

history = model.fit(
    ds_train_im,
    epochs=EPOCHS,
    validation_data=ds_test,
)

产出：

Epoch 55/60
loss: 0.0065 - sparse_categorical_accuracy: 0.9977 - val_loss: 0.8658 - val_sparse_categorical_accuracy: 0.8487

实验二：用 cGAN (Conditional GAN) 跑 Oversampling 後的不平衡资料

有鉴於实验一生成器效果不彰，所以这次实验我尝试将少量样本透过 Oversampling 的方式放大50倍，再拿去做 GAN 来生成样本。

idx_we_want = []
for idx in [0,1,2,3,4,5,7]: 
  idx_we_want += list(range(sum(counts[:idx]) ,sum(counts[:idx])+5000))
train_label_imbalanced = train_labels_sorted[idx_we_want]
train_images_imbalanced = train_images_sorted[idx_we_want]

idx_we_want = list(range(sum(counts[:6]),sum(counts[:6])+100)) + list(range(sum(counts[:8]),sum(counts[:8])+100)) + list(range(sum(counts[:9]),sum(counts[:9])+100))
train_label_689 = train_labels_sorted[idx_we_want]
train_images_689 = train_images_sorted[idx_we_want]

train_label_689 = train_label_689.repeat(50)
train_images_689 = train_images_689.repeat(50, axis=0)

train_label_imbalanced = np.concatenate((train_label_imbalanced, train_label_689))
train_images_imbalanced = np.concatenate((train_images_imbalanced, train_images_689), axis=0)

train_images_imbalanced, train_label_imbalanced = shuffle(train_images_imbalanced, train_label_imbalanced)

资料分布，[0,1,2,3,4,5,7]拿前面的5000张，而[6,8,9]则是用前100张 repeat 50次，来产生等量的5000张。

{0: 5000,
 1: 5000,
 2: 5000,
 3: 5000,
 4: 5000,
 5: 5000,
 6: 5000,
 7: 5000,
 8: 5000,
 9: 5000}

GAN 训练结果：

Epoch 40/40
g_loss: 0.8425 - d_loss: 0.6364

生成器产生的1,2,3:

生成器产生的6,8,9:

看起来6,8,9比实验一更完整，我们一样将此产出4900个合成样本来训练模型。

生成：

Epoch 57/60
loss: 0.0060 - sparse_categorical_accuracy: 0.9980 - val_loss: 0.8995 - val_sparse_categorical_accuracy: 0.8453

准确度84.5%，可惜并没有比实验一来的高。

以上就是这次使用 GAN 来产生合成样本的实验。