Few-Shot 学习与 Reptile

import os

os.environ["KERAS_BACKEND"] = "tensorflow"

import keras
from keras import layers

import matplotlib.pyplot as plt
import numpy as np
import random
import tensorflow as tf
import tensorflow_datasets as tfds

learning_rate = 0.003
meta_step_size = 0.25

inner_batch_size = 25
eval_batch_size = 25

meta_iters = 2000
eval_iters = 5
inner_iters = 4

eval_interval = 1
train_shots = 20
shots = 5
classes = 5

class Dataset:
    # 这个类将便于从Omniglot数据集中创建一个少样本数据集
    # 可以快速抽样，同时也允许同时创建新标签。
    def __init__(self, training):
        # 下载包含omniglot数据的tfrecord文件并转换为
        # 数据集。
        split = "train" if training else "test"
        ds = tfds.load("omniglot", split=split, as_supervised=True, shuffle_files=False)
        # 遍历数据集以获取每个单独的图像及其类别，
        # 并将数据放入字典中。
        self.data = {}

        def extraction(image, label):
            # 此函数将缩小Omniglot图像到所需大小，
            # 缩放像素值并将RGB图像转换为灰度图像
            image = tf.image.convert_image_dtype(image, tf.float32)
            image = tf.image.rgb_to_grayscale(image)
            image = tf.image.resize(image, [28, 28])
            return image, label

        for image, label in ds.map(extraction):
            image = image.numpy()
            label = str(label.numpy())
            if label not in self.data:
                self.data[label] = []
            self.data[label].append(image)
        self.labels = list(self.data.keys())

    def get_mini_dataset(
        self, batch_size, repetitions, shots, num_classes, split=False
    ):
        temp_labels = np.zeros(shape=(num_classes * shots))
        temp_images = np.zeros(shape=(num_classes * shots, 28, 28, 1))
        if split:
            test_labels = np.zeros(shape=(num_classes))
            test_images = np.zeros(shape=(num_classes, 28, 28, 1))

        # 从整个标签集中获取一随机子集的标签。
        label_subset = random.choices(self.labels, k=num_classes)
        for class_idx, class_obj in enumerate(label_subset):
            # 使用枚举索引值作为少样本学习中
            # 小批量的临时标签。
            temp_labels[class_idx * shots : (class_idx + 1) * shots] = class_idx
            # 如果创建用于测试的拆分数据集，从每个标签选择一个额外样本
            # 以创建测试数据集。
            if split:
                test_labels[class_idx] = class_idx
                images_to_split = random.choices(
                    self.data[label_subset[class_idx]], k=shots + 1
                )
                test_images[class_idx] = images_to_split[-1]
                temp_images[
                    class_idx * shots : (class_idx + 1) * shots
                ] = images_to_split[:-1]
            else:
                # 对于随机选择的label_subset中的每个索引，抽样
                # 必要数量的图像。
                temp_images[
                    class_idx * shots : (class_idx + 1) * shots
                ] = random.choices(self.data[label_subset[class_idx]], k=shots)

        dataset = tf.data.Dataset.from_tensor_slices(
            (temp_images.astype(np.float32), temp_labels.astype(np.int32))
        )
        dataset = dataset.shuffle(100).batch(batch_size).repeat(repetitions)
        if split:
            return dataset, test_images, test_labels
        return dataset


import urllib3

urllib3.disable_warnings()  # 禁用在下载过程中可能发生的SSL警告。
train_dataset = Dataset(training=True)
test_dataset = Dataset(training=False)

_, axarr = plt.subplots(nrows=5, ncols=5, figsize=(20, 20))

sample_keys = list(train_dataset.data.keys())

for a in range(5):
    for b in range(5):
        temp_image = train_dataset.data[sample_keys[a]][b]
        temp_image = np.stack((temp_image[:, :, 0],) * 3, axis=2)
        temp_image *= 255
        temp_image = np.clip(temp_image, 0, 255).astype("uint8")
        if b == 2:
            axarr[a, b].set_title("类别 : " + sample_keys[a])
        axarr[a, b].imshow(temp_image, cmap="gray")
        axarr[a, b].xaxis.set_visible(False)
        axarr[a, b].yaxis.set_visible(False)
plt.show()

def conv_bn(x):
    x = layers.Conv2D(filters=64, kernel_size=3, strides=2, padding="same")(x)
    x = layers.BatchNormalization()(x)
    return layers.ReLU()(x)


inputs = layers.Input(shape=(28, 28, 1))
x = conv_bn(inputs)
x = conv_bn(x)
x = conv_bn(x)
x = conv_bn(x)
x = layers.Flatten()(x)
outputs = layers.Dense(classes, activation="softmax")(x)
model = keras.Model(inputs=inputs, outputs=outputs)
model.compile()
optimizer = keras.optimizers.SGD(learning_rate=learning_rate)

training = []
testing = []
for meta_iter in range(meta_iters):
    frac_done = meta_iter / meta_iters
    cur_meta_step_size = (1 - frac_done) * meta_step_size
    # 暂时保存模型的权重。
    old_vars = model.get_weights()
    # 从完整数据集中获取一个样本。
    mini_dataset = train_dataset.get_mini_dataset(
        inner_batch_size, inner_iters, train_shots, classes
    )
    for images, labels in mini_dataset:
        with tf.GradientTape() as tape:
            preds = model(images)
            loss = keras.losses.sparse_categorical_crossentropy(labels, preds)
        grads = tape.gradient(loss, model.trainable_weights)
        optimizer.apply_gradients(zip(grads, model.trainable_weights))
    new_vars = model.get_weights()
    # 对元步骤执行SGD。
    for var in range(len(new_vars)):
        new_vars[var] = old_vars[var] + (
            (new_vars[var] - old_vars[var]) * cur_meta_step_size
        )
    # 在元学习步骤后，将新训练的权重重新加载到模型中。
    model.set_weights(new_vars)
    # 评估循环
    if meta_iter % eval_interval == 0:
        accuracies = []
        for dataset in (train_dataset, test_dataset):
            # 从完整数据集中抽取一个小数据集。
            train_set, test_images, test_labels = dataset.get_mini_dataset(
                eval_batch_size, eval_iters, shots, classes, split=True
            )
            old_vars = model.get_weights()
            # 在样本上训练并获取结果准确率。
            for images, labels in train_set:
                with tf.GradientTape() as tape:
                    preds = model(images)
                    loss = keras.losses.sparse_categorical_crossentropy(labels, preds)
                grads = tape.gradient(loss, model.trainable_weights)
                optimizer.apply_gradients(zip(grads, model.trainable_weights))
            test_preds = model.predict(test_images)
            test_preds = tf.argmax(test_preds).numpy()
            num_correct = (test_preds == test_labels).sum()
            # 获取评估准确率后重置权重。
            model.set_weights(old_vars)
            accuracies.append(num_correct / classes)
        training.append(accuracies[0])
        testing.append(accuracies[1])
        if meta_iter % 100 == 0:
            print(
                "batch %d: train=%f test=%f" % (meta_iter, accuracies[0], accuracies[1])
            )

# 首先，对训练和测试数组进行一些预处理，以便于展示。
window_length = 100
train_s = np.r_[
    training[window_length - 1 : 0 : -1],
    training,
    training[-1:-window_length:-1],
]
test_s = np.r_[
    testing[window_length - 1 : 0 : -1], testing, testing[-1:-window_length:-1]
]
w = np.hamming(window_length)
train_y = np.convolve(w / w.sum(), train_s, mode="valid")
test_y = np.convolve(w / w.sum(), test_s, mode="valid")

# 显示训练准确性。
x = np.arange(0, len(test_y), 1)
plt.plot(x, test_y, x, train_y)
plt.legend(["test", "train"])
plt.grid()

train_set, test_images, test_labels = dataset.get_mini_dataset(
    eval_batch_size, eval_iters, shots, classes, split=True
)
for images, labels in train_set:
    with tf.GradientTape() as tape:
        preds = model(images)
        loss = keras.losses.sparse_categorical_crossentropy(labels, preds)
    grads = tape.gradient(loss, model.trainable_weights)
    optimizer.apply_gradients(zip(grads, model.trainable_weights))
test_preds = model.predict(test_images)
test_preds = tf.argmax(test_preds).numpy()

_, axarr = plt.subplots(nrows=1, ncols=5, figsize=(20, 20))

sample_keys = list(train_dataset.data.keys())

for i, ax in zip(range(5), axarr):
    temp_image = np.stack((test_images[i, :, :, 0],) * 3, axis=2)
    temp_image *= 255
    temp_image = np.clip(temp_image, 0, 255).astype("uint8")
    ax.set_title(
        "标签 : {}, 预测 : {}".format(int(test_labels[i]), test_preds[i])
    )
    ax.imshow(temp_image, cmap="gray")
    ax.xaxis.set_visible(False)
    ax.yaxis.set_visible(False)
plt.show()