LSTMによる異常検知の練習

%matplotlib inline
import numpy as np
import chainer
import chainer.functions as F
import chainer.links as L
from chainer import training
from chainer.training import extensions
from chainer.utils import type_check
import os
import logging
import copy
from chainer import reporter as reporter_module
import matplotlib.pyplot as plt

class Accuracy(chainer.function.Function):
    def __init__(self, eps=0.1):
        super(Accuracy, self).__init__()
        self.eps = eps

    def forward(self, inputs):
        xp = chainer.cuda.get_array_module(*inputs)
        y, t = inputs
        diff = np.abs(y - t)
        return xp.asarray((diff < self.eps).mean(dtype=y.dtype)),

def accuracy(y, t, eps=0.1):
    return Accuracy(eps=eps)(y, t)

class ForcePredictor(chainer.Chain):
    def __init__(self, n_input, n_units, n_layers=3):
        """次の瞬間の加圧力を推定するネットワーク

        :param int n_input: 入力要素数(電流・電圧・加圧力の3要素ないし
            シリ－ズの場合加圧力が2つなので4要素
        :param int n_units: 隠れ層の要素数
        :param n_layers: 隠れ層の数
        """
        super(ForcePredictor, self).__init__()
        with self.init_scope():
            self.lstm = L.NStepLSTM(n_layers, n_input, n_units,
                                    dropout=0.2)
            self.output_layer = L.Linear(n_units, 1)
        self.loss = None
        self.accuracy = None
        self.hx = None
        self.cx = None

    def clear(self):
        self.loss = None
        self.accuracy = None

    def reset_state(self):
        self.hx = None
        self.cx = None

    def chain(self, x):
        """推定する

        :param x: 過去のデータ
            複数の推定を同時に行うためリストを入力とする。
            リストの各項目は[n_input, t]の配列であり、
            tはリストの項目間で異なっていてもいい。
        :type x: list of :py:class:`chainer.variable.Variable`
        :return: 次の瞬間の推定値
            複数の推定結果を出力するため[n, 1]の配列。
            nは入力した項目数と等しい
        :rtype: :py:class:`chainer.variable.Variable`
        """
        assert isinstance(x, list), print(type(x))
        batch_size = len(x)
        hy, cy, ys = self.lstm(self.hx, self.cx, x)
        assert hy.shape[1] == batch_size, print(hy.shape)
        self.hx = hy
        self.cx = cy
        h = hy[-1]
        y = self.output_layer(h)
        assert y.shape[0] == batch_size, print(y.shape)
        return y

    def __call__(self, x, t):
        y = self.chain(x)
        if not isinstance(t, self.xp.ndarray):
            t = self.xp.asarray(t).reshape(y.shape)
        self.loss = F.mean_squared_error(y, t)
        self.accuracy = accuracy(y, t, 0.01)
        acc05 = accuracy(y, t, 0.05)
        acc10 = accuracy(y, t, 0.10)
        chainer.report({"loss": self.loss,
                        "accuracy": self.accuracy,
                        "acc05": acc05,
                        "acc10": acc10}, self)
        return self.loss

class SequencePredictIterator(chainer.dataset.Iterator):
    def __init__(self, dataset, batch_size, min_length=1,
                 repeat=True):
        """
        :param dataset: 2つ以上の要素を含んだタプルのデータセット、
            タプル内の最初の要素が予測対象
        :type dataset: :py.class:`chainer.datasets.TupleDataset`
        :param int batch_size: ミニバッチのサイズ
        :param int min_length: バッチ内のシーケンスの最小長さ
        :param bool repeat: 入力データセットを繰り返すかどうか
        """
        self.dataset = dataset
        self.batch_size = batch_size
        self.repeat = repeat
        self.epoch = 0
        self.is_new_epoch = False
        self._previous_epoch_detail = -1
        self.min_length = max(1, min_length)
        assert len(dataset) > self.min_length

    def __next__(self):
        if not self.repeat and self.epoch > 0:
            raise StopIteration

        self._previous_epoch_detail = self.epoch_detail
        self.epoch += 1

        if self.batch_size == 1:
            xs = self.dataset[:-1]
            t = self.dataset[-1][0]
            return xs, t

        total_length = len(self.dataset)
        min_length = self.min_length
        offsets = np.random.randint(
            0, min(int(total_length / 3 * 2),
                   total_length - min_length - 1),
            size=self.batch_size)
        lengths = [np.random.randint(self.min_length,
                                     total_length - x)
                   for x in offsets]
        indexes = list(zip(lengths, offsets))
        # indexes.sort()

        xs = [self.dataset[i:i + l] for l, i in indexes]
        t = [self.dataset[i + l][0] for l, i in indexes]
        return list(zip(xs, t))

    @property
    def epoch_detail(self):
        return self.epoch

    @property
    def previous_epoch_detail(self):
        if self._previous_epoch_detail < 0:
            return None
        return self._previous_epoch_detail

    def serialize(self, serializer):
        self.epoch = serializer('epoch', self.epoch)
        self.is_new_epoch = serializer('is_new_epoch',
                                       self.is_new_epoch)
        self._previous_epoch_detail = serializer(
            'previous_epoch_detail',
            self._previous_epoch_detail)

    def reset(self):
        self.epoch = 0
        self.is_new_epoch = False
        self._previous_epoch_detail = -1

def test_data_create(n_data=100, n_loop=3, ratio=0.8):
    """試験用に三相交流波形を模擬したデータセットを用意する

    :param int n_data: データ数
    :param int n_loop: 何周期分の波形にするか
    :param ratio: 訓練データとテストデータの比率
    :return: データセットのペア
    """
    t = np.linspace(0., 2 * np.pi * n_loop, num=n_data)
    x_train = np.array(np.sin(t)).astype(np.float32)
    y_train = np.array(np.sin(t + 2 / 3 * np.pi)).astype(
        np.float32)
    z_train = np.array(np.sin(t + 4 / 3 * np.pi)).astype(
        np.float32)
    n_train = int(n_data * ratio)
    train = chainer.datasets.TupleDataset(x_train[:n_train],
                                          y_train[:n_train],
                                          z_train[:n_train])
    test = chainer.datasets.TupleDataset(x_train[n_train:],
                                         y_train[n_train:],
                                         z_train[n_train:])
    return train, test

def convert(batch, device):
    if device is None:
        def to_device(x):
            return x
    elif device < 0:
        to_device = chainer.cuda.to_cpu
    else:
        def to_device(x):
            return chainer.cuda.to_gpu(x, device,
                                       chainer.cuda.Stream.null)

    def to_device_batch(batch_):
        if device is None:
            return batch_
        elif device < 0:
            return [to_device(np.asarray(x)) for x in batch_]
        else:
            xp = chainer.cuda.cupy.get_array_module(*batch_)
            concat = xp.concatenate(batch_, axis=0)
            sections = np.cumsum([len(x) for x in batch_[:-1]],
                                 dtype='i')
            concat_dev = to_device(concat)
            batch_dev = chainer.cuda.cupy.split(concat_dev,
                                                sections)
            return batch_dev

    x_out = to_device_batch([x for x, _ in batch])
    y_out = to_device_batch([y for _, y in batch])
    return tuple([x_out, y_out])


class Updater(training.StandardUpdater):

    def __init__(self, train_iter, optimizer, device=-1):
        super(Updater, self).__init__(
            train_iter, optimizer, converter=convert,
            device=device)

    def update_core(self):
        train_iter = self.get_iterator('main')
        optimizer = self.get_optimizer('main')

        batch = train_iter.__next__()

        xs, ts = self.converter(batch, self.device)

        x_in = [chainer.Variable(x) for x in xs]
        t_in = np.asarray(ts).reshape((-1, 1))
        loss = optimizer.target(x_in, t_in)

        optimizer.target.cleargrads()
        loss.backward()
        loss.unchain_backward()
        optimizer.update()

class Evaluator(training.extensions.Evaluator):
    def __init__(self, iterator, target, device=-1, **kwargs):
        super(Evaluator, self).__init__(iterator, target,
                                        converter=convert,
                                        device=device, **kwargs)

    def evaluate(self):
        iterator = self._iterators['main']
        target = self._targets['main']
        eval_func = self.eval_func or target

        if self.eval_hook:
            self.eval_hook(self)
        it = copy.copy(iterator)
        summary = reporter_module.DictSummary()

        for batch in it:
            observation = {}
            with reporter_module.report_scope(observation):
                xs, ts = self.converter(batch, self.device)
                x_in = [chainer.Variable(x) for x in xs]
                t_in = np.asarray(ts).reshape((-1, 1))
                eval_func(x_in, t_in)

            summary.add(observation)

        return summary.compute_mean()

# 3つの要素を入力し、100個のパラメータを持つ隠れ層を3つ重ねたネットワーク
model = ForcePredictor(n_input=3, n_units=100, n_layers=3)
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.GradientClipping(5))

# 訓練データの準備
t_dataset, v_dataset = test_data_create(500)
train_iter = SequencePredictIterator(t_dataset, batch_size=20,
                                     min_length=50)
val_iter = SequencePredictIterator(v_dataset, batch_size=10,
                                   min_length=50, repeat=False)

# 訓練の設定
updater = Updater(train_iter, optimizer)
trainer = training.Trainer(updater, (100, "epoch"))
eval_model = model.copy()
log_interval = (10, "epoch")
trainer.extend(Evaluator(val_iter, eval_model,
        eval_hook=lambda _: eval_model.reset_state()),
               trigger=log_interval)
trainer.extend(chainer.training.extensions.LogReport(
    trigger=log_interval))
trainer.extend(extensions.PrintReport(
    ['epoch', 'main/loss', "validation/main/loss"]),
               trigger=log_interval)

# 訓練
trainer.run()

epoch       main/loss   validation/main/loss
10          0.134788    0.0844831             
20          0.0223034   0.0225472             
30          0.0090264   0.0099782             
40          0.00516568  0.00232969            
50          0.00252769  0.000137683           
60          0.00143399  0.00344819            
70          0.00153117  8.24596e-06           
80          0.00122108  0.000140406           
90          0.000937449  5.31944e-05           
100         0.000845839  9.81018e-05           

# 推定精度確認
log = trainer.get_extension("LogReport").log.pop()
print(log["main/accuracy"], log["validation/main/accuracy"],
      log["validation/main/acc05"], log["validation/main/acc10"])

0.23500000312924385 0.6000000238418579 1.0 1.0

# 訓練結果確認
# 確認用データ作成
t = np.linspace(0., 2*np.pi*3, num=500)
X = np.sin(t, dtype=np.float32)
Y = np.sin(t + 2/3 * np.pi, dtype=np.float32)
Z = np.sin(t + 4/3 * np.pi, dtype=np.float32)
plt.plot(t, X)
plt.plot(t, Y)
plt.plot(t, Z)
plt.show()

pre_skip = 20 # 推定を始めるため最小の長さ
X_data = chainer.datasets.TupleDataset(X, Y, Z)
X_in = [np.asarray(X_data[:i+pre_skip], dtype=np.float32)
        for i in range(len(X) - pre_skip - 1)]
X_in = [chainer.variable.Variable(v) for v in X_in]
model.reset_state()
with chainer.configuration.using_config('train', False):
    X_pred = model.chain(X_in)
plt.plot(t[pre_skip + 1:], X_pred.data.reshape(-1), color="r")  # 推定結果(赤)
plt.plot(t, X, color="b")  # 正解データ(青)
plt.show()

# 正解データとの差を確認
# 完全一致は難しいので0.05以下の差は誤差として
# それ以上ずれたデータ個数の割合を算出
real_data = list()
input_data = list()
for i in range(pre_skip, len(X_data)):
    input_data.append(np.asarray(X_data[:i], dtype=np.float32))
    real_data.append(X_data[i][0])
real_data = np.asarray(real_data, dtype=np.float32)
input_data = list(map(chainer.variable.Variable, input_data))
model.reset_state()
with chainer.configuration.using_config('train', False):
    pred_data = model.chain(input_data)
diff = np.abs(pred_data.data.ravel() - real_data, dtype=np.float32)
(diff > 0.05).mean(dtype=np.float32)

0.0

# 異常データの入力を模擬して波形の一部を0に
X_err = X.copy()
X_err[100:120]=0
X_data_err = chainer.datasets.TupleDataset(X_err, Y, Z)
X_in_err = [np.asarray(X_data_err[:i+pre_skip], dtype=np.float32)
        for i in range(len(X) - pre_skip - 1)]
X_in_err = [chainer.variable.Variable(v) for v in X_in_err]
model.reset_state()
with chainer.configuration.using_config('train', False):
    X_pred_err = model.chain(X_in_err)
plt.plot(t[pre_skip + 1:], X_pred_err.data.reshape(-1), color="r")  # 推定結果(赤)
plt.plot(t, X_err, color="b")  #異常模擬データ(青)
plt.show()

# 異常データが入力されたので推定結果と入力データに差異が出ていることを確認
real_data = list()
input_data = list()
for i in range(pre_skip, len(X_data_err)):
    input_data.append(np.asarray(X_data_err[:i], dtype=np.float32))
    real_data.append(X_data_err[i][0])
real_data = np.asarray(real_data, dtype=np.float32)
input_data = list(map(chainer.variable.Variable, input_data))
model.reset_state()
with chainer.configuration.using_config('train', False):
    pred_data = model.chain(input_data)
diff = np.abs(pred_data.data.ravel() - real_data, dtype=np.float32)
(diff > 0.05).mean(dtype=np.float32)

0.058333334