第13章 ディープラーニングのためのデータウィンドウとベースライン

In [28]:

import pandas as pd
import numpy as np
import datetime as dt
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras import Model
from tensorflow.keras.losses import MeanSquaredError
from tensorflow.keras.metrics import MeanAbsoluteError

import pandas as pd import numpy as np import datetime as dt import matplotlib.pyplot as plt import tensorflow as tf from tensorflow.keras import Model from tensorflow.keras.losses import MeanSquaredError from tensorflow.keras.metrics import MeanAbsoluteError

データウィンドウを作成する¶

In [2]:

url_train = 'https://raw.githubusercontent.com/marcopeix/TimeSeriesForecastingInPython/master/data/train.csv'
url_val = 'https://raw.githubusercontent.com/marcopeix/TimeSeriesForecastingInPython/master/data/val.csv'
url_test = 'https://raw.githubusercontent.com/marcopeix/TimeSeriesForecastingInPython/master/data/test.csv'

df_train = pd.read_csv(url_train, index_col=0)
df_val = pd.read_csv(url_val, index_col=0)
df_test = pd.read_csv(url_test, index_col=0)

url_train = 'https://raw.githubusercontent.com/marcopeix/TimeSeriesForecastingInPython/master/data/train.csv' url_val = 'https://raw.githubusercontent.com/marcopeix/TimeSeriesForecastingInPython/master/data/val.csv' url_test = 'https://raw.githubusercontent.com/marcopeix/TimeSeriesForecastingInPython/master/data/test.csv' df_train = pd.read_csv(url_train, index_col=0) df_val = pd.read_csv(url_val, index_col=0) df_test = pd.read_csv(url_test, index_col=0)

In [30]:

class DataWindow:
    def __init__(self, input_width, label_width, shift, df_train, df_val, df_test, label_columns=None):
        # window size
        self.input_width = input_width
        self.label_width = label_width
        self.shift = shift
        self.total_window_size = input_width + shift
        
        # データ
        self.df_train = df_train
        self.df_val = df_val
        self.df_test = df_test
        
        # ラベル
        self.label_columns = label_columns
        if label_columns is not None:
            self.label_columns_indices = {name: i for i, name in enumerate(label_columns)}
        self.column_indices = {name: i for i, name in enumerate(self.df_train.columns)}
        
        # スライス
        self.input_slice = slice(0, input_width)
        self.input_indices = np.arange(self.total_window_size)[self.input_slice]
        
        # ラベル開始位置
        self.label_start = self.total_window_size - self.label_width
        self.labels_slice = slice(self.label_start, None)
        self.label_indices = np.arange(self.total_window_size)[self.labels_slice]
    
    def split_to_inputs_labels(self, features):
        inputs = features[:, self.input_slice, :]
        labels = features[:, self.labels_slice, :]
        if self.label_columns is not None:
            labels = tf.stack([labels[:, :, self.column_indices[name]] for name in self.label_columns], axis=-1)
        inputs.set_shape([None, self.input_width, None])
        labels.set_shape([None, self.label_width, None])
        return inputs, labels
    
    def plot(self, plot_col: str, model=None, max_subplots=3):
        inputs, labels = self.sample_batch
        plt.figure(figsize=(12, 8))
        plot_col_index = self.column_indices[plot_col]
        n_max = min(max_subplots, len(inputs))
        
        for n in range(n_max):
            plt.subplot(n_max, 1, n+1)
            plt.ylabel(f'{plot_col} [scaled]')
            plt.plot(self.input_indices, inputs[n, :, plot_col_index], label='Inputs', marker='.', zorder=-10)
            
            if self.label_columns:
                label_col_index = self.label_columns_indices.get(plot_col, None)
            else:
                label_col_index = plot_col_index
            
            if label_col_index is None:
                continue
            
            plt.scatter(self.label_indices, labels[n, :, label_col_index], edgecolors='k', label='Labels', c='tab:green', s=64)
            
            if model is not None:
                predictions = model(inputs)
                plt.scatter(self.label_indices, predictions[n, :, label_col_index], marker='X', edgecolors='k', label='Predictions', c='tab:red', s=64)
            
            if n == 0:
                plt.legend()
        plt.xlabel('Time (h)')
    
    def make_dataset(self, data):
        data = np.array(data, dtype=np.float32)
        ds = tf.keras.utils.timeseries_dataset_from_array(
            data=data,
            targets=None,
            sequence_length=self.total_window_size,
            sequence_stride=1,
            shuffle=True,
            batch_size=32,
        )
        ds = ds.map(self.split_to_inputs_labels)
        return ds
    
    @property
    def train(self):
        return self.make_dataset(self.df_train)
    
    @property
    def val(self):
        return self.make_dataset(self.df_val)
    
    @property
    def test(self):
        return self.make_dataset(self.df_test)
    
    @property
    def sample_batch(self):
        """Get and cache an example batch of `inputs, labels` for plotting."""
        result = getattr(self, '_sample_batch', None)
        if result is None:
            result = next(iter(self.train))
            self._sample_batch = result
        return result

class DataWindow: def __init__(self, input_width, label_width, shift, df_train, df_val, df_test, label_columns=None): # window size self.input_width = input_width self.label_width = label_width self.shift = shift self.total_window_size = input_width + shift # データ self.df_train = df_train self.df_val = df_val self.df_test = df_test # ラベル self.label_columns = label_columns if label_columns is not None: self.label_columns_indices = {name: i for i, name in enumerate(label_columns)} self.column_indices = {name: i for i, name in enumerate(self.df_train.columns)} # スライス self.input_slice = slice(0, input_width) self.input_indices = np.arange(self.total_window_size)[self.input_slice] # ラベル開始位置 self.label_start = self.total_window_size - self.label_width self.labels_slice = slice(self.label_start, None) self.label_indices = np.arange(self.total_window_size)[self.labels_slice] def split_to_inputs_labels(self, features): inputs = features[:, self.input_slice, :] labels = features[:, self.labels_slice, :] if self.label_columns is not None: labels = tf.stack([labels[:, :, self.column_indices[name]] for name in self.label_columns], axis=-1) inputs.set_shape([None, self.input_width, None]) labels.set_shape([None, self.label_width, None]) return inputs, labels def plot(self, plot_col: str, model=None, max_subplots=3): inputs, labels = self.sample_batch plt.figure(figsize=(12, 8)) plot_col_index = self.column_indices[plot_col] n_max = min(max_subplots, len(inputs)) for n in range(n_max): plt.subplot(n_max, 1, n+1) plt.ylabel(f'{plot_col} [scaled]') plt.plot(self.input_indices, inputs[n, :, plot_col_index], label='Inputs', marker='.', zorder=-10) if self.label_columns: label_col_index = self.label_columns_indices.get(plot_col, None) else: label_col_index = plot_col_index if label_col_index is None: continue plt.scatter(self.label_indices, labels[n, :, label_col_index], edgecolors='k', label='Labels', c='tab:green', s=64) if model is not None: predictions = model(inputs) plt.scatter(self.label_indices, predictions[n, :, label_col_index], marker='X', edgecolors='k', label='Predictions', c='tab:red', s=64) if n == 0: plt.legend() plt.xlabel('Time (h)') def make_dataset(self, data): data = np.array(data, dtype=np.float32) ds = tf.keras.utils.timeseries_dataset_from_array( data=data, targets=None, sequence_length=self.total_window_size, sequence_stride=1, shuffle=True, batch_size=32, ) ds = ds.map(self.split_to_inputs_labels) return ds @property def train(self): return self.make_dataset(self.df_train) @property def val(self): return self.make_dataset(self.df_val) @property def test(self): return self.make_dataset(self.df_test) @property def sample_batch(self): """Get and cache an example batch of `inputs, labels` for plotting.""" result = getattr(self, '_sample_batch', None) if result is None: result = next(iter(self.train)) self._sample_batch = result return result

ベースラインモデルを実装する¶

In [49]:

# まずは次のステップを前のステップで予測結果とする
single_step_window = DataWindow(input_width=1, label_width=1, shift=1, df_train=df_train, df_val=df_val, df_test=df_test, label_columns=['traffic_volume'])
wide_window = DataWindow(input_width=24, label_width=24, shift=1, df_train=df_train, df_val=df_val, df_test=df_test, label_columns=['traffic_volume'])

# まずは次のステップを前のステップで予測結果とする single_step_window = DataWindow(input_width=1, label_width=1, shift=1, df_train=df_train, df_val=df_val, df_test=df_test, label_columns=['traffic_volume']) wide_window = DataWindow(input_width=24, label_width=24, shift=1, df_train=df_train, df_val=df_val, df_test=df_test, label_columns=['traffic_volume'])

In [50]:

class Baseline(Model):
    def __init__(self, label_index=None):
        super().__init__()
        self.label_index = label_index
        
    def call(self, inputs):
        if self.label_index is None:
            return inputs
        elif isinstance(self.label_index, list):
            tensors = []
            for index in self.label_index:
                res = inputs[:, :, index]
                res = res[:, :, tf.newaxis]
                tensors.append(res)
            return tf.concat(tensors, axis=-1)
        else:
            res = inputs[:, :, self.label_index]
            return res[:, :, tf.newaxis]

class Baseline(Model): def __init__(self, label_index=None): super().__init__() self.label_index = label_index def call(self, inputs): if self.label_index is None: return inputs elif isinstance(self.label_index, list): tensors = [] for index in self.label_index: res = inputs[:, :, index] res = res[:, :, tf.newaxis] tensors.append(res) return tf.concat(tensors, axis=-1) else: res = inputs[:, :, self.label_index] return res[:, :, tf.newaxis]

In [51]:

column_indices = {name: i for i, name in enumerate(df_train.columns)}
baseline_last = Baseline(column_indices['traffic_volume'])
baseline_last.compile(loss=MeanSquaredError(), metrics=[MeanAbsoluteError()])

column_indices = {name: i for i, name in enumerate(df_train.columns)} baseline_last = Baseline(column_indices['traffic_volume']) baseline_last.compile(loss=MeanSquaredError(), metrics=[MeanAbsoluteError()])

In [52]:

performane_val = {}
performance_test = {}

performane_val['Baseline - Last'] = baseline_last.evaluate(single_step_window.val)
performance_test['Baseline - Last'] = baseline_last.evaluate(single_step_window.test, verbose=0)

performane_val = {} performance_test = {} performane_val['Baseline - Last'] = baseline_last.evaluate(single_step_window.val) performance_test['Baseline - Last'] = baseline_last.evaluate(single_step_window.test, verbose=0)

110/110 [==============================] - 0s 2ms/step - loss: 684054.1250 - mean_absolute_error: 595.2283

In [53]:

# single_step_window.plot('traffic_volume', baseline_last)だとあまり状況が見えないのでwide_windowにしていく
wide_window.plot('traffic_volume', baseline_last)

# single_step_window.plot('traffic_volume', baseline_last)だとあまり状況が見えないのでwide_windowにしていく wide_window.plot('traffic_volume', baseline_last)

In [60]:

# マルチステップの場合
multi_window = DataWindow(input_width=24, label_width=24, shift=24, df_train=df_train, df_val=df_val, df_test=df_test, label_columns=['traffic_volume'])

class MultiStepLastBaseline(Model):
    def __init__(self, label_index=None):
        super().__init__()
        self.label_index = label_index
    
    def call(self, inputs):
        if self.label_index is None:
            return tf.tile(inputs[:, -1:, :], [1, 24, 1])
        return tf.tile(inputs[:, -1:, self.label_index:], [1, 24, 1])
    
class RepeatBaseline(Model):
    def __init__(self, label_index=None):
        super().__init__()
        self.label_index = label_index
    
    def call(self, inputs):
        return inputs[:, :, self.label_index:]

# マルチステップの場合 multi_window = DataWindow(input_width=24, label_width=24, shift=24, df_train=df_train, df_val=df_val, df_test=df_test, label_columns=['traffic_volume']) class MultiStepLastBaseline(Model): def __init__(self, label_index=None): super().__init__() self.label_index = label_index def call(self, inputs): if self.label_index is None: return tf.tile(inputs[:, -1:, :], [1, 24, 1]) return tf.tile(inputs[:, -1:, self.label_index:], [1, 24, 1]) class RepeatBaseline(Model): def __init__(self, label_index=None): super().__init__() self.label_index = label_index def call(self, inputs): return inputs[:, :, self.label_index:]

In [61]:

ms_baseline_last = MultiStepLastBaseline(label_index=column_indices['traffic_volume'])
ms_baseline_last.compile(loss=MeanSquaredError(), metrics=[MeanAbsoluteError()])
ms_baseline_repeat = RepeatBaseline(label_index=column_indices['traffic_volume'])
ms_baseline_repeat.compile(loss=MeanSquaredError(), metrics=[MeanAbsoluteError()])

ms_val_performance = {}
ms_test_performance = {}

ms_val_performance['Baseline - Last'] = ms_baseline_last.evaluate(multi_window.val)
ms_test_performance['Baseline - Last'] = ms_baseline_last.evaluate(multi_window.test)
ms_val_performance['Baseline - Repeat'] = ms_baseline_repeat.evaluate(multi_window.val)
ms_test_performance['Baseline - Repeat'] = ms_baseline_repeat.evaluate(multi_window.test)


multi_window.plot('traffic_volume', ms_baseline_last)
multi_window.plot('traffic_volume', ms_baseline_repeat)

ms_baseline_last = MultiStepLastBaseline(label_index=column_indices['traffic_volume']) ms_baseline_last.compile(loss=MeanSquaredError(), metrics=[MeanAbsoluteError()]) ms_baseline_repeat = RepeatBaseline(label_index=column_indices['traffic_volume']) ms_baseline_repeat.compile(loss=MeanSquaredError(), metrics=[MeanAbsoluteError()]) ms_val_performance = {} ms_test_performance = {} ms_val_performance['Baseline - Last'] = ms_baseline_last.evaluate(multi_window.val) ms_test_performance['Baseline - Last'] = ms_baseline_last.evaluate(multi_window.test) ms_val_performance['Baseline - Repeat'] = ms_baseline_repeat.evaluate(multi_window.val) ms_test_performance['Baseline - Repeat'] = ms_baseline_repeat.evaluate(multi_window.test) multi_window.plot('traffic_volume', ms_baseline_last) multi_window.plot('traffic_volume', ms_baseline_repeat)

109/109 [==============================] - 1s 5ms/step - loss: 12769637.0000 - mean_absolute_error: 2974.6951
54/54 [==============================] - 0s 5ms/step - loss: 12498208.0000 - mean_absolute_error: 2964.0754
109/109 [==============================] - 2s 3ms/step - loss: 10588796.0000 - mean_absolute_error: 2431.1047
54/54 [==============================] - 0s 3ms/step - loss: 10346166.0000 - mean_absolute_error: 2403.8494

In [63]:

# 多出力のベースラインモデル
col_names = ['temp', 'traffic_volume']
mo_single_step_window = DataWindow(input_width=1, label_width=1, shift=1, df_train=df_train, df_val=df_val, df_test=df_test, label_columns=col_names)
mo_wide_window = DataWindow(input_width=24, label_width=24, shift=1, df_train=df_train, df_val=df_val, df_test=df_test, label_columns=col_names)

mo_baseline_last = Baseline(label_index=[column_indices[col] for col in col_names])
mo_baseline_last.compile(loss=MeanSquaredError(), metrics=[MeanAbsoluteError()])

mo_val_performance = {}
mo_test_performance = {}

# テキストだとmo_wide_windowを使ってevaluateをしているが本来的にはmo_single_step_windowを使う必要がある
mo_val_performance['Baseline - Last'] = mo_baseline_last.evaluate(mo_single_step_window.val)
mo_test_performance['Baseline - Last'] = mo_baseline_last.evaluate(mo_single_step_window.val)

for col in col_names:
    mo_wide_window.plot(col, mo_baseline_last)

# 多出力のベースラインモデル col_names = ['temp', 'traffic_volume'] mo_single_step_window = DataWindow(input_width=1, label_width=1, shift=1, df_train=df_train, df_val=df_val, df_test=df_test, label_columns=col_names) mo_wide_window = DataWindow(input_width=24, label_width=24, shift=1, df_train=df_train, df_val=df_val, df_test=df_test, label_columns=col_names) mo_baseline_last = Baseline(label_index=[column_indices[col] for col in col_names]) mo_baseline_last.compile(loss=MeanSquaredError(), metrics=[MeanAbsoluteError()]) mo_val_performance = {} mo_test_performance = {} # テキストだとmo_wide_windowを使ってevaluateをしているが本来的にはmo_single_step_windowを使う必要がある mo_val_performance['Baseline - Last'] = mo_baseline_last.evaluate(mo_single_step_window.val) mo_test_performance['Baseline - Last'] = mo_baseline_last.evaluate(mo_single_step_window.val) for col in col_names: mo_wide_window.plot(col, mo_baseline_last)

110/110 [==============================] - 0s 2ms/step - loss: 342027.7812 - mean_absolute_error: 298.0399
110/110 [==============================] - 0s 2ms/step - loss: 342027.8438 - mean_absolute_error: 298.0400

In [ ]: