Code for CNN-TS method

CNN-TS/environment.txt

@@ -0,0 +1,163 @@
+absl-py==0.2.0
+alabaster==0.7.10
+anaconda-client==1.6.3
+anaconda-navigator==1.6.2
+anaconda-project==0.6.0
+asn1crypto==0.22.0
+astor==0.6.2
+astroid==1.4.9
+astropy==1.3.2
+Babel==2.4.0
+backports.shutil-get-terminal-size==1.0.0
+beautifulsoup4==4.6.0
+bitarray==0.8.1
+blaze==0.10.1
+bleach==1.5.0
+bokeh==0.12.5
+boto==2.46.1
+Bottleneck==1.2.1
+cffi==1.10.0
+chardet==3.0.3
+click==6.7
+cloudpickle==0.2.2
+clyent==1.2.2
+colorama==0.3.9
+comtypes==1.1.2
+contextlib2==0.5.5
+cryptography==1.8.1
+cycler==0.10.0
+Cython==0.25.2
+cytoolz==0.8.2
+dask==0.14.3
+datashape==0.5.4
+decorator==4.0.11
+distributed==1.16.3
+docutils==0.13.1
+entrypoints==0.2.2
+et-xmlfile==1.0.1
+fastcache==1.0.2
+Flask==0.12.2
+Flask-Cors==3.0.2
+gast==0.2.0
+gevent==1.2.1
+greenlet==0.4.12
+grpcio==1.11.0
+h5py==2.7.0
+HeapDict==1.0.0
+html5lib==0.9999999
+idna==2.5
+imagesize==0.7.1
+ipykernel==4.6.1
+ipython==5.3.0
+ipython-genutils==0.2.0
+ipywidgets==6.0.0
+isort==4.2.5
+itsdangerous==0.24
+jdcal==1.3
+jedi==0.10.2
+Jinja2==2.9.6
+jsonschema==2.6.0
+jupyter==1.0.0
+jupyter-client==5.0.1
+jupyter-console==5.1.0
+jupyter-core==4.3.0
+Keras==2.1.6
+lazy-object-proxy==1.2.2
+lightgbm==2.1.1
+llvmlite==0.18.0
+locket==0.2.0
+lxml==3.7.3
+Markdown==2.6.11
+MarkupSafe==0.23
+matplotlib==2.0.2
+menuinst==1.4.7
+mistune==0.7.4
+mpmath==0.19
+msgpack-python==0.4.8
+multipledispatch==0.4.9
+navigator-updater==0.1.0
+nbconvert==5.1.1
+nbformat==4.3.0
+networkx==1.11
+nltk==3.2.3
+nose==1.3.7
+notebook==5.0.0
+numba==0.33.0
+numexpr==2.6.2
+numpy==1.14.3
+numpydoc==0.6.0
+odo==0.5.0
+olefile==0.44
+openpyxl==2.4.7
+packaging==16.8
+pandas==0.20.1
+pandocfilters==1.4.1
+partd==0.3.8
+path.py==10.3.1
+pathlib2==2.2.1
+patsy==0.4.1
+pep8==1.7.0
+pickleshare==0.7.4
+Pillow==4.1.1
+ply==3.10
+prompt-toolkit==1.0.14
+protobuf==3.5.2.post1
+psutil==5.2.2
+py==1.4.33
+pycosat==0.6.2
+pycparser==2.17
+pycrypto==2.6.1
+pycurl==7.43.0
+pyflakes==1.5.0
+Pygments==2.2.0
+pylint==1.6.4
+pyodbc==4.0.16
+pyOpenSSL==17.0.0
+pyparsing==2.1.4
+pytest==3.0.7
+python-dateutil==2.6.0
+pytz==2017.2
+PyWavelets==0.5.2
+pywin32==220
+PyYAML==3.12
+pyzmq==16.0.2
+QtAwesome==0.4.4
+qtconsole==4.3.0
+QtPy==1.2.1
+requests==2.14.2
+rope-py3k==0.9.4.post1
+scikit-image==0.13.0
+scikit-learn==0.18.1
+scipy==0.19.0
+seaborn==0.7.1
+simplegeneric==0.8.1
+singledispatch==3.4.0.3
+six==1.11.0
+snowballstemmer==1.2.1
+sortedcollections==0.5.3
+sortedcontainers==1.5.7
+sphinx==1.5.6
+spyder==3.1.4
+SQLAlchemy==1.1.9
+statsmodels==0.8.0
+sympy==1.0
+tables==3.2.2
+tblib==1.3.2
+tensorboard==1.8.0
+tensorflow==1.8.0
+termcolor==1.1.0
+testpath==0.3
+toolz==0.8.2
+tornado==4.5.1
+traitlets==4.3.2
+unicodecsv==0.14.1
+wcwidth==0.1.7
+Werkzeug==0.14.1
+widgetsnbextension==2.0.0
+win-unicode-console==0.5
+wrapt==1.10.10
+xlrd==1.0.0
+XlsxWriter==0.9.6
+xlwings==0.10.4
+xlwt==1.2.0
+zict==0.1.2

CNN-TS/metrics.py

@@ -0,0 +1,38 @@
+"""Evaluation metrics, taken from the competition Github repo
+https://github.com/M4Competition/M4-methods
+The factor of 2 in SMAPE has been changed to 200 to be consistent with the R code.
+"""
+import numpy as np
+
+
+
+def smape(a, b):
+    """
+    Calculates sMAPE
+
+    :param a: actual values
+    :param b: predicted values
+    :return: sMAPE
+    """
+    a = np.reshape(a, (-1,))
+    b = np.reshape(b, (-1,))
+    return np.mean(200.0 * np.abs(a - b) / (np.abs(a) + np.abs(b))).item()   # BT changed 2 to 200
+
+
+def mase(insample, y_test, y_hat_test, freq):
+    """
+    Calculates MAsE
+
+    :param insample: insample data
+    :param y_test: out of sample target values
+    :param y_hat_test: predicted values
+    :param freq: data frequency
+    :return:
+    """
+    y_hat_naive = []
+    for i in range(freq, len(insample)):
+        y_hat_naive.append(insample[(i - freq)])
+
+    masep = np.mean(abs(insample[freq:] - y_hat_naive))
+
+    return np.mean(abs(y_test - y_hat_test)) / masep

CNN-TS/model_definitions.py

@@ -0,0 +1,179 @@
+
+import keras as ks
+import pandas as pd
+import numpy as np
+import metrics
+from collections import namedtuple
+import tensorflow as tf
+import datetime as dt
+import os
+from utilities import *
+
+
+def yearly_model(series_length):
+    input = ks.layers.Input((series_length,))
+    yearly_input = ks.layers.Reshape((series_length, 1))(input)
+    yearly_hidden1 = ks.layers.Dense(units=50, activation='relu')(ks.layers.Flatten()(yearly_input))
+    yearly_hidden2 = ks.layers.Dense(units=50, activation='relu')(yearly_hidden1)
+    yearly_output = ks.layers.Dense(units=1, activation='linear')(yearly_hidden2)
+    est = ks.Model(inputs=input, outputs=yearly_output)
+    est.compile(optimizer=ks.optimizers.Adam(lr=0.001), loss='mse', metrics=['mse'])
+    epochs = 250
+    batch_size = 1000
+    return est, epochs, batch_size
+
+
+def quarterly_model(series_length):
+    input = ks.layers.Input((series_length,))
+    yearly_input = ks.layers.Reshape((series_length, 1))(input)
+    yearly_avg = ks.layers.AveragePooling1D(pool_size=4, strides=4, padding='valid')(yearly_input)
+    yearly_hidden1 = ks.layers.Dense(units=50, activation='relu')(ks.layers.Flatten()(yearly_avg))
+    yearly_hidden2 = ks.layers.Dense(units=50, activation='relu')(yearly_hidden1)
+    yearly_output = ks.layers.Dense(units=1, activation='linear')(yearly_hidden2)
+
+    yearly_avg_up = ks.layers.UpSampling1D(size=4)(yearly_avg)
+    periodic_diff = ks.layers.Subtract()([input, ks.layers.Flatten()(yearly_avg_up)])
+    periodic_input = ks.layers.Reshape((series_length // 4, 4, 1))(periodic_diff)
+    periodic_conv = ks.layers.Conv2D(filters=3, kernel_size=(1, 4), strides=(1, 4),
+                                     padding='valid')(periodic_input)
+    periodic_hidden1 = ks.layers.Dense(units=50, activation='relu')(ks.layers.Flatten()(periodic_conv))
+    periodic_hidden2 = ks.layers.Dense(units=50, activation='relu')(periodic_hidden1)
+    periodic_output = ks.layers.Dense(units=1, activation='linear')(periodic_hidden2)
+    output = ks.layers.Add()([yearly_output, periodic_output])
+    est = ks.Model(inputs=input, outputs=output)
+    est.compile(optimizer=ks.optimizers.Adam(lr=0.001), loss='mse', metrics=['mse'])
+    epochs = 250
+    batch_size = 1000
+    return est, epochs, batch_size
+
+
+def monthly_model(series_length):
+    input = ks.layers.Input((series_length,))
+    yearly_input = ks.layers.Reshape((series_length, 1))(input)
+    yearly_avg = ks.layers.AveragePooling1D(pool_size=12, strides=12, padding='valid')(yearly_input)
+    yearly_hidden1 = ks.layers.Dense(units=20, activation='relu')(ks.layers.Flatten()(yearly_avg))
+    yearly_hidden2 = ks.layers.Dense(units=20, activation='relu')(yearly_hidden1)
+    yearly_output = ks.layers.Dense(units=1, activation='linear')(yearly_hidden2)
+
+    yearly_avg_up = ks.layers.UpSampling1D(size=12)(yearly_avg)
+    periodic_diff = ks.layers.Subtract()([input, ks.layers.Flatten()(yearly_avg_up)])
+    periodic_input = ks.layers.Reshape((series_length // 12, 12, 1))(periodic_diff)
+    periodic_conv = ks.layers.Conv2D(filters=3, kernel_size=(1, 12), strides=(1, 12),
+                                     padding='valid')(periodic_input)
+    periodic_hidden1 = ks.layers.Dense(units=20, activation='relu')(ks.layers.Flatten()(periodic_conv))
+    periodic_hidden2 = ks.layers.Dense(units=20, activation='relu')(periodic_hidden1)
+    periodic_output = ks.layers.Dense(units=1, activation='linear')(periodic_hidden2)
+    output = ks.layers.Add()([yearly_output, periodic_output])
+    est = ks.Model(inputs=input, outputs=output)
+    est.compile(optimizer=ks.optimizers.Adam(lr=0.001), loss='mse', metrics=['mse'])
+    epochs = 250
+    batch_size = 1000
+    return est, epochs, batch_size
+
+
+def weekly_model(series_length):
+    if series_length == 52:
+        input = ks.layers.Input((series_length,))
+        hidden1 = ks.layers.Dense(units=20, activation='relu')(input)
+        hidden2 = ks.layers.Dense(units=20, activation='relu')(hidden1)
+        output = ks.layers.Dense(units=1, activation='linear')(hidden2)
+        est = ks.Model(inputs=input, outputs=output)
+        est.compile(optimizer=ks.optimizers.Adam(lr=0.001), loss='mse', metrics=['mse'])
+        epochs = 250
+        batch_size = 1000
+    else:
+        input = ks.layers.Input((series_length,))
+        yearly_input = ks.layers.Reshape((series_length, 1))(input)
+        yearly_avg = ks.layers.AveragePooling1D(pool_size=52, strides=52, padding='valid')(yearly_input)
+        yearly_hidden1 = ks.layers.Dense(units=20, activation='relu')(ks.layers.Flatten()(yearly_avg))
+        yearly_hidden2 = ks.layers.Dense(units=20, activation='relu')(yearly_hidden1)
+        yearly_output = ks.layers.Dense(units=1, activation='linear')(yearly_hidden2)
+
+        yearly_avg_up = ks.layers.UpSampling1D(size=52)(yearly_avg)
+        periodic_diff = ks.layers.Subtract()([input, ks.layers.Flatten()(yearly_avg_up)])
+        periodic_input = ks.layers.Reshape((series_length // 52, 52, 1))(periodic_diff)
+        periodic_conv = ks.layers.Conv2D(filters=3, kernel_size=(1, 52), strides=(1, 52),
+                                         padding='valid')(periodic_input)
+        periodic_hidden1 = ks.layers.Dense(units=20, activation='relu')(ks.layers.Flatten()(periodic_conv))
+        periodic_hidden2 = ks.layers.Dense(units=20, activation='relu')(periodic_hidden1)
+        periodic_output = ks.layers.Dense(units=1, activation='linear')(periodic_hidden2)
+        output = ks.layers.Add()([yearly_output, periodic_output])
+        est = ks.Model(inputs=input, outputs=output)
+        est.compile(optimizer=ks.optimizers.Adam(lr=0.001), loss='mse', metrics=['mse'])
+        epochs = 250
+        batch_size = 1000
+    return est, epochs, batch_size
+
+
+def daily_model(series_length):
+    input = ks.layers.Input((series_length,))
+    weekly_input = ks.layers.Reshape((series_length, 1))(input)
+    weekly_avg = ks.layers.AveragePooling1D(pool_size=7, strides=7, padding='valid')(weekly_input)
+    weekly_hidden1 = ks.layers.Dense(units=20, activation='relu')(ks.layers.Flatten()(weekly_avg))
+    weekly_hidden2 = ks.layers.Dense(units=20, activation='relu')(weekly_hidden1)
+    weekly_output = ks.layers.Dense(units=1, activation='linear')(weekly_hidden2)
+
+    weekly_avg_up = ks.layers.UpSampling1D(size=7)(weekly_avg)
+    periodic_diff = ks.layers.Subtract()([input, ks.layers.Flatten()(weekly_avg_up)])
+    periodic_input = ks.layers.Reshape((series_length // 7, 7, 1))(periodic_diff)
+    periodic_conv = ks.layers.Conv2D(filters=3, kernel_size=(1, 7), strides=(1, 7),
+                                     padding='valid')(periodic_input)
+    periodic_hidden1 = ks.layers.Dense(units=20, activation='relu')(ks.layers.Flatten()(periodic_conv))
+    periodic_hidden2 = ks.layers.Dense(units=20, activation='relu')(periodic_hidden1)
+    periodic_output = ks.layers.Dense(units=1, activation='linear')(periodic_hidden2)
+    output = ks.layers.Add()([weekly_output, periodic_output])
+    est = ks.Model(inputs=input, outputs=output)
+    est.compile(optimizer=ks.optimizers.Adam(lr=0.001), loss='mse', metrics=['mse'])
+    epochs = 250
+    batch_size = 1000
+    return est, epochs, batch_size
+
+
+def hourly_model(series_length):
+    input = ks.layers.Input((series_length,))
+    weekly_input = ks.layers.Reshape((series_length, 1))(input)
+    weekly_avg = ks.layers.AveragePooling1D(pool_size=168, strides=168, padding='valid')(weekly_input)
+    weekly_hidden1 = ks.layers.Dense(units=20, activation='relu')(ks.layers.Flatten()(weekly_avg))
+    weekly_hidden2 = ks.layers.Dense(units=20, activation='relu')(weekly_hidden1)
+    weekly_output = ks.layers.Dense(units=1, activation='linear')(weekly_hidden2)
+    weekly_avg_up = ks.layers.UpSampling1D(size=168)(weekly_avg)
+
+    daily_diff = ks.layers.Subtract()([input, ks.layers.Flatten()(weekly_avg_up)])
+    daily_diff_input = ks.layers.Reshape((series_length // 7, 7, 1))(daily_diff)
+    daily_diff_conv = ks.layers.Conv2D(filters=3, kernel_size=(1, 7), strides=(1, 7),
+                                       padding='valid')(daily_diff_input)
+    daily_diff_hidden1 = ks.layers.Dense(units=20, activation='relu')(ks.layers.Flatten()(daily_diff_conv))
+    daily_diff_hidden2 = ks.layers.Dense(units=20, activation='relu')(daily_diff_hidden1)
+    daily_diff_output = ks.layers.Dense(units=1, activation='linear')(daily_diff_hidden2)
+
+    daily_avg = ks.layers.AveragePooling1D(pool_size=24, strides=24, padding='valid')(weekly_input)
+    daily_avg_up = ks.layers.UpSampling1D(size=24)(daily_avg)
+    hourly_diff = ks.layers.Subtract()([input, ks.layers.Flatten()(daily_avg_up)])
+    hourly_diff_input = ks.layers.Reshape((series_length // 24, 24, 1))(hourly_diff)
+    hourly_diff_conv = ks.layers.Conv2D(filters=3, kernel_size=(1, 24), strides=(1, 24),
+                                        padding='valid')(hourly_diff_input)
+    hourly_diff_hidden1 = ks.layers.Dense(units=20, activation='relu')(ks.layers.Flatten()(hourly_diff_conv))
+    hourly_diff_hidden2 = ks.layers.Dense(units=20, activation='relu')(hourly_diff_hidden1)
+    hourly_diff_output = ks.layers.Dense(units=1, activation='linear')(hourly_diff_hidden2)
+
+    output = ks.layers.Add()([weekly_output, daily_diff_output, hourly_diff_output])
+
+    est = ks.Model(inputs=input, outputs=output)
+    est.compile(optimizer=ks.optimizers.Adam(lr=0.001), loss='mse', metrics=['mse'])
+    epochs = 250
+    batch_size = 1000
+    return est, epochs, batch_size
+
+
+def get_model_params():
+    # define the model parameters
+    Model = namedtuple('Model', ['freq_name', 'horizon', 'freq', 'model_constructor', 'training_lengths',
+                                 'cycle_length', 'augment'])
+    yearly = Model('Yearly', 6, 1, yearly_model, [10, 20, 30], 1, False)
+    quarterly = Model('Quarterly', 8, 4, quarterly_model, [20, 48, 100], 4, False)
+    monthly = Model('Monthly', 18, 12, monthly_model, [48, 120, 240], 12, False)
+    weekly = Model('Weekly', 13, 1, weekly_model, [52, 520, 1040], 52, True)
+    daily = Model('Daily', 14, 1, daily_model, [98], 7, True)
+    hourly = Model('Hourly', 48, 24, hourly_model, [672], 7*24, True)
+    return [yearly, quarterly, monthly, weekly, daily, hourly]
+