Difan Deng: Time series forecasting (#434)

Author: Github Actions

development/_downloads/b4aa13e52c27833de473ca84d042e32d/example_time_series_forecasting.ipynb

@@ -0,0 +1,90 @@
+{
+  "cells": [
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "collapsed": false
+      },
+      "outputs": [],
+      "source": [
+        "%matplotlib inline"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "\n# Time Series Forecasting\n\nThe following example shows how to fit a sample forecasting model\nwith AutoPyTorch. This is only a dummmy example because of the limited size of the dataset.\nThus, it could be possible that the AutoPyTorch model does not perform as well as a dummy predictor\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "collapsed": false
+      },
+      "outputs": [],
+      "source": [
+        "import os\nimport tempfile as tmp\nimport warnings\nimport copy\n\nos.environ['JOBLIB_TEMP_FOLDER'] = tmp.gettempdir()\nos.environ['OMP_NUM_THREADS'] = '1'\nos.environ['OPENBLAS_NUM_THREADS'] = '1'\nos.environ['MKL_NUM_THREADS'] = '1'\n\nwarnings.simplefilter(action='ignore', category=UserWarning)\nwarnings.simplefilter(action='ignore', category=FutureWarning)\n\nfrom sktime.datasets import load_longley\ntargets, features = load_longley()\n\nforecasting_horizon = 3\n\n# each series represent an element in the List\n# we take the last forecasting_horizon  as test targets. The itme before that as training targets\n# Normally the value to be forecasted should follow the training sets\ny_train = [targets[: -forecasting_horizon]]\ny_test = [targets[-forecasting_horizon:]]\n\n# same for features. For uni-variant models, X_train, X_test can be omitted\nX_train = [features[: -forecasting_horizon]]\n# Here x_test indicates the 'known future features': they are the features known previously, features that are unknown\n# could be replaced with NAN or zeros (which will not be used by our networks). If no feature is known beforehand,\n# we could also omit X_test\nknown_future_features = list(features.columns)\nX_test = [features[-forecasting_horizon:]]\n\nstart_times = [targets.index.to_timestamp()[0]]\nfreq = '1Y'\n\nfrom autoPyTorch.api.time_series_forecasting import TimeSeriesForecastingTask"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## Build and fit a forecaster\n\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "collapsed": false
+      },
+      "outputs": [],
+      "source": [
+        "api = TimeSeriesForecastingTask()"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## Search for an ensemble of machine learning algorithms\n\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "collapsed": false
+      },
+      "outputs": [],
+      "source": [
+        "api.search(\n    X_train=X_train,\n    y_train=copy.deepcopy(y_train),\n    X_test=X_test,\n    optimize_metric='mean_MASE_forecasting',\n    n_prediction_steps=forecasting_horizon,\n    memory_limit=16 * 1024,  # Currently, forecasting models need much more memories than it actually requires\n    freq=freq,\n    start_times=start_times,\n    func_eval_time_limit_secs=50,\n    total_walltime_limit=60,\n    min_num_test_instances=1000,  # proxy validation sets. This only works for the tasks with more than 1000 series\n    known_future_features=known_future_features,\n)\n\n\nfrom autoPyTorch.datasets.time_series_dataset import TimeSeriesSequence\n\ntest_sets = []\n\n# We could construct test sets from scratch\nfor feature, future_feature, target, start_time in zip(X_train, X_test,y_train, start_times):\n    test_sets.append(\n        TimeSeriesSequence(X=feature.values,\n                           Y=target.values,\n                           X_test=future_feature.values,\n                           start_time=start_time,\n                           is_test_set=True,\n                           # additional information required to construct a new time series sequence\n                           **api.dataset.sequences_builder_kwargs\n                           )\n    )\n# Alternatively, if we only want to forecast the value after the X_train, we could directly ask datamanager to\n# generate a test set:\n# test_sets2 = api.dataset.generate_test_seqs()\n\npred = api.predict(test_sets)"
+      ]
+    }
+  ],
+  "metadata": {
+    "kernelspec": {
+      "display_name": "Python 3",
+      "language": "python",
+      "name": "python3"
+    },
+    "language_info": {
+      "codemirror_mode": {
+        "name": "ipython",
+        "version": 3
+      },
+      "file_extension": ".py",
+      "mimetype": "text/x-python",
+      "name": "python",
+      "nbconvert_exporter": "python",
+      "pygments_lexer": "ipython3",
+      "version": "3.8.12"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
+}

development/_downloads/bc82bea3a5dd7bdba60b65220891d9e5/examples_python.zip

@@ -0,0 +1,90 @@
+"""
+======================
+Time Series Forecasting
+======================
+
+The following example shows how to fit a sample forecasting model
+with AutoPyTorch. This is only a dummmy example because of the limited size of the dataset.
+Thus, it could be possible that the AutoPyTorch model does not perform as well as a dummy predictor
+"""
+import os
+import tempfile as tmp
+import warnings
+import copy
+
+os.environ['JOBLIB_TEMP_FOLDER'] = tmp.gettempdir()
+os.environ['OMP_NUM_THREADS'] = '1'
+os.environ['OPENBLAS_NUM_THREADS'] = '1'
+os.environ['MKL_NUM_THREADS'] = '1'
+
+warnings.simplefilter(action='ignore', category=UserWarning)
+warnings.simplefilter(action='ignore', category=FutureWarning)
+
+from sktime.datasets import load_longley
+targets, features = load_longley()
+
+forecasting_horizon = 3
+
+# each series represent an element in the List
+# we take the last forecasting_horizon  as test targets. The itme before that as training targets
+# Normally the value to be forecasted should follow the training sets
+y_train = [targets[: -forecasting_horizon]]
+y_test = [targets[-forecasting_horizon:]]
+
+# same for features. For uni-variant models, X_train, X_test can be omitted
+X_train = [features[: -forecasting_horizon]]
+# Here x_test indicates the 'known future features': they are the features known previously, features that are unknown
+# could be replaced with NAN or zeros (which will not be used by our networks). If no feature is known beforehand,
+# we could also omit X_test
+known_future_features = list(features.columns)
+X_test = [features[-forecasting_horizon:]]
+
+start_times = [targets.index.to_timestamp()[0]]
+freq = '1Y'
+
+from autoPyTorch.api.time_series_forecasting import TimeSeriesForecastingTask
+############################################################################
+# Build and fit a forecaster
+# ==========================
+api = TimeSeriesForecastingTask()
+
+############################################################################
+# Search for an ensemble of machine learning algorithms
+# =====================================================
+api.search(
+    X_train=X_train,
+    y_train=copy.deepcopy(y_train),
+    X_test=X_test,
+    optimize_metric='mean_MASE_forecasting',
+    n_prediction_steps=forecasting_horizon,
+    memory_limit=16 * 1024,  # Currently, forecasting models need much more memories than it actually requires
+    freq=freq,
+    start_times=start_times,
+    func_eval_time_limit_secs=50,
+    total_walltime_limit=60,
+    min_num_test_instances=1000,  # proxy validation sets. This only works for the tasks with more than 1000 series
+    known_future_features=known_future_features,
+)
+
+
+from autoPyTorch.datasets.time_series_dataset import TimeSeriesSequence
+
+test_sets = []
+
+# We could construct test sets from scratch
+for feature, future_feature, target, start_time in zip(X_train, X_test,y_train, start_times):
+    test_sets.append(
+        TimeSeriesSequence(X=feature.values,
+                           Y=target.values,
+                           X_test=future_feature.values,
+                           start_time=start_time,
+                           is_test_set=True,
+                           # additional information required to construct a new time series sequence
+                           **api.dataset.sequences_builder_kwargs
+                           )
+    )
+# Alternatively, if we only want to forecast the value after the X_train, we could directly ask datamanager to
+# generate a test set:
+# test_sets2 = api.dataset.generate_test_seqs()
+
+pred = api.predict(test_sets)

development/_downloads/e2b09b5dc272261aeb82b46c2e2d6ce6/example_time_series_forecasting.py

@@ -134,7 +134,7 @@ Search for an ensemble of machine learning algorithms
  .. code-block:: none
 
 
-    <autoPyTorch.api.tabular_classification.TabularClassificationTask object at 0x7fdc1b398100>
+    <autoPyTorch.api.tabular_classification.TabularClassificationTask object at 0x7f70872c5490>
 
 
 
@@ -166,22 +166,21 @@ Print the final ensemble performance
  .. code-block:: none
 
     {'accuracy': 0.8670520231213873}
-    |    | Preprocessing                                                                                    | Estimator                                                       |   Weight |
-    |---:|:-------------------------------------------------------------------------------------------------|:----------------------------------------------------------------|---------:|
-    |  0 | None                                                                                             | CBLearner                                                       |     0.32 |
-    |  1 | SimpleImputer,Variance Threshold,MinorityCoalescer,OneHotEncoder,QuantileTransformer,KitchenSink | no embedding,ShapedMLPBackbone,FullyConnectedHead,nn.Sequential |     0.2  |
-    |  2 | SimpleImputer,Variance Threshold,NoCoalescer,OneHotEncoder,StandardScaler,SRC                    | embedding,MLPBackbone,FullyConnectedHead,nn.Sequential          |     0.2  |
-    |  3 | SimpleImputer,Variance Threshold,MinorityCoalescer,OneHotEncoder,NoScaler,KitchenSink            | embedding,ResNetBackbone,FullyConnectedHead,nn.Sequential       |     0.12 |
-    |  4 | SimpleImputer,Variance Threshold,MinorityCoalescer,OneHotEncoder,QuantileTransformer,KitchenSink | no embedding,ShapedMLPBackbone,FullyConnectedHead,nn.Sequential |     0.08 |
-    |  5 | SimpleImputer,Variance Threshold,MinorityCoalescer,OneHotEncoder,QuantileTransformer,KitchenSink | no embedding,ShapedMLPBackbone,FullyConnectedHead,nn.Sequential |     0.08 |
+    |    | Preprocessing                                                                         | Estimator                                                       |   Weight |
+    |---:|:--------------------------------------------------------------------------------------|:----------------------------------------------------------------|---------:|
+    |  0 | SimpleImputer,Variance Threshold,NoCoalescer,OneHotEncoder,MinMaxScaler,FastICA       | no embedding,MLPBackbone,FullyConnectedHead,nn.Sequential       |     0.56 |
+    |  1 | SimpleImputer,Variance Threshold,MinorityCoalescer,OneHotEncoder,Normalizer,KernelPCA | embedding,ShapedResNetBackbone,FullyConnectedHead,nn.Sequential |     0.38 |
+    |  2 | SimpleImputer,Variance Threshold,NoCoalescer,NoEncoder,StandardScaler,PCA             | no embedding,MLPBackbone,FullyConnectedHead,nn.Sequential       |     0.02 |
+    |  3 | None                                                                                  | CBLearner                                                       |     0.02 |
+    |  4 | None                                                                                  | SVMLearner                                                      |     0.02 |
     autoPyTorch results:
             Dataset name: Australian
             Optimisation Metric: accuracy
             Best validation score: 0.8713450292397661
-            Number of target algorithm runs: 21
-            Number of successful target algorithm runs: 19
+            Number of target algorithm runs: 28
+            Number of successful target algorithm runs: 27
             Number of crashed target algorithm runs: 0
-            Number of target algorithms that exceeded the time limit: 2
+            Number of target algorithms that exceeded the time limit: 1
             Number of target algorithms that exceeded the memory limit: 0
 
 
@@ -191,7 +190,7 @@ Print the final ensemble performance
 
 .. rst-class:: sphx-glr-timing
 
-   **Total running time of the script:** ( 5 minutes  20.003 seconds)
+   **Total running time of the script:** ( 5 minutes  37.043 seconds)
 
 
 .. _sphx_glr_download_examples_20_basics_example_tabular_classification.py:

development/_downloads/fb625db3c50d423b1b7881136ffdeec8/examples_jupyter.zip

@@ -125,7 +125,7 @@ Search for an ensemble of machine learning algorithms
  .. code-block:: none
 
 
-    <autoPyTorch.api.tabular_regression.TabularRegressionTask object at 0x7fdca57c8b50>
+    <autoPyTorch.api.tabular_regression.TabularRegressionTask object at 0x7f7113dcb220>
 
 
 
@@ -167,12 +167,12 @@ Print the final ensemble performance
     |  2 | SimpleImputer,Variance Threshold,NoCoalescer,OneHotEncoder,StandardScaler,NoFeaturePreprocessing | no embedding,ShapedMLPBackbone,FullyConnectedHead,nn.Sequential |     0.1  |
     |  3 | None                                                                                             | LGBMLearner                                                     |     0.04 |
     autoPyTorch results:
-            Dataset name: 7a5ffe66-f075-11ec-8806-a30cbc8a0bb8
+            Dataset name: a964bba4-f6c8-11ec-87fd-b1d4bc580917
             Optimisation Metric: r2
             Best validation score: 0.8670098636440993
-            Number of target algorithm runs: 24
+            Number of target algorithm runs: 23
             Number of successful target algorithm runs: 22
-            Number of crashed target algorithm runs: 1
+            Number of crashed target algorithm runs: 0
             Number of target algorithms that exceeded the time limit: 1
             Number of target algorithms that exceeded the memory limit: 0
 
@@ -183,7 +183,7 @@ Print the final ensemble performance
 
 .. rst-class:: sphx-glr-timing
 
-   **Total running time of the script:** ( 5 minutes  35.570 seconds)
+   **Total running time of the script:** ( 5 minutes  42.728 seconds)
 
 
 .. _sphx_glr_download_examples_20_basics_example_tabular_regression.py: