✨ Add new modules (#66)

* ✨ Add new model

* ✨ Add new images

* ✨ Add new utility function

* ✨ Add new utility function

* ✨ Add new utility function

* 🍎 Adjust default cfg values

README.md

@@ -2,7 +2,7 @@
   <img src="img/tsts-logo.png" width="600"/>
 </div>
 
-[![pypi](https://img.shields.io/pypi/v/tsts?style=flat)](https://pypi.org/project/tsts/0.4.0/)
+[![pypi](https://img.shields.io/pypi/v/tsts?style=flat)](https://pypi.org/project/tsts/0.6.0/)
 [![license](https://img.shields.io/github/license/TakuyaShintate/tsts?style=flat)](https://github.com/TakuyaShintate/tsts/blob/main/LICENSE)
 
 ([Docs](https://takuyashintate.github.io/tsts/))([Benchmark](https://github.com/TakuyaShintate/tsts/tree/main/benchmark/))
@@ -61,14 +61,18 @@ Following modules are supported.
     </tr>
     <tr>
       <td>
+        <li>AR</li>
+        <li>HistricalInertia</li>
         <li>Informer</li>
         <li>NBeats</li>
         <li>Seq2Seq</li>
       </td>
       <td>
         <li>DILATE</li>
+        <li>MAE</li>
         <li>MAPE</li>
         <li>MSE</li>
+        <li>SmoothMAE</li>
       </td>
       <td>
         <li>MAE</li>
@@ -78,6 +82,8 @@ Following modules are supported.
       </td>
       <td>
         <li>Adam</li>
+        <li>AdamW</li>
+        <li>SAM</li>
         <li>SGD</li>
       </td>
       <td>
@@ -131,13 +137,16 @@ For inference, it needs to load parameters from a log directory generated in tra
 import torch
 from tsts.scalers import StandardScaler
 from tsts.solvers import TimeSeriesForecaster
+from tsts.utils import plot
 
 # Initialize scaler with training dataset
 sin_dataset = torch.sin(torch.arange(0.0, 100.0, 0.1))
 sin_dataset = sin_dataset.unsqueeze(-1)
-X_scaler = StandardScaler()
 # NOTE: 0.75 is default training validation dataset ratio (training: 0.75, validation: 0.25)
 num_train_samples = int(0.75 * len(sin_dataset))
+sin_dataset = sin_dataset[:num_train_samples]
+
+X_scaler = StandardScaler()
 X_scaler.fit(sin_dataset)
 
 # Define test dataset
@@ -150,14 +159,19 @@ forecaster = TimeSeriesForecaster("cfg.yml")
 Z = forecaster.predict(X)
 
 # Initialize target to compare prediction with it
-y = torch.sin(torch.arange(110.0, 110.8, 0.1)).unsqueeze(1)
+y = torch.sin(torch.arange(110.0, 114.8, 0.1)).unsqueeze(1)
 y = X_scaler.transform(y)
 
 # Result
-print(Z)
-print(y)
+plot(Z, y, xticks=torch.arange(110.0, 114.8, 0.1))
 ```
 
+Result:
+
+<div align="center">
+  <img src="img/result-getting-started.png" width="700"/>
+</div>
+
 ## Examples
 
 See [Benchmark](https://github.com/TakuyaShintate/tsts/tree/main/benchmark/) for advanced usage.

benchmark/histricalinertia/README.md

@@ -0,0 +1,22 @@
+# Historical Inertia: A Neglected but Powerful Baseline for Long Sequence Time-series Forecasting
+
+## Citation
+
+```
+@article{DBLP:journals/corr/abs-2103-16349,
+  author    = {Yue Cui and
+               Jiandong Xie and
+               Kai Zheng},
+  title     = {Historical Inertia: An Ignored but Powerful Baseline for Long Sequence
+               Time-series Forecasting},
+  journal   = {CoRR},
+  volume    = {abs/2103.16349},
+  year      = {2021},
+  url       = {https://arxiv.org/abs/2103.16349},
+  eprinttype = {arXiv},
+  eprint    = {2103.16349},
+  timestamp = {Wed, 07 Apr 2021 15:31:46 +0200},
+  biburl    = {https://dblp.org/rec/journals/corr/abs-2103-16349.bib},
+  bibsource = {dblp computer science bibliography, https://dblp.org}
+}
+```

benchmark/histricalinertia/hi-ett-h1.yml

@@ -0,0 +1,30 @@
+DEVICE: "cuda:0"
+
+IO:
+  LOOKBACK: 720
+  HORIZON: 48
+
+TRAINING:
+  TRAIN_DATA_SPLIT: "col"
+  TRAIN_DATA_RATIO: 0.75
+  NUM_EPOCHS: 1
+
+MODEL:
+  NAME: "HistricalInertia"
+
+DATASET:
+  BASE_START_INDEX: 48
+  NORM_PER_DATASET: False
+
+DATALOADER:
+  BATCH_SIZE_TRAIN: 32
+
+METRICS:
+  NAMES: ["MAE", "MSE"]
+  ARGS: [{}, {}]
+
+SCALER:
+  NAME: "StandardScaler"
+
+LOGGER:
+  LOG_DIR: "hi-ett-h1-sam"

benchmark/histricalinertia/test_hi_ett_h1.py

@@ -0,0 +1,45 @@
+# TODO: Add unnormalized results
+
+import pandas as pd
+import torch
+from tqdm import tqdm
+from tsts.metrics import MAE, MSE
+from tsts.scalers import StandardScaler
+from tsts.solvers import TimeSeriesForecaster
+
+lookback = 720
+horizon = 48
+start = 12 * 30 * 24 + 4 * 30 * 24 + lookback
+end = 12 * 30 * 24 + 8 * 30 * 24 + lookback - horizon
+
+X = pd.read_csv("/path/to/ETTh1.csv")
+X = X[["OT"]]
+X = X.values
+X = torch.tensor(X, dtype=torch.float32)
+
+solver = TimeSeriesForecaster("./hi-ett-h1.yml")
+
+# Initialize scalers with training dataset
+num_train_samples = int(0.75 * (12 * 30 * 24 + 4 * 30 * 24))
+X_scaler = StandardScaler()
+X_scaler.fit(X[:num_train_samples])
+y_scaler = StandardScaler()
+y_scaler.fit(X[:num_train_samples])
+
+metric1 = MSE()
+metric2 = MAE()
+
+for i in tqdm(range(start, end)):
+    x = X[i - lookback : i]
+    y = X[i : i + horizon]
+    y_mask = torch.ones_like(y)
+    x = X_scaler.transform(x)
+    y = y_scaler.transform(y)
+    Z = solver.predict(x)
+    metric1.update(Z, y, y_mask)
+    metric2.update(Z, y, y_mask)
+
+score1 = metric1()
+score2 = metric2()
+print(f"MSE: {score1}")
+print(f"MAE: {score2}")

benchmark/histricalinertia/train_hi_ett_h1.py

@@ -0,0 +1,15 @@
+import pandas as pd
+import torch
+from tsts.solvers import TimeSeriesForecaster
+
+start = 0
+end = 12 * 30 * 24 + 4 * 30 * 24
+
+X = pd.read_csv("/path/to/ETTh1.csv")
+X = X[["OT"]]
+X = X.values
+X = X[start:end]
+X = torch.tensor(X, dtype=torch.float32)
+
+solver = TimeSeriesForecaster("./hi-ett-h1.yml")
+solver.fit([X])

docs/.buildinfo

@@ -1,4 +1,4 @@
 # Sphinx build info version 1
 # This file hashes the configuration used when building these files. When it is not found, a full rebuild will be done.
-config: 74b06218c3307010ebd142fcaea93044
+config: 293831d798069392d74282626242223e
 tags: 645f666f9bcd5a90fca523b33c5a78b7

docs/_sources/index.rst.txt

@@ -4,9 +4,10 @@ Tutorials
 
 .. toctree::
     :maxdepth: 1
-    :caption: Models
+    :caption: Basic
 
     tutorials/training
+    tutorials/inference
 
 ====
 Docs