Merge pull request #415 from WenjieDu/(feat)add_stemgnn

Add StemGNN modules and implement it as an imputation model

pypots/imputation/__init__.py

@@ -31,6 +31,7 @@
 from .koopa import Koopa
 from .micn import MICN
 from .tide import TiDE
+from .stemgnn import StemGNN
 
 # naive imputation methods
 from .locf import LOCF
@@ -64,6 +65,7 @@
     "Koopa",
     "MICN",
     "TiDE",
+    "StemGNN",
     # naive imputation methods
     "LOCF",
     "Mean",

pypots/imputation/stemgnn/__init__.py

@@ -0,0 +1,24 @@
+"""
+The package of the partially-observed time-series imputation model StemGNN.
+
+Refer to the paper
+`Defu Cao, Yujing Wang, Juanyong Duan, Ce Zhang, Xia Zhu, Congrui Huang, Yunhai Tong, Bixiong Xu, Jing Bai, Jie Tong, Qi Zhang.
+"Spectral Temporal Graph Neural Network for Multivariate Time-series Forecasting".
+In Advances in Neural Information Processing Systems, 2020.
+<https://proceedings.neurips.cc/paper_files/paper/2020/file/cdf6581cb7aca4b7e19ef136c6e601a5-Paper.pdf>`_
+
+Notes
+-----
+This implementation is inspired by the official one https://github.com/microsoft/StemGNN
+
+"""
+
+# Created by Wenjie Du <wenjay.du@gmail.com>
+# License: BSD-3-Clause
+
+
+from .model import StemGNN
+
+__all__ = [
+    "StemGNN",
+]

pypots/imputation/stemgnn/core.py

@@ -0,0 +1,82 @@
+"""
+The core wrapper assembles the submodules of StemGNN imputation model
+and takes over the forward progress of the algorithm.
+"""
+
+# Created by Wenjie Du <wenjay.du@gmail.com>
+# License: BSD-3-Clause
+
+import torch.nn as nn
+
+from ...nn.modules.stemgnn import BackboneStemGNN
+from ...nn.modules.saits import SaitsLoss, SaitsEmbedding
+
+
+class _StemGNN(nn.Module):
+    def __init__(
+        self,
+        n_steps,
+        n_features,
+        n_layers,
+        n_stacks,
+        d_model,
+        dropout_rate=0.5,
+        leaky_rate=0.2,
+        ORT_weight: float = 1,
+        MIT_weight: float = 1,
+    ):
+        super().__init__()
+
+        self.n_steps = n_steps
+
+        self.saits_embedding = SaitsEmbedding(
+            n_features * 2,
+            d_model,
+            with_pos=False,
+        )
+        self.backbone = BackboneStemGNN(
+            units=d_model,
+            stack_cnt=n_stacks,
+            time_step=n_steps,
+            multi_layer=n_layers,
+            horizon=n_steps,
+            dropout_rate=dropout_rate,
+            leaky_rate=leaky_rate,
+        )
+
+        # for the imputation task, the output dim is the same as input dim
+        self.output_projection = nn.Linear(d_model, n_features)
+        self.saits_loss_func = SaitsLoss(ORT_weight, MIT_weight)
+
+    def forward(self, inputs: dict, training: bool = True) -> dict:
+        X, missing_mask = inputs["X"], inputs["missing_mask"]
+
+        # WDU: the original StemGNN paper isn't proposed for imputation task. Hence the model doesn't take
+        # the missing mask into account, which means, in the process, the model doesn't know which part of
+        # the input data is missing, and this may hurt the model's imputation performance. Therefore, I apply the
+        # SAITS embedding method to project the concatenation of features and masks into a hidden space, as well as
+        # the output layers to project back from the hidden space to the original space.
+        enc_out = self.saits_embedding(X, missing_mask)
+
+        # StemGNN encoder processing
+        enc_out, _ = self.backbone(enc_out)
+        # project back the original data space
+        reconstruction = self.output_projection(enc_out)
+
+        imputed_data = missing_mask * X + (1 - missing_mask) * reconstruction
+        results = {
+            "imputed_data": imputed_data,
+        }
+
+        # if in training mode, return results with losses
+        if training:
+            X_ori, indicating_mask = inputs["X_ori"], inputs["indicating_mask"]
+            loss, ORT_loss, MIT_loss = self.saits_loss_func(
+                reconstruction, X_ori, missing_mask, indicating_mask
+            )
+            results["ORT_loss"] = ORT_loss
+            results["MIT_loss"] = MIT_loss
+            # `loss` is always the item for backward propagating to update the model
+            results["loss"] = loss
+
+        return results

pypots/imputation/stemgnn/data.py

@@ -0,0 +1,24 @@
+"""
+Dataset class for StemGNN.
+"""
+
+# Created by Wenjie Du <wenjay.du@gmail.com>
+# License: BSD-3-Clause
+
+from typing import Union
+
+from ..saits.data import DatasetForSAITS
+
+
+class DatasetForStemGNN(DatasetForSAITS):
+    """Actually StemGNN uses the same data strategy as SAITS, needs MIT for training."""
+
+    def __init__(
+        self,
+        data: Union[dict, str],
+        return_X_ori: bool,
+        return_y: bool,
+        file_type: str = "hdf5",
+        rate: float = 0.2,
+    ):
+        super().__init__(data, return_X_ori, return_y, file_type, rate)