try: replicate og vanilla ae

Signed-off-by: Avik Basu <[email protected]>

numalogic/models/autoencoder/variants/vanilla.py

@@ -59,7 +59,7 @@ def _construct_layers(self, layersizes: Sequence[int]) -> nn.ModuleList:
  layers.extend(
  [
  nn.Linear(start_layersize, lsize),
- nn.BatchNorm1d(self.n_features),
+ # nn.BatchNorm1d(self.n_features),
  nn.Tanh(),
  nn.Dropout(p=self.dropout_p),
  ]
@@ -69,8 +69,10 @@ def _construct_layers(self, layersizes: Sequence[int]) -> nn.ModuleList:
  layers.extend(
  [
  nn.Linear(start_layersize, layersizes[-1]),
- nn.BatchNorm1d(self.n_features),
- nn.ReLU(),
+ # nn.BatchNorm1d(self.n_features),
+ # nn.ReLU(),
+ nn.Tanh(),
+ nn.Dropout(p=self.dropout_p),
  ]
  )
  return layers
@@ -120,7 +122,7 @@ def _construct_layers(self, layersizes: Sequence[int]) -> nn.ModuleList:
  layers.extend(
  [
  nn.Linear(layersizes[idx], layersizes[idx + 1]),
- nn.BatchNorm1d(self.n_features),
+ # nn.BatchNorm1d(self.n_features),
  nn.Tanh(),
  nn.Dropout(p=self.dropout_p),
  ]
@@ -190,14 +192,14 @@ def forward(self, batch: Tensor) -> tuple[Tensor, Tensor]:
  decoded = self.decoder(encoded)
  return encoded, torch.swapdims(decoded, 1, 2)
 
- def _get_reconstruction_loss(self, batch: Tensor):
+ def _get_reconstruction_loss(self, batch: Tensor, reduction="mean") -> Tensor:
  _, recon = self.forward(batch)
- return self.criterion(batch, recon)
+ return 0.5 * self.criterion(batch, recon, reduction=reduction)
 
  def predict_step(self, batch: Tensor, batch_idx: int, dataloader_idx: int = 0):
  """Returns reconstruction for streaming input."""
  recon = self.reconstruction(batch)
- return self.criterion(batch, recon, reduction="none")
+ return 0.5 * self.criterion(batch, recon, reduction="none")
 
 
 class SparseVanillaAE(VanillaAE):

numalogic/tools/data.py

@@ -191,7 +191,7 @@ def __getitem__(self, idx: Union[int, slice]) -> npt.NDArray[float]:
  return np.stack(output)
  if idx >= len(self):
  raise IndexError(f"{idx} out of bound!")
- return self._data[idx : idx + self._seq_len]
+ return self._data[(idx * self._stride) : (idx * self._stride) + self._seq_len]
 
 
 class StreamingDataLoader(DataLoader):

numalogic/udfs/inference.py

@@ -84,8 +84,7 @@ def compute(cls, model: artifact_t, input_: npt.NDArray[float], **_) -> npt.NDAr
  model.eval()
  try:
  with torch.no_grad():
- _, out = model.forward(x)
- recon_err = model.criterion(out, x, reduction="none")
+ recon_err = model._get_reconstruction_loss(x, reduction="none")
  except Exception as err:
  raise RuntimeError("Model forward pass failed!") from err
  return np.ascontiguousarray(recon_err).squeeze(0)

numalogic/udfs/trainer/_base.py

@@ -33,6 +33,7 @@
 )
 from numalogic.udfs.entities import TrainerPayload
 from numalogic.udfs.tools import TrainMsgDeduplicator
+import torch
 
 _struct_log = configure_logger()
 
@@ -120,8 +121,12 @@ def compute(
  model, train_dataloaders=DataLoader(train_ds, batch_size=trainer_cfg.batch_size)
  )
  train_reconerr = trainer.predict(
- model, dataloaders=DataLoader(train_ds, batch_size=trainer_cfg.batch_size)
- ).numpy()
+ model,
+ dataloaders=DataLoader(train_ds, batch_size=trainer_cfg.batch_size),
+ unbatch=False,
+ )
+ train_reconerr = torch.mean(train_reconerr, dim=1).numpy()
+
  dict_artifacts["inference"] = KeyedArtifact(
  dkeys=[numalogic_cfg.model.name], artifact=model, stateful=numalogic_cfg.model.stateful
  )
@@ -233,12 +238,12 @@ def exec(self, keys: list[str], datum: Datum) -> Messages:
  _add_summary(
  summary=NAN_SUMMARY,
  labels=_metric_label_values,
- data=nan_counter,
+ data=np.sum(nan_counter),
  )
  _add_summary(
  summary=INF_SUMMARY,
  labels=_metric_label_values,
- data=inf_counter,
+ data=np.sum(inf_counter),
  )
 
  # Initialize artifacts
@@ -363,7 +368,7 @@ def get_feature_arr(
  raw_df: pd.DataFrame,
  metrics: list[str],
  fill_value: float = 0.0,
- ) -> tuple[npt.NDArray[float], float, float]:
+ ) -> tuple[npt.NDArray[float], pd.Series, pd.Series]:
  """
  Get feature array from the raw dataframe.
 
@@ -378,14 +383,15 @@ def get_feature_arr(
  nan_counter: Number of nan values
  inf_counter: Number of inf values
  """
- nan_counter = 0
- for col in metrics:
+ nan_counter = np.zeros(len(metrics), dtype=int)
+ inf_counter = np.zeros(len(metrics), dtype=int)
+ for idx, col in enumerate(metrics):
  if col not in raw_df.columns:
  raw_df[col] = fill_value
- nan_counter += len(raw_df)
+ nan_counter[idx] += len(raw_df)
  feat_df = raw_df[metrics]
- nan_counter += raw_df.isna().sum().all()
- inf_counter = np.isinf(feat_df).sum().all()
+ nan_counter += feat_df.isna().sum()
+ inf_counter = np.isinf(feat_df).sum()
  feat_df = feat_df.fillna(fill_value).replace([np.inf, -np.inf], fill_value)
  return feat_df.to_numpy(dtype=np.float32), nan_counter, inf_counter