out_dim or n_features

[ZhixuanLiu]

mtad-gat-pytorch/mtad_gat.py

Line 62 in 8f907cc

self.recon_model = ReconstructionModel(window_size, gru_hid_dim, recon_hid_dim, out_dim, recon_n_layers, dropout)

self.recon_model = ReconstructionModel(window_size, gru_hid_dim, recon_hid_dim, out_dim, recon_n_layers, dropout)
should the "out_dim" be changed to "n_features" to match the shape of input x while the loss is calculated by MSELoss(recons, x) ??

[axeloh]

We do it like this because we do not always want to forecast and reconstruct all input features.
For example, for SMAP and MSL datasets, the input consists of one-hot encoded data plus one (continuous) telemetry value. So here we do not want to forecast and reconstruct the one-hot encoded features, only the telemetry value.

mtad-gat-pytorch/utils.py

Lines 40 to 53 in 8f907cc

	def get_target_dims(dataset):
	"""
	:param dataset: Name of dataset
	:return: index of data dimension that should be modeled (forecasted and reconstructed),
	returns None if all input dimensions should be modeled
	"""
	if dataset == "SMAP":
	return [0]
	elif dataset == "MSL":
	return [0]
	elif dataset == "SMD":
	return None
	else:
	raise ValueError("unknown dataset " + str(dataset))

That is why we for MSL and SMAP use [0] as target dim (dim of the telemetry value).

In train.py the out_dim is set based on this:

mtad-gat-pytorch/train.py

Lines 54 to 70 in 8f907cc

	n_features = x_train.shape[1]
	target_dims = get_target_dims(dataset)
	if target_dims is None:
	out_dim = n_features
	print(f"Will forecast and reconstruct all {n_features} input features")
	elif type(target_dims) == int:
	print(f"Will forecast and reconstruct input feature: {target_dims}")
	out_dim = 1
	else:
	print(f"Will forecast and reconstruct input features: {target_dims}")
	out_dim = len(target_dims)
	train_dataset = SlidingWindowDataset(x_train, window_size, target_dims)
	test_dataset = SlidingWindowDataset(x_test, window_size, target_dims)
	train_loader, val_loader, test_loader = create_data_loaders(

To get correct loss during training (forecastings and recons are compared against the correct input features) we must also check for this:

mtad-gat-pytorch/training.py

Lines 111 to 115 in 8f907cc

	preds, recons = self.model(x)
	if self.target_dims is not None:
	x = x[:, :, self.target_dims]
	y = y[:, :, self.target_dims].squeeze(-1)

[ZhixuanLiu]

Thanks for the explanation. I did miss that part in train.py. It helps a lot.