Pytorch Forecasting: Understanding 'target' and 'time_varying_unknown_reals' in TimeSeriesDataSet

Question

I made some changes to the code in the tutorial:

from typing import Dict
import torch
from torch import nn
from pytorch_forecasting.models import BaseModel
import numpy as np
import pandas as pd
from pytorch_forecasting import TimeSeriesDataSet
from sklearn.preprocessing import StandardScaler

v1 = np.arange(10)
v2 = np.arange(10)*2
test_data = pd.DataFrame(
    dict(
        value1=v1,
        value2=v2,
        value3=v1 + v2,
        group=np.repeat(np.arange(2), 5),
        time_idx=np.tile(np.arange(5), 2),
    )
)
transfer = StandardScaler()
data_new = transfer.fit_transform(v1.reshape(-1,1))

dataset = TimeSeriesDataSet(
    test_data,
    group_ids=["group"],
    target="value3",
    time_idx="time_idx",
    min_encoder_length=2,
    max_encoder_length=2,
    min_prediction_length=2,
    max_prediction_length=2,
    time_varying_unknown_reals=["value1", "value2", "value3"],
)

print(dataset.get_parameters())

dataloader = dataset.to_dataloader(batch_size=4)

x, y = next(iter(dataloader))
print("x =", x)
print("\ny =", y)
print("\nsizes of x =")
for key, value in x.items():
    print(f"\t{key} = {value.size()}")

The test_data and data_new is:

id	value1	value2	value3	group	time_idx	data_new
0	0	0	0	0	0	-1.5666989
1	1	2	3	0	1	-1.21854359
2	2	4	6	0	2	-0.87038828
3	3	6	9	0	3	-0.52223297
4	4	8	12	0	4	-0.17407766
5	5	10	15	1	0	0.17407766
6	6	12	18	1	1	0.52223297
7	7	14	21	1	2	0.87038828
8	8	16	24	1	3	1.21854359
9	9	18	27	1	4	1.5666989

The output is:

{'time_idx': 'time_idx', 'target': 'value3', 'group_ids': ['group'], 
'weight': None, 'max_encoder_length': 2, 'min_encoder_length': 2, 
'min_prediction_idx': 0, 'min_prediction_length': 2, 'max_prediction_length': 2,
'time_varying_unknown_reals': ['value1', 'value2', 'value3'], 
'scalers': {'value1': StandardScaler(), 'value2': StandardScaler()}}


x = {'encoder_cat': tensor([], size=(4, 2, 0), dtype=torch.int64), 'encoder_cont': tensor([[[ 0.5222,  0.5222,  6.0000],
         [ 0.8704,  0.8704,  7.0000]],

        [[-1.2185, -1.2185,  1.0000],
         [-0.8704, -0.8704,  2.0000]],

        [[-1.5667, -1.5667,  0.0000],
         [-1.2185, -1.2185,  1.0000]],

        [[ 0.1741,  0.1741,  5.0000],
         [ 0.5222,  0.5222,  6.0000]]]), 'encoder_target': tensor([[6, 7],
        [1, 2],
        [0, 1],
        [5, 6]]), 'encoder_lengths': tensor([2, 2, 2, 2]), 'decoder_cat': tensor([], size=(4, 2, 0), dtype=torch.int64), 'decoder_cont': tensor([[[ 1.2185,  1.2185,  8.0000],
         [ 1.5667,  1.5667,  9.0000]],

        [[-0.5222, -0.5222,  3.0000],
         [-0.1741, -0.1741,  4.0000]],

        [[-0.8704, -0.8704,  2.0000],
         [-0.5222, -0.5222,  3.0000]],

        [[ 0.8704,  0.8704,  7.0000],
         [ 1.2185,  1.2185,  8.0000]]]), 'decoder_target': tensor([[8, 9],
        [3, 4],
        [2, 3],
        [7, 8]]), 'decoder_lengths': tensor([2, 2, 2, 2]), 'decoder_time_idx': tensor([[3, 4],
        [3, 4],
        [2, 3],
        [2, 3]]), 'groups': tensor([[1],
        [0],
        [0],
        [1]]), 'target_scale': tensor([[0., 0.],
        [0., 0.],
        [0., 0.],
        [0., 0.]])}

y = (tensor([[8, 9],
        [3, 4],
        [2, 3],
        [7, 8]]), None)

sizes of x =
    encoder_cat = torch.Size([4, 2, 0])
    encoder_cont = torch.Size([4, 2, 3])
    encoder_target = torch.Size([4, 2])
    encoder_lengths = torch.Size([4])
    decoder_cat = torch.Size([4, 2, 0])
    decoder_cont = torch.Size([4, 2, 3])
    decoder_target = torch.Size([4, 2])
    decoder_lengths = torch.Size([4])
    decoder_time_idx = torch.Size([4, 2])
    groups = torch.Size([4, 1])
    target_scale = torch.Size([4, 2])

The problem is 'target': 'value3' and 'time_varying_unknown_reals': ['value1', 'value2', 'value3']

In my opinion, this is an autoregressive problem: Covariant1, Covariant2, X(t) -> X(t+1).

If value3 is removed in time_varying_unknown_reals, it is a prediction problem: Covariant1, Covariant2 -> X

the problem is the output of the encoder_cont, its shape is (4, 2, 3) for (batch_size, encoder_len, unknown_reals_dim).

It seems that the first and second columns are the result of value1 and value2 standardization, but the third column is not value3, it is id of test_data.

I don't understand. In my opinion, the output should be:

'encoder_cont': tensor([[[ 0.5222, 0.5222, 18], [ 0.8704, 0.8704, 21]], ...

'encoder_target': tensor([[18, 21] ...

Decoder and y are the same problem, I want to predict value3 instead of id