'tuple' object has no attribute 'items' in models

[pjbhaumik]

• PyTorch-Forecasting version: 0.10.3
• PyTorch version: 1.10.2+cpu
• Python version: 3.8
• Operating System: jupyterlab notebook

Expected behavior

I executed code Baseline().predict(val_dataloader) in order to get a baseline model and expected to get result baseline predictions

Actual behavior

However, result was

pytorch lightning returns a tuple from get_init_args(), so it won't have .items() attribute

Code to reproduce the problem

training = TimeSeriesDataSet(
df[lambda x: x.time_idx <= training_cutoff],
time_idx="time_idx",
target="NPD_LTR_nps",
group_ids=["Label"],
min_encoder_length=max_encoder_length // 2, # keep encoder length long (as it is in the validation set)
max_encoder_length=max_encoder_length,
min_prediction_length=1,
max_prediction_length=max_prediction_length,
static_categoricals=["Label"],
time_varying_known_reals=["time_idx"],
time_varying_unknown_reals= df.drop(columns = ['time_idx', 'Label', 'FLT_LEG_SCHD_DPRT_LDT']).columns.to_list(),
add_relative_time_idx=True,
add_target_scales=True,
add_encoder_length=True,
)

validation = TimeSeriesDataSet.from_dataset(training, df, predict=True, stop_randomization=True)

batch_size = 128 # set this between 32 to 128
train_dataloader = training.to_dataloader(train=True, batch_size=batch_size, num_workers=0)
val_dataloader = validation.to_dataloader(train=False, batch_size=batch_size * 10, num_workers=0)

actuals = torch.cat([y for x, (y, weight) in iter(val_dataloader)])
baseline_predictions = Baseline().predict(val_dataloader)
(actuals - baseline_predictions).abs().mean().item()

[Songloading]

I am getting same error from DeepAR.from_dataset

[echolql]

I got same error from TemporalFusionTransformer.from_dataset

PyTorch-Forecasting version: 0.10.3
PyTorch version: 1.13.1+gpu
Python version: 3.10
Operating System: jupyterlab notebook
pytorch-lightning-1.9.1

[bgrey001]

I am getting the same error for DeepAR.from_set and NBeats.from_dataset

[Da-byte]

I am getting the same error on Baseline, DeepAR, NHiTS and NBeats models. The error seems to have begun occurring 24h ago, whilst the same code worked perfectly 30h ago for me.

[robinkoppe]

I have the same error with TemporalFusionTransformer.from_dataset.

PyTorch-Forecasting V 0.10.3
Python V 3.10.9
Torch V 1.13.1

[robinkoppe]

Seems like pytorch_lightning had an update yesterday (see link https://pypi.org/project/pytorch-lightning/#history). Reverting to version 1.9.0 worked for me.

use pip install pytorch_lightning==1.9.0

[bgrey001]

Seems like pytorch_lightning had an update yesterday (see link https://pypi.org/project/pytorch-lightning/#history). Reverting to version 1.9.0 worked for me.

use pip install pytorch_lightning==1.9.0

Can confirm this solves this issue seemingly, but I am faced with a myriad of other errors only trying to run the tutorial code. Has anyone been facing numerous bugs?

[anoetzli]

Seems like pytorch_lightning had an update yesterday (see link https://pypi.org/project/pytorch-lightning/#history). Reverting to version 1.9.0 worked for me.
use pip install pytorch_lightning==1.9.0

Can confirm this solves this issue seemingly, but I am faced with a myriad of other errors only trying to run the tutorial code. Has anyone been facing numerous bugs?

Can confirm that downgrading pytorch_lightning to 1.9.0 resolved the issue, but I also get a myriad of other issues just running through the tutorial code, as the previous user also observed.

Yesterday I spent hours setting up different environments (conda, virtualenv) and attempted to setup an environment using different package managers (pip, conda), and could not proceed with the tutorials.

It would be helpful if someone could share a requirements file of an environment that actually works!

[pjbhaumik]

Hi, I am closing this issue, since the fix is to use the previous version of pytorch lightning per @robinkoppe.

However, I am facing several new issues as are others. Like @Da-byte, my code was working successfully before last Friday....

I am suggesting the use of a new module called darts (https://pypi.org/project/darts/) which implements TFT, DeepAR, NBEATS, NHiiTS and other methods such as AARIMA, ES etc...

[l-cdc]

This should be fixed by Lightning-AI/pytorch-lightning#16851, which was included in pytorch-lightning 1.9.4 and 2.0.0. I could not test either yet but maybe that's useful.