How to properly fix random seed with TFT?

Expected behavior

Hello guys I wonder how to fix seed to get reproducibility of my experiments

Right now I’m using this function before the start of the training pl.seed_everything(42)

Actual behavior

However, it doesn’t work. It is caused by dropout because when I set dropout = 0 it had worked.

# configure network and trainer
pl.seed_everything(42)
trainer = pl.Trainer(
    deterministic=True,
    gpus=[0],
    gradient_clip_val=0.1,
)


tft = TemporalFusionTransformer.from_dataset(
    training,
    learning_rate=0.03,
    hidden_size=128,
    attention_head_size=2,
    dropout=0.3,  
    hidden_continuous_size=32,  
    output_size=7,  
    loss=QuantileLoss(),
    reduce_on_plateau_patience=4,
)

trainer.fit(
    tft,
    train_dataloader=train_dataloader,
    val_dataloaders=val_dataloader,
)

So, how do I do to fix this issue?

Thanks,

This is my colab notebook.

[nessessence]

Hi, could you try this?: pl.seed_everything(42,workers=True)
// It works for me

[monkiravn]

What about adding shuffle=False in to_dataloader()?

Hi Luming Sun, Thanks for your suggestion, I have tried but it hasn’t worked too.