@Cameronwood611 thanks for this; it seems like this fix is breaking the behaviour of models, so I'm not sure it's the right way to go. Is there any way you could share the model you're using for #2947, for debugging purposes? Or even a simpler, smaller dummy model that just runs into the issue. What kind of model are we talking about?

@Cameronwood611 thanks for this; it seems like this fix is breaking the behaviour of models, so I'm not sure it's the right way to go. Is there any way you could share the model you're using for #2947, for debugging purposes? Or even a simpler, smaller dummy model that just runs into the issue. What kind of model are we talking about?

Yes! I updated the issue linked but I'll also post the example here --sorry about that. I noticed I had to raise the for loop above 1000 iterations (around the default recursion depth) for it to break; not sure if that's related to different model params, not loading in the model, etc. I'm not sure of the side effects this change has but maybe we could look into a certain filter or adding SelectFields into a set? Would be willing to help with this effort.

import random
import numpy as np
import pandas as pd
from gluonts.torch import PyTorchPredictor
from gluonts.dataset.common import ListDataset, DataEntry
from gluonts.torch import TemporalFusionTransformerEstimator


estimator = TemporalFusionTransformerEstimator(
freq='S',
context_length=10,
prediction_length=5,
trainer_kwargs={'max_epochs': 5}
)
train = ListDataset([{'target': np.array([random.uniform(-1, 1) for _ in range(50)]), 'start': pd.Period('01-01-2023', freq='S')} for _ in range(20)], freq='S')

pred = ListDataset([{'target': np.array([random.uniform(-1, 1) for _ in range(50)]), 'start': pd.Period('01-01-2023', freq='S')} for _ in range(20)], freq='S')
model = estimator.train(train, pred)

x, y, z = np.array([x for x in range(10)]), np.array([y for y in range (10)]), np.array([z for z in range(10)]) #for example
start = pd.Period('01-01-2023', freq='S')
_input = ListDataset([{'target': x, 'start': start},
                          {'target': y, 'start': start},
                          {'target': z, 'start': start}], freq='S')
for i in range(5000): #used to be 1000 for it to break..
    pred = model.predict(_input)
    output = [sig.mean for sig in pred]  #this calls on a generator which causes the recursion error 

@Cameronwood611 thanks for the bug hunting and the fix, if you update it as suggested then it should work fine, and we'll backport it and release it soon

@Cameronwood611 thanks for the bug hunting and the fix, if you update it as suggested then it should work fine, and we'll backport it and release it soon

Thank you!