Chronos-2: Add support for specifying `freq` strings directly

[GGA-PERSO]

I know the freq of my short time series (with less than 3 points) but I assume you are trying to infer the freq in convert_df_input_to_list_of_dicts_input
could you please add a freq parameter to avoid to infer the freq when we know it ?

[abdulfatir]

@GGA-PERSO chronos-forecasting==2.2 adds an option to disable dataframe validation using validate_inputs=False. Does this address your use case? If not, I would recommend using the predict_quantiles API instead. That does not need timestamps and accepts numpy/torch arrays.

[GGA-PERSO]

@abdulfatir no it doesn't as you also infer the freq when validate_inputs=False
freq = pd.infer_freq(timestamps)

[abdulfatir]

@GGA-PERSO Are all time series shorter than 3 timesteps?

[GGA-PERSO]

yes as I do inference on only one time serie and this time serie has less than 3 historical data point wich is not enough to infer the freq even if I know it

[abdulfatir]

@GGA-PERSO In that case, please use predict_quantiles. Here is an example, if you are working with pandas dataframes.

import torch
import pandas as pd
from chronos import Chronos2Pipeline

LENGTH = 2
pipeline = Chronos2Pipeline.from_pretrained("amazon/chronos-2", device_map="cpu")
df = pd.DataFrame(
    {"item_id": "T1", "target": torch.rand(LENGTH), "timestamp": pd.date_range("2025", freq="h", periods=LENGTH)}
)

# The following will raise an error
# pipeline.predict_df(df, prediction_length=2, quantile_levels=[0.1, 0.5, 0.9], validate_inputs=False)

inputs = [
    {
        "target": group_df["target"].to_numpy().reshape(1, -1),  # Add leading "variate" dimension
    }
    for _, group_df in df.groupby("item_id")
]
quantiles, _ = pipeline.predict_quantiles(inputs, prediction_length=2, quantile_levels=[0.1, 0.5, 0.9])
quantiles = torch.stack(quantiles, dim=0)  # (batch_size=1, num_variates=1, prediction_length=2, num_quantiles=3)

That said, I am not so sure about using the model for a timeseries with a context length of 2. What would be a "good prediction" here?

[GGA-PERSO]

I am building an app where end user can select and display a product before to run a forecast.

there is no good prediction indeed for this specific case and to be honest I don't really care about the forecast here.
I just wanted an unique solution (predict_df) without a crash that could work for all my products individually

I thought that passing the freq parameter when it's known was a good compromise
But you don't seem to want to add it.
So let's close the discussion.

[abdulfatir]

@GGA-PERSO Just to provide additional context: Dealing with freq strings (properly) is often a can of worms, given the significant differences in pandas versions. That's the reason we are hesitant to add this functionality for an edge case. That said, we take your feedback and might consider this in the next minor release after discussing internally. cc @shchur

[GGA-PERSO]

If it could help you here what nixtla does
https://github.com/Nixtla/nixtla/blob/289fdd6636fb80892a348242adfab3af7b75d105/nixtla/nixtla_client.py#L228