Accept optional date in process_parquet for more precise subsetting

[kvantricht]

Currently we take end_date of training data and go back one year to subset training time series:

presto-worldcereal/presto/inference.py

Lines 397 to 410 in e8d5bbc

	# df['start_date'] = seasons.get_season_start(df[['lat','lon']])
	# For now, in absence of a relevant start_date, we get time difference with respect
	# to end_date so we can take 12 months counted back from end_date
	df["valid_date_ind"] = (
	(((df["timestamp"] - df["end_date"]).dt.days + 365) / 30).round().astype(int)
	)
	# Now reassign start_date to the actual subset counted back from end_date
	df["start_date"] = df["end_date"] - pd.DateOffset(years=1) + pd.DateOffset(days=1)
	df_pivot = df[(df["valid_date_ind"] >= 0) & (df["valid_date_ind"] < 12)].pivot(
	index=index_columns, columns="valid_date_ind", values=feature_columns
	)

However, if we train a dedicated CatBoost on a subset of data for a small AOI, we may benefit from subsetting based on the requested start_date and end_date (has to be one year) by the user. Can we adapt the method to accept an optional argument, e.g. end_date which - if given - dictates the subsetting of the timeseries?

We then need to be careful to be resilient to different years of the training data, and also drop samples that don't fall entirely within the requested time frame (adapted for the year of the sample).

[kvantricht]

@cbutsko I think we can close this? It's tackled on worldcereal-classification side?

[cbutsko]

This will be tackled in this issue