[BUG] Check pytorch-forecasting models with freq > "1D"

[martins0n]

🐛 Bug Report

PytorchForecastingTransform._calculate_freq_unit not work correctly with 'MS' freq for example

Expected behavior

It should work correctly for any defualt pandas frequency.

How To Reproduce

import pandas as pd

from etna.datasets.tsdataset import TSDataset
from etna.pipeline import Pipeline
from etna.transforms import DateFlagsTransform
from etna.transforms import LagTransform
from etna.metrics import SMAPE
from etna.transforms import PytorchForecastingTransform
from pytorch_forecasting.data import GroupNormalizer
from etna.models.nn import DeepARModel
from datetime import datetime

import warnings

current_time_start = datetime.now().time()
original_df = pd.read_csv(
    "examples/data/example_dataset.csv", sep=','
)

print(original_df)
df = TSDataset.to_dataset(original_df)
df.index = pd.Index(pd.date_range("2021-01-01", freq="MS", periods=len(df)), name="timestamp")

ts = TSDataset(df, freq="MS")

print(ts)

HORIZON = 1

transform_date = DateFlagsTransform(month_number_in_year=True, day_number_in_week=False, day_number_in_month=False, out_column="regressor_dateflag")
num_lags = 10
transform_lag = LagTransform(in_column="target", lags=[HORIZON+i for i in range(num_lags)], out_column="regressor_target_lag")
lag_columns = [f"regressor_target_lag_{HORIZON+i}" for i in range(num_lags)]

transform_deepar = PytorchForecastingTransform(
    max_encoder_length=HORIZON,
    max_prediction_length=HORIZON,
    time_varying_known_reals=["time_idx"]+lag_columns,
    time_varying_unknown_reals=["target"],
    time_varying_known_categoricals=["regressor_dateflag_month_number_in_year"],
    target_normalizer=GroupNormalizer(groups=["segment"]),
)

model_deepar = DeepARModel(max_epochs=3, learning_rate=[0.01], gpus=0, batch_size=64)
metrics = [SMAPE()]

pipeline_deepar = Pipeline(
    model=model_deepar,
    horizon=HORIZON,
    transforms=[transform_lag, transform_date, transform_deepar]
)

# metrics_deepar, forecast_deepar, fold_info_deepar = pipeline_deepar.backtest(ts, metrics=metrics, n_folds=3, n_jobs=1)
pipeline_deepar.fit(ts=ts)

forecasted = pipeline_deepar.forecast()
print(forecasted)

Environment

No response

Additional context

No response

Checklist

• Bug appears at the latest library version

[martins0n]

Possible workaround -- to change timestamp index with dummy timestamp index with daily frequency