[BUG] EXPERIMENTAL PR: Solve the bug in data_module

[phoeenniixx]

This PR solves the bug in data_module where the static_categorical_features and static_continuous_features were not correctly calculated in __getitem__ of nested class

[phoeenniixx]

This solves the bug that was found while writing the integration tests for tft model in #1812

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@5685c59). Learn more about missing BASE report.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1834   +/-   ##
=======================================
  Coverage        ?   86.82%           
=======================================
  Files           ?       51           
  Lines           ?     5668           
  Branches        ?        0           
=======================================
  Hits            ?     4921           
  Misses          ?      747           
  Partials        ?        0           

Flag	Coverage Δ
cpu	86.82% <100.00%> (?)
pytest	86.82% <100.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[fkiraly]

It fixes the bug, so it is approved.

But, the new code seems unnecessarily complicated?
Is it not possible to avoid the loops by simply using array operations smartly?

[phoeenniixx]

Sure I'll try to find a better way

[fkiraly]

the "basic principle" is to replace sth like

my_list = [] * len(my_vector)
for i, x in enumerate(my_vector):
    my_list[i] = 42*x

by something like

my_list = list(42*my_vector)

[phoeenniixx]

the "basic principle" is to replace sth like

my_list = [] * len(my_vector)
for i, x in enumerate(my_vector):
    my_list[i] = 42*x

by something like

my_list = list(42*my_vector)

so we can use this approach by creating a mask right? then this mask can easily be mutiplied to get the final st_cat_values_for_item?
Like

is_categorical_mask = [
                self.data_module.time_series_metadata["col_type"].get(col_name) == "C"
                for col_name in static_col_names
            ]
is_continuous_mask = [not b for b in is_categorical_mask]
st_cat_values_for_item = raw_st_tensor[is_categorical_mask] 
st_cont_values_for_item = raw_st_tensor[is_continuous_mask]

(similar to what we do in timeseries to get the groups)

Thinking of which, we can use tensors here, then in as in place of

is_continuous_mask = [not b for b in is_categorical_mask]

it will just be

is_continuous_mask = ~is_categorical_mask

that is simpler?

[fkiraly]

yes, exactly

[fkiraly]

plus, naively, I would expect something like

is_categorical_mask =  self.data_module.time_series_metadata["col_type"] == "C"

to work.

[phoeenniixx]

plus, naively, I would expect something like

is_categorical_mask =  self.data_module.time_series_metadata["col_type"] == "C"

to work.

We just need the static_categorical in this case, not all categorical
and i am not sure this comparison works with dict, I am sure similar way is possible with pandas, but with dict?

[fkiraly]

Yes, I think this is better!

[agobbifbk]

Hello! I have a comment on the current situation: we are assuming that the data are already sorted and there is no gap in the temporal index. Since the output is taken and windowing in the d2 layer, we need to be sure that all the data is equally spaced.
Maybe it is better to check it somehow during the getitem, and in a second moment to enlarge the dataset and put nans in the holes. What do you think?

[phoeenniixx]

Hello! I have a comment on the current situation: we are assuming that the data are already sorted and there is no gap in the temporal index. Since the output is taken and windowing in the d2 layer, we need to be sure that all the data is equally spaced. Maybe it is better to check it somehow during the getitem, and in a second moment to enlarge the dataset and put nans in the holes. What do you think?

Yes! ig we should do this, but I think we could add it in next iterations? First complete a basic e2e prototype? We have added the warning, so the users know the code still has loop holes.

[agobbifbk]

Ok it was just a comment, we need to remember this :-)