from_dict / to_dict should accept same orient parameter

[femtotrader]

Hello,

pd.DataFrame.to_dict accepts orient parameter in
['dict', 'list', 'series', 'split', 'records', 'index'] (default being 'dict')

http://pandas.pydata.org/pandas-docs/version/0.17.1/generated/pandas.DataFrame.to_dict.html

from_dict only accepts orient parameter in ['columns', 'index'] (default being 'columns')
I think that from_dict should accept same parameter orient for API consistency

http://pandas.pydata.org/pandas-docs/version/0.17.1/generated/pandas.DataFrame.from_dict.html

Same for Panel
http://pandas.pydata.org/pandas-docs/version/0.17.1/generated/pandas.Panel.from_dict.html
to_dict: missing method for Panel

Same for Series
http://pandas.pydata.org/pandas-docs/version/0.17.0/generated/pandas.Series.to_dict.html
from_dict: missing method for Series

Kind regards

[jreback]

you would have to show a specific example here of why this API change would actually be useful. Don't just list lots of code references.

[femtotrader]

Use case is to be able to save to MongoDB Series, DataFrame, Panel and also retrieve them.

An other feature to add to to_dict will be to output dict of NumPy arrays instead of dict of Python lists.

[jreback]

pls show a specific example
and a use case - you are describing a very general case

[femtotrader]

https://bitbucket.org/djcbeach/monary/issues/19/use-pandas-series-dataframe-and-panel-with

[jreback]

still not sure WHY this would be a good thing to add to the API nor what you are actually asking

[bluenote10]

what you are actually asking

I guess he is referring to the problem that to_dict has more modes than from_dict, which means that some to_dict conversions don't have an inverse operation.

It's an API inconsistency, and a consistent API is a "good thing". Here is an example of a problem that arises from this inconsistency:

When converting a DataFrame to JSON we are using the orient="split" option, because it seems to be only option that preserves a DataFrame entirely (order of columns and index). In the inverse operation we would need the equivalent. We cannot use to_json/read_json directly (which both support orient="split"), because the DataFrame is only a nested element in a bigger JSON structure, which requires operating on dicts. But Pandas lacks the inverse operation for dicts... We tried to temporarily convert the dict to string, allowing to use the read_json on the string with the orient split option, but the double conversion makes it prohibitively slow.

[xysscn]

what you are actually asking

I guess he is referring to the problem that to_dict has more modes than from_dict, which means that some to_dict conversions don't have an inverse operation.

It's an API inconsistency, and a consistent API is a "good thing". Here is an example of a problem that arises from this inconsistency:

When converting a DataFrame to JSON we are using the orient="split" option, because it seems to be only option that preserves a DataFrame entirely (order of columns and index). In the inverse operation we would need the equivalent. We cannot use to_json/read_json directly (which both support orient="split"), because the DataFrame is only a nested element in a bigger JSON structure, which requires operating on dicts. But Pandas lacks the inverse operation for dicts... We tried to temporarily convert the dict to string, allowing to use the read_json on the string with the orient split option, but the double conversion makes it prohibitively slow.

I have the same problem with bluenote10, the double conversion is too slow for a frequently called api.Does anyone have a workaround?

[sjrl]

I'm running into the same problem here. My current workaround is to use from_records to reverse to_dict(orient="split")

# Initial dataframe
df = pd.DataFrame.from_records([{"col1": "text_1", "col2": 1}, {"col1": "text_2", "col2": 2}])

# to_dict and reverse with from_records
pd.DataFrame.from_records(**df.to_dict(orient="split"))