input_output.load_time_series() -- allow "value" column to support strings? about mir_eval

[justinsalamon]

At the moment the code specifically loads both columns as floats:
https://github.com/craffel/mir_eval/blob/master/mir_eval/input_output.py#L229

Can you use np.loadtxt to load columns of different datatypes? It seems as though you have to specify a type and all columns are expected to be of that type.

I wrote this loader with melody in mind - it acts as a first check to ensure the data is of the correct type before proceeding. If you make this function return values that are not necessarily floats, you'd have to add a new check for the melody eval code or it could break...

[ejhumphrey]

I believe that yes, you can load different datatypes using that function. The docstring has this example:

>>> d = StringIO("M 21 72\nF 35 58")
>>> np.loadtxt(d, dtype={'names': ('gender', 'age', 'weight'),
...                      'formats': ('S1', 'i4', 'f4')})
array([('M', 21, 72.0), ('F', 35, 58.0)],
      dtype=[('gender', '|S1'), ('age', '<i4'), ('weight', '<f4')])

One, it's overkill for what I'm asking, but I think you could massage it to behave as intended. Two, while convenient, there's nothing that strictly says it's necessary to use this function (np.loadtxt) versus iterating over a file handle; it's what the rest of the loaders do anyways.

Sure, format assertions are always good, and we'd need a different one if the implementation changes.

[craffel]

I think lists of strings are saner than np.arrays of strings, so I'd prefer that it just dealt with list-like objects of any type. Feel free to change at will.

[craffel]

So, currently there are @bmcfee 's functions for loading annotations (ranges)/events: mir_eval.io.load_events and mir_eval.io.load_intervals. load_events loads in one- or two-column files; the first column always gets read in as floats (times), the second (if it exists) gets read in as a list of strings (labels). Similarly, load_annotation loads in two- or three-; first two are floats specifying the intervals (ranges), last column if it exists is a list of string labels. @justinsalamon's load_time_series does exactly the same thing as load_events except that the second column reads in floats. It seems like these functions should be merged, and the last column should be either returned as a np.ndarray or a list of strings.

Similarly, mir_eval.util.adjust_intervals is a useful function across many tasks but currently is only really suitable for segments. It'd be useful to use for chords, but the "padding" label should be just 'N' (no chord), I would argue; and the function only allows for setting a prefix. Similarly, it'd be useful for time series, but the padding label is always a string. It seems like for both of these cases it'd be nice if it could handle labels as a string or float list-like, and that the padding label could be set arbitrarily.

Is everyone comfortable with me trying to merge all of this functionality?

[bmcfee]

load_time_series does exactly the same thing as load_events except that the second column reads in floats. It seems like these functions should be merged, and the last column should be either returned as a np.ndarray or a list of strings.

Sure. load_events allows one argument (converter) to be passed in to specify the event index type (defaults to float). Seems like the quick fix here is to replicate this functionality for label parsing, so we have event_converter and label_converter (defaults to str) in https://github.com/craffel/mir_eval/blob/master/mir_eval/input_output.py#L60-L61 .

It'd be useful to use for chords, but the "padding" label should be just 'N' (no chord), I would argue; and the function only allows for setting a prefix.

Agreed, but I'm not sure of an elegant way to do this right now. The __%s notation is a crutch following the dummy label generation in the loader, where you want synthetic labels to be unique (numbered). I figured we should have consistent keying for synthetic labels. But, in adjust_*, we only ever use __T_MIN and __T_MAX. I really wish python supported the %*s formatter right about now...

[craffel]

We still need to merge the loaders. load_time_series loads two float columns. load_events loads one float column and optionally one string column. load_intervals loads two float columns and optionally one string column.

[craffel]

OK, the changes proposed above
#35 (comment)
have been implemented and all code/example usage has been updated to reflect this change.

@bmcfee @ejhumphrey @urinieto Important - io.load_intervals now does not return labels; if you want to load labeled intervals you need to use io.load_labeled_intervals. Same is true for load_events vs load_labeled_events but AFAIK no one was using load_events for loading labels. Also, none of these functions will generate synthetic labels themselves. If you want synthetic labels like these functions used to return when no labels were present, use util.generate_labels. Take a look at the evaluators/example usage for updated usage examples.