compose Obs and Exp snipper classes -> ObsExpSnipper

[sergpolly]

thinking about the baseclass for snippers #227 (comment) ... not obvious how would that look like - i.e. how one would unify Obs and Exp...

However, one can compose Obs and Exp to give rise to ObsExp one like so:

class ObsExpSnipper:
    def __init__(
        self,
        clr,
        expected,
        view_df,
        cooler_opts=None,
        min_diag=2,
        expected_value_col="balanced.avg",
    ):
        """Class for generating expected-normalised snips from a cooler"""
        self.exp_snipper = ExpectedSnipper(
            clr=clr,
            expected=expected,
            view_df=view_df,
            min_diag=min_diag,
            expected_value_col=expected_value_col,
        )
        self.clr_snipper = CoolerSnipper(
            clr=clr,
            view_df=view_df,
            cooler_opts=cooler_opts,
            min_diag=min_diag,
        )

    def select(self, region1, region2):
        """Select a portion of the cooler for snipping based on two regions in the view"""
        self._expected = self.exp_snipper.select(region1, region2)
        matrix = self.clr_snipper.select(region1, region2)
        return matrix

    def snip(self, matrix, region1, region2, tup):
        """Extract an expected-normalised snippet from the matrix"""
        obs_snip = self.clr_snipper.snip(matrix, region1, region2, tup)
        exp_snip = self.exp_snipper.snip(self._expected, region1, region2, tup)

        return obs_snip / exp_snip

this might look inefficient, but i think it's mostly duplicated work in snip itself, select one probably is fine as is ...
Some of that could be mitigated, i think, by making snippers work on pre-calculated bin_id-s instead of recalculating nucleotides into bins several times like it is now (i.e. taking advantage of those lo1, hi1, lo2, hi2 - perhaps keeping them absolute + keeping region offsets along the way). just a thought anyhow

[sergpolly]

here is a better version i believe:

class ObsExpSnipper:
    def __init__(
        self,
        clr,
        expected,
        view_df,
        cooler_opts=None,
        min_diag=2,
        expected_value_col="balanced.avg",
    ):
        """Class for generating expected-normalised snips from a cooler"""
        self.exp_snipper = ExpectedSnipper(
            clr=clr,
            expected=expected,
            view_df=view_df,
            min_diag=min_diag,
            expected_value_col=expected_value_col,
        )
        self.clr_snipper = CoolerSnipper(
            clr=clr,
            view_df=view_df,
            cooler_opts=cooler_opts,
            min_diag=min_diag,
        )

    def select(self, region1, region2):
        """Select a portion of the cooler based on two regions in the view, normalize by expected"""
        observed = self.clr_snipper.select(region1, region2)  # CSR sparse matrix
        expected = self.exp_snipper.select(region1, region2)  # LazyToeplitz subscriptable object
        # generating O/E right away in a CSR format:
        expected_data = array_to_csr_data(expected, observed)
        observed.data = observed.data / expected_data
        return observed

    def snip(self, matrix, region1, region2, tup):
        """Extract an expected-normalised snippet from the matrix"""
        s1, e1, s2, e2 = tup
        offset1 = self.clr_snipper.offsets[region1]
        offset2 = self.clr_snipper.offsets[region2]
        binsize = self.clr_snipper.binsize
        # bins relative to start of respective chromosomes
        lo1, hi1 = s1 // binsize - offset1, e1 // binsize - offset1
        lo2, hi2 = s2 // binsize - offset2, e2 // binsize - offset2

        return _snip(
            matrix,
            offset = (offset1, offset2),
            bad_bins_masks = (self.clr_snipper._isnan1, self.clr_snipper._isnan2),
            min_diag = self.clr_snipper.min_diag,
            bbox = (lo1, hi1, lo2, hi2),
        )

definitions of _snip and array_to_csr_data are in a gist, along with another helper function that @nvictus suggested.
Cool thing is that now snip method of every snipper class, can be rewritten using this universal _snip function.
The reason we're not hiding the lo/hi calculation logic into _snip because I believe snipping should operate on bins right away instead of genomic coordinates (eventually).