Update rel tol computation in netcdf_comparison

Denominator is now the column max abs value, rather than pointwise absolute
value... we were getting failures in mks near sign changes in some fields,
where absolute difference was roughly the same throughout the column but values
near zero were three or four orders of magnitude smaller than the rest of the
column.

MARBL_tools/netcdf_comparison.py

@@ -206,8 +206,8 @@ def _variable_check_loose(ds_base, ds_new, rtol, atol, thres):
  conversion_factor = _get_conversion_factor(ds_base, ds_new, var)
 
  # (1) Are NaNs in the same place?
- mask = ~np.isnan(ds_base[var].data)
- if np.any(mask ^ ~np.isnan(ds_new[var].data)):
+ mask = np.isfinite(ds_base[var].data)
+ if np.any(mask ^ np.isfinite(ds_new[var].data)):
  error_checking['messages'].append('NaNs are not in same place')
 
  # (2) compare everywhere that baseline is 0
@@ -237,8 +237,13 @@ def _variable_check_loose(ds_base, ds_new, rtol, atol, thres):
  new_data = np.where(np.abs(ds_base[var].data[mask]) > thres,
  conversion_factor*ds_new[var].data[mask],
  0)
+ # denominator for relative error is column max value
+ # note the assumption that column is first dimension
+ col_max = np.nanmax(np.abs(ds_base[var].data), axis=tuple(np.arange(1,len(ds_base[var].data.shape))))
+ rel_denom = (np.ones(ds_base[var].data.transpose().shape)*col_max).transpose()
+ rel_denom = rel_denom[mask]
  rel_err = np.where(base_data != 0, np.abs(new_data - base_data), 0) / \
- np.where(base_data != 0, np.abs(base_data), 1)
+ np.where(rel_denom != 0, rel_denom, 1)
  if np.any(rel_err > rtol):
  if rtol == 0:
  abs_err = np.abs(new_data - base_data)