Minor adjust to the ROQ weight generation

1. Convert the light travel time explicitly calculated value: uses 1.1 * R_earth / c (i.e. 10% more than the expected maximum)
2. Use a single fixed time_space: previously, the time resolution was set, then recalculated from the difference in the first two elements of time_samples. This could lead to buggy behaviour since that is no longer an evenly spaced array after the subtraction from the start time is made.
3. Add info statements
4. Adds a check if the seglen is shorter than the duration
5. Apply a safety factor of 5 to ensure the time step is short enough. This will increase the size of the ROQ array basis by ~ a factor of 5, but in testing I found it helped remove bias from examples just searching over RA, DEC, and geocent_time.
6. Add explicit clean up of memory

Closes #372 (closed)

bilby/gw/likelihood.py

@@ -875,7 +875,7 @@ class ROQGravitationalWaveTransient(GravitationalWaveTransient):
 
         d_inner_h = interp1d(
             self.weights['time_samples'][indices],
-            d_inner_h_tc_array, kind='cubic')(ifo_time)
+            d_inner_h_tc_array, kind='cubic', assume_sorted=True)(ifo_time)
 
         optimal_snr_squared = \
             np.vdot(np.abs(h_plus_quadratic + h_cross_quadratic)**2,
@@ -960,21 +960,9 @@ class ROQGravitationalWaveTransient(GravitationalWaveTransient):
                     ifo.name, len(overlap_frequencies),
                     min(overlap_frequencies), max(overlap_frequencies)))
 
-            # array to be filled with data, shifted by discrete time_samples
-            tc_shifted_data = np.zeros([
-                len(self.weights['time_samples']), len(overlap_frequencies)],
-                dtype=complex)
-
-            # shift data to beginning of the prior increment by the time step
-            shifted_data =\
-                ifo.frequency_domain_strain[ifo.frequency_mask][ifo_idxs] * \
-                np.exp(2j * np.pi * overlap_frequencies *
-                       self.weights['time_samples'][0])
-            single_time_shift = np.exp(
-                2j * np.pi * overlap_frequencies * time_space)
-            for j in range(len(self.weights['time_samples'])):
-                tc_shifted_data[j] = shifted_data
-                shifted_data *= single_time_shift
+            data = ifo.frequency_domain_strain[ifo.frequency_mask][ifo_idxs]
+            tc_shifted_data = data * np.exp(
+                2j * np.pi * overlap_frequencies * time_samples[:, np.newaxis])
 
             # to not kill all computers this minimises the memory usage of the
             # required inner products
@@ -987,7 +975,7 @@ class ROQGravitationalWaveTransient(GravitationalWaveTransient):
                 linear_matrix[roq_idxs],
                 max_elements) * 4 / ifo.strain_data.duration
 
-            del tc_shifted_data, shifted_data, single_time_shift, overlap_frequencies
+            del tc_shifted_data, overlap_frequencies
             gc.collect()
 
             self.weights[ifo.name + '_quadratic'] = build_roq_weights(