Fix recall for Hamming distance

[ankane]

Both sift-256-hamming and word2bits-800-hamming always report 0 recall, as the distances in the HDF5 files are floats between 0 and 1 rather than the Hamming distance. Multiplying by the dataset dimensions fixes it.

This issue is likely the cause of #420.

[ankane]

Another approach would be to change Hamming distance to use mean.

 metrics = {
     "hamming": Metric(
-        distance=lambda a, b: np.sum(a.astype(np.bool_) ^ b.astype(np.bool_)),
+        distance=lambda a, b: np.mean(a.astype(np.bool_) ^ b.astype(np.bool_)),
         distance_valid=lambda a: True
     ),

[maumueller]

Thanks!

[ankane]

Thanks @maumueller!