|
col = neighs[:,1:].reshape(-1,) |
Current code:
row = (np.repeat(idx0.reshape(-1,1), k, axis=1)).reshape(-1,)
col = neighs[:,1:].reshape(-1,)
Proposed replacement:
row = (np.repeat(idx0.reshape(-1,1), k, axis=1)).ravel()
col = neighs[:,1:].ravel()
Both reshape(-1,) and ravel() can be used to flatten arrays into 1-D.
Benchmarks show that ravel() consistently outperforms reshape(-1,), because it is optimized for flattening and avoids extra shape compatibility checks.
Additionally, ravel() is more robust: it automatically returns a view when possible and only falls back to a copy if required, while reshape(-1,) requires the array to be contiguous in memory.
GARNET/hnsw.py
Line 14 in 0ce9b4d
Current code:
Proposed replacement:
Both reshape(-1,) and ravel() can be used to flatten arrays into 1-D.
Benchmarks show that ravel() consistently outperforms reshape(-1,), because it is optimized for flattening and avoids extra shape compatibility checks.
Additionally, ravel() is more robust: it automatically returns a view when possible and only falls back to a copy if required, while reshape(-1,) requires the array to be contiguous in memory.