pure numpy rmsd_qcp(): 2056.56 rmsds / s
│ │ │ │ +
pure numpy rmsd_qcp(): 815.07 rmsds / s
│ │ │ │
│ │ │ │
│ │ │ │
│ │ │ │
│ │ │ │
│ │ │ │
│ │ │ │ ├── html2text {}
│ │ │ │ │ @@ -59,15 +59,15 @@
│ │ │ │ │ start = time.time()
│ │ │ │ │
│ │ │ │ │ for i in range(100):
│ │ │ │ │ md.rmsd(t, t, 0)
│ │ │ │ │
│ │ │ │ │ print('md.rmsd(): %.2f rmsds / s' % ((t.n_frames * 100) / (time.time() -
│ │ │ │ │ start)))
│ │ │ │ │ -md.rmsd(): 276734.84 rmsds / s
│ │ │ │ │ +md.rmsd(): 30880.44 rmsds / s
│ │ │ │ │ But for some applications like clustering, we want to run many rmsd()
│ │ │ │ │ calculations per trajectory frame. Under these circumstances, the centering of
│ │ │ │ │ the trajectories is going to be done many times, which leads to a slight
│ │ │ │ │ slowdown. If we manually center the trajectory and then inform the rmsd()
│ │ │ │ │ function that the centering has been precentered, we can achieve ~2x speedup,
│ │ │ │ │ depending on your machine and the number of atoms.
│ │ │ │ │ In [4]:
│ │ │ │ │ @@ -75,28 +75,28 @@
│ │ │ │ │ start = time.time()
│ │ │ │ │
│ │ │ │ │ for i in range(100):
│ │ │ │ │ md.rmsd(t, t, 0, precentered=True)
│ │ │ │ │
│ │ │ │ │ print('md.rmsd(precentered=True): %.2f rmsds / s' % ((t.n_frames * 100) /
│ │ │ │ │ (time.time() - start)))
│ │ │ │ │ -md.rmsd(precentered=True): 1022681.70 rmsds / s
│ │ │ │ │ +md.rmsd(precentered=True): 61870.09 rmsds / s
│ │ │ │ │ Just for fun, let's compare this code to the straightforward numpy
│ │ │ │ │ implementation of the same algorithm, which mdtraj has (mostly for testing) in
│ │ │ │ │ the mdtraj.geometry.alignment subpackage
│ │ │ │ │ In [5]:
│ │ │ │ │ from mdtraj.geometry.alignment import rmsd_qcp
│ │ │ │ │
│ │ │ │ │ start = time.time()
│ │ │ │ │ for k in range(t.n_frames):
│ │ │ │ │ rmsd_qcp(t.xyz[0], t.xyz[k])
│ │ │ │ │
│ │ │ │ │ print('pure numpy rmsd_qcp(): %.2f rmsds / s' % (t.n_frames / (time.time() -
│ │ │ │ │ start)))
│ │ │ │ │ -pure numpy rmsd_qcp(): 2056.56 rmsds / s
│ │ │ │ │ +pure numpy rmsd_qcp(): 815.07 rmsds / s
│ │ │ │ │ md.rmsd() code is a lot faster. If you go look at the rmsd_qcp source code,
│ │ │ │ │ you'll see that it's not because that code is particularly slow or unoptimized.
│ │ │ │ │ It's about as good as you can do with numpy. The reason for the speed
│ │ │ │ │ difference is that an inordinate amount of time was put into hand-optimizing an
│ │ │ │ │ SSE3 implementation in C for the md.rmsd() code.
│ │ │ │ │ (rmsd-benchmark.ipynb.gz; rmsd-benchmark_eval.ipynb.gz; rmsd-benchmark.py)
│ │ │ │ │ Next Previous
│ │ │ ├── ./usr/share/doc/python-mdtraj-doc/html/examples/two-pass-clustering.html
│ │ │ │ @@ -458,17 +458,17 @@
│ │ │ │
│ │ │ │