Ase interface - close #133

[YuuuXie]

fixed the ASE logger issue reported in #133

[smheidrich]

This PR seems to fix my issue (thanks!) but makes the code run insanely slowly compared to having trajectory=None in my tests. I think allow_calculation should be actually handled here the way ASE does it:

Suggested change

	if name not in self.results.keys():
	if name not in self.results.keys():
	if not allow_calculation:
	return None

⇒ Fast again

[YuuuXie]

Thanks for finding this issue and giving the suggestion. A new commit has been made on this allow_calculation change.

[codecov-io]

Codecov Report

Merging #136 into master will increase coverage by 0.01%.
The diff coverage is 100%.

@@            Coverage Diff            @@
##           master    #136      +/-   ##
=========================================
+ Coverage   59.29%   59.3%   +0.01%     
=========================================
  Files          31      31              
  Lines        4761    4760       -1     
=========================================
  Hits         2823    2823              
+ Misses       1938    1937       -1

Impacted Files	Coverage Δ
flare/ase/logger.py	95% <ø> (ø)	⬆️
flare/ase/calculator.py	81.72% <100%> (+0.4%)	⬆️
flare/ase/otf.py	60.34% <100%> (-0.11%)	⬇️
flare/ase/otf_md.py	98.14% <100%> (ø)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 7afbd4b...6615cfa. Read the comment docs.

[smheidrich]

Thanks again for fixing this, I think there is one more issue with this PR: The resulting trajectory file has about twice as many frames as there were MD steps - I guess this is because, as far as I understand, call_observers is called a second time whenever there is a DFT calculation, whereas in ASE, it's called only once per MD step.

[YuuuXie]

Thanks for pointing out this. Actually we were not using ASE trajectory for doing things, so we did not consider the compatibility with it. Since we have our own observer/logger, we have to record the positions and forces from DFT calculations, so as to collect training data, that's why call_observers is also used in DFT call.

I will open a new issue for this enhancement, i.e. to make our log file better compatible with ASE primitive log files, or even merge some of them.

For your ASE trajectory file, a temporary solution is to remove the DFT frames in post processing. You can look at the folder otf_data generated by your otf run. In the file dft_positions.xyz all the frames for DFT calculation are collected, you will see lines like Frame: x meaning at step x, GP gives a force prediction, and then DFT is also called to make a force prediction for the same frame.

For example, in the toy example you gave (in #133 ), you can run the following to remove the DFT frames in your trajectory file (the new_trj is the new trajectory file with DFT frames removed):

from ase.io.trajectory import Trajectory

# extract dft frame number from file
dft_file = open('otf_data/dft_positions.xyz')
dft_frames = []
for line in dft_file.readlines():
    if 'Frame' in line:
        dft_frames.append(int(line.split()[-1]))

# ase trajectory
trj = Trajectory('otf_md.traj')
new_trj = Trajectory('otf_md_new.traj', mode='w')
j = 0
for i in range(len(trj)):
    md_step = len(new_trj)
    if md_step == dft_frames[j]: # the current step is dft 
        j += 1
        continue
    else:
        new_trj.write(trj[i])

Let me know if you find any other issues

[smheidrich]

@YuuuuXie Thanks a lot for going to the trouble of writing a workaround! I'll use this for the time being.

Making FLARE's OTF-MD stuff behave more like ASE's MD classes in general would be great, too, of course.