GitLab

* cuda port for aev computation

* benchmark, setup and import

* fix bug

* fix flake8 and jit

* fix collections.abc for python>3.7

* setup gitingore

* format code with formatter

* ignore more

* Use torch API to set streams

* Use pytorch's caching allocator

* empty line

* fix

* cuaev correntness testr, instruction of install on readme

* readme

* fix readme

* fix readme

* fix readme

* fix readme

* add usage in readme

* fix readme

* add test in readme

* fix readme

* -std=c++14

* bug fix - add async data copy

* bug fix - add missing stream sync

* code refactor and cosmetic changes

* aev benchmark for big protein

* remove mdtraj

* remove print

* move pdb to dataset folder

* cosmetic changes

* Move torchani/extensions -> torchani/cuaev

* clang-format -i

* cleanup

* return aev from cuComputeAEV

* Update aev.py

* Update aev.py

* fix flake8

* fix LGTM unused local variable

* clang-format

* fix

* save

* install change
Co-authored-by: richard <yueyericardo@gmail.com>
Co-authored-by: Xiang Gao <qasdfgtyuiop@gmail.com>

* Make training benchmark print a warning if not synchronizing, and add a sync before function also

* Take triu_index out since it does not run

* slightly increase default precision

* flake8
Co-authored-by: Ignacio Pickering <ign.pickering@gmail.com>

* modifications to make tests faster

* Now ani1x is not loaded for any aev_computer necessary

* flake*

* Avoid test warnings by loading from json and setting initial velocities to a small number

* Avoid loading from cif in ManualMirror and Repeat also

* Add comments to clarify test_energies and delete unused variable

* Fix inaccurate comment

* Delete unused variable in test_ensemble

* Add comments and clarifications to test_ase

* flake8
Co-authored-by: Ignacio Pickering <ign.pickering@gmail.com>

ANI1x NNs are updated to be consistent with what we used in the paper

* Remove everything about chunking

* aev.py

* neurochem trainer

* training-benchmark-nsys-profile.py

* fix eval

* training-benchmark.py

* nnp_training.py

* flake8

* nnp_training_force.py

* fix dtype of species

* fix

* flake8

* requires_grad_

* git ignore

* fix

* original

* fix

* fix

* fix

* fix

* save

* save

* save

* save

* save

* save

* save

* save

* save

* collate

* fix

* save

* fix

* save

* save

* fix

* save

* fix

* fix

* no len

* float

* save

* save

* save

* save

* save

* save

* save

* save

* save

* fix

* save

* save

* save

* save

* fix

* fix

* fix

* fix mypy

* don't remove outliers

* save

* save

* save

* fix

* flake8

* save

* fix

* flake8

* docs

* more docs

* fix test_data

* remove test_data_new

* fix

* massive training time improvement

* flake8

* improve validation also

* improve training-benchmark.py

* fix devices

* send num_atoms to GPU

* rerun

* Add an argument for inserting synchronization

* Update training-benchmark.py

* Update training-benchmark.py
Co-authored-by: Farhad Ramezanghorbani <farhadrgh@users.noreply.github.com>

* Add warning to ChemicalSymbolsToInts

* Change ocurrences of ChemicalSymbolsToInts in the code to reflect safe usage

* flake8

* Add warning to call also

* Add clarification

* docs dont fetch magic functions so documentation is moved to the class docstring

* fix clarification

* flake8

* bypass slow code

* delete reference to all_atoms which was unused

* delete whitespace

* make flake8 happy

* rerun

* actual change was off

* typo did not turn on bypassing code

* Add CODATA and IUPAC unit conversion factors

* Add units to main module

* change neurochem and training-benchmark to use the new hartree2kcalmol conversion

* change training-benchmark to use hartree2kcalmol conversion

* change comp6 to use the new hartree2kcalmol conversion

* change training-benchmark without nsys to also use hartree2kcalmol conversion

* change examples to use hartree2kcalmol conversion

* Rename ase to ase_ani to avoid naming conflicts that come from naming internal module the same as an external one we are using

* fix import

* Add ASE's hartree_to_ev conversion to compare with codata 2018

* Standarize our units to be essentially the same as ASE and CODATA 2014

* Change vibration_analysis to use the consistent constants

* Fix explanation in docstring

* Add millieV conversion factor

* Add millieV conversion factor to vibrational analysis

* Add clarification

* flake8

* revert name change

* Fix typo in function name

* Add units to __all__'

* Improve units docstring and comments in preparation for docs

* delete pad from docs

* Add units to api.rst

* minor formatting of docstrings and comments

* Add functions into the API

* final formatting changes

* flake8

* try to make mypy happy

* Checking if mypy prefers this

* rerun

* nvidia profiling for md

* 1 model

* hook functions

* Update md-benchmark-nsys-profile.py

* nvidia profiling for md

* 1 model

Add helper module to convert species from element id in periodic table to 0, 1, 2, 3, ... format (#396)

* Init

* fix test

* flake8

* try fix

* Fix stupidity of len(self)

* Simplify triple_by_molecule

* fix

* fix

* fix

* Remove ignite

* remove neurochem-test

* remove cache-aev

* Update utils.py

* Update utils.py

* Update utils.py

* Update utils.py

* remove float32 dtypes from comp6.py

* cached

* typo and comments

* easy to read

* change some names

* fix unit test

* empty line

* fix

* fix

* add docs and add whether include_energies

* docs

* other properties for shuffled dataset

* docs

* dtype for benchmark

* add properties to test

* style