aiqm / torchani Goto Github PK

• remove tqdm #151
• remove MaxAE #150

• Add 3D structures from NIST #146
• Add more off-equilibrium structures, reactions
• Test structure optimization #153

pytorch/ignite#300

Waiting for:

• pytorch/pytorch#12199

There: https://github.com/pytorch/pytorch/blob/master/torch/tensor.py#L291

• Test if structure minimization works
• Print the distances to see if they get too close
• Visualize neighbor list to see if it actually works.

• Modify setup.py to allow install on machine that do not have sphinx
• Manually manage docstring as in pytorch
• Change the style of docstring into the same as pytorch
• Use the script that reproduce ANI-1x NeuroChem result to replace the code in examples/
• Add helpers to allow converting from list of strings like ['H', 'H', 'C'] into tensor like tensor([0, 0, 1])
• Update readme
• Determine which license to use for opensource (GPL, MIT, BSD?)
• Clean up old tags
• Make a final review on API and the whole code to see if it can be improved
• Read through all comments and docstring to make sure it is complete enough but not too verbose.
• Write a loader to load NeuroChem training setup
• Write tutorial
• Allow caching AEV during training

Would it be possible to create a __version__ attribute or similar so that individual runs can be reproduced? We usually recommend versioneer or similar for this kind of operation.

Edge branch is PyTorch's master with the following PRs merged:

• torch.combinations: pytorch/pytorch#9393
• tensor.device, etc. in JIT: pytorch/pytorch#12363
• cosine_similarity: pytorch/pytorch#12199
• outplace ops: pytorch/pytorch#12413

Outdated list:

Wait for pytorch/pytorch#10978 to be resolved and remove pad and pad_coordinates in https://github.com/aiqm/torchani/blob/master/torchani/utils.py

Wait for pytorch/pytorch#12206 to be resolved to make utils.present_species JITable

Hi @zasdfgbnm

I get a different energy every time I use the energy_force.py example in the qrefine branch of torchani.

('Energy:', -40.30905532836914)
('Energy:', -40.434627532958984)
('Energy:', -40.549381256103516)
('Energy:', -40.627140045166016)
('Energy:', -40.549705505371094)
('Energy:', -40.49227523803711)

This is using phenix.python which is built on python 2.7.15

To ensure phenix.python was not causing the problem, I used a conda environment with python 2.7.3

('Energy:', -40.51961135864258)
('Energy:', -40.46982955932617)

In comparison, the latest version of torchani running on python 3.7.2 leads to consistent energies when I run energy_force.py:

Energy: -40.425621032714844
Energy: -40.425621032714844
Energy: -40.425621032714844

• doc on github pages
• pypi release build

Visualizing AEV parameter changes

When trying to install torchANI as instructed in the installation notes everything goes OK when installing pytorch, but the pip part to install torchANI fails:

(torchani) [henrique@cpd08 ~] $ conda install pytorch-nightly -c pytorch
Solving environment: done

## Package Plan ##

  environment location: /home/henrique/bin/anaconda3/envs/torchani

  added / updated specs: 
    - pytorch-nightly


The following NEW packages will be INSTALLED:

    blas:            1.0-mkl                                                 
    ca-certificates: 2018.03.07-0                                            
    certifi:         2018.11.29-py37_0                                       
    cffi:            1.11.5-py37he75722e_1                                   
    intel-openmp:    2019.1-144                                              
    libedit:         3.1.20170329-h6b74fdf_2                                 
    libffi:          3.2.1-hd88cf55_4                                        
    libgcc-ng:       8.2.0-hdf63c60_1                                        
    libgfortran-ng:  7.3.0-hdf63c60_0                                        
    libstdcxx-ng:    8.2.0-hdf63c60_1                                        
    mkl:             2019.1-144                                              
    mkl_fft:         1.0.6-py37hd81dba3_0                                    
    mkl_random:      1.0.2-py37hd81dba3_0                                    
    ncurses:         6.1-he6710b0_1                                          
    ninja:           1.8.2-py37h6bb024c_1                                    
    numpy:           1.15.4-py37h7e9f1db_0                                   
    numpy-base:      1.15.4-py37hde5b4d6_0                                   
    openssl:         1.1.1a-h7b6447c_0                                       
    pip:             18.1-py37_0                                             
    pycparser:       2.19-py37_0                                             
    python:          3.7.1-h0371630_7                                        
    pytorch-nightly: 1.0.0.dev20190101-py3.7_cuda9.0.176_cudnn7.4.1_0 pytorch
    readline:        7.0-h7b6447c_5                                          
    setuptools:      40.6.3-py37_0                                           
    sqlite:          3.26.0-h7b6447c_0                                       
    tk:              8.6.8-hbc83047_0                                        
    wheel:           0.32.3-py37_0                                           
    xz:              5.2.4-h14c3975_4                                        
    zlib:            1.2.11-h7b6447c_3                                       

Proceed ([y]/n)? y

Preparing transaction: done
Verifying transaction: done
Executing transaction: done
(torchani) [henrique@cpd08 ~] $ pip install torchani
Collecting torchani
  Using cached https://files.pythonhosted.org/packages/26/49/b7b92a8e7467164fdbf1c6850b57be66ed0144aeb4b4a492ece9cada89c0/torchani-0.2.1-py3-none-any.whl
Collecting pytorch-ignite-nightly (from torchani)
  Using cached https://files.pythonhosted.org/packages/89/85/804ec40744ccba56e5a2b061e0e2f1c8c90289c33a878f0ce9fce35422ef/pytorch_ignite_nightly-20190101-py2.py3-none-any.whl
Collecting torch-nightly (from torchani)
  Could not find a version that satisfies the requirement torch-nightly (from torchani) (from versions: )
No matching distribution found for torch-nightly (from torchani)

Waiting for: pytorch/ignite#306

Dependencies:

• #141
• #140
• ASE interface and PBC support
• #102

hi @zasdfgbnm,

I think the dataset .h5 files:

• https://github.com/aiqm/torchani/blob/master/dataset/ani_gdb_s01.h5
• https://github.com/aiqm/torchani/blob/master/dataset/ani_gdb_s02.h5
• https://github.com/aiqm/torchani/blob/master/dataset/ani_gdb_s03.h5
• https://github.com/aiqm/torchani/blob/master/dataset/ani_gdb_s04.h5

are corrupted when I use:

git clone https://github.com/aiqm/torchani.git

I then installed torchani, and tested the energy_force.py script successfully.

But when I run:

python nnp_training.py

I get an error:

Error:
 File "h5py/h5f.pyx", line 85, in h5py.h5f.open
OSError: Unable to open file (file signature not found)

To solve the problem:
I download the four individual .h5 files manually from the Github website using the download button.
I then copied the files to the dataset folder.
Then the training script works as expected.

I am using:
- Mac OS X 10.14.2
- anaconda3
- Python 3.7.2
- torchani.version = '0.2.2'
- torch.version = '1.0.0.dev20190214'
- h5py.version = '2.9.0'

pytorch/pytorch#17652

pytorch/pytorch#17662 pytorch/pytorch#27629

Wait for pytorch/pytorch#9393 to be merged and remove AEVComputer. _combinations from TorchANI

Hey. I love your project :) Just wanted to point out that if there are no atoms with Rca distance the code crashes with the following stack trace because of empty array leading to n=0.
I guess it might make sense to handle this somehow?

/shared/sdoerr/Software/miniconda3/envs/torchani/lib/python3.6/site-packages/torch/nn/modules/container.py in forward(self, input)
     90     def forward(self, input):
     91         for module in self._modules.values():
---> 92             input = module(input)
     93         return input
     94 

/shared/sdoerr/Software/miniconda3/envs/torchani/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    475             result = self._slow_forward(*input, **kwargs)
    476         else:
--> 477             result = self.forward(*input, **kwargs)
    478         for hook in self._forward_hooks.values():
    479             hook_result = hook(self, input, result)

/shared/sdoerr/Software/miniconda3/envs/torchani/lib/python3.6/site-packages/torchani/aev.py in forward(self, species_coordinates)
    294 
    295         radial_terms, angular_terms, indices_r, indices_a = \
--> 296             self._terms_and_indices(species, coordinates)
    297         mask_r = self._compute_mask_r(species_, indices_r)
    298         mask_a = self._compute_mask_a(species_, indices_a, present_species)

/shared/sdoerr/Software/miniconda3/envs/torchani/lib/python3.6/site-packages/torchani/aev.py in _terms_and_indices(self, species, coordinates)
    180         vec = vec.gather(-2, _indices_a)
    181 
--> 182         vec = self._combinations(vec, -2)
    183         angular_terms = self._angular_subaev_terms(*vec)
    184 

/shared/sdoerr/Software/miniconda3/envs/torchani/lib/python3.6/site-packages/torchani/aev.py in _combinations(self, tensor, dim)
    198         index1 = grid_y.masked_select(
    199             torch.triu(torch.ones(n, n, device=tensor.device),
--> 200                        diagonal=1) == 1)
    201         index2 = grid_x.masked_select(
    202             torch.triu(torch.ones(n, n, device=tensor.device),

RuntimeError: invalid argument 1: expected a matrix at /opt/conda/conda-bld/pytorch-nightly_1540805525195/work/aten/src/TH/generic/THTensorMoreMath.cpp:1270

File "/home/clean/phenix-1.14rc1-3177/build/../modules/qrefine/command_line/refine.py", line 116, in
run(args=sys.argv[1:], log=log)
File "/home/clean/phenix-1.14rc1-3177/build/../modules/qrefine/command_line/refine.py", line 111, in run
log = log)
File "/home/clean/phenix-1.14rc1-3177/modules/qrefine/qr.py", line 368, in run
model = model)
File "/home/clean/phenix-1.14rc1-3177/modules/qrefine/qr.py", line 243, in create_restraints_manager
clustering = params.cluster.clustering)
File "/home/clean/phenix-1.14rc1-3177/modules/qrefine/restraints.py", line 101, in init
self.qm_engine = self.create_qm_engine()
File "/home/clean/phenix-1.14rc1-3177/modules/qrefine/restraints.py", line 127, in create_qm_engine
calculator = TorchAni()
File "/home/clean/phenix-1.14rc1-3177/modules/qrefine/plugin/ase/torchani_qr.py", line 36, in init
from_nc=network_dir, ensemble=8)
File "/home/clean/phenix-1.14rc1-3177/base/lib/python2.7/site-packages/torch/jit/init.py", line 555, in init_then_register
original_init(self, *args, **kwargs)
File "/home/clean/phenix-1.14rc1-3177/base/lib/python2.7/site-packages/torchani/nn.py", line 399, in init
self.aev_computer.dtype, self.aev_computer.device, filename)
File "/home/clean/phenix-1.14rc1-3177/base/lib/python2.7/site-packages/torch/jit/init.py", line 555, in init_then_register
original_init(self, *args, **kwargs)
File "/home/clean/phenix-1.14rc1-3177/base/lib/python2.7/site-packages/torchani/nn.py", line 57, in init
layer_setups = self._parse(buffer)
File "/home/clean/phenix-1.14rc1-3177/base/lib/python2.7/site-packages/torchani/nn.py", line 123, in _parse
tree = parser.parse(nnf_file)
File "/home/clean/phenix-1.14rc1-3177/base/lib/python2.7/site-packages/lark/lark.py", line 197, in parse
return self.parser.parse(text)
File "/home/clean/phenix-1.14rc1-3177/base/lib/python2.7/site-packages/lark/parser_frontends.py", line 137, in parse
return self.parser.parse(text)
File "/home/clean/phenix-1.14rc1-3177/base/lib/python2.7/site-packages/lark/parsers/xearley.py", line 126, in parse
column = scan(i, column)
File "/home/clean/phenix-1.14rc1-3177/base/lib/python2.7/site-packages/lark/parsers/xearley.py", line 115, in scan
raise UnexpectedInput(stream, i, text_line, text_column, {item.expect for item in to_scan}, set(to_scan))
lark.lexer.UnexpectedInput: No token defined for: '-' in u'-3671' at line 47 col 11

Expecting: set(['INT', 'FLOAT', u'__FILE3'])

model:ani-1ccx_8x
same model running with ANI not got error

Dataset outputs of dict, with keys like "coordinates", "species", "energies", the output of dataset is chunk, which is different conformations of the same molecule.

Dataloader samples a batch of several chunks, and concat fields that are concatable (those that has shape independent on number of atoms in a molecule)

torchani.ignite.Container accepts a TODO

Wait for pytorch/pytorch#9532 to be merged and remove list operations for the shape in https://github.com/aiqm/torchani/blob/master/torchani/aev.py