nvnmd#1707
Conversation
Codecov Report
@@ Coverage Diff @@
## devel #1707 +/- ##
==========================================
+ Coverage 76.08% 77.10% +1.01%
==========================================
Files 96 118 +22
Lines 7927 9411 +1484
==========================================
+ Hits 6031 7256 +1225
- Misses 1896 2155 +259
Continue to review full report at Codecov.
|
njzjz
left a comment
njzjz
left a comment
There was a problem hiding this comment.
Before further review, I have the following general comments:
(1) Translate all Chinese comments into English;
(2) Format codes under deepmd/nvnmd with pep8;
(3) Add doc to each new class and methods to explain the role of them;
(4) Reduce data in the examples diretory. As an example, one data set is enough.
(5) Add unit tests for new codes;
(6) Use os.path.join to combine paths instead of /.
wanghan-iapcm
left a comment
wanghan-iapcm
left a comment
There was a problem hiding this comment.
I notice that all examples are removed. What is the reason?
Please follow @njzjz 's comment :
(4) Reduce data in the examples diretory. As an example, one data set is enough.
wanghan-iapcm
left a comment
wanghan-iapcm
left a comment
There was a problem hiding this comment.
All C++ code implementing an OP (can be tested without TF C++ interface), should be moved from source/op to source/lib/src.
UTs of the OP implementation should be added.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
| .Input("dv: T") | ||
| .Input("grad_v: T") | ||
| .Input("grad_dv: T") | ||
| .Attr("prec: float") |
There was a problem hiding this comment.
Here prec is float but x may be either double or float. I think this cannot improve accuracy to double when you multiply x by prec.
| self.quantize_descriptor = True | ||
| self.quantize_fitting_net = True | ||
| else: | ||
| pass |
There was a problem hiding this comment.
raise RuntimeError if it's not expected behavior
This comment was marked as resolved.
This comment was marked as resolved.
| log = logging.getLogger(__name__) | ||
|
|
||
|
|
||
| class Encode(): |
There was a problem hiding this comment.
A lot of codes don't have the unit test, for example this class. See details of converge:
https://app.codecov.io/gh/deepmodeling/deepmd-kit/compare/1707/tree/deepmd/nvnmd
|
Code quality is acceptable. ping @wanghan-iapcm for more comments |
wanghan-iapcm
left a comment
wanghan-iapcm
left a comment
There was a problem hiding this comment.
The size of the data files coord.npy and force.npy are 1M. Could you please reduce the number of frames to make the files as small as possible? I would say 0.1M would be ideal. They present only for the purpose of showing the code is working.
|
|
||
|
|
||
| def descrpt2r4(inputs, natoms): | ||
| """ replace :math:`r_{ji} \rightarrow r'_{ji}` |
There was a problem hiding this comment.
| """ replace :math:`r_{ji} \rightarrow r'_{ji}` | |
| r""" replace :math:`r_{ji} \rightarrow r'_{ji}` |
otherwise \r will be recognized as the line separator.
Co-authored-by: LiuGroupHNU <liujie123@HNU> Co-authored-by: MoPinghui <mopinghui1020@gmail.com> Co-authored-by: Han Wang <92130845+wanghan-iapcm@users.noreply.github.com>
5 participants
Our training procedure consists of not only the continuous neural network (CNN) training, but also the quantized neural network (QNN) training which uses the results of CNN as inputs, to help users train machine-learning models that are compatible with the unique non von Neumann computer.