Deepmd-kit

- Training parameters: Several training parameters have been updated. Original training data is splited into training data and validation data. Please read the document to apply the changes. Old styles can still work but are not recommended.
- Model inference: Old models trained by v1 will not work in v2. Run `dp convert-from` to convert old models to v2.
- Python interface: `deepmd.DeepPot` has been moved to `deepmd.infer.DeepPot`. <!-- add an alias? -->
- C++ interface: `NNPInter` has been renamed to `deepmd::DeepPot` and `NNPInter.h` has been renamed to `DeepPot.h`. Use `-ldeepmd_cc` to link instead.

New features
- [Model compression](https://arxiv.org/abs/2107.02103) (#350 586 610 921 948 956 1000 1008 1020 1043)
- Parallel training (892 905 913 1030 1032) (Bytedance)
- ROCm device support (656 )
- New descriptor: three body embedding (`se_e3`)
- Hybridization of descriptors (`hybrid`)
- Type embedding
- Training and inference the dipole (vector) and polarizability (matrix). (495 538 927)
- Support derivatives of the tensor properties. (805)
- Split of training and validation dataset.
- Model deviation for virial
- Add subcommand and python interface to calculate model-deviation (715)
- Automatically determine the sel from the training data. (831)
- Building with lammps with plugin mode (930 945)

Performance improvement:
- More efficient training: all customized OPs are implemented with GPU.
- MPI support for atomic model deviation 628
- speedup ROCm kernels which use atomicAdd (809 815 ) (from ByteDance)
- speedup CUDA kernels (use atomicAdd inside) by reducing the global memory write (811)
- speedup tabulate cuda kernel by reducing shm using (830) (Bytedance)
- speedup `format_nlist_b` (832 845)
- speedup `scan_nlist` kernel (1028)

Enhancements
- Strict argument check in the input script.
- Auto conversion of input file to v2.0 compatibility
- Append out_file when lammps restarts 640
- Document and examples for the C++ interface 652 663
- Instructions for the i-pi 660
- Document for the network size and sel 657
- Use fmod to wrap the coord of atoms (solve slow PBC) (741)
- bit operations to encode neighbor information
- add CUDA/ROCM buidling documents (739)
- add type-embedding developer doc (762 967)
- add model compression support for models with exclude_types feature (754)
- improve the doc and user interface of model compression (772)
- support converting models generated in v1.3 to 2.0 compatibility (725)
- give a default value to T and convert models from v1.2 to 2.0 compatibility (789)
- improved documents for conda (798 925)
- throw a message if tf runtime is incompatible (797)
- capture OOM and print debug message (801)
- add message for DecodeError raised when using model compression (839)
- Passing error to TF instead of exit (918)
- refactor docs (952)
- add an example of `nopbc` and related docs (994)
- add `dp --version` (995)
- add the argument `tensorboard_freq` to control sampling ratio during training. (996)
- add sphinx plugins `viewcode` and `intersphinx` (997)
- generate Python API document automatically (998)
- give a clear message if `model.get_ntypes()<data.get_ntypes()` (1016)
- add docstring for `descrpt/se_e2_a` (1017)
- add docstring for `fit/ener` (1024)
- add `InputNlist` into API doc (1009)
- save checkpoint files with step and keep recent files (1031)


Improvement of the code for developers
- Support version of the model. Easily check model compatability
- Clear and pythonic python interface
- C++ lib that can be tested independently
- C++ API that can be tested independently
- OP supports multi-device.
- Added `deepmd` namespace for the C++ API
- UT for Cuda/ROCm code (569)
- UT for model compression (586)
- UT for prod_force/virial ops (703 741)
- CI test Lammps build (600)
- allow c++ tests to run without internet (785)
- build low and high precision at the same time (879)
- support to specify CUDA/ROCm root in python pkg building (834) (Bytedance)
- use cached Session to speed up py tests (833)
- remove cub include for CUDA>=11 (866)
- Add Errcheck after every kernel function runs And merge redundant code (855)
- adapt changes to auditwheel directory in manylinux (889)
- enhance the cli to generate doc json file (891)
- raise warning before training if `sel` is not enough (914)
- make native MD compatible with v2.0 (950)
- fix type hints and add doc for `exclude_types` (1005)
- use TF's built-in method to get numpy dtype (1035)


Bug fixings:
- Remove `using namespace std`. Solve compiling compatability problem.
- `cuda` memory access error 566
- Relative force model deviation is not copied back at single precision 599
- Correct way of allocating memory in float precision 612
- Fix TB logdir remove bug 617
- Illegal nlist 680
- Bug in `prod_virial_grad` that causes wrong results when training with virials 685
- Uniform random seed 691
- Illegal nlist 680
- Bug in `prod_virial_grad` that causes wrong results when training with virials 685
- Uniform random seed 691
- fix bug of adding int to a None random seed (705)
- reuse the zero layer rather than building a new one (714)
- fix bug in CI (739)
- fix bug 824 and Synchronize updates to CUDA cod (828)
- Fix the empty neighbor distance array in neighbor_stat.py (882)
- fix InvalidArgumentError caused by zero sel and optimize zero matrix (900)
- fix 'NoneType' has no len() in auto_sel (911)
- set input `DeepmdData.type_map` to input `type_map` (924)
- Fix member declartion of `deepmd` and `deepmd.entrypoints`. (922)
- add aliases to Arguments (933)
- fix bug of gelu activation function (939)
- convert `decay_rate` to `stop_lr` from old inputs (949)
- only enable link what you use on GNU compilers (962)
- Do not find protobuf for python (963)
- fix an error in stress by ase interface (964)
- remove bare `except` and limit the `try` clause (977)
- fix python cmake error (976)
- Instantiate RunOptions first when training. (1019)
- Fix complier type in cmake: `CMAKE_COMPILER_IS_GNUCXX` (1038)
- other cleanups of the code (968 970 975 999 1004 1002 1001 1010 1014 1012 1011 1021 1036 1037)


Contributors
* Han Bao
* Roberto Car
* Junhan Chang
* Yixiao Chen
* Ye Ding
* Weinan E
* Jiequn Han
* Li'ang Huang
* Weile Jia
* Zeyu Li
* Ziyao Li
* Yinnian Lin
* Yihao Liu
* Xinzijian Liu
* Denghui Lu
* Marián Rynik
* Shaochen Shi
* Ping Tuo
* Bo Wang
* Haidi Wang
* Han Wang
* Yingze Wang
* Yu Xia
* Fengbo Yuan
* Jiabin Yang
* Haotian Ye
* Jinzhe Zeng
* Duo Zhang
* Linfeng Zhang
* Yuzhi Zhang

New features:
- Support input script in yaml format 271
- Enable setting test size individually for each system 267

Improvements:
- Support building with TF2.3
- Improved documentation. Automatically generated doc page. Explained keys in the training input script. 297
- Compile with latest stable release of lammps (stable_29Oct2020). Old releases are not supported anymore. 302
- Multi-device and multi-implementation support for OPs. 254

Bug fixings:
- Allow ntypes_model > ntypes_data (fix 261) 296
- Bug in setting `coeff_ener` when `coeff_atom_ener` is given 290

Bug fixings
- Fix several compiling issues of v1.2.3

New feature:
- added tool to convert a v1.2 compatible model to v1.3 compatible model.

Bug fixing:
- neighbor list bug in the GPU implementation.

New features:
- fast gelu implementation
- relax the restriction on the maximal number of neighbors to 4096
- one-sided embedding net

Bug fixings:
- check if the cell volume is positive.
- bugs of dplr
- loading coord when data has only one frame
- ase calculator None-PBC system
- details printing when testing on multiple systems.
- bugs in transform

New features of dp train:
- Polarizability and dipole fitting
- If provided with `stop_lr`, the `decay_rate` will be computed automatically
- Support non-pbc system: add an empty file named nopbc to the data system.
- Use envs `TF_INTRA_OP_PARALLELISM_THREADS` and `TF_INTER_OP_PARALLELISM_THREADS` to control the multi-threading of tf, clean up command line options.
- When the key `systems` is provided with a string, all possible systems will be recursively searched within the dir given by `systems`
- User specific atomic energy.
- Exclude types when building descriptors.
- User specific activation function and network precision.
- Transform neural network parameters.

New features of dp test:
- Recursively search all possible systems in a directory.
- Weighted average of test results

Bug fixings:
- Cuda version compatibility
- raw_to_set when nframe == 1
- Slow test when testing multiple systems
- Potentially wrong neighbor list when number of neighbors is larger than 256.
- Fix bug 163 : broken lib link of dp_ipi
- Infering number of types from data