v2.0.0

[amcadmus]

Breaking changes to v1.3

• 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.
• 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 (model compression #350 add ut for model compression #586 optimize UT for dp transfer and dp compress #610 Reconstruct the interface of model compression #921 add model compression support for models without training script #948 speedup model compression UTs #956 add model compression training support for deepmd-kit #1000 rm 'dp transfer' from the model compression workflow #1008 decouple activation function's type from model compression's process in SE_A, now tanh & gelu is both available. #1020 decouple activation function's type from model compression's process in SE_A, now tanh & gelu is both available. #1020 fix model compression error #1043)
• Parallel training (Replace PS-Worker mode with multi-worker one. #892 Find available GPUs in an elegant way. #905 Fix the expanding logic of SLURM_JOB_NODELIST and add unit tests for parallel training. #913 Use a local Numpy random generator. #1030 Avoid the same batch sequence among workers. #1032) (Bytedance)
• ROCm device support (Add ROCM support for DEEPMD #656 )
• New descriptor: three body embedding (se_e3)
• Hybridization of descriptors (hybrid)
• Type embedding
• Training and inference the dipole (vector) and polarizability (matrix). (API modification: support combined dipole training mode and global dipole test mode. #495 Support loss normalization #538 add lammps compute style for atomic deep tensor #927)
• Support derivatives of the tensor properties. (make deep tensor deriv again #805)
• Split of training and validation dataset.
• Model deviation for virial
• Add subcommand and python interface to calculate model-deviation (Add subcommand to calculate model-deviation #715)
• Automatically determine the sel from the training data. (automatically determine the sel from the training data. #831)
• Building with lammps with plugin mode (clean library and header dependencies #930 add LAMMPS plugin #945)

Performance improvement:

• More efficient training: all customized OPs are implemented with GPU.
• MPI support for atomic model deviation support MPI and other atom_styles for LAMMPS atomic keyword #628
• speedup ROCm kernels which use atomicAdd (speedup ROCm kernel which uses atomicAdd #809 Synchronize CUDA _r modifications to ROCM #815 ) (from ByteDance)
• speedup CUDA kernels (use atomicAdd inside) by reducing the global memory write (speedup CUDA kernels(use atomicAdd inside) by reducing the global memory write #811)
• speedup tabulate cuda kernel by reducing shm using (speedup tabulate cuda kernel by reducing shm using #830) (Bytedance)
• speedup format_nlist_b (speed up format_nlist_b #832 Synchronize format_nlist_b in CUDA with ROCm #845)
• speedup scan_nlist kernel (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 append to out_file when LAMMPS restarts #640
• Document and examples for the C++ interface add an example of C++ inference to doc #652 Doc api cc #663
• Instructions for the i-pi Add instructions for i-PI #660
• Document for the network size and sel Added doc for netsize setting, num_nodes specification, and sel setting in doc/troubleshooting/ #657
• Use fmod to wrap the coord of atoms (solve slow PBC) (Use fmod to wrap the coord of atoms. UT for force/virial ops #741)
• bit operations to encode neighbor information
• add CUDA/ROCM buidling documents (add CUDA/ROCM documents & fix CI (fix #730) #739)
• add type-embedding developer doc (add type-embedding developer doc #762 update type embedding doc #967)
• add model compression support for models with exclude_types feature (add model compression support for models with exclude_types feature #754)
• improve the doc and user interface of model compression (Add a more detailed introduction for model compression #772)
• support converting models generated in v1.3 to 2.0 compatibility (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 (give a default value to T and convert models from v1.2 #789)
• improved documents for conda (add documents for conda #798 update conda install documents #925)
• throw a message if tf runtime is incompatible (throw a message if tf runtime is incompatible #797)
• capture OOM and print debug message (add OOM message #801)
• add message for DecodeError raised when using model compression (add message for DecodeError #839)
• Passing error to TF instead of exit (Passing error to TF instead of exit #918)
• refactor docs (refactor docs #952)
• add an example of nopbc and related docs (add an example of nopbc and related docs #994)
• add dp --version (add dp --version #995)
• add the argument tensorboard_freq to control sampling ratio during training. (Add the argument tensorboard_freq to control sampling ratio during training. #996)
• add sphinx plugins viewcode and intersphinx (add sphinx plugins viewcode and intersphinx #997)
• generate Python API document automatically (generate Python API document automatically #998)
• give a clear message if model.get_ntypes()<data.get_ntypes() (give a clear message if model.get_ntypes()<data.get_ntypes() #1016)
• add docstring for descrpt/se_e2_a (add docstring to DescrptSeA #1017)
• add docstring for fit/ener (add docstring for EnerFitting #1024)
• add InputNlist into API doc (add InputNlist into API doc #1009)
• save checkpoint files with step and keep recent files (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 (Add lib coord unittest #569)
• UT for model compression (add ut for model compression #586)
• UT for prod_force/virial ops (Add gpu op unittest in source/tests #703 Use fmod to wrap the coord of atoms. UT for force/virial ops #741)
• CI test Lammps build (test lammps builds #600)
• allow c++ tests to run without internet (allow c++ tests to run without internet #785)
• build low and high precision at the same time (build low and high precision at the same time #879)
• support to specify CUDA/ROCm root in python pkg building (support to specify CUDA/ROCm root in python pkg building #834) (Bytedance)
• use cached Session to speed up py tests (use cached Session to speed up py tests #833)
• remove cub include for CUDA>=11 (remove cub include for CUDA>=11 #866)
• Add Errcheck after every kernel function runs And merge redundant code (Add Errcheck after every kernel function runs And merge redundant code #855)
• adapt changes to auditwheel directory in manylinux (adapt changes to auditwheel directory in manylinux #889)
• enhance the cli to generate doc json file (enhance the cli to generate doc json file #891)
• raise warning before training if sel is not enough (raise warning before training if sel is not enough #914)
• make native MD compatible with v2.0 (make native MD compatible with v2.0 #950)
• fix type hints and add doc for exclude_types (fix type hints and add doc for exclude_types #1005)
• use TF's built-in method to get numpy dtype (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 Fix bug in coord.cu #566
• Relative force model deviation is not copied back at single precision fix issue #598 #599
• Correct way of allocating memory in float precision Fix bug in tf allocate_temp #612
• Fix TB logdir remove bug fix TB logdir remove bug #617
• Illegal nlist fix issue 668 #680
• Bug in prod_virial_grad that causes wrong results when training with virials fix bug in op definition prod_virial_grad_multi_device.cc.  #685
• Uniform random seed Fix uniform random seed #691
• Illegal nlist fix issue 668 #680
• Bug in prod_virial_grad that causes wrong results when training with virials fix bug in op definition prod_virial_grad_multi_device.cc.  #685
• Uniform random seed Fix uniform random seed #691
• fix bug of adding int to a None random seed (Fix bug of adding int to a None random seed #705)
• reuse the zero layer rather than building a new one (fix bug in ener fitting: zero net should share the parameters with the working net #714)
• fix bug in CI (add CUDA/ROCM documents & fix CI (fix #730) #739)
• fix bug 824 and Synchronize updates to CUDA cod (fix bug 824 and Synchronize updates to CUDA code #828)
• Fix the empty neighbor distance array in neighbor_stat.py (Fix the empty neighbor distance array in neighbor_stat.py  #882)
• fix InvalidArgumentError caused by zero sel and optimize zero matrix (fix InvalidArgumentError caused by zero sel and optimize zero matrix #900)
• fix 'NoneType' has no len() in auto_sel (fix 'NoneType' has no len() in auto_sel #911)
• set input DeepmdData.type_map to input type_map (set input DeepmdData.type_map to input type_map #924)
• Fix member declartion of deepmd and deepmd.entrypoints. (Fix member declartion of deepmd and deepmd.entrypoints. #922)
• add aliases to Arguments (add aliases to Arguments #933)
• fix bug of gelu activation function (fix bug of gelu activation function #939)
• convert decay_rate to stop_lr from old inputs (convert decay_rate to stop_lr from old inputs #949)
• only enable link what you use on GNU compilers (cherry pick #536: only enable link what you use on GNU compilers #962)
• Do not find protobuf for python (cherry pick #535: Do not find protobuf for python #963)
• fix an error in stress by ase interface (cherry pick #338: fix an error in stress by ase interface  #964)
• remove bare except and limit the try clause (remove bare except and limit the try clause #977)
• fix python cmake error (fix python cmake error #976)
• Instantiate RunOptions first when training. (Instantiate RunOptions first when training. #1019)
• Fix complier type in cmake: CMAKE_COMPILER_IS_GNUCXX (Fix typo: CMAKE_COMPILER_IS_GNUCXX #1038)
• other cleanups of the code (add code comment for InputNlist struct #968 [ImgBot] Optimize images #970 update contributors #975 unpin tensorflow version #999 add a note for DP-GUI #1004 fix a bug in setup.py #1002 fix incorrect type hints and docstrings #1001 add examples to docstings #1010 bump sphinx_rtd_theme version #1014 show citation in the GitHub #1012 move parameters docstring from constructor method to class itself #1011 Fix typo: easy-install.md #1021 get library extension suffix from built-in method #1036 clean up the unnecessary including files of device source files #1037)

Contributors

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

---

This discussion was created from the release v2.0.0.