Add Deep Potential plugin for CP2K by Cloudac7 · Pull Request #1620 · cp2k/cp2k

Cloudac7 · 2021-08-09T10:05:00Z

Deep Potential (DP) is a machine learning potential based on deep neural network, first proposed by Zhang L. et al. (Doi: 10.1016/j.cpc.2018.03.016) By learning from coordinates as well as energy and forces of different structures from DFT calculations, a potential describing the whole system could be frozen to a file. By reading the file, program could yield energy, forces etc. from structures, and then MD simulation with DFT accuracy could be fastened greatly. This PR add Deep Potential support for CP2K, with potential file (e.g. graph.pb) supplied, CP2K could run MD simulation.

Only DeePMD-kit >= 2.0.0 is supportted.

To enable the DeePMD-kit feature, --with-deepmd as well as --with-tfcc flag in toolchain should not be set to no. The following ways installing has been added.

Please set --deepmd-mode to gpu if GPU support is needed, and load CUDA environment before installing.
From exist DeePMD-kit C++ interface, with$deepmd_root and $tensorflow_root (Described here in detail), by setting --with-deepmd=$deepmd_root and --with-tfcc=$tensorflow_root.
From source code and library from Github and Anaconda Cloud, by setting the both flags to install. For now, DeePMD-kit v2.0.0.b1 and Tensorflow C++ interface 2.3.0 will be installed, and only CUDA Toolkit 10.1 is OK from this way.
From system, by adding the header files and library to where ld command could read directly.

The result from CP2K has been tested to be the same with which from Lammps, supportted by Deepmodeling.

oschuett · 2021-08-09T10:29:18Z

Thanks for the PR. This is certainly impressive work!

As a first step, I'd ask you to remove any merge commits from your git history, ie. make it linear.

Furthermore, I'd suggest you squash the git history to a single (or few) commit(s).

oschuett

Thanks for fixing the git history. I left a couple of comments. Once everything is up and running it would also be good to add some regression tests.

oschuett · 2021-08-10T07:30:53Z

tools/toolchain/scripts/stage6/install_deepmd.sh

+      echo "libtensorflow_cc-${tfcc_ver}.tar.bz2 is found"
+    else
+      download_pkg ${DOWNLOADER_FLAGS} ${deepmd_sha256} \
+        https://anaconda.org/deepmodeling/libtensorflow_cc/2.3.0/download/linux-64/libtensorflow_cc-${tfcc_ver}.tar.bz2


Installing Tensorflow via Conda is a nice solution given how painful it is to build from source. However, I'm wondering why are you not using the "official" package from conda forge?

Installing Tensorflow via Conda is a nice solution given how painful it is to build from source. However, I'm wondering why are you not using the "official" package from conda forge?

The packages used here are provided by Deep Modeling, the main developer group of DeePMD-kit. I will also try the package on conda-forge.

The packages used here are provided by Deep Modeling, the main developer group of DeePMD-kit. I will also try the package on conda-forge.

Failed because of the lost of include/google/protobuf/type.pb.h. The package on conda-forge seems not include some headers needed by DeePMD-kit.

That header file is probably part of libprotobuf. So, I guess installing Conda packages by hand is not a very robust solution after all.

I think, then we should keep Tensorflow out of the toolchain for now. It's actually not that hard to install via pip and most HPC centers also provide it.

I've opened #1622 to make Tensorflow available in the CI. Once it's merged you can rebase this PR and continue testing.

oschuett · 2021-08-10T07:36:46Z

tools/toolchain/scripts/stage6/install_deepmd.sh

+    ;;
+esac
+
+cub_link="https://github.com/NVlabs/cub/archive/c3cceac115c072fb63df1836ff46d8c60d9eb304.zip"


I believe CUB is now included in the regular CUDA Toolkit. So, there is no need to install it separately anymore.

Well, in DeePMD-kit repository, they just use the very fork of CUB as submodule, while removing it error would be raised when compiling. So here I tried to add it to DeePMD-kit code by downloading code archive file directly. According to your description, it might be OK building without it, and I will have a try.

I believe CUB is now included in the regular CUDA Toolkit. So, there is no need to install it separately anymore.

In their latest commits, they remove extra CUB requirement for CUDA>=11, shown here, as it has been included in CUDA Toolkit. However, for former CUDA versions, it might still be needed.

Ok, then let's keep it for a year or so until everybody has upgrade to CUDA >= 11.

oschuett · 2021-08-10T07:42:29Z

src/force_fields_input.F

+!> \param nonbonded ...
+!> \param section ...
+!> \param start ...
+!> \author teo


You can put your own name here or remove the \author tag as you prefer.

oschuett · 2021-08-10T07:44:11Z

src/manybody_deepmd.F

+      dcoord => dpmd_coord
+      dbox => dpmd_box
+      datype => dpmd_atype
+!      IF (do_parallel) THEN


Left overs copied from QUIP? See also remainder of this routine.

oschuett · 2021-08-10T07:52:46Z

tools/toolchain/scripts/arch_base.tmpl

 #
 FCDEBFLAGS  = ${FCDEBFLAGS}
 CFLAGS      = ${CFLAGS}
+CXXFLAGS    = ${CXXFLAGS}


Does DeePMD-kit also offer a C API or would it be easy to add? So, far we've managed to keep C++ out of CP2K and I'd really like to keep it that way.

dev-zero · 2021-08-10T08:51:26Z

src/deepmd_c_wrapper.h

+ * \brief A minimal wrapper for DeepMD-kit C++ interface.
+ * \author Yongbin Zhuang and Yunpei Liu
+ ******************************************************************************/
+


include guard #ifndef DEEPMOD_C_WRAPPER_H ... missing

dev-zero · 2021-08-10T08:56:17Z

src/deepmd_wrapper.F

+      END DO
+      c_model(N + 1) = C_NULL_CHAR
+#if(__DEEPMD)
+      create_nnp%ptr = create_nnp_c(c_model)


you should be able to simply pass along TRIM(model) // C_NULL_CHAR directly without having to copy, see for example src/pw/pw_fpga.F

dev-zero · 2021-08-10T08:58:20Z

src/deepmd_wrapper.F

+                               dener, dforce, dvirial, datom_ener, &
+                               datom_virial, dcoord, datype, dbox) BIND(C, name="compute_nnp")
+         USE ISO_C_BINDING
+      TYPE(C_PTR), INTENT(IN), VALUE                     :: nnp, vecsize, dener, dforce, dvirial, &


Can you declare the proper types rather than simply declare everything as a C_PTR?
Also, I somehow doubt that everything is INTENT(IN).

dev-zero · 2021-08-10T09:04:39Z

src/deepmd_wrapper.F

+#if(__DEEPMD)
+      CALL compute_nnp_c(pot, C_LOC(vecsize), C_LOC(dener), C_LOC(dforce(1)), &
+                         C_LOC(dvirial(1)), C_LOC(datom_ener(1)), C_LOC(datom_virial(1)), &
+                         C_LOC(dcoord(1)), C_LOC(datype(1)), C_LOC(dbox(1)))


Wouldn't C_LOC work on the arrays directly rather than taking the first argument first?
Also, it should be possible to drop POINTER but they should likely to be declared CONTIGUOUS instead.

dev-zero · 2021-08-10T09:08:11Z

src/manybody_deepmd.F

+      TYPE(deepmd_data_type), POINTER            :: deepmd_data
+      TYPE(deepmd_pot_type), POINTER             :: deepmd
+      TYPE(nnp), SAVE                                :: deepmd_tmp_pot1, &
+                                                        deepmd_tmp_pot2


No. If you have to preserve data you have to add it to a suitable env. No global vars or shared state.

dev-zero · 2021-08-10T09:08:26Z

src/manybody_deepmd.F

+
+   PRIVATE
+   LOGICAL, SAVE, PRIVATE :: do_create(2) = .TRUE.
+   INTEGER, SAVE, PRIVATE :: do_count = 0


No, see further down below.

Cloudac7 · 2021-08-11T03:44:12Z

Thanks for reviewing. We will check our code and modify then.

oschuett · 2021-08-16T11:59:38Z

Tensorflow is now readily available within the CI containers (#1622). So, please rebase your branch and adopt install_deepmd.sh.

oschuett · 2021-10-17T08:10:08Z

Gentle ping, is this PR still being worked on?

oschuett · 2022-01-27T11:32:14Z

@deepmodeling, does anybody know if there is still any interest in this PR?

oschuett · 2022-03-13T22:08:15Z

I'm going to close this PR now. Personally, I think it's a really pity because obviously a lot of work went into it. But I'm afraid this wouldn't be useful to anybody without further support from the authors.

robinzyb · 2022-08-13T12:43:59Z

I'm going to close this PR now. Personally, I think it's a really pity because obviously a lot of work went into it. But I'm afraid this wouldn't be useful to anybody without further support from the authors.

hi, @oschuett,
@Cloudac7 and me are the contributors to this pr. I will make a new pr since my colleagues have developed relevant FEP methods based on DPMD(https://aip.scitation.org/doi/10.1063/5.0098330) and we are planning to reorganize these codes in a more pretty format.

robinzyb · 2022-08-13T12:46:21Z

@oschuett is it ok for opening a new subsection named DPNN/DPMD under FORCE_EVAL or nest the section in FORCE_EVAL/NNP

oschuett · 2022-08-13T20:40:37Z

Hi @robinzyb, this is great news!

is it ok for opening a new subsection named DPNN/DPMD under FORCE_EVAL or nest the section in

Yes, please create new subsections below FORCE_EVAL, MOTION/MD, and MOTION/FREE_ENERGY as you see fit.

Add DeePMD-kit C++ Interface

c06d0f9

Cloudac7 force-pushed the deepmd_latest branch from 0849dc8 to c06d0f9 Compare August 10, 2021 03:33

Make code Pretty

fcb0915

oschuett reviewed Aug 10, 2021

View reviewed changes

dev-zero reviewed Aug 10, 2021

View reviewed changes

dev-zero requested changes Aug 10, 2021

View reviewed changes

oschuett closed this Mar 13, 2022

Cloudac7 mentioned this pull request Dec 6, 2023

Deep potential plugin for CP2K implemented from C interface #3145

Merged

Conversation

Cloudac7 commented Aug 9, 2021

Uh oh!

oschuett commented Aug 9, 2021

Uh oh!

oschuett left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

oschuett Aug 10, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Cloudac7 commented Aug 11, 2021

Uh oh!

oschuett commented Aug 16, 2021

Uh oh!

oschuett commented Oct 17, 2021

Uh oh!

oschuett commented Jan 27, 2022

Uh oh!

oschuett commented Mar 13, 2022

Uh oh!

robinzyb commented Aug 13, 2022

Uh oh!

robinzyb commented Aug 13, 2022

Uh oh!

oschuett commented Aug 13, 2022

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

oschuett Aug 10, 2021 •

edited

Loading