[BUG] _check-whether-"sel"-is-big-enough not working | kit-2.0-beta3_

[Vibsteamer]

Summary
Checking for whether the number of "sel" is big enough in training is not working. Users can train with extremely small values without "WARNING".
But the performance gain in training keeps increasing when decreasing "sel" to rather small value (may mean "sel" did work) .
The "rcut" in test is larger as 12.0, not sure if there is some limit in counting neighbor atoms that was broken and led to the problem.

Deepmd-kit version, installation way, input file, running commands, error log, etc.

DEEPMD INFO    installed to:         /tmp/pip-req-build-qqv2ggzp/_skbuild/linux-x86_64-3.9/cmake-install
DEEPMD INFO    source :              v2.0.0.b3
DEEPMD INFO    source brach:         HEAD
DEEPMD INFO    source commit:        de428e3
DEEPMD INFO    source commit at:     2021-07-04 22:12:13 +0800
DEEPMD INFO    build float prec:     double
DEEPMD INFO    build with tf inc:    /opt/deepmd-kit-2.0.0.b3/lib/python3.9/site-packages/tensorflow/include;/opt/deepmd-kit-2.0.0.b3/lib/python3.9/site-packages/tensorflow/include
DEEPMD INFO    build with tf lib:

Platform : ALI - ehpc - beijing

Steps to Reproduce
please use the attached input("rcut":12.0, "sel": [10,10,10]) and data to train.

Further Information, Files, and Links
input

{
    "model": {
        "descriptor": {
            "type": "se_e2_a",
            "sel": [
                10,
                10,
                10
            ],
            "rcut_smth": 2.0,
            "rcut": 12.0,
            "neuron": [
                25,
                50,
                100
            ],
            "resnet_dt": false,
            "axis_neuron": 12,
            "type_one_side": true,
            "seed": 1801819940,
            "_activation_function": "tanh"
        },
        "fitting_net": {
            "neuron": [
                240,
                240,
                240
            ],
            "resnet_dt": true,
            "_coord_norm": true,
            "_type_fitting_net": false,
            "seed": 2375417769,
            "_activation_function": "tanh"
        },
        "type_map": [
            "Mg",
            "Al",
            "Cu"
        ]
    },
    "learning_rate": {
        "type": "exp",
        "start_lr": 0.001,
        "decay_steps": 80000,
        "stop_lr": 5e-8,
        "_decay_rate": 0.95
    },
    "loss": {
        "start_pref_e": 0.02,
        "limit_pref_e": 2,
        "start_pref_f": 1000,
        "limit_pref_f": 1,
        "start_pref_v": 0.01,
        "limit_pref_v": 1
    },
    "training": {
        "training_data": {
           "systems": "../../data/",
           "batch_size": "auto"
           },
       "validation_data":{
            "systems":          "../../data/",
            "batch_size":       1,
            "numb_btch":        4,
            "_comment":         "that's all"
        },
        "numb_steps": 16000000,
        "seed": 3982377700,
        "_comment": "that's all",
        "disp_file": "lcurve.out",
        "disp_freq": 2000,
        "numb_test": 1,
        "save_freq": 2000,
        "save_ckpt": "model.ckpt",
        "disp_training": true,
        "time_training": true,
        "profiling": false,
        "profiling_file": "timeline.json"
    }
}

data
https://pan.baidu.com/s/139Rsf7qHF6x750xMekAyhQ
uaxz

[njzjz]

The warning was added back in #914, available in v2.0.0.b4. The check will do once before training.

For runtime checking, I may ask @denghuilu if it is effective. If not, we may not need it during training. But for MD simulations, we can consider to add an extra option to check it every N time step.

[Vibsteamer]

Thank you, may not need during training.
Solved.
Best~