Large energy and force Errors when using Atom type embedding (test for deepmd-2.0beta)

[Manyi-Yang]

>Summary

The energy and force Errors for training using Atom type embedding are large.

>Deepmd-kit version input.json lcurve.out

2.0.0.b0, conda

>input.json

 {
    "model": {
        "type_map": [
            "S"
        ],

	"type_embedding":{
	    "neuron":		[2, 4, 8],
	    "resnet_dt":	false,
	    "seed":		1
	},	    
        "descriptor": {
            "type": "se_a_tpe",
            "sel": [
                128
            ],
            "rcut_smth": 0.5,
            "rcut": 6.5,
            "neuron": [
                25,
                50,
                100
            ],
            "resnet_dt": false,
            "axis_neuron": 16,
            "seed": 722586222
        },
        
        "fitting_net": {
            "neuron": [
                240,
                240,
                240,
                240
            ],
            "resnet_dt": true,
            "seed": 711230366
        }
    },
    "learning_rate": {
        "type": "exp",
        "start_lr": 0.001,
        "decay_steps": 5000,
        "stop_lr": 3.51e-8
    },
    "loss": {
	"type":"ener",
        "start_pref_e": 0.02,
        "limit_pref_e": 1,
        "start_pref_f": 1000,
        "limit_pref_f": 1,
        "start_pref_v": 0.0,
        "limit_pref_v": 0.0
    },
    "training": {

	"seed":		10,
	"disp_file":	"lcurve.out",
        "stop_batch": 1000000,
        "disp_freq": 1000,
        "numb_test": 5,
        "save_freq": 1000,
        "save_ckpt": "model.ckpt",
        "profiling": false,
        "profiling_file": "timeline.json",
        "training_data": {
             "systems": ["../data.iters/iter.000000/02.fp/data.002",
                 "../data.iters/iter.000000/02.fp/data.032",
                 "../data.iters/iter.000000/02.fp/data.000",
                 "../data.iters/iter.000000/02.fp/data.034",
                 "../data.iters/iter.000001/02.fp/data.004",
                 "../data.iters/iter.000001/02.fp/data.036",
                 "../data.iters/iter.000001/02.fp/data.006",
                 "../data.iters/iter.000001/02.fp/data.038",
                 "../data.iters/iter.000002/02.fp/data.016",
                 "../data.iters/iter.000002/02.fp/data.048",
                 "../data.iters/iter.000002/02.fp/data.018",
                 "../data.iters/iter.000002/02.fp/data.050",
                 "../data.iters/iter.000003/02.fp/data.022",
                 "../data.iters/iter.000003/02.fp/data.054",
                 "../data.iters/iter.000003/02.fp/data.052",
                 "../data.iters/iter.000003/02.fp/data.020",
                 "../data.iters/iter.000004/02.fp/data.058",
                 "../data.iters/iter.000004/02.fp/data.024",
                 "../data.iters/iter.000004/02.fp/data.026",
                 "../data.iters/iter.000004/02.fp/data.056",
                 "../data.iters/iter.000005/02.fp/data.060",
                 "../data.iters/iter.000005/02.fp/data.028",
                 "../data.iters/iter.000005/02.fp/data.030",
                 "../data.iters/iter.000005/02.fp/data.062",
                 "../data.iters/iter.000006/02.fp/data.010",
                 "../data.iters/iter.000006/02.fp/data.008",
                 "../data.iters/iter.000006/02.fp/data.042",
                 "../data.iters/iter.000006/02.fp/data.040",
                 "../data.iters/iter.000007/02.fp/data.044",
                 "../data.iters/iter.000007/02.fp/data.014",
                 "../data.iters/iter.000007/02.fp/data.046",
                 "../data.iters/iter.000007/02.fp/data.012"],
             "batch_size": 1
	},
	     "validation_data":{
             "systems": ["../data.iters/iter.000000/02.fp/data.002",
                 "../data.iters/iter.000000/02.fp/data.032",
                 "../data.iters/iter.000000/02.fp/data.000",
                 "../data.iters/iter.000000/02.fp/data.034",
                 "../data.iters/iter.000001/02.fp/data.004",
                 "../data.iters/iter.000001/02.fp/data.036",
                 "../data.iters/iter.000001/02.fp/data.006",
                 "../data.iters/iter.000001/02.fp/data.038",
                 "../data.iters/iter.000002/02.fp/data.016",
                 "../data.iters/iter.000002/02.fp/data.048",
                 "../data.iters/iter.000002/02.fp/data.018",
                 "../data.iters/iter.000002/02.fp/data.050",
                 "../data.iters/iter.000003/02.fp/data.022",
                 "../data.iters/iter.000003/02.fp/data.054",
                 "../data.iters/iter.000003/02.fp/data.052",
                 "../data.iters/iter.000003/02.fp/data.020",
                 "../data.iters/iter.000004/02.fp/data.058",
                 "../data.iters/iter.000004/02.fp/data.024",
                 "../data.iters/iter.000004/02.fp/data.026",
                 "../data.iters/iter.000004/02.fp/data.056",
                 "../data.iters/iter.000005/02.fp/data.060",
                 "../data.iters/iter.000005/02.fp/data.028",
                 "../data.iters/iter.000005/02.fp/data.030",
                 "../data.iters/iter.000005/02.fp/data.062",
                 "../data.iters/iter.000006/02.fp/data.010",
                 "../data.iters/iter.000006/02.fp/data.008",
                 "../data.iters/iter.000006/02.fp/data.042",
                 "../data.iters/iter.000006/02.fp/data.040",
                 "../data.iters/iter.000007/02.fp/data.044",
                 "../data.iters/iter.000007/02.fp/data.014",
                 "../data.iters/iter.000007/02.fp/data.046",
                 "../data.iters/iter.000007/02.fp/data.012"],
	    "batch_size":	1,
	    "numb_btch":	3,
	    "_comment":		"that's all"
	},
	"_comment":	"that's all"
    }
}

lcurve.out

step rmse_val rmse_trn rmse_e_val rmse_e_trn rmse_f_val rmse_f_trn lr
0 4.64e+02 4.64e+02 2.90e+02 2.90e+02 9.26e-01 8.04e-01 1.0e-03
1000 2.75e+01 3.01e+01 4.40e-02 4.74e-02 8.69e-01 9.51e-01 1.0e-03
2000 2.86e+01 3.14e+01 2.69e-02 4.51e-02 9.03e-01 9.93e-01 1.0e-03
..................
163000 1.29e+01 1.28e+01 1.86e-02 2.95e-02 9.28e-01 9.14e-01 1.9e-04
164000 1.27e+01 1.42e+01 4.62e-02 4.76e-02 9.07e-01 1.02e+00 1.9e-04
165000 1.33e+01 1.14e+01 2.72e-02 1.08e-02 9.79e-01 8.40e-01 1.8e-04
166000 1.16e+01 1.25e+01 9.23e-02 7.24e-02 8.52e-01 9.21e-01 1.8e-04
167000 1.23e+01 1.14e+01 2.35e-02 6.55e-02 9.04e-01 8.38e-01 1.8e-04
168000 1.20e+01 1.41e+01 4.60e-02 1.97e-02 8.82e-01 1.04e+00 1.8e-04
169000 1.17e+01 1.18e+01 2.20e-02 3.75e-02 8.63e-01 8.65e-01 1.8e-04
170000 1.13e+01 1.00e+01 3.51e-02 9.29e-02 8.55e-01 7.52e-01 1.7e-04
171000 1.22e+01 1.06e+01 3.65e-02 1.23e-02 9.17e-01 7.98e-01 1.7e-04
172000 1.18e+01 1.10e+01 5.21e-02 2.40e-02 8.92e-01 8.29e-01 1.7e-04
173000 1.17e+01 1.13e+01 5.59e-02 2.00e-02 8.83e-01 8.54e-01 1.7e-04
174000 1.24e+01 1.10e+01 7.28e-02 7.33e-02 9.30e-01 8.31e-01 1.7e-04
175000 1.14e+01 1.25e+01 7.05e-02 5.81e-02 8.83e-01 9.64e-01 1.7e-04
176000 1.07e+01 1.23e+01 5.51e-02 2.85e-02 8.24e-01 9.48e-01 1.7e-04
177000 1.13e+01 1.18e+01 7.20e-02 4.19e-02 8.71e-01 9.11e-01 1.7e-04
178000 1.12e+01 1.11e+01 5.49e-02 1.91e-02 8.63e-01 8.60e-01 1.7e-04
179000 1.20e+01 1.04e+01 4.85e-02 3.96e-02 9.28e-01 8.04e-01 1.7e-04
180000 1.19e+01 1.19e+01 6.74e-02 1.78e-02 9.37e-01 9.48e-01 1.6e-04
181000 1.15e+01 1.17e+01 8.61e-02 1.21e-03 9.07e-01 9.26e-01 1.6e-04
182000 1.13e+01 1.04e+01 3.48e-02 1.01e-02 8.94e-01 8.22e-01 1.6e-04
183000 1.05e+01 1.10e+01 3.94e-02 6.62e-02 8.35e-01 8.75e-01 1.6e-04
184000 1.16e+01 1.14e+01 2.73e-02 8.72e-02 9.24e-01 9.01e-01 1.6e-04

[amcadmus]

Could you please follow the example for the type embedding? Especially the type of the descriptor.
https://github.com/deepmodeling/deepmd-kit/blob/devel/examples/water/se_e2_a_tebd/input.json

For the case of water, one write

	"type_embedding":{
	    "neuron":		[2, 4, 8],
	    "resnet_dt":	false,
	    "seed":		1
	},	    
	"descriptor" :{
	    "type":		"se_e2_a",
	    "sel":		[46, 92],
	    "rcut_smth":	0.50,
	    "rcut":		6.00,
	    "neuron":		[25, 50, 100],
	    "resnet_dt":	false,
	    "axis_neuron":	16,
	    "type_one_side":	true,
	    "seed":		1,
	    "_comment":		" that's all"
	},

The type_embedding will automatically be used in the descriptor with type "se_e2_a", which is equivalent to the old one "se_a"

[njzjz]

Hi Manyi, could you patch #1592 and see if it works for you?