[BUG] input v0 converting to input v2 bug

[yhliu918]

attribute error when converting input versions
When I run input.json v0 under environment deepmd-kit v2.0.0.b3, it would first convert it into input_v2_compat.json. However, it still have key error like decay_rate/ n_neuron is not allowed in strict mode.
I attached my input file below:

{
    "model": {
        "descriptor": {
            "type": "se_a",
            "sel": [
                500
            ],
            "rcut_smth": 2.0,
            "rcut": 9.0,
            "neuron": [
                25,
                50,
                100
            ],
            "resnet_dt": false,
            "axis_neuron": 12,
            "seed": 1084011120
        },
        "fitting_net": {
            "n_neuron": [
                240,
                240,
                240
            ],
            "resnet_dt": true,
            "coord_norm": true,
            "type_fitting_net": false,
            "seed": 3151311092
        },
        "type_map": [
            "Re"
        ]
    },
    "learning_rate": {
        "type": "exp",
        "start_lr": 0.0001,
        "decay_steps": 25000,
        "decay_rate": 0.95
    },
    "loss": {
        "start_pref_e": 100,
        "limit_pref_e": 100,
        "start_pref_f": 1,
        "limit_pref_f": 1,
        "start_pref_v": 2.0,
        "limit_pref_v": 2.0
    },
    "training": {
        "systems": ['systems'],
        "set_prefix": "set",
        "stop_batch": 5000000,
        "batch_size": "auto",
        "seed": 813129136,
        "_comment": "that's all",
        "disp_file": "lcurve.out",
        "disp_freq": 2000,
        "numb_test": 2,
        "save_freq": 2000,
        "save_ckpt": "model.ckpt",
        "load_ckpt": "model.ckpt",
        "disp_training": true,
        "time_training": true,
        "profiling": false,
        "profiling_file": "timeline.json"
    }
}

The error log is like

[njzjz]

load_ckpt is also deprecated. See #907.

[yhliu918]

yeah, but n_neuron is not automatically converted into neuron? in other words, when I'm using v0 input, I should change the input file first by myself?

[amcadmus]

yeah, but n_neuron is not automatically converted into neuron? in other words, when I'm using v0 input, I should change the input file first by myself?

The format of your input is of v1.

What is the version of your dargs? Have you updated it to the latest release 0.2.6?

[yhliu918]

it's the latest version 0.2.6

[njzjz]

• n_neuron was fixed in add aliases to Arguments #932 and add aliases to Arguments #933, but it does not work from v1.3.0 to v1.3.3
• decay_rate and load_ckpt should be processed in compat.py. I think load_ckpt has already deprecated in v1, but we never remove it in examples and documents?
• coord_norm and type_fitting_net are keys in v0 that are deprecated in v1? But we never processed it in compat.py

[amcadmus]

* `n_neuron` was fixed in [add aliases to Arguments #932](https://github.com/deepmodeling/deepmd-kit/pull/932) and [add aliases to Arguments #933](https://github.com/deepmodeling/deepmd-kit/pull/933), but it does not work from v1.3.0 to v1.3.3

* `decay_rate` and `load_ckpt` should be processed in `compat.py`. I think `load_ckpt` has already deprecated in v1, but we never remove it in examples and documents?

* `coord_norm` and `type_fitting_net` are keys in v0 that are deprecated in v1? But we never processed it in `compat.py`

I will rm load_ckpt from examples.

coord_norm and type_fitting_net are deprecated keys