File size: 3,656 Bytes
d2f661a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
import gc

from fire import Fire
import torch
from omegaconf import OmegaConf

from ldcast.models.autoenc import autoenc, encoder
from ldcast.models.genforecast import analysis, training, unet

from train_nowcaster import setup_data


def setup_model(

    num_timesteps=5,

    model_dir="../models/test/",

    autoenc_weights_fn="../models/autoenc/autoenc-32-0.01.pt",

    use_obs=True,

    use_nwp=False,

    nwp_input_patches=4,

    num_nwp_vars=9,

    lr=1e-4

):
    enc = encoder.SimpleConvEncoder()
    dec = encoder.SimpleConvDecoder()    
    autoencoder_obs = autoenc.AutoencoderKL(enc, dec)
    autoencoder_obs.load_state_dict(torch.load(autoenc_weights_fn))    

    autoencoders = []
    input_patches = []
    input_size_ratios = []
    embed_dim = []
    analysis_depth = []
    if use_obs:
        autoencoders.append(autoencoder_obs)
        input_patches.append(1)
        input_size_ratios.append(1)
        embed_dim.append(128)
        analysis_depth.append(4)
    if use_nwp:
        autoencoder_nwp = autoenc.DummyAutoencoder(width=num_nwp_vars)
        autoencoders.append(autoencoder_nwp)
        input_patches.append(nwp_input_patches)
        input_size_ratios.append(2)
        embed_dim.append(32)
        analysis_depth.append(2)

    analysis_net = analysis.AFNONowcastNetCascade(
        autoencoders,
        input_patches=input_patches,
        input_size_ratios=input_size_ratios,
        train_autoenc=False,
        output_patches=num_timesteps,
        cascade_depth=3,
        embed_dim=embed_dim,
        analysis_depth=analysis_depth
    )

    model = unet.UNetModel(in_channels=autoencoder_obs.hidden_width,
        model_channels=256, out_channels=autoencoder_obs.hidden_width,
        num_res_blocks=2, attention_resolutions=(1,2), 
        dims=3, channel_mult=(1, 2, 4), num_heads=8,
        num_timesteps=num_timesteps, context_ch=analysis_net.cascade_dims
    )

    (ldm, trainer) = training.setup_genforecast_training(
        model, autoencoder_obs, context_encoder=analysis_net,
        model_dir=model_dir, lr=lr
    )
    gc.collect()
    return (ldm, trainer)


def train(

    future_timesteps=8,

    use_obs=True,

    use_nwp=False,

    sample_shape=(4,4),

    batch_size=8,

    sampler=None,

    ckpt_path=None,

    initial_weights=None,

    strict_weights=True,

    model_dir=None,

    lr=1e-4

):
    if sampler is None:
        sampler_file = None
    else:
        sampler_file = {
            s: f"{sampler}_{s}.pkl" for s in ["test", "train", "valid"]
        }

    print("Loading data...")
    datamodule = setup_data(
        future_timesteps=future_timesteps,
        use_obs=use_obs,
        use_nwp=use_nwp,
        sampler_file=sampler_file,
        batch_size=batch_size,
        sample_shape=sample_shape
    )

    print("Setting up model...")
    (model, trainer) = setup_model(
        num_timesteps=future_timesteps//4,
        use_obs=use_obs,
        use_nwp=use_nwp,
        model_dir=model_dir,
        lr=lr
    )
    if initial_weights is not None:
        print(f"Loading weights from {initial_weights}...")
        model.load_state_dict(
            torch.load(initial_weights, map_location=model.device),
            strict=strict_weights
        )

    print("Starting training...")
    trainer.fit(model, datamodule=datamodule, ckpt_path=ckpt_path)


def main(config=None, **kwargs):
    config = OmegaConf.load(config) if (config is not None) else {}
    config.update(kwargs)
    train(**config)


if __name__ == "__main__":
    Fire(main)