poke-vae/inference.py at main · tiboitel/poke-vae · GitHub

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
131
import json
import torch
import numpy as np
from sklearn.preprocessing import StandardScaler

from model import VAE
from data import encode_types, encode_talent
from utils import print_pokemon_entries


class PokeVAEGenerator:
    """
    Inference-only wrapper around a trained PokeVAE.
    Owns model, scaler, config, and vocabularies.
    """

    def __init__(
            self,
            model: VAE,
            cfg: dict,
            scaler: StandardScaler,
            type_to_idx: dict,
            talent_to_idx: dict,
            device: torch.device,
            ):
        self.model = model
        self.cfg = cfg
        self.scaler = scaler
        self.type_to_idx = type_to_idx
        self.talent_to_idx = talent_to_idx
        self.idx_to_type = {i: t for t, i in type_to_idx.items()}
        self.idx_to_talent = {i: t for t, i in talent_to_idx.items()}
        self.device = device

        self.model.eval()

    # ------------------------------------------------------------------
    # Construction
    # ------------------------------------------------------------------

    @classmethod
    def from_checkpoint(cls, path: str, device: torch.device | None = None):
        ckpt = torch.load(path, map_location="cpu")

        cfg = ckpt["config"]
        device = device or torch.device(
                "cuda" if torch.cuda.is_available() else "cpu"
                )

        model = VAE(cfg).to(device)
        model.load_state_dict(ckpt["model"])

        scaler = StandardScaler()
        scaler.mean_ = ckpt["scaler_mean"].numpy()
        scaler.scale_ = ckpt["scaler_scale"].numpy()
        scaler.n_features_in_ = scaler.mean_.shape[0]

        return cls(
                model=model,
                cfg=cfg,
                scaler=scaler,
                type_to_idx=ckpt["type_to_idx"],
                talent_to_idx=ckpt["talent_to_idx"],
                device=device,
                )

    # ------------------------------------------------------------------
    # Encoding
    # ------------------------------------------------------------------

    def encode_pokemon(self, pokemon):
        stats = np.array(
                [[pokemon["base"][s] for s in self.cfg["stats_order"]]],
                dtype=np.float32,
                )
        stats = self.scaler.transform(stats)

        types = encode_types(pokemon, self.type_to_idx)[None, :]
        talent_idx = encode_talent(pokemon, self.talent_to_idx)

        talents = np.zeros((1, self.cfg["talent_dim"]), dtype=np.float32)
        talents[0, talent_idx] = 1.0

        x = np.concatenate([stats, types, talents], axis=1)
        x = torch.tensor(x, dtype=torch.float32).to(self.device)

        with torch.no_grad():
            mu, logvar = self.model.encode(x)
            z = self.model.reparameterize(mu, logvar)

        return z

    # ------------------------------------------------------------------
    # Sampling
    # ------------------------------------------------------------------

    def sample_random(self, n: int = 1):
        z = torch.randn(n, self.cfg["latent_dim"], device=self.device)

        with torch.no_grad():
            stats, types, talents = self.model.decode(z)

        stats = self.scaler.inverse_transform(stats.cpu().numpy())

        types = (torch.sigmoid(types) > 0.5).int()
        for i in range(types.size(0)):
            if types[i].sum() == 0:
                types[i, torch.randint(0, self.cfg["type_dim"], (1,))] = 1

        talents = talents.argmax(dim=1)

        print_pokemon_entries(
                stats,
                types,
                talents,
                self.idx_to_type,
                self.idx_to_talent,
                self.cfg,
                )

def main():
    with open("pokemons.json", encoding="utf-8") as f:
        pokemons = json.load(f)

    model = PokeVAEGenerator.from_checkpoint("model.pt")

    print("Random:")
    model.sample_random(n=15)

if __name__ == "__main__":
    main()