File size: 4,611 Bytes
2e1dc7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
132
133
134
135
136
137
138
from functools import partial
from typing import Dict

# from models.musicgen_lm import MusicgenLm


def transfer(src, dest, srcname, destname):
    dest[destname].copy_(src[srcname])


def transfer_weights(af_dict: Dict, lag_dict: Dict, n_layers: int,
                     has_xatt: bool) -> Dict:

    # af_dict = audiocraft_lm.state_dict()
    # lag_dict = lag_lm.state_dict()

    t = partial(transfer, af_dict, lag_dict)

    hop = 3 if has_xatt else 2

    # transfer conditioner
    for s in ("weight", "bias"):
        lag_dict[f"conditioner.output_proj.{s}"] = af_dict[
            f"condition_provider.conditioners."
            f"description.output_proj.{s}"].clone()
        # t(
        #     f"condition_provider.conditioners.description.output_proj.{s}",
        #     f"conditioner.output_proj.{s}",
        # )

    # transfer embeddings
    for i in range(4):
        t(f"emb.{i}.weight", f"decoder.token_emb.emb.{i}.weight")

    # transfer decoder layers
    for layeridx in range(n_layers):

        lamsi_prefix = "decoder.attn_layers.layers"
        af_prefix = f"transformer.layers.{layeridx}"

        # ---- LAMSI SUBLAYER 0: self attention ----
        lamsi_prefix_layer = f"{lamsi_prefix}.{layeridx * hop}"

        # att norm
        for s in ("weight", "bias"):
            t(f"{af_prefix}.norm1.{s}", f"{lamsi_prefix_layer}.0.0.{s}")

        # attention q, k, v, out
        af_qkv = af_dict[f"{af_prefix}.self_attn.in_proj_weight"]
        assert (af_qkv.shape[0] % 3) == 0
        dim = af_qkv.shape[0] // 3
        att_q = af_qkv[:dim]
        att_k = af_qkv[dim:dim * 2]
        att_v = af_qkv[dim * 2:]
        lag_dict[f"{lamsi_prefix_layer}.1.to_q.weight"].copy_(att_q)
        lag_dict[f"{lamsi_prefix_layer}.1.to_k.weight"].copy_(att_k)
        lag_dict[f"{lamsi_prefix_layer}.1.to_v.weight"].copy_(att_v)
        t(
            f"{af_prefix}.self_attn.out_proj.weight",
            f"{lamsi_prefix_layer}.1.to_out.weight",
        )

        # ---- LAMSI SUBLAYER 1: cross attention ----
        if has_xatt:
            lamsi_prefix_layer = f"{lamsi_prefix}.{layeridx * hop + 1}"

            # xatt norm
            for s in ("weight", "bias"):
                t(f"{af_prefix}.norm_cross.{s}",
                  f"{lamsi_prefix_layer}.0.0.{s}")

            # xatt q, k, v, out
            af_xqkv = af_dict[f"{af_prefix}.cross_attention.in_proj_weight"]
            assert (af_xqkv.shape[0] % 3) == 0
            dim = af_xqkv.shape[0] // 3
            xatt_q = af_xqkv[:dim]
            xatt_k = af_xqkv[dim:dim * 2]
            xatt_v = af_xqkv[dim * 2:]
            lag_dict[f"{lamsi_prefix_layer}.1.to_q.weight"].copy_(xatt_q)
            lag_dict[f"{lamsi_prefix_layer}.1.to_k.weight"].copy_(xatt_k)
            lag_dict[f"{lamsi_prefix_layer}.1.to_v.weight"].copy_(xatt_v)
            t(
                f"{af_prefix}.cross_attention.out_proj.weight",
                f"{lamsi_prefix_layer}.1.to_out.weight",
            )

        # ---- LAMSI SUBLAYER 2: feed forward ----
        lamsi_prefix_layer = f"{lamsi_prefix}.{layeridx * hop + (hop - 1)}"

        # ff norm
        for s in ("weight", "bias"):
            t(f"{af_prefix}.norm2.{s}", f"{lamsi_prefix_layer}.0.0.{s}")

        # ff layers
        t(f"{af_prefix}.linear1.weight",
          f"{lamsi_prefix_layer}.1.ff.0.0.weight")

        t(f"{af_prefix}.linear2.weight", f"{lamsi_prefix_layer}.1.ff.2.weight")

    # transfer final norm
    for s in ("weight", "bias"):
        t(f"out_norm.{s}", f"decoder.attn_layers.final_norm.{s}")

    # transfer output linears
    for i in range(4):
        t(f"linears.{i}.weight", f"decoder.to_logits.linears.{i}.weight")

    # lag_lm.load_state_dict(lag_dict)

    return lag_dict


if __name__ == "__main__":
    import torch
    import config as cfg
    from models.musicgen_lm import MusicgenLm
    from hyperparameters import PretrainedMelodyLmParams

    lag_lm = MusicgenLm(PretrainedMelodyLmParams())

    lag_state_dict = lag_lm.state_dict()
    af_state_dict = torch.load(cfg.weights_dir() / "audiocraft-melody.pt")

    print(
        sum(p.numel()
            for k, p in af_state_dict.items()
            if "conditioner" not in k))
    print(
        sum(p.numel()
            for k, p in lag_state_dict.items()
            if "conditioner" not in k))
    print(list((k, p.numel()) for k, p in af_state_dict.items()))
    print(list((k, p.numel()) for k, p in lag_state_dict.items()))

    new_lag_dict = transfer_weights(af_state_dict, lag_state_dict, 48, False)

    torch.save(new_lag_dict, cfg.weights_dir() / "lm-melody-weights.pt")