File size: 7,707 Bytes
fd012a7
 
86f7f0a
 
 
 
 
 
 
fd012a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
86f7f0a
 
 
fd012a7
86f7f0a
fd012a7
 
 
 
86f7f0a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fd012a7
86f7f0a
 
 
 
 
 
 
 
 
 
 
 
 
 
fd012a7
86f7f0a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fd012a7
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
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
from typing import Union

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import tqdm
from transformers import LlamaModel, LlamaConfig

from midi_tokenizer import MIDITokenizerV1, MIDITokenizerV2, MIDITokenizer

config_name_list = ["tv1-medium", "tv2-medium", "tv2o-medium", "tv2-large", "tv2o-large"]


class MIDIModelConfig:
    def __init__(self, tokenizer: Union[MIDITokenizerV1, MIDITokenizerV2],
                 net_config: LlamaConfig, net_token_config: LlamaConfig):
        self.tokenizer = tokenizer
        self.net_config = net_config
        self.net_token_config = net_token_config
        self.n_embd = net_token_config.hidden_size

    @staticmethod
    def get_config(tokenizer_ver="v2", optimise_midi=True, n_layer=12, n_head=16, n_embd=1024, n_inner=4096):
        tokenizer = MIDITokenizer(tokenizer_ver)
        tokenizer.set_optimise_midi(optimise_midi)
        net_config = LlamaConfig(vocab_size=tokenizer.vocab_size,
                                 hidden_size=n_embd, num_attention_heads=n_head,
                                 num_hidden_layers=n_layer, intermediate_size=n_inner,
                                 pad_token_id=tokenizer.pad_id, max_position_embeddings=4096)
        net_token_config = LlamaConfig(vocab_size=tokenizer.vocab_size,
                                       hidden_size=n_embd, num_attention_heads=n_head // 4,
                                       num_hidden_layers=n_layer // 4, intermediate_size=n_inner // 4,
                                       pad_token_id=tokenizer.pad_id, max_position_embeddings=4096)
        return MIDIModelConfig(tokenizer, net_config, net_token_config)

    @staticmethod
    def from_name(name="tv2o-medium"):
        tv, size = name.split("-")
        tv = tv[1:]
        if tv[-1] == "o":
            o = True
            tv = tv[:-1]
        else:
            o = False
        if tv not in ["v1", "v2"]:
            raise ValueError(f"Unknown tokenizer version {tv}")
        if size == "medium":
            return MIDIModelConfig.get_config(tokenizer_ver=tv, optimise_midi=o,
                                              n_layer=12, n_head=16, n_embd=1024, n_inner=4096)
        elif size == "large":
            return MIDIModelConfig.get_config(tokenizer_ver=tv, optimise_midi=o,
                                              n_layer=24, n_head=16, n_embd=1024, n_inner=4096)
        else:
            raise ValueError(f"Unknown model size {size}")


class MIDIModel(nn.Module):
    def __init__(self, config: MIDIModelConfig, *args, **kwargs):
        super(MIDIModel, self).__init__()
        self.tokenizer = config.tokenizer
        self.net = LlamaModel(config.net_config)
        self.net_token = LlamaModel(config.net_token_config)
        self.lm_head = nn.Linear(config.n_embd, self.tokenizer.vocab_size, bias=False)
        self.device = "cpu"

    def to(self, *args, **kwargs):
        if "device" in kwargs:
            self.device = kwargs["device"]
        return super(MIDIModel, self).to(*args, **kwargs)

    def forward_token(self, hidden_state, x=None):
        """

        :param hidden_state: (batch_size, n_embd)
        :param x: (batch_size, token_sequence_length)
        :return: (batch_size, 1 + token_sequence_length, vocab_size)
        """
        hidden_state = hidden_state.unsqueeze(1)  # (batch_size, 1, n_embd)
        if x is not None:
            x = self.net_token.embed_tokens(x)
            hidden_state = torch.cat([hidden_state, x], dim=1)
        hidden_state = self.net_token.forward(inputs_embeds=hidden_state).last_hidden_state
        return self.lm_head(hidden_state)

    def forward(self, x):
        """
        :param x: (batch_size, midi_sequence_length, token_sequence_length)
        :return: hidden (batch_size, midi_sequence_length, n_embd)
        """

        # merge token sequence
        x = self.net.embed_tokens(x)
        x = x.sum(dim=-2)
        x = self.net.forward(inputs_embeds=x)
        return x.last_hidden_state

    def sample_top_p_k(self, probs, p, k, generator=None):
        probs_sort, probs_idx = torch.sort(probs, dim=-1, descending=True)
        probs_sum = torch.cumsum(probs_sort, dim=-1)
        mask = probs_sum - probs_sort > p
        probs_sort[mask] = 0.0
        mask = torch.zeros(probs_sort.shape[-1], device=probs_sort.device)
        mask[:k] = 1
        probs_sort = probs_sort * mask
        probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True))
        shape = probs_sort.shape
        next_token = torch.multinomial(probs_sort.reshape(-1, shape[-1]),
                                       num_samples=1, generator=generator).reshape(*shape[:-1], 1)
        next_token = torch.gather(probs_idx, -1, next_token).reshape(*shape[:-1])
        return next_token

    @torch.inference_mode()
    def generate(self, prompt=None, max_len=512, temp=1.0, top_p=0.98, top_k=20, generator=None):
        tokenizer = self.tokenizer
        max_token_seq = tokenizer.max_token_seq
        if prompt is None:
            input_tensor = torch.full((1, max_token_seq), tokenizer.pad_id, dtype=torch.long, device=self.device)
            input_tensor[0, 0] = tokenizer.bos_id  # bos
        else:
            prompt = prompt[:, :max_token_seq]
            if prompt.shape[-1] < max_token_seq:
                prompt = np.pad(prompt, ((0, 0), (0, max_token_seq - prompt.shape[-1])),
                                mode="constant", constant_values=tokenizer.pad_id)
            input_tensor = torch.from_numpy(prompt).to(dtype=torch.long, device=self.device)
        input_tensor = input_tensor.unsqueeze(0)
        cur_len = input_tensor.shape[1]
        bar = tqdm.tqdm(desc="generating", total=max_len - cur_len)
        with bar:
            while cur_len < max_len:
                end = False
                hidden = self.forward(input_tensor)[0, -1].unsqueeze(0)
                next_token_seq = None
                event_name = ""
                for i in range(max_token_seq):
                    mask = torch.zeros(tokenizer.vocab_size, dtype=torch.int64, device=self.device)
                    if i == 0:
                        mask[list(tokenizer.event_ids.values()) + [tokenizer.eos_id]] = 1
                    else:
                        param_name = tokenizer.events[event_name][i - 1]
                        mask[tokenizer.parameter_ids[param_name]] = 1

                    logits = self.forward_token(hidden, next_token_seq)[:, -1:]
                    scores = torch.softmax(logits / temp, dim=-1) * mask
                    sample = self.sample_top_p_k(scores, top_p, top_k, generator=generator)
                    if i == 0:
                        next_token_seq = sample
                        eid = sample.item()
                        if eid == tokenizer.eos_id:
                            end = True
                            break
                        event_name = tokenizer.id_events[eid]
                    else:
                        next_token_seq = torch.cat([next_token_seq, sample], dim=1)
                        if len(tokenizer.events[event_name]) == i:
                            break
                if next_token_seq.shape[1] < max_token_seq:
                    next_token_seq = F.pad(next_token_seq, (0, max_token_seq - next_token_seq.shape[1]),
                                           "constant", value=tokenizer.pad_id)
                next_token_seq = next_token_seq.unsqueeze(1)
                input_tensor = torch.cat([input_tensor, next_token_seq], dim=1)
                cur_len += 1
                bar.update(1)
                if end:
                    break
        return input_tensor[0].cpu().numpy()