File size: 6,317 Bytes
e2b7617
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import math
import logging

import torch
import torch.nn as nn
from torch.nn import functional as F

logger = logging.getLogger(__name__)
from SCMG.config import varables

# class ModelConfig():
#     rate_dropout_embedding = 0.1
#     rate_dropout_residue = 0.1
#     rate_dropout_attention = 0.1
#     block_size=125
#     def __init__(self, size_vocab, **kwargs):
#         self.size_vocab = size_vocab
#         for k,v in kwargs.items():
#             setattr(self, k, v)

class CausalSelfAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        assert config[varables.DIM_ATTENTION] % config[varables.NUM_HEADS] == 0
        self.key = nn.Linear(config[varables.DIM_EMBEDDING], config[varables.DIM_ATTENTION])
        self.query = nn.Linear(config[varables.DIM_EMBEDDING], config[varables.DIM_ATTENTION])
        self.value = nn.Linear(config[varables.DIM_EMBEDDING], config[varables.DIM_ATTENTION])
        self.dropout_attention = nn.Dropout(config[varables.RATE_DROPOUT])
        self.dropout_residue = nn.Dropout(config[varables.RATE_DROPOUT])
        self.projection = nn.Linear(config[varables.DIM_ATTENTION], config[varables.DIM_EMBEDDING])
        self.register_buffer("mask", torch.tril(torch.ones(config[varables.SIZE_BLOCK], config[varables.SIZE_BLOCK]))
                                     .view(1, 1, config[varables.SIZE_BLOCK], config[varables.SIZE_BLOCK]))
        self.n_head = config[varables.NUM_HEADS]
        self.single_head_dim = config[varables.DIM_ATTENTION] // self.n_head
        self.attention_features = config[varables.DIM_ATTENTION]

    def forward(self, x, layer_past=None,boundary=None):
        B, T, C = x.size()
        k =   self.key(x).view(B, T, self.n_head,self.single_head_dim).transpose(1, 2)
        q = self.query(x).view(B, T, self.n_head,self.single_head_dim).transpose(1, 2)
        v = self.value(x).view(B, T, self.n_head,self.single_head_dim).transpose(1, 2)
        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
        if boundary is None:
            att = att.masked_fill(self.mask[:,:,:T,:T] == 0, float('-inf'))
        else:
            mask = torch.tril(torch.ones(config[varables.SIZE_BLOCK], config[varables.SIZE_BLOCK]))\
                            .view(1, 1, config[varables.SIZE_BLOCK], config[varables.SIZE_BLOCK])\
                            .repeat(B,1,1,1)
            for i in range(len(boundary)):
                mask[i,0,:boundary[i],::boundary[i]] = 1
            att = att.masked_fill(mask[:,:,:T,:T] == 0, float('-inf'))
        att = F.softmax(att, dim=-1)
        att = self.dropout_attention(att)
        y = att @ v
        y = y.transpose(1, 2).contiguous().view(B, T, self.attention_features)
        y = self.dropout_residue(self.projection(y))
        return y

class Block(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.ln1 = nn.LayerNorm(config[varables.DIM_EMBEDDING])
        self.ln2 = nn.LayerNorm(config[varables.DIM_EMBEDDING])
        self.attn = CausalSelfAttention(config)
        self.mlp = nn.Sequential(
            nn.Linear(config[varables.DIM_EMBEDDING], config[varables.DIM_FEEDFORWARD]),
            nn.GELU(),
            nn.Linear(config[varables.DIM_FEEDFORWARD], config[varables.DIM_EMBEDDING]),
            nn.Dropout(config[varables.RATE_DROPOUT]),
        )

    def forward(self, x, boundary):
         # = y_input
        x = x + self.attn(self.ln1(x),boundary)
        x = x + self.mlp(self.ln2(x))
        return x

class Model(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.tok_emb = nn.Embedding(config[varables.SIZE_VOCAB], config[varables.DIM_EMBEDDING])
        self.pos_emb = nn.Parameter(torch.zeros(1, config[varables.SIZE_BLOCK], config[varables.DIM_EMBEDDING]))
        self.drop = nn.Dropout(config[varables.RATE_DROPOUT])
        self.blocks = nn.ModuleList([Block(config) for _ in range(config[varables.NUM_LAYERS])])
        # self.blocks = nn.Sequential(*[Block(config) for _ in range(config[varables.NUM_LAYERS])])
        self.ln_f = nn.LayerNorm(config[varables.DIM_EMBEDDING])
        self.head = nn.Linear(config[varables.DIM_EMBEDDING], config[varables.SIZE_VOCAB], bias=False)
        self.block_size = config[varables.SIZE_BLOCK]
        self.apply(self._init_weights)
        logger.info("number of parameters: %e", sum(p.numel() for p in self.parameters()))
        self.optimizer = None

    def get_block_size(self):
        return self.block_size

    def _init_weights(self, module):
        if isinstance(module, (nn.Linear, nn.Embedding)):
            module.weight.data.normal_(mean=0.0, std=0.02)
            if isinstance(module, nn.Linear) and module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)
    def init_optimizers(self,train_config):
        optimizer = torch.optim.Adam(self.parameters(), lr=train_config[varables.RATE_LEARNING])
        return optimizer
    def init_scheduler(self,train_config):
        scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer, step_size=train_config[varables.SIZE_STEP], gamma=train_config[varables.GAMMA])
        return scheduler
    def get_collate_fn(self, vocab,vocab2):
        def collate(results):
            x_in = None
            y_in = [a[0] + [vocab[varables.TOKEN_SEP]] + a[1] for a in results]
            boundary = [a[2] for a in results]
            max_len = max([len(a) for a in y_in])
            y = torch.tensor([(a+[vocab[varables.TOKEN_PAD]]*(max_len-len(a))) for a in y_in],dtype=torch.long)
            return x_in,y,boundary
        return collate

    def forward(self, x_in, y_in, y_out=None,boundary=None):
        b, t = y_in.size()
        assert t <= self.block_size
        token_embeddings = self.tok_emb(y_in)
        position_embeddings = self.pos_emb[:, :t, :]
        x = self.drop(token_embeddings + position_embeddings)
        for block in self.blocks:
            x = block(x,boundary)
        x = self.ln_f(x)
        logits = self.head(x)
        loss = None
        if y_out is not None:
            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), y_out.view(-1))
        return logits, loss