whisper-at/src/whisper_at_train/models.py at main · YuanGongND/whisper-at · 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
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
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
# -*- coding: utf-8 -*-
# @Time    : 1/26/23 11:13 PM
# @Author  : Yuan Gong
# @Affiliation  : Massachusetts Institute of Technology
# @Email   : yuangong@mit.edu
# @File    : high_mdls.py

# high models

import numpy as np
import torch
import math
import torch.nn.functional as F
from torch import Tensor
from torch import nn
from whisper.model import ResidualAttentionBlock, Linear

def _no_grad_trunc_normal_(tensor, mean, std, a, b):
    def norm_cdf(x):
        # Computes standard normal cumulative distribution function
        return (1. + math.erf(x / math.sqrt(2.))) / 2.

    with torch.no_grad():
        # Values are generated by using a truncated uniform distribution and
        # then using the inverse CDF for the normal distribution.
        # Get upper and lower cdf values
        l = norm_cdf((a - mean) / std)
        u = norm_cdf((b - mean) / std)

        # Uniformly fill tensor with values from [l, u], then translate to
        # [2l-1, 2u-1].
        tensor.uniform_(2 * l - 1, 2 * u - 1)

        # Use inverse cdf transform for normal distribution to get truncated
        # standard normal
        tensor.erfinv_()

        # Transform to proper mean, std
        tensor.mul_(std * math.sqrt(2.))
        tensor.add_(mean)

        # Clamp to ensure it's in the proper range
        tensor.clamp_(min=a, max=b)
        return tensor

def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
    return _no_grad_trunc_normal_(tensor, mean, std, a, b)

class TLTR(nn.Module):
    def __init__(self, label_dim=527, n_layer=33, rep_dim=1280, mode='basic'):
        super().__init__()
        self.mode = mode
        self.n_layer = n_layer
        self.rep_dim = rep_dim
        self.label_dim = label_dim

        # (baseline) mean pool over time and layer, and mlp head
        if mode == 'mean_mlp' or mode == 'last_mlp':
            self.mlp_layer = nn.Sequential(nn.LayerNorm(self.rep_dim), nn.Linear(self.rep_dim, self.label_dim))

        # (baseline) mean pool over time, and weight average over layers, and mlp head
        if mode == 'wa_mlp':
            self.mlp_layer = nn.Sequential(nn.LayerNorm(self.rep_dim), nn.Linear(self.rep_dim, self.label_dim))
            self.layer_weight = torch.nn.Parameter(torch.tensor([1 / self.n_layer] * self.n_layer))

        # (baseline) mean pool over layer, and apply a original rep_dim transformer
        if 'mean_tr' in mode or 'last_tr' in mode:
            self.num_att_head = int(mode.split('_')[-1])
            self.time_tr = ResidualAttentionBlock(self.rep_dim, self.num_att_head)
            self.mlp_layer = nn.Sequential(nn.LayerNorm(self.rep_dim), nn.Linear(self.rep_dim, self.label_dim))

        # (baseline) weight average over layers, and apply a original rep_dim transformer
        if 'wa_tr' in mode:
            self.num_att_head = int(mode.split('_')[-1])
            self.layer_weight = torch.nn.Parameter(torch.tensor([1 / self.n_layer] * self.n_layer))
            self.time_tr = ResidualAttentionBlock(self.rep_dim, self.num_att_head)
            self.mlp_layer = nn.Sequential(nn.LayerNorm(self.rep_dim), nn.Linear(self.rep_dim, self.label_dim))

        # (baseline) weight average over layers, and apply a low-dimensional transformer
        if 'wa_down_tr' in mode: # 512_1
            self.inter_rep_dim = int(mode.split('_')[-2])
            self.num_att_head = int(mode.split('_')[-1])

            self.down_layer = nn.Sequential(nn.LayerNorm(self.rep_dim), nn.Linear(self.rep_dim, self.inter_rep_dim))
            self.layer_weight = torch.nn.Parameter(torch.tensor([1 / self.n_layer] * self.n_layer))
            self.time_tr = ResidualAttentionBlock(self.inter_rep_dim, self.num_att_head)
            self.mlp_layer = nn.Sequential(nn.LayerNorm(self.inter_rep_dim), nn.Linear(self.inter_rep_dim, self.label_dim))

        # (proposed), tl-tr, weight average over layers, and apply a original rep_dim transformer
        if 'lw_tr' in mode:
            self.num_tatt_head = int(mode.split('_')[-2])
            self.num_latt_head = int(mode.split('_')[-1])
            self.time_tr = ResidualAttentionBlock(self.rep_dim, self.num_tatt_head)
            self.layer_tr = ResidualAttentionBlock(self.rep_dim, self.num_latt_head)
            self.mlp_layer = nn.Sequential(nn.LayerNorm(self.rep_dim), nn.Linear(self.rep_dim, self.label_dim))

        # (proposed), tl-tr with low-dimension projection, lower the dimension of the transformer # lw_down_tr_512_1_8
        if 'lw_down_tr' in mode:
            self.inter_rep_dim = int(mode.split('_')[-3])
            self.num_tatt_head = int(mode.split('_')[-2])
            self.num_latt_head = int(mode.split('_')[-1])

            self.down_layer = nn.Sequential(nn.LayerNorm(self.rep_dim), nn.Linear(self.rep_dim, self.inter_rep_dim))
            self.time_tr = ResidualAttentionBlock(self.inter_rep_dim, self.num_tatt_head)
            self.layer_tr = ResidualAttentionBlock(self.inter_rep_dim, self.num_latt_head)
            self.mlp_layer = nn.Sequential(nn.LayerNorm(self.inter_rep_dim), nn.Linear(self.inter_rep_dim, self.label_dim))

    def forward(self, audio_rep):
        # audio_rep in shape (# batch size, #whisper_enc_layer, time length after (20x) pooling, whisper_enc_dim)
        # e.g., (B, 32, 25, 1280) for whisper large-v1

        # (baseline)
        if self.mode == 'mean_mlp':
            audio_rep = torch.mean(audio_rep, dim=1)
            audio_rep = torch.mean(audio_rep, dim=1)
            audio_rep = self.mlp_layer(audio_rep)
            return audio_rep

        # (baseline)
        elif self.mode == 'last_mlp':
            audio_rep = audio_rep[:, -1, :, :] # get the last layer
            audio_rep = torch.mean(audio_rep, dim=1)
            audio_rep = self.mlp_layer(audio_rep)
            return audio_rep

        # (baseline)
        elif self.mode == 'wa_mlp':
            audio_rep = torch.mean(audio_rep, dim=2) # [B, 32 1280]
            audio_rep = torch.permute(audio_rep, (0, 2, 1)) # (B, 1280, 32)
            audio_rep = (audio_rep @ self.layer_weight) / self.layer_weight.sum()
            audio_rep = self.mlp_layer(audio_rep)
            return audio_rep

        # (baseline)
        elif 'mean_tr' in self.mode:
            audio_rep = torch.mean(audio_rep, dim=1) # [B, 25, 1280]
            audio_rep = self.time_tr(audio_rep) # [B, 25, 1280]
            audio_rep = torch.mean(audio_rep, dim=1)  # [B*32, 1280]
            audio_rep = self.mlp_layer(audio_rep)
            return audio_rep

        # (baseline) time transformer on the last layer representation
        elif 'last_tr' in self.mode:
            audio_rep = audio_rep[:, -1, :, :]  # [B, 25, 1280]
            audio_rep = self.time_tr(audio_rep) # [B, 25, 1280]
            audio_rep = torch.mean(audio_rep, dim=1)  # [B*32, 1280]
            audio_rep = self.mlp_layer(audio_rep)
            return audio_rep

        # (baseline) time transformer on the layer-wise weight-average representation
        elif 'wa_tr' in self.mode:
            audio_rep = torch.permute(audio_rep, (0, 2, 3, 1)) # (B, 25, 1280, 32)
            audio_rep = (audio_rep @ self.layer_weight) / self.layer_weight.sum() # [B, 25, 1280]
            audio_rep = self.time_tr(audio_rep) # [B, 25, 1280]
            audio_rep = torch.mean(audio_rep, dim=1)  # [B*25, 1280]
            audio_rep = self.mlp_layer(audio_rep)
            return audio_rep

        # (baseline) weight average with low-dimension projection
        elif 'wa_down_tr' in self.mode:
            audio_rep = torch.permute(audio_rep, (0, 2, 3, 1)) # (B, 25, 1280, 32)
            audio_rep = (audio_rep @ self.layer_weight) / self.layer_weight.sum() # [B, 25, 1280]
            audio_rep = self.down_layer(audio_rep)
            audio_rep = self.time_tr(audio_rep) # [B, 25, 1280]
            audio_rep = torch.mean(audio_rep, dim=1)  # [B*32, 1280]
            audio_rep = self.mlp_layer(audio_rep)
            return audio_rep

        # (proposed) tl-tr
        elif 'lw_tr' in self.mode:
            B = audio_rep.shape[0]
            audio_rep = audio_rep.reshape(B*self.n_layer, audio_rep.shape[2], audio_rep.shape[3]) # [B*32, 25, 1280]
            audio_rep = self.time_tr(audio_rep) # [B*32, 25, 1280]
            audio_rep = torch.mean(audio_rep, dim=1) # [B*32, 1280]
            audio_rep = audio_rep.reshape(B, self.n_layer, audio_rep.shape[1]) # [B, 32, 1280]
            audio_rep = self.layer_tr(audio_rep) # [B, 32, 1280]
            audio_rep = torch.mean(audio_rep, dim=1)  # [B, 1280]
            audio_rep = self.mlp_layer(audio_rep)
            return audio_rep

        #(proposed)  tl-tr with low-dimensional projection
        elif 'lw_down_tr' in self.mode:
            B = audio_rep.shape[0]
            audio_rep = self.down_layer(audio_rep)
            audio_rep = audio_rep.reshape(B*self.n_layer, audio_rep.shape[2], audio_rep.shape[3]) # [B*32, 25, 1280]
            audio_rep = self.time_tr(audio_rep) # [B*32, 25, 1280]
            audio_rep = torch.mean(audio_rep, dim=1) # [B*32, 1280]
            audio_rep = audio_rep.reshape(B, self.n_layer, audio_rep.shape[1]) # [B, 32, 1280]
            audio_rep = self.layer_tr(audio_rep) # [B, 32, 1280]
            audio_rep = torch.mean(audio_rep, dim=1)  # [B, 1280]
            audio_rep = self.mlp_layer(audio_rep)
            return audio_rep