forked from facebookresearch/vicreg
-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
174 lines (146 loc) · 6.31 KB
/
models.py
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
import transformers
from torch import nn, optim
import torch.nn.functional as F
import torch
from utils import off_diagonal, FullGatherLayer
transformers.logging.set_verbosity_error()
class VICReg(nn.Module):
def __init__(self, args):
super().__init__()
self.args = args
self.num_features = int(self.args.mlp.split("-")[-1])
self.backbone = transformers.AutoModel.from_pretrained(args.arch)
self.backbone.train()
self.embedding = self.backbone.config.hidden_size
self.projector = Projector(args, self.embedding)
if self.args.use_param_weights:
self.param = nn.ParameterDict({'sim_coeff':nn.Parameter(torch.rand(1) + self.args.sim_coeff),
'std_coeff':nn.Parameter(torch.rand(1) + self.args.std_coeff),
'cov_coeff':nn.Parameter(torch.rand(1) + self.args.cov_coeff),
})
def forward(self, batch, only_inference = False):
if only_inference:
return self.projector(self.backbone(**batch).pooler_output)
else:
x = self.projector(self.backbone(**batch).pooler_output)
y = self.projector(self.backbone(**batch).pooler_output)
repr_loss = F.mse_loss(x, y)
x = torch.cat(FullGatherLayer.apply(x), dim=0)
y = torch.cat(FullGatherLayer.apply(y), dim=0)
x = x - x.mean(dim=0)
y = y - y.mean(dim=0)
std_x = torch.sqrt(x.var(dim=0) + 0.0001)
std_y = torch.sqrt(y.var(dim=0) + 0.0001)
std_loss = torch.mean(F.relu(1 - std_x)) / 2 + torch.mean(F.relu(1 - std_y)) / 2
cov_x = (x.T @ x) / (self.args.batch_size - 1)
cov_y = (y.T @ y) / (self.args.batch_size - 1)
cov_loss = off_diagonal(cov_x).pow_(2).sum().div(self.num_features) + off_diagonal(cov_y).pow_(2).sum().div(self.num_features)
if self.args.use_param_weights:
loss = (
self.param.sim_coeff * repr_loss
+ self.param.std_coeff * std_loss
+ self.param.cov_coeff * cov_loss
)
else:
loss = (
self.args.sim_coeff * repr_loss
+ self.args.std_coeff * std_loss
+ self.args.cov_coeff * cov_loss
)
return {'loss':loss,
'repr_loss':repr_loss,
'std_loss':std_loss,
'cov_loss':cov_loss}
class BarlowModule(nn.Module):
def __init__(self, args):
super().__init__()
self.bert = transformers.BertModel.from_pretrained('bert-base-uncased',
hidden_dropout_prob=args.dropout_prob,
attention_probs_dropout_prob=args.dropout_prob)
self.bert.train()
sizes = [self.bert.config.hidden_size] + list(map(int, args.projector.split('-')))
layers = []
for i in range(len(sizes) - 2):
layers.append(MLP(input_dim=sizes[i],hidden_size=sizes[i+1],output_dim=None))
layers.append(nn.Linear(sizes[-2], sizes[-1], bias=False))
self.projector = nn.Sequential(*layers)
# normalization layer for the representations z1 and z2
self.bn = nn.BatchNorm1d(sizes[-1], affine=False)
def forward(self,x):
bert_output = self.bert(**x)
embedding = bert_output.pooler_output
projection = self.projector(embedding)
return self.bn(projection)
class MLP(torch.nn.Module):
def __init__(self, input_dim=2048, hidden_size=4096, output_dim=256):
super().__init__()
self.output_dim = output_dim
self.input_dim = input_dim
self.model = nn.Sequential(
nn.Linear(input_dim, hidden_size, bias=False),
nn.BatchNorm1d(hidden_size),
nn.ReLU(inplace=True),
nn.Linear(hidden_size, output_dim, bias=True) if output_dim is not None else nn.Identity(),
)
def forward(self, x):
x = self.model(x)
return x
def Projector(args, embedding):
mlp_spec = f"{embedding}-{args.mlp}"
layers = []
f = list(map(int, mlp_spec.split("-")))
for i in range(len(f) - 2):
layers.append(nn.Linear(f[i], f[i + 1]))
layers.append(nn.BatchNorm1d(f[i + 1]))
layers.append(nn.ReLU(True))
layers.append(nn.Linear(f[-2], f[-1], bias=False))
return nn.Sequential(*layers)
class LARS(optim.Optimizer):
def __init__(
self,
params,
lr,
weight_decay=0,
momentum=0.9,
eta=0.001,
weight_decay_filter=None,
lars_adaptation_filter=None,
):
defaults = dict(
lr=lr,
weight_decay=weight_decay,
momentum=momentum,
eta=eta,
weight_decay_filter=weight_decay_filter,
lars_adaptation_filter=lars_adaptation_filter,
)
super().__init__(params, defaults)
@torch.no_grad()
def step(self):
for g in self.param_groups:
for p in g["params"]:
dp = p.grad
if dp is None:
continue
if g["weight_decay_filter"] is None or not g["weight_decay_filter"](p):
dp = dp.add(p, alpha=g["weight_decay"])
if g["lars_adaptation_filter"] is None or not g[
"lars_adaptation_filter"
](p):
param_norm = torch.norm(p)
update_norm = torch.norm(dp)
one = torch.ones_like(param_norm)
q = torch.where(
param_norm > 0.0,
torch.where(
update_norm > 0, (g["eta"] * param_norm / update_norm), one
),
one,
)
dp = dp.mul(q)
param_state = self.state[p]
if "mu" not in param_state:
param_state["mu"] = torch.zeros_like(p)
mu = param_state["mu"]
mu.mul_(g["momentum"]).add_(dp)
p.add_(mu, alpha=-g["lr"])