|
a |
|
b/algorithms/simclr.py |
|
|
1 |
import math |
|
|
2 |
import torch |
|
|
3 |
from torch import nn, optim |
|
|
4 |
|
|
|
5 |
from algorithms.arch.resnet import loadResnetBackbone |
|
|
6 |
import utilities.runUtils as rutl |
|
|
7 |
|
|
|
8 |
def device_as(t1, t2): |
|
|
9 |
""" |
|
|
10 |
Moves t1 to the device of t2 |
|
|
11 |
""" |
|
|
12 |
return t1.to(t2.device) |
|
|
13 |
|
|
|
14 |
##==================== Model =============================================== |
|
|
15 |
|
|
|
16 |
class ContrastiveLoss(nn.Module): |
|
|
17 |
""" |
|
|
18 |
Vanilla Contrastive loss, also called InfoNceLoss as in SimCLR paper |
|
|
19 |
""" |
|
|
20 |
def __init__(self, batch_size, temperature=0.5): |
|
|
21 |
super().__init__() |
|
|
22 |
self.batch_size = batch_size |
|
|
23 |
self.temperature = temperature |
|
|
24 |
self.mask = (~torch.eye(batch_size * 2, batch_size * 2, dtype=bool)).float() |
|
|
25 |
|
|
|
26 |
def calc_similarity_batch(self, a, b): |
|
|
27 |
representations = torch.cat([a, b], dim=0) |
|
|
28 |
return nn.functional.cosine_similarity(representations.unsqueeze(1), representations.unsqueeze(0), dim=2) |
|
|
29 |
|
|
|
30 |
def forward(self, proj_1, proj_2): |
|
|
31 |
""" |
|
|
32 |
proj_1 and proj_2 are batched embeddings [batch, embedding_dim] |
|
|
33 |
where corresponding indices are pairs |
|
|
34 |
z_i, z_j in the SimCLR paper |
|
|
35 |
""" |
|
|
36 |
batch_size = proj_1.shape[0] |
|
|
37 |
z_i = nn.functional.normalize(proj_1, p=2, dim=1) |
|
|
38 |
z_j = nn.functional.normalize(proj_2, p=2, dim=1) |
|
|
39 |
|
|
|
40 |
similarity_matrix = self.calc_similarity_batch(z_i, z_j) |
|
|
41 |
|
|
|
42 |
sim_ij = torch.diag(similarity_matrix, batch_size) |
|
|
43 |
sim_ji = torch.diag(similarity_matrix, -batch_size) |
|
|
44 |
|
|
|
45 |
positives = torch.cat([sim_ij, sim_ji], dim=0) |
|
|
46 |
|
|
|
47 |
nominator = torch.exp(positives / self.temperature) |
|
|
48 |
|
|
|
49 |
denominator = device_as(self.mask, similarity_matrix) * torch.exp(similarity_matrix / self.temperature) |
|
|
50 |
|
|
|
51 |
all_losses = -torch.log(nominator / torch.sum(denominator, dim=1)) |
|
|
52 |
loss = torch.sum(all_losses) / (2 * self.batch_size) |
|
|
53 |
return loss |
|
|
54 |
|
|
|
55 |
class SimCLR(nn.Module): |
|
|
56 |
def __init__(self, featx_arch, projector_sizes, |
|
|
57 |
batch_size, temp, pretrained=None): |
|
|
58 |
super().__init__() |
|
|
59 |
rutl.START_SEED() |
|
|
60 |
|
|
|
61 |
mlp_dim = projector_sizes[0] |
|
|
62 |
embedding_size = projector_sizes[1] |
|
|
63 |
|
|
|
64 |
self.batch_size = batch_size |
|
|
65 |
self.temp = temp |
|
|
66 |
self.backbone, outfeatx_size = loadResnetBackbone(arch=featx_arch, |
|
|
67 |
torch_pretrain=pretrained) |
|
|
68 |
# add mlp projection head |
|
|
69 |
self.projector = nn.Sequential( |
|
|
70 |
nn.Linear(in_features=outfeatx_size, out_features=mlp_dim), |
|
|
71 |
nn.BatchNorm1d(mlp_dim), |
|
|
72 |
nn.ReLU(), |
|
|
73 |
nn.Linear(in_features=mlp_dim, out_features=embedding_size), |
|
|
74 |
# nn.BatchNorm1d(embedding_size), |
|
|
75 |
) |
|
|
76 |
def forward(self, y1, y2): |
|
|
77 |
z1 = self.projector(self.backbone(y1)) |
|
|
78 |
z2 = self.projector(self.backbone(y2)) |
|
|
79 |
loss = ContrastiveLoss(self.batch_size, self.temp) |
|
|
80 |
return loss(z1, z2) |
|
|
81 |
|
|
|
82 |
##==================== OPTIMISER =============================================== |
|
|
83 |
|
|
|
84 |
class LARS(optim.Optimizer): |
|
|
85 |
def __init__( |
|
|
86 |
self, |
|
|
87 |
params, |
|
|
88 |
lr, |
|
|
89 |
momentum=0, |
|
|
90 |
dampening=0, |
|
|
91 |
weight_decay=0, |
|
|
92 |
nesterov=False, |
|
|
93 |
trust_coefficient=0.001, |
|
|
94 |
eps=1e-8, |
|
|
95 |
): |
|
|
96 |
|
|
|
97 |
defaults = dict( |
|
|
98 |
lr=lr, |
|
|
99 |
momentum=momentum, |
|
|
100 |
dampening=dampening, |
|
|
101 |
weight_decay=weight_decay, |
|
|
102 |
nesterov=nesterov, |
|
|
103 |
trust_coefficient=trust_coefficient, |
|
|
104 |
eps=eps, |
|
|
105 |
) |
|
|
106 |
if nesterov and (momentum <= 0 or dampening != 0): |
|
|
107 |
raise ValueError("Nesterov momentum requires a momentum and zero dampening") |
|
|
108 |
|
|
|
109 |
super().__init__(params, defaults) |
|
|
110 |
|
|
|
111 |
def __setstate__(self, state): |
|
|
112 |
super().__setstate__(state) |
|
|
113 |
|
|
|
114 |
for group in self.param_groups: |
|
|
115 |
group.setdefault("nesterov", False) |
|
|
116 |
|
|
|
117 |
@torch.no_grad() |
|
|
118 |
def step(self, closure=None): |
|
|
119 |
"""Performs a single optimization step. |
|
|
120 |
|
|
|
121 |
Args: |
|
|
122 |
closure (callable, optional): A closure that reevaluates the model |
|
|
123 |
and returns the loss. |
|
|
124 |
""" |
|
|
125 |
loss = None |
|
|
126 |
if closure is not None: |
|
|
127 |
with torch.enable_grad(): |
|
|
128 |
loss = closure() |
|
|
129 |
|
|
|
130 |
# exclude scaling for params with 0 weight decay |
|
|
131 |
for group in self.param_groups: |
|
|
132 |
weight_decay = group["weight_decay"] |
|
|
133 |
momentum = group["momentum"] |
|
|
134 |
dampening = group["dampening"] |
|
|
135 |
nesterov = group["nesterov"] |
|
|
136 |
|
|
|
137 |
for p in group["params"]: |
|
|
138 |
if p.grad is None: |
|
|
139 |
continue |
|
|
140 |
|
|
|
141 |
d_p = p.grad |
|
|
142 |
p_norm = torch.norm(p.data) |
|
|
143 |
g_norm = torch.norm(p.grad.data) |
|
|
144 |
|
|
|
145 |
# lars scaling + weight decay part |
|
|
146 |
if weight_decay != 0: |
|
|
147 |
if p_norm != 0 and g_norm != 0: |
|
|
148 |
lars_lr = p_norm / (g_norm + p_norm * weight_decay + group["eps"]) |
|
|
149 |
lars_lr *= group["trust_coefficient"] |
|
|
150 |
|
|
|
151 |
d_p = d_p.add(p, alpha=weight_decay) |
|
|
152 |
d_p *= lars_lr |
|
|
153 |
|
|
|
154 |
# sgd part |
|
|
155 |
if momentum != 0: |
|
|
156 |
param_state = self.state[p] |
|
|
157 |
if "momentum_buffer" not in param_state: |
|
|
158 |
buf = param_state["momentum_buffer"] = torch.clone(d_p).detach() |
|
|
159 |
else: |
|
|
160 |
buf = param_state["momentum_buffer"] |
|
|
161 |
buf.mul_(momentum).add_(d_p, alpha=1 - dampening) |
|
|
162 |
if nesterov: |
|
|
163 |
d_p = d_p.add(buf, alpha=momentum) |
|
|
164 |
else: |
|
|
165 |
d_p = buf |
|
|
166 |
|
|
|
167 |
p.add_(d_p, alpha=-group["lr"]) |
|
|
168 |
|
|
|
169 |
return loss |