-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathauto_lambda_callback.py
165 lines (133 loc) · 6.63 KB
/
auto_lambda_callback.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
# adapted from https://github.com/lorenmt/auto-lambda
import copy
import logging
import torch
from torch.optim import Adam
from src.trainer.base_trainer import BaseTrainer
from src.utils.callbacks.callback import Callback
class AutoLambdaCallback(Callback):
def __init__(self, meta_lr, weight_init=0.1):
logging.info("Initializing AutoLambdaCallback.")
self.weight_init = weight_init
self.meta_lr = meta_lr
def connect(self, trainer: BaseTrainer, *args, **kwargs):
self.num_tasks = trainer.benchmark.num_tasks
self.device = trainer.device
self.model = trainer.model
self.model_ = copy.deepcopy(trainer.model)
self.meta_weights = torch.tensor([self.weight_init] * self.num_tasks, requires_grad=True, device=self.device)
self.meta_optimizer = Adam([self.meta_weights], lr=self.meta_lr)
def get_lr(self, trainer: BaseTrainer):
if trainer.scheduler is None:
return trainer.optimizer.param_groups[0]["lr"]
else:
return trainer.scheduler.get_last_lr()[0]
def virtual_step(self, trainer, train_x, train_y, alpha, model_optim):
"""
Compute unrolled network theta' (virtual step)
"""
# forward & compute loss
if type(train_x) == list: # multi-domain setting [many-to-many]
train_pred = [self.model(x, t) for t, x in enumerate(train_x)]
else: # single-domain setting [one-to-many]
train_pred = self.model(train_x)
train_loss = self.model_fit(trainer, train_pred, train_y)
loss = sum([w * train_loss[i] for i, w in enumerate(self.meta_weights)])
# compute gradient
gradients = torch.autograd.grad(loss, self.model.parameters())
# do virtual step (update gradient): theta' = theta - alpha * sum_i lambda_i * L_i(f_theta(x_i), y_i)
with torch.no_grad():
for weight, weight_, grad in zip(self.model.parameters(), self.model_.parameters(), gradients):
if "momentum" in model_optim.param_groups[0].keys(): # used in SGD with momentum
m = model_optim.state[weight].get("momentum_buffer", 0.0) * model_optim.param_groups[0]["momentum"]
else:
m = 0
weight_.copy_(weight - alpha * (m + grad + model_optim.param_groups[0]["weight_decay"] * weight))
def unrolled_backward(self, trainer: BaseTrainer, train_x, train_y, val_x, val_y, alpha, model_optim):
"""
Compute un-rolled loss and backward its gradients
"""
# do virtual step (calc theta`)
self.virtual_step(trainer, train_x, train_y, alpha, model_optim)
# define weighting for primary tasks (with binary weights)
# TODO: no primary tasks atm.
# pri_weights = [1,]
# for t in self.train_tasks:
# if t in self.pri_tasks:
# pri_weights += [1.0]
# else:
# pri_weights += [0.0]
# compute validation data loss on primary tasks
if type(val_x) == list:
val_pred = [self.model_(x, t) for t, x in enumerate(val_x)]
else:
val_pred = self.model_(val_x)
val_loss = self.model_fit(trainer, val_pred, val_y)
loss = sum(val_loss)
# compute hessian via finite difference approximation
model_weights_ = tuple(self.model_.parameters())
d_model = torch.autograd.grad(loss, model_weights_, allow_unused=True)
hessian = self.compute_hessian(trainer, d_model, train_x, train_y)
# update final gradient = - alpha * hessian
with torch.no_grad():
for mw, h in zip([self.meta_weights], hessian):
mw.grad = -alpha * h
def compute_hessian(self, trainer, d_model, train_x, train_y):
norm = torch.cat([w.view(-1) for w in d_model]).norm()
eps = 0.01 / norm
# \theta+ = \theta + eps * d_model
with torch.no_grad():
for p, d in zip(self.model.parameters(), d_model):
p += eps * d
if type(train_x) == list:
train_pred = [self.model(x, t) for t, x in enumerate(train_x)]
else:
train_pred = self.model(train_x)
train_loss = self.model_fit(trainer, train_pred, train_y)
loss = sum([w * train_loss[i] for i, w in enumerate(self.meta_weights)])
d_weight_p = torch.autograd.grad(loss, self.meta_weights)
# \theta- = \theta - eps * d_model
with torch.no_grad():
for p, d in zip(self.model.parameters(), d_model):
p -= 2 * eps * d
if type(train_x) == list:
train_pred = [self.model(x, t) for t, x in enumerate(train_x)]
else:
train_pred = self.model(train_x)
train_loss = self.model_fit(trainer, train_pred, train_y)
loss = sum([w * train_loss[i] for i, w in enumerate(self.meta_weights)])
d_weight_n = torch.autograd.grad(loss, self.meta_weights)
# recover theta
with torch.no_grad():
for p, d in zip(self.model.parameters(), d_model):
p += eps * d
hessian = [(p - n) / (2.0 * eps) for p, n in zip(d_weight_p, d_weight_n)]
return hessian
def model_fit(self, trainer: BaseTrainer, pred, targets):
return trainer.loss_fn(pred, targets)
def on_before_training_epoch(self, trainer: BaseTrainer, *args, **kwargs):
logging.info(f"Meta-weights: {self.meta_weights}")
logging.info("Setting meta dataloader")
self.meta_dataloader = iter(trainer.benchmark.train_dataloader())
def on_before_forward(self, trainer: BaseTrainer, *args, **kwargs):
if trainer.model.training:
batch = next(self.meta_dataloader)
if len(batch) == 3:
# TODO: hack for cityscapes. the api is not consistent
x_meta, y_meta = batch[0], batch[1:]
else:
x_meta, y_meta = batch[0], batch[1]
x_meta = x_meta.to(self.device)
y_meta = [yy.to(self.device) for yy in y_meta]
# update meta-weights with Auto-Lambda
last_lr = self.get_lr(trainer)
self.meta_optimizer.zero_grad()
x, y = trainer.x, trainer.y
self.unrolled_backward(trainer, x, y, x_meta, y_meta, last_lr, trainer.optimizer)
self.meta_optimizer.step()
# register meta-weights to trainer
trainer.meta_weights = self.meta_weights
# log meta-weights per task at each iteration
meta_weights = self.meta_weights.detach().cpu().numpy()
for t, w in enumerate(meta_weights):
trainer.log(key=f"meta-weights/task-{t}", value=w)