-
Notifications
You must be signed in to change notification settings - Fork 17
/
cagrad.py
303 lines (265 loc) · 10.8 KB
/
cagrad.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
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
from copy import deepcopy
from typing import Iterable, List, Optional, Tuple
import numpy as np
import time
import torch
from omegaconf import OmegaConf
from mtrl.agent import grad_manipulation as grad_manipulation_agent
from mtrl.utils.types import ConfigType, TensorType
#from mtrl.agent.mgda import MinNormSolver
from scipy.optimize import minimize, Bounds, minimize_scalar
def _check_param_device(param: TensorType, old_param_device: Optional[int]) -> int:
"""This helper function is to check if the parameters are located
in the same device. Currently, the conversion between model parameters
and single vector form is not supported for multiple allocations,
e.g. parameters in different GPUs, or mixture of CPU/GPU.
The implementation is taken from: https://github.com/pytorch/pytorch/blob/22a34bcf4e5eaa348f0117c414c3dd760ec64b13/torch/nn/utils/convert_parameters.py#L57
Args:
param ([TensorType]): a Tensor of a parameter of a model.
old_param_device ([int]): the device where the first parameter
of a model is allocated.
Returns:
old_param_device (int): report device for the first time
"""
# Meet the first parameter
if old_param_device is None:
old_param_device = param.get_device() if param.is_cuda else -1
else:
warn = False
if param.is_cuda: # Check if in same GPU
warn = param.get_device() != old_param_device
else: # Check if in CPU
warn = old_param_device != -1
if warn:
raise TypeError(
"Found two parameters on different devices, "
"this is currently not supported."
)
return old_param_device
def apply_vector_grad_to_parameters(
vec: TensorType, parameters: Iterable[TensorType], accumulate: bool = False
):
"""Apply vector gradients to the parameters
Args:
vec (TensorType): a single vector represents the gradients of a model.
parameters (Iterable[TensorType]): an iterator of Tensors that are the
parameters of a model.
"""
# Ensure vec of type Tensor
if not isinstance(vec, torch.Tensor):
raise TypeError(
"expected torch.Tensor, but got: {}".format(torch.typename(vec))
)
# Flag for the device where the parameter is located
param_device = None
# Pointer for slicing the vector for each parameter
pointer = 0
for param in parameters:
# Ensure the parameters are located in the same device
param_device = _check_param_device(param, param_device)
# The length of the parameter
num_param = param.numel()
# Slice the vector, reshape it, and replace the old grad of the parameter
if accumulate:
param.grad = (
param.grad + vec[pointer : pointer + num_param].view_as(param).data
)
else:
param.grad = vec[pointer : pointer + num_param].view_as(param).data
# Increment the pointer
pointer += num_param
class Agent(grad_manipulation_agent.Agent):
def __init__(
self,
env_obs_shape: List[int],
action_shape: List[int],
action_range: Tuple[int, int],
device: torch.device,
agent_cfg: ConfigType,
multitask_cfg: ConfigType,
cfg_to_load_model: Optional[ConfigType] = None,
should_complete_init: bool = True,
):
"""Regularized gradient algorithm."""
agent_cfg_copy = deepcopy(agent_cfg)
del agent_cfg_copy['cagrad_c']
del agent_cfg_copy['cagrad_method']
OmegaConf.set_struct(agent_cfg_copy, False)
agent_cfg_copy.cfg_to_load_model = None
agent_cfg_copy.should_complete_init = False
agent_cfg_copy.loss_reduction = "none"
OmegaConf.set_struct(agent_cfg_copy, True)
super().__init__(
env_obs_shape=env_obs_shape,
action_shape=action_shape,
action_range=action_range,
multitask_cfg=multitask_cfg,
agent_cfg=agent_cfg_copy,
device=device,
)
self.agent._compute_gradient = self._compute_gradient
self._rng = np.random.default_rng()
self.cagrad_c = agent_cfg['cagrad_c']
self.cagrad_method = agent_cfg['cagrad_method']
fn_maps = {
"cagrad": self.cagrad,
"cagrad_exact": self.cagrad_exact,
}
for k in range(2, 50):
fn_maps[f"cagrad_fast{k}"] = self.cagrad_fast
fn_names = ", ".join(fn_maps.keys())
assert self.cagrad_method in fn_maps, \
f"[error] unrealized fn {self.cagrad_method}, currently we have {fn_names}"
self.cagrad_fn = fn_maps[self.cagrad_method]
self.wi_map = {}
self.num_param_block = -1
self.conflicts = []
self.last_w = None
self.save_target = 500000
if "fast" in self.cagrad_method:
num_tasks = multitask_cfg['num_envs']
self.fast_n = int(self.cagrad_method[self.cagrad_method.find("fast")+4:])
self.fast_w = torch.zeros((self.fast_n)).cuda()
self.fast_w[:-1] = 1/num_tasks
self.fast_w[-1] = 1 - (self.fast_n-1)/num_tasks
self.fast_w_numpy = self.fast_w.cpu().numpy()
if should_complete_init:
self.complete_init(cfg_to_load_model=cfg_to_load_model)
def _compute_gradient(
self,
loss: TensorType, # batch x 1
parameters: List[TensorType],
step: int,
component_names: List[str],
env_metadata: grad_manipulation_agent.EnvMetadata,
retain_graph: bool = False,
allow_unused: bool = False,
) -> None:
#t0 = time.time()
task_loss = self._convert_loss_into_task_loss(
loss=loss, env_metadata=env_metadata
)
num_tasks = task_loss.shape[0]
grad = []
if "fast" in self.cagrad_method:
# 2 losses approximation
n = self.fast_n
idx = np.random.permutation(num_tasks)
losses = [0] * n
for j in range(n-1):
losses[j] = task_loss[idx[j]]
for j in range(n, num_tasks):
losses[-1] += task_loss[idx[j]]
losses[-1] /= (num_tasks - n + 1)
for loss in losses:
grad.append(
tuple(
_grad.contiguous()
for _grad in torch.autograd.grad(
loss,
parameters,
retain_graph=True,
allow_unused=allow_unused,
)
)
)
else:
for index in range(num_tasks):
grad.append(
tuple(
_grad.contiguous()
for _grad in torch.autograd.grad(
task_loss[index],
parameters,
retain_graph=(retain_graph or index != num_tasks - 1),
allow_unused=allow_unused,
)
)
)
grad_vec = torch.cat(
list(
map(lambda x: torch.nn.utils.parameters_to_vector(x).unsqueeze(0), grad)
),
dim=0,
) # num_tasks x dim
regularized_grad = self.cagrad_fn(grad_vec, num_tasks)
apply_vector_grad_to_parameters(regularized_grad, parameters)
def cagrad(self, grad_vec, num_tasks):
"""
grad_vec: [num_tasks, dim]
"""
grads = grad_vec
GG = grads.mm(grads.t()).cpu()
scale = (torch.diag(GG)+1e-4).sqrt().mean()
GG = GG / scale.pow(2)
Gg = GG.mean(1, keepdims=True)
gg = Gg.mean(0, keepdims=True)
w = torch.zeros(num_tasks, 1, requires_grad=True)
if num_tasks == 50:
w_opt = torch.optim.SGD([w], lr=50, momentum=0.5)
else:
w_opt = torch.optim.SGD([w], lr=25, momentum=0.5)
c = (gg+1e-4).sqrt() * self.cagrad_c
w_best = None
obj_best = np.inf
for i in range(21):
w_opt.zero_grad()
ww = torch.softmax(w, 0)
obj = ww.t().mm(Gg) + c * (ww.t().mm(GG).mm(ww) + 1e-4).sqrt()
if obj.item() < obj_best:
obj_best = obj.item()
w_best = w.clone()
if i < 20:
obj.backward()
w_opt.step()
ww = torch.softmax(w_best, 0)
gw_norm = (ww.t().mm(GG).mm(ww)+1e-4).sqrt()
lmbda = c.view(-1) / (gw_norm+1e-4)
g = ((1/num_tasks + ww * lmbda).view(
-1, 1).to(grads.device) * grads).sum(0) / (1 + self.cagrad_c**2)
return g
def cagrad_exact(self, grad_vec, num_tasks):
grads = grad_vec / 100.
g0 = grads.mean(0)
GG = grads.mm(grads.t())
x_start = np.ones(num_tasks)/num_tasks
bnds = tuple((0,1) for x in x_start)
cons=({'type':'eq','fun':lambda x:1-sum(x)})
A = GG.cpu().numpy()
b = x_start.copy()
c = (self.cagrad_c*g0.norm()).cpu().item()
def objfn(x):
return (x.reshape(1,num_tasks).dot(A).dot(b.reshape(num_tasks, 1)) + \
c * np.sqrt(x.reshape(1,num_tasks).dot(A).dot(x.reshape(num_tasks,1))+1e-8)).sum()
res = minimize(objfn, x_start, bounds=bnds, constraints=cons)
w_cpu = res.x
ww= torch.Tensor(w_cpu).to(grad_vec.device)
gw = (grads * ww.view(-1, 1)).sum(0)
gw_norm = gw.norm()
lmbda = c / (gw_norm+1e-4)
g = (g0 + lmbda * gw) / (1 + lmbda)
return g * 100
def cagrad_fast(self, grad_vec, num_tasks):
n = self.fast_n
scale = 100.
grads = grad_vec / scale
GG = grads.mm(grads.t())
g0_norm = (self.fast_w.view(1, -1).mm(GG).mm(self.fast_w.view(-1, 1))+1e-8).sqrt().item()
x_start = np.ones(n) / n
bnds = tuple((0,1) for x in x_start)
cons=({'type':'eq','fun':lambda x:1-sum(x)})
A = GG.cpu().numpy()
c = self.cagrad_c*g0_norm
def objfn(x):
return (x.reshape(1,n).dot(A).dot(self.fast_w_numpy.reshape(n,1)) + \
c * np.sqrt(x.reshape(1,n).dot(A).dot(x.reshape(n,1))+1e-8)).sum()
res = minimize(objfn, x_start, bounds=bnds, constraints=cons)
w_cpu = res.x
ww= torch.Tensor(w_cpu).to(grad_vec.device)
gw = (grads * ww.view(-1, 1)).sum(0)
gw_norm = np.sqrt(w_cpu.reshape(1,n).dot(A).dot(w_cpu.reshape(n,1))+1e-8).item()
lmbda = c / (gw_norm+1e-4)
g = ((self.fast_w.view(-1,1)+ww.view(-1,1)*lmbda)*grads).sum(0)
g = g / (1 + self.cagrad_c) * scale
return g