-
Notifications
You must be signed in to change notification settings - Fork 0
/
metalearner_ppo2.py
263 lines (202 loc) · 10.8 KB
/
metalearner_ppo2.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
import torch
import torch.nn as nn
from torch.distributions.kl import kl_divergence
from torch.nn.utils.convert_parameters import (vector_to_parameters,
parameters_to_vector)
from rl_utils.optimization import conjugate_gradient
from rl_utils.torch_utils import (weighted_mean, detach_distribution, weighted_normalize)
class MetaLearnerPPO(object):
"""Meta-learner
The meta-learner is responsible for sampling the trajectories/episodes
(before and after the one-step adaptation), compute the inner loss, compute
the updated parameters based on the inner-loss, and perform the meta-update.
[1] Chelsea Finn, Pieter Abbeel, Sergey Levine, "Model-Agnostic
Meta-Learning for Fast Adaptation of Deep Networks", 2017
(https://arxiv.org/abs/1703.03400)
[2] Richard Sutton, Andrew Barto, "Reinforcement learning: An introduction",
2018 (http://incompleteideas.net/book/the-book-2nd.html)
[3] John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan,
Pieter Abbeel, "High-Dimensional Continuous Control Using Generalized
Advantage Estimation", 2016 (https://arxiv.org/abs/1506.02438)
[4] John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan,
Pieter Abbeel, "Trust Region Policy Optimization", 2015
(https://arxiv.org/abs/1502.05477)
"""
def __init__(self, sampler, policy, baseline, gamma=0.95,
fast_lr=0.5, tau=1.0, device='cpu'):
self.sampler = sampler
self.policy = policy
self.baseline = baseline
self.gamma = gamma
self.fast_lr = fast_lr
self.tau = tau
self.to(device)
def inner_loss(self, episodes, params=None):
"""Compute the inner loss for the one-step gradient update. The inner
loss is REINFORCE with baseline [2], computed on advantages estimated
with Generalized Advantage Estimation (GAE, [3]).
"""
values = self.baseline(episodes)
advantages = episodes.gae(values, tau=self.tau)
advantages = weighted_normalize(advantages, weights=episodes.mask)
pi = self.policy(episodes.observations, params=params)
log_probs = pi.log_prob(episodes.actions)
if log_probs.dim() > 2:
log_probs = torch.sum(log_probs, dim=2)
loss = -weighted_mean(log_probs * advantages, dim=0, weights=episodes.mask) # negative as update rule has gradient ascent while optimizer uses gradient descent
return loss
def adapt(self, episodes, first_order=False, params=None, lr=None):
"""Adapt the parameters of the policy network to a new task, from
sampled trajectories `episodes`, with a one-step gradient update [1].
"""
if lr is None:
lr = self.fast_lr
# Fit the baseline to the training episodes
self.baseline.fit(episodes)
# Get the loss on the training episodes
loss = self.inner_loss(episodes, params=params)
# Get the new parameters after a one-step gradient update
params = self.policy.update_params(loss, step_size=lr, first_order=first_order, params=params)
return params, loss
def sample(self, tasks, first_order=False):
"""Sample trajectories (before and after the update of the parameters)
for all the tasks `tasks`.
"""
episodes = []
losses = []
for task in tasks:
self.sampler.reset_task(task)
self.policy.reset_context()
train_episodes = self.sampler.sample(self.policy, gamma=self.gamma)
# inner loop (for CAVIA, this only updates the context parameters)
params, loss = self.adapt(train_episodes, first_order=first_order)
# rollouts after inner loop update
valid_episodes = self.sampler.sample(self.policy, params=params, gamma=self.gamma)
episodes.append((train_episodes, valid_episodes))
losses.append(loss.item())
return episodes, losses
def test(self, tasks, num_steps, batch_size, halve_lr):
"""Sample trajectories (before and after the update of the parameters)
for all the tasks `tasks`.batchsize
"""
episodes_per_task = []
for task in tasks:
# reset context params (for cavia) and task
self.policy.reset_context()
self.sampler.reset_task(task)
# start with blank params
params = None
# gather some initial experience and log performance
test_episodes = self.sampler.sample(self.policy, gamma=self.gamma, params=params, batch_size=batch_size)
# initialise list which will log all rollouts for the current task
curr_episodes = [test_episodes]
for i in range(1, num_steps + 1):
# lower learning rate after first update (for MAML, as described in their paper)
if i == 1 and halve_lr:
lr = self.fast_lr / 2
else:
lr = self.fast_lr
# inner-loop update
params, loss = self.adapt(test_episodes, first_order=True, params=params, lr=lr)
# get new rollouts
test_episodes = self.sampler.sample(self.policy, gamma=self.gamma, params=params, batch_size=batch_size)
curr_episodes.append(test_episodes)
episodes_per_task.append(curr_episodes)
self.policy.reset_context()
return episodes_per_task
def surrogate_loss(self, episodes, eps_clip=0.1,critic_weight=0.5,entropy_wt=0.01,old_pis=None):
losses, kls, pis = [], [], []
if old_pis is None:
old_pis = [None] * len(episodes)
critic_criterion = nn.MSELoss()
for (train_episodes, valid_episodes), old_pi in zip(episodes, old_pis):
# do inner-loop update
self.policy.reset_context()
params, _ = self.adapt(train_episodes)
with torch.set_grad_enabled(old_pi is None):
# get action values after inner-loop update
pi = self.policy(valid_episodes.observations, params=params)
pis.append(detach_distribution(pi))
if old_pi is None:
old_pi = detach_distribution(pi)
returns = valid_episodes.returns
returns = (returns - returns.mean(dim=0)) / (returns.std(dim=0) + 1e-5)
values = self.baseline(valid_episodes)
print('Returns and Values shape',returns.shape,values.squeeze(2).shape)
advantages = valid_episodes.gae(values, tau=self.tau)
advantages = weighted_normalize(advantages, weights=valid_episodes.mask)
log_ratio = (pi.log_prob(valid_episodes.actions)
- old_pi.log_prob(valid_episodes.actions))
if log_ratio.dim() > 2:
log_ratio = torch.sum(log_ratio, dim=2)
ratio = torch.exp(log_ratio)
# loss1 = weighted_mean(ratio * advantages, dim=0, weights=valid_episodes.mask)
# loss2 = weighted_mean(torch.clamp(ratio, 1-eps_clip, eps_clip) * advantages,dim=0, weights=valid_episodes.mask)
loss1 = ratio * advantages
loss2 = torch.clamp(ratio, 1-eps_clip, eps_clip) * advantages
print('Loss1',loss1)
print('Loss2',loss2)
actor_loss = -torch.min(loss1,loss2).mean()
print('Actor Loss',actor_loss)
critic_loss = critic_criterion(values.squeeze(2),returns)
print('Critic Loss',critic_loss)
entropy = torch.mean(- torch.exp(pi.log_prob(valid_episodes.actions))*pi.log_prob(valid_episodes.actions))
print('Entropy',entropy)
loss = actor_loss + critic_weight*critic_loss - entropy_wt*entropy
print('Total Loss',loss)
losses.append(loss)
mask = valid_episodes.mask
if valid_episodes.actions.dim() > 2:
mask = mask.unsqueeze(2)
# kl = weighted_mean(kl_divergence(pi, old_pi), dim=0, weights=mask)
# kls.append(kl)
return torch.mean(torch.stack(losses, dim=0)), pis
def step(self,episodes,meta_lr,eps_clip=0.1,critic_weight=0.5,entropy_wt=0.01):
"""Meta Optimization Step for PPO"""
# meta_test_steps=3
# for i in range(meta_test_steps):
old_loss, old_pis = self.surrogate_loss(episodes,eps_clip=eps_clip,critic_weight=critic_weight,entropy_wt=entropy_wt)
# this part will take higher order gradients through the inner loop:
grads = torch.autograd.grad(old_loss, self.policy.parameters())
grads = parameters_to_vector(grads)
old_params = parameters_to_vector(self.policy.parameters())
vector_to_parameters(old_params - meta_lr * grads, self.policy.parameters())
# meta_optim.zero_grad()
# old_loss.backward()
# meta_optim.step()
return old_loss
# def step(self, episodes, max_kl=1e-3, cg_iters=10, cg_damping=1e-2,
# ls_max_steps=10, ls_backtrack_ratio=0.5):
# """Meta-optimization step (ie. update of the initial parameters), based
# on Trust Region Policy Optimization (TRPO, [4]).
# """
# old_loss, _, old_pis = self.surrogate_loss(episodes)
# # this part will take higher order gradients through the inner loop:
# grads = torch.autograd.grad(old_loss, self.policy.parameters())
# grads = parameters_to_vector(grads)
# # Compute the step direction with Conjugate Gradient
# hessian_vector_product = self.hessian_vector_product(episodes, damping=cg_damping)
# stepdir = conjugate_gradient(hessian_vector_product, grads, cg_iters=cg_iters)
# # Compute the Lagrange multiplier
# shs = 0.5 * torch.dot(stepdir, hessian_vector_product(stepdir))
# lagrange_multiplier = torch.sqrt(shs / max_kl)
# step = stepdir / lagrange_multiplier
# # Save the old parameters
# old_params = parameters_to_vector(self.policy.parameters())
# # Line search
# step_size = 1.0
# for _ in range(ls_max_steps):
# vector_to_parameters(old_params - step_size * step, self.policy.parameters())
# loss, kl, _ = self.surrogate_loss(episodes, old_pis=old_pis)
# improve = loss - old_loss
# if (improve.item() < 0.0) and (kl.item() < max_kl):
# break
# step_size *= ls_backtrack_ratio
# else:
# print('no update?')
# vector_to_parameters(old_params, self.policy.parameters())
# return loss
def to(self, device, **kwargs):
self.policy.to(device, **kwargs)
self.baseline.to(device, **kwargs)
self.device = device