-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdilo.py
194 lines (164 loc) · 9.5 KB
/
dilo.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
import copy
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim.lr_scheduler import CosineAnnealingLR
from dilo_utils import DEFAULT_DEVICE, update_exponential_moving_average
EXP_ADV_MAX = 100.
def asymmetric_l2_loss(u, tau):
return torch.mean(torch.abs(tau - (u < 0).float()) * u**2)
def f_prime_inverse(residual, name='Pearson_chi_square', temperatrue=3.0):
if name == "Reverse_KL":
return torch.exp(residual * temperatrue)
elif name == "Pearson_chi_square":
return torch.max(residual, torch.zeros_like(residual))
class DILO(nn.Module):
def __init__(self, qf,vf,policy, optimizer_factory,
lamda, maximizer, beta,ita,tau, gradient_type,use_twinV=False, lr=3e-4, discount=0.99, alpha=0.005, max_steps=int(1e6)):
super().__init__()
self.qf = qf.to(DEFAULT_DEVICE)
self.q_target = copy.deepcopy(qf).requires_grad_(False).to(DEFAULT_DEVICE)
self.vf = vf.to(DEFAULT_DEVICE)
self.v_target = copy.deepcopy(vf).requires_grad_(False).to(DEFAULT_DEVICE)
self.policy = policy.to(DEFAULT_DEVICE)
self.v_optimizer = optimizer_factory(self.vf.parameters(), lr = lr)
self.q_optimizer = optimizer_factory(self.qf.parameters(), lr = lr)
self.policy_optimizer = optimizer_factory(self.policy.parameters(), lr = 1e-4)
self.policy_lr_schedule = CosineAnnealingLR(self.policy_optimizer, max_steps)
self.lamda = lamda # 0.8
self.maximizer = maximizer
self.beta = beta # 0.5
self.ita = ita # 0.5
self.tau = tau # 3.0/50
self.discount = discount
self.alpha = alpha
self.gradient_type = gradient_type
self.update_steps = 0
self.use_twinV = use_twinV
def f_star(self, residual, type='chi_square'):
if type=='chi_square':
omega_star = torch.max(residual / 2 + 1, torch.zeros_like(residual))
return residual * omega_star - (omega_star - 1)**2
else:
raise NotImplementedError("f star for divergence not implemented")
def update_full(self, obs,acts,next_obs,next_next_obs,terminals, is_expert, expert_obs, expert_acts, expert_next_obs, expert_next_next_obs, expert_terminals):
v_loss_val = 0.0
metrics = {}
if self.use_twinV:
v_curr = self.qf.both(obs, next_obs)
v_next = self.qf.both(next_obs, next_next_obs)
gt_v_curr = self.qf.both(expert_obs, expert_next_obs)
gt_v_next = self.qf.both(expert_next_obs, expert_next_next_obs)
v_curr = torch.stack(v_curr,dim=1)
v_next = torch.stack(v_next,dim=1)
gt_v_curr = torch.stack(gt_v_curr,dim=1)
gt_v_next = torch.stack(gt_v_next,dim=1)
else:
v_curr = self.qf(obs)
v_next = self.qf(next_obs)
gt_v_curr = self.qf(expert_obs)
v_curr_target = self.q_target(obs, next_obs).detach()
v_next_target = self.q_target(next_obs,next_next_obs).detach()
# Update value function
if self.maximizer == 'smoothed_chi':
if self.use_twinV:
gt_v_next = self.qf.both(expert_next_obs, expert_next_next_obs)
gt_v_next = torch.stack(gt_v_next,dim=1)
backward_residual = ((1. - terminals.view(-1,1).float()) * self.discount * v_next - v_curr_target.view(-1,1))
forward_residual = ((1. - terminals.view(-1,1).float()) * self.discount * v_next_target.view(-1,1) - v_curr)
gt_v_curr_target = self.q_target(expert_obs,expert_next_obs).detach()
gt_v_next_target = self.q_target(expert_next_obs,expert_next_next_obs).detach()
gt_backward_residual = (1. - expert_terminals.view(-1,1).float()) * self.discount * gt_v_next - gt_v_curr_target.view(-1,1)
gt_forward_residual = (1. - expert_terminals.view(-1,1).float()) * self.discount * gt_v_next_target.view(-1,1) - gt_v_curr
else:
gt_v_next = self.qf(expert_next_obs,expert_next_next_obs)
backward_residual = ((1. - terminals.float()) * self.discount * v_next - v_curr_target)
forward_residual = ((1. - terminals.float()) * self.discount * v_next_target - v_curr)
gt_v_curr_target = self.q_target(expert_obs,expert_next_obs).detach()
gt_v_next_target = self.q_target(expert_next_obs,expert_next_next_obs).detach()
gt_backward_residual = (1. - expert_terminals.float()) * self.discount * gt_v_next - gt_v_curr_target
gt_forward_residual = (1. - expert_terminals.float()) * self.discount * gt_v_next_target - gt_v_curr
backward_dual_loss = (self.lamda)*self.ita*(self.beta * self.f_star(gt_backward_residual) + (1-self.beta) * self.f_star(backward_residual) - (1-self.beta)*backward_residual).mean() # First iteartion that worked decently
forward_dual_loss = (self.lamda)* (self.beta * self.f_star(gt_forward_residual) + (1-self.beta) * self.f_star(forward_residual) - (1-self.beta)*forward_residual).mean() # First iteartion that worked decently
else:
raise NotImplementedError('Unavailable divergence for full gradient update')
# For logging
expert_v_val = gt_v_curr.mean().item()
replay_v_val = v_curr.mean().item()
if is_expert.sum()>0:
unseen_expert_v_val = gt_v_curr[is_expert.bool()].mean().item()
else:
unseen_expert_v_val = -1
if (1-is_expert).sum()>0:
unseen_replay_v_val = v_curr[(1-is_expert).bool()].mean().item()
else:
unseen_replay_v_val = -1
if self.use_twinV:
pi_residual = forward_residual[:,0].clone().detach()
else:
pi_residual = forward_residual.clone().detach()
v_loss_val += 0.5*(forward_dual_loss.item() + backward_dual_loss.item())
self.q_optimizer.zero_grad(set_to_none=True)
forward_grad_list, backward_grad_list = [], []
forward_dual_loss.backward(retain_graph=True)
for param in list(self.qf.parameters()):
forward_grad_list.append(param.grad.clone().detach().reshape(-1))
backward_dual_loss.backward()
for i, param in enumerate(list(self.qf.parameters())):
backward_grad_list.append(param.grad.clone().detach().reshape(-1) - forward_grad_list[i])
forward_grad, backward_grad = torch.cat(forward_grad_list), torch.cat(backward_grad_list)
parallel_coef = (torch.dot(forward_grad, backward_grad) / max(torch.dot(forward_grad, forward_grad),
1e-10)).item() # avoid zero grad caused by f*
forward_grad = (1 - parallel_coef) * forward_grad + backward_grad
param_idx = 0
for i, grad in enumerate(forward_grad_list):
forward_grad_list[i] = forward_grad[param_idx: param_idx + grad.shape[0]]
param_idx += grad.shape[0]
# reset gradient and calculate
self.q_optimizer.zero_grad(set_to_none=True)
if self.maximizer == 'smoothed_chi':
v_loss = (1-self.lamda)*v_curr.mean()
else:
v_loss = (gt_v_curr-200).pow(2).mean()
v_loss.backward()
for i, param in enumerate(list(self.qf.parameters())):
param.grad += forward_grad_list[i].reshape(param.grad.shape)
self.q_optimizer.step()
v_loss_val += v_loss.item()
# Update policy network
temperature = self.tau
pi_residual=self.qf(obs, next_obs).detach()
weight = torch.exp(temperature*pi_residual)
weight = torch.clamp_max(weight, EXP_ADV_MAX).detach()
policy_out = self.policy(obs)
bc_losses = (acts-policy_out).pow(2).sum(1)
# bc_losses = -policy_out.log_prob(acts)
policy_loss = torch.mean(weight * bc_losses)
if is_expert.sum()>0:
unseen_expert_pol_weight = weight[is_expert.bool()].mean().item()
else:
unseen_expert_pol_weight = -1
if (1-is_expert).sum()>0:
unseen_replay_pol_weight = weight[(1-is_expert).bool()].mean().item()
else:
unseen_replay_pol_weight = -1
self.policy_optimizer.zero_grad(set_to_none=True)
policy_loss.backward()
self.policy_optimizer.step()
self.policy_lr_schedule.step()
# # Update target Q network
update_exponential_moving_average(self.q_target, self.qf, self.alpha)
metrics['v_loss'] = v_loss_val
metrics['policy_loss'] = policy_loss.item()
metrics['expert_v_val'] = expert_v_val
metrics['replay_v_val'] = replay_v_val
metrics['unseen_expert_v_val'] = unseen_expert_v_val
metrics['unseen_replay_v_val'] = unseen_replay_v_val
metrics['unseen_expert_pol_weight'] = unseen_expert_pol_weight
metrics['unseen_replay_pol_weight'] = unseen_replay_pol_weight
return metrics
def update(self, obs,acts,next_obs,next_next_obs, terminals, is_expert, expert_obs, expert_acts, expert_next_obs, expert_next_next_obs, expert_terminals):
if self.gradient_type == 'semi':
return self.update_semi( obs,acts,next_obs,next_next_obs,terminals, is_expert, expert_obs, expert_acts, expert_next_obs, expert_next_next_obs, expert_terminals)
elif self.gradient_type == 'full':
return self.update_full(obs,acts,next_obs,next_next_obs,terminals, is_expert, expert_obs, expert_acts, expert_next_obs, expert_next_next_obs, expert_terminals)