-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathrelation_net.py
132 lines (126 loc) · 4.71 KB
/
relation_net.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
from tianshou.data import Batch, ReplayBuffer, to_numpy, to_torch, to_torch_as
import stable_baselines3.common.logger as L
import functools
import gym
import numpy as np
from torch.nn import functional as F
from einops.layers.torch import Rearrange
from encoder import *
import einops
class RNEncoder(nn.Module):
def __init__(self, obs_space, act_space, cfg):
super().__init__()
self.cfg = cfg
obs_space = gym.spaces.Box(low=-1, high=1000, shape=cfg.obs_shape)
self.enc = ImpalaEncoder(obs_space, channels=cfg.filters, flatten=False)
c, h, w = self.enc.final_shape
self.pred_z_cat = create_mlp(cfg.filters[-1], cfg.obj_cat_num, [cfg.filters[-1]], return_seq=True)
self.output_shape = (h, w, c + cfg.obj_cat_num)
def split_obs(self, o):
shape = o.shape
obs_shape = self.cfg.obs_shape
mask_shape = (8, 8, self.cfg.obj_cat_num)
obs = o[...,:np.prod(obs_shape)].reshape(*shape[:-1], *obs_shape)
mask = o[...,np.prod(obs_shape):].reshape(*shape[:-1], *mask_shape)
return obs, mask.detach()
def forward(self, x, ret_latent=False):
if isinstance(x, dict):
x = x['obs']
obs, obj_cat = self.split_obs(x)
out0 = self.enc(obs).permute(0,2,3,1) # (h, w, c)
out = torch.cat([out0, obj_cat], dim=-1)
if ret_latent:
return out, out0
else:
return out
def enc_loss(self, b, latent=None):
if self.cfg.enc_coeff <= 0:
pred_loss = torch.Tensor([0]).to(b.obs.device).sum()
else:
obs, obj_cat = self.split_obs(b.obs)
if latent is None:
latent = self.enc(obs)
pred_z_cat = self.pred_z_cat(latent)
pred_z_cat_loss = -(F.log_softmax(pred_z_cat, dim=-1) * obj_cat).sum(-1)
pred_z_cat_loss = (pred_z_cat_loss).sum([1,2]).mean()
L.record_mean('encoder/pred_loss', pred_z_cat_loss.item())
pred_loss = self.cfg.enc_coeff * pred_z_cat_loss
return pred_loss
class AddSInfo(nn.Module):
def __init__(self, h, w, c, cout=32, channel_first=False, use_mlp=True):
super().__init__()
identity = torch.tensor([[[1.0, 0.0, 1.0], [0.0, 1.0, 1.0]]], dtype=torch.float32)
grid = F.affine_grid(identity, [1, 1, h, w])
grid = grid.permute(0, 3, 1, 2).contiguous()
# (1, 2, h, w)
self.register_buffer('grid', grid)
assert channel_first == False
if not channel_first:
# (1, h, w, 2)
self.grid = grid.permute(0,2,3,1)
self.use_mlp = use_mlp
if self.use_mlp:
self.mlp = nn.Linear(c+2, cout)
def forward(self, x):
x = torch.cat([x, self.grid.to(x.device).expand(x.shape[0], -1, -1, -1)], dim=-1)
if self.use_mlp:
x = self.mlp(x)
return x
class ObjSummary(nn.Module):
def __init__(self, c, obj_cat_num):
super().__init__()
self.head = 4
self.query_atten = QueryMultiHeadAttention(obj_cat_num, c, self.head,
to_q_net=[32], to_k_net=[32], to_v_net=[32], to_out_net=[])
self.out_dim = c * obj_cat_num
"""
x: (N, B, E)
obj_cat: (N, B, S)
out: (B, S*E)
"""
def forward(self, x, obj_cat):
mask = einops.repeat(obj_cat, 'n b s -> b h s n', h=self.head)
out = self.query_atten(x, mask=mask)
out = einops.rearrange(out, 's n e -> n (s e)')
return out
class RNModule(nn.Module):
def __init__(self, input_shape, action_space, cfg):
super().__init__()
self.cfg = cfg
h, w, c = input_shape
obj_cat_num = c - 32
self.obj_cat_num = c - 32
self.add_sinfo = AddSInfo(h, w, c, cout=32)
self.trans = Rearrange('n h w c -> (h w) n c')
self.atten = nn.MultiheadAttention(32, 4)
if not cfg.use_sep_mlp:
create_layer = nn.Linear
else:
create_layer = functools.partial(MultiLinear, num_linears=self.obj_cat_num)
fdim = 32
self.mlp = create_mlp(64, fdim, [64], create_layer=create_layer, return_seq=True)
self.ac = nn.Linear(fdim, action_space.n + 1)
def forward(self, x, ret_atten_wts=False, mask_out = None):
obj_cat = x[...,-self.obj_cat_num:] # B, H, W, S
atten_wts = None
x = self.add_sinfo(x)
x = self.trans(x)
atten_out, atten_wts = self.atten(x, x, x)
x0 = x
x = torch.cat([x, atten_out], dim=-1) # (N, B, 64)
if self.cfg.use_sep_mlp:
x = x.unsqueeze(-2).expand(-1, -1, self.obj_cat_num, -1) # (N, B, S, 64)
out = self.mlp(x)
if self.cfg.use_sep_mlp:
obj_cat = einops.repeat(obj_cat, 'b h w s -> (h w) b s k', k=1) # n, b, s, k
if mask_out is not None:
obj_cat = obj_cat * einops.repeat(to_torch_as(mask_out, obj_cat), 's -> s k', k=1)
if True:
obj_cat[...,-1,:] += 1e-4
obj_cat = obj_cat / obj_cat.sum(-2, keepdim=True)
out = (out * obj_cat).sum(-2) # N, B, 64
out = out.amax(0) # (n, 64)
out = self.ac(out)
if ret_atten_wts:
return out, atten_wts
return out