-
Notifications
You must be signed in to change notification settings - Fork 3
/
model_rn.py
143 lines (116 loc) · 4.9 KB
/
model_rn.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
import torch
from torch import nn
from torch.nn.init import kaiming_uniform_, normal_
import torch.nn.functional as F
from utils import *
import wandb
class RelationNetworks(nn.Module):
''' Code modified from https://github.com/rosinality/relation-networks-pytorch '''
def __init__(
self,
channels_out=64,
embed_size=32,
mlp_hidden=64,
n_vocab = 10,
embed_dim = 3,
latents_dim = 64,
classes=1,
use_wandb = False,
):
super().__init__()
self.conv = nn.Sequential(
nn.Conv2d( 3 , channels_out, [3, 3], 1, 0, bias=False),
nn.BatchNorm2d(channels_out),
nn.ReLU(),
nn.Conv2d(channels_out, channels_out, [3, 3], 1, 0, bias=False),
nn.BatchNorm2d(channels_out),
nn.ReLU(),
nn.Conv2d(channels_out, channels_out, [3, 3], 2, 0, bias=False),
nn.BatchNorm2d(channels_out),
nn.ReLU(),
nn.Conv2d(channels_out, channels_out, [3, 3], 2, 0, bias=False),
nn.BatchNorm2d(channels_out),
nn.ReLU(),
)
self.embed = nn.Embedding(n_vocab, embed_dim)
self.n_concat = channels_out * 2 + latents_dim + 2 * 2 # the 2*2 is for coordinates
self.g = nn.Sequential(
nn.Linear(self.n_concat, mlp_hidden),
nn.ReLU(),
nn.Linear(mlp_hidden, mlp_hidden),
nn.ReLU(),
# nn.Linear(mlp_hidden, mlp_hidden),
# nn.ReLU(),
nn.Linear(mlp_hidden, mlp_hidden),
nn.ReLU(),
)
self.f_cat_dim = mlp_hidden * 2 + latents_dim * 4
self.f = nn.Sequential(
nn.Linear(self.f_cat_dim, mlp_hidden),
nn.ReLU(),
nn.Linear(mlp_hidden, int(mlp_hidden/2)),
nn.ReLU(),
nn.Dropout(),
nn.Linear(int(mlp_hidden/2), classes),
)
self.channels_out = channels_out
self.latents_dim = latents_dim
self.mlp_hidden = mlp_hidden
self.use_wandb = use_wandb
coords_dim = 5
coords = torch.linspace(-int(coords_dim/2), int(coords_dim/2), coords_dim)
x = coords.unsqueeze(0).repeat(coords_dim, 1)
y = coords.unsqueeze(1).repeat(1, coords_dim)
coords = torch.stack([x, y]).unsqueeze(0)
self.register_buffer('coords', coords)
def rn_embed(self, image, latents):
conv = self.conv(image)
batch_size, n_channel, conv_h, conv_w = conv.size()
n_pair = conv_h * conv_w
w, a = latents
w_dim = w.size()[-1]
a = softmax_this(a)
conv = torch.mul(conv, a.unsqueeze(2).unsqueeze(2))
w_tile = w.unsqueeze(1).expand(batch_size, n_pair * n_pair, w_dim)
conv = torch.cat([conv, self.coords.expand(batch_size, 2, conv_h, conv_w)], 1)
n_channel += 2
conv_tr = conv.view(batch_size, n_channel, -1).permute(0, 2, 1)
conv1 = conv_tr.unsqueeze(1).expand(batch_size, n_pair, n_pair, n_channel)
conv2 = conv_tr.unsqueeze(2).expand(batch_size, n_pair, n_pair, n_channel)
conv1 = conv1.contiguous().view(-1, n_pair * n_pair, n_channel)
conv2 = conv2.contiguous().view(-1, n_pair * n_pair, n_channel)
concat_vec = torch.cat([conv1, conv2, w_tile], 2).view(-1, self.n_concat)
g = self.g(concat_vec)
g = g.view(-1, n_pair * n_pair, self.mlp_hidden).sum(1).squeeze()
return g
def forward(self, input, latents):
x0, x1 = input
wx, wa, a0, a1 = latents.permute([1,0,2])
g0 = self.rn_embed(x0, (wx, a0))
g1 = self.rn_embed(x1, (wa, a1))
g = torch.cat([g0, g1, latents.view([x0.size()[0], -1])], dim=1)
f = self.f(g)
# stats(f, 'f before')
# f = torch.sigmoid(f)
# stats(f, 'f after ')
# ef = torch.exp(f)
# print('check for inf values in f: finite? {}'.format(torch.isfinite(f).any()))
# print('check for inf values in g1: finite? {}'.format(torch.isfinite(g1).any()))
# print('check for inf values in g0: finite? {}'.format(torch.isfinite(g0).any()))
# if self.use_wandb and torch.isfinite(f).any():
# wandb.log ({
# 'g0': wandb.Histogram (g0.detach().cpu().numpy()),
# 'g1': wandb.Histogram (g1.detach().cpu().numpy()),
# 'f' : wandb.Histogram (f.detach().cpu().numpy() ),
# # 'exp_f': ef.detach().cpu().numpy(),
# }, commit=False)
return f
class Ereason():
def __init__(self, channels_out=64, latents_dim = 64, use_wandb = False):
self.rn = RelationNetworks(channels_out=64, latents_dim = 64, use_wandb = False)
def forward(self, input, latents):
return self.rn(self, input, latents)
def softmax_this(input, beta = 1.):
z = torch.sum( torch.exp(beta*input) )
input = torch.exp(beta*input) / (z + + 1e-9 )
return (input)