tunable-priors/inv_class_8bit.py at main · decandido/tunable-priors · GitHub

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
import torch
import torch.nn as nn
from torch import optim

from nflows import flows, distributions

from transform import create_transform
import json
import numpy as np
from tqdm import tqdm
from torch.nn import functional as F
import math
from utils import *


class NFOptimizer_cs(nn.Module):
	def __init__(self,args,perc_k):
		super().__init__()
		self.args=args
		self.perc_k=perc_k
		self.img_dim=3*64*64
		self.latent_dim=3*64*64
		device=args.device
		with open(self.args.flow_config) as fp:
			flow_config = json.load(fp)
		distribution = distributions.StandardNormal((self.img_dim,))
		transform = create_transform(3, 64, 64,
	                                 num_bits=8,
	                                 **flow_config)
		gen = flows.Flow(transform, distribution)
		gen.load_state_dict(torch.load(args.ckpt, map_location=device))
		gen.eval()

		self.gen=gen.to(device)
		self.init_state()

	def init_state(self):
		self.best=None

		self.perc_k=self.perc_k/100


		k_dims=int(self.perc_k*self.latent_dim)

		self.latent_z = torch.zeros((self.args.batchsize,self.img_dim),
	                        dtype=torch.float,
	                        requires_grad=True, device=self.args.device)

		mask_indices = torch.tensor([k_dims])[:, None]
		mask = (reversed(torch.arange(self.img_dim)).expand(mask_indices.shape[0], -1)
                < mask_indices).float()
		self.mask = mask.to(self.args.device)

	def invert_(self,y,img,A,gamma):
		self.img=img
		self.y=y
		self.A=A
		optimizer=optim.Adam([self.latent_z],self.args.lr)
		pbar = tqdm(range(self.args.steps))
		mse_min = np.inf
		mse_loss = 0
		reference_loss = 0


		for i in pbar:

			loss=0
			latent_z = (self.latent_z * self.mask[None, ...]).permute([1, 0, 2]).reshape(-1, self.img_dim)


			img_gen,log_det=self.gen._transform.inverse(latent_z)
			log_det=torch.abs(log_det)

			gen_obsv=torch.matmul(img_gen.view([-1,self.img_dim]),self.A)
			mse_loss = F.mse_loss(self.y,gen_obsv).mean()

			post_loss=(latent_z.norm(dim=1)**2  + log_det)
			post_loss=gamma*nats_to_bits_per_dim(post_loss,3,64,64)


			reference_vector = self.img

			loss+=mse_loss + post_loss.mean()

			reference_loss = F.mse_loss(reference_vector,img_gen).mean()
			optimizer.zero_grad()
			loss.backward()
			optimizer.step()

			if  reference_loss < mse_min:
				mse_min = reference_loss
				self.best = img_gen.detach().cpu()

			pbar.set_description(
                (
                    f" mse: {mse_loss:.4f};"
                    f" post: {post_loss.mean():.4f};"
                    f" mse_ref: {reference_loss:.4f};"
                )
            )

	def invert(self,y,img,A,gamma):
		self.invert_(y,img,A,gamma)
		return self.img,self.best


def make_measurements_noise(args,img,perc_m):
	img_dim=64*64*3
	n=64*64*3

	test_img=img.view([-1,img_dim])

	perc=perc_m/100

	m=int(perc*(img_dim))


	A = np.random.normal(0,1/np.sqrt(m), size=(n,m))
	A = torch.tensor(A, dtype=torch.float, requires_grad=False, device=args.device)

	noise = np.random.normal(0,1,size=(args.batchsize,m))
	noise = noise * 0.1/np.sqrt(m)

	noise*=255
	noise = torch.tensor(noise, dtype=torch.float, requires_grad=False, device=args.device)

	y=torch.matmul(test_img,A) + noise

	return y,A