File size: 5,646 Bytes
bc2cdff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import mitsuba as mi
mi.set_variant("cuda_ad_rgb")
import matplotlib.patches as patches
import drjit as dr
import matplotlib.pyplot as plt
from matplotlib.patches import Polygon
from PDE2D.Coefficient import *
from PDE2D.utils import *
from PDE2D.BoundaryShape import *
from PDE2D.Solver import *
from PDE2D.Solver.constant.wos_constant import Particle
from PDE2D import GreenSampling, Split, PATH
import argparse

parser = argparse.ArgumentParser(description='''Optimization Sphere''')
parser.add_argument('--spp', default = 8, type=int)
parser.add_argument('--resprimal', default = 5, type = int)
parser.add_argument('--restensor', default = 16, type = int)
parser.add_argument('--fdstep', default = 1e-2, type = float)
args = parser.parse_args()

green = GreenSampling.Polynomial
conf_numbers = [UInt32(0), UInt32(1)]
conf_vis = 0
epsilon = 1e-4
use_accel = True
bbox = [[-1, -1], [1, 1]]
resolution_image = [2 ** args.resprimal, 2 ** args.resprimal]
spp_image = 2 ** args.spp

split = Split.Normal
fd_step = args.fdstep
res_tensor = args.restensor


def boundary(points, parameters):
    angle = dr.atan2(points[0], points[1])
    return parameters["scale"] * dr.sin(angle * parameters["freq"]) + parameters["bias"]
parameters1_d = {}
parameters1_d["freq"] = 1
parameters1_d["bias"] = 6
parameters1_d["scale"] = 12
parameters2_d = {}
parameters2_d["freq"] = 8
parameters2_d["bias"] = 4
parameters2_d["scale"] = 8

dirichlet1 = FunctionCoefficient("dirichlet", parameters1_d, boundary)
dirichlet2 = FunctionCoefficient("dirichlet", parameters2_d, boundary)

shape = load_bunny(scale = 1, dirichlet = [dirichlet1, dirichlet2], neumann = [ConstantCoefficient("neumann", 10)], epsilon = epsilon)

out_val = 1
image = (np.arange(res_tensor, dtype = np.float32)) / res_tensor
image = np.tile(image, (res_tensor, 1))
image *= 2
image += out_val

f = ConstantCoefficient("source", 0)
σ = ConstantCoefficient("screening", 2)
#grad_zero_points = shape.create_neumann_points(resolution = 64, spp = 2)
grad_zero_points = None
α = TextureCoefficient("diffusion", bbox = bbox, tensor_np = image, out_val=out_val, grad_zero_points=grad_zero_points)
data_holder = DataHolder(shape = shape, α = α, σ = σ, f=f)

wos = WostVariable(data_holder, green_sampling=green, use_accelaration=use_accel, opt_params= ["diffusion.texture.tensor"])
opt = Adam(lr = 0.1, params = wos.opt_params)
wos.update(opt)

points = create_image_points(bbox, resolution_image, spp_image)

L, p = wos.solve(points_in = points, conf_numbers = conf_numbers, split = split)
image_0, tensor = create_image_from_result(L, resolution_image)
fig, (ax1) = plt.subplots(1, 1, figsize=[5, 5])
plot_image(image_0[conf_vis], ax1)
ax1.set_title("Primal Result")

loss_grad = compute_loss_grad_image(result = tensor)
dL = compute_dL_image(loss_grad=loss_grad, spp = spp_image)

fig, ax = plt.subplots(1,1, figsize = (5,5))
plot_image(loss_grad[0].numpy(), ax)

dL_image,_ = create_image_from_result(dL, resolution_image)
fig, ax = plt.subplots(1,1, figsize = (5,5))
plot_image(dL_image[0] * spp_image, ax)

L_grad, p = wos.solve_grad(points_in = points, split = split, dL = dL, conf_numbers=conf_numbers, verbose = True)
grad_prb = dr.grad(α.tensor).numpy()

def step_texture(texture : Coefficient, i, j, fd_step):
    tex1 = texture.copy()
    tex2 = texture.copy()
    index = i * tex1.tensor.shape[0] + j
    dr.scatter_add(tex1.tensor.array, +fd_step, index)
    dr.scatter_add(tex2.tensor.array, -fd_step, index)
    dr.make_opaque(tex1.tensor)
    dr.make_opaque(tex2.tensor)
    tex1.update_texture()
    tex2.update_texture()
    return tex1, tex2

fd_grad = np.zeros_like(image)
for i in range(image.shape[0]):
    for j in range(image.shape[1]):
        α1, α2 = step_texture(α, i, j, fd_step)
        data_holder1 = DataHolder(shape = shape, α = α1, α_split = α, σ = σ, σ_split = σ, f = f)
        data_holder2 = DataHolder(shape = shape, α = α2, α_split = α, σ = σ, σ_split = σ, f = f)
        wos1 = WostVariable(data_holder1, use_accelaration = use_accel, green_sampling = green)
        wos2 = WostVariable(data_holder2, use_accelaration = use_accel, green_sampling = green)
        L1, _ = wos1.solve(points, split = split, conf_numbers=conf_numbers, fd_forward=True, verbose = False)
        L2, _ = wos2.solve(points, split = split, conf_numbers=conf_numbers, fd_forward=True, verbose = False)
        image1, tensor1 = create_image_from_result(L1, resolution_image)
        image2, tensor2 = create_image_from_result(L2, resolution_image)
        val1 = np.sum(MSE_numpy(image1))
        val2 = np.sum(MSE_numpy(image2))
        print(val1, val2)
        fd_grad[i,j] = (val1 - val2) / (2 * fd_step)
        print(f"{i},{j}")

fig, (ax1, ax2, ax3) = plt.subplots(1,3, figsize=  (14,5))
maxval = max(np.max(fd_grad), np.max(grad_prb))
minval = min(np.min(fd_grad), np.min(grad_prb))
max_range = max(maxval, -minval)
plot_image(fd_grad, ax1, input_range=(-max_range, max_range), cmap = "coolwarm")
plot_image(grad_prb, ax2, input_range=(-max_range, max_range), cmap = 'coolwarm')
plot_image(np.abs(fd_grad.squeeze()-grad_prb.squeeze()), ax3, cmap = 'coolwarm')
ax1.set_title("FD")
ax2.set_title("PRB")

import os
def create_path(path):
    if not os.path.exists(path):
        os.makedirs(path)
path = os.path.join(PATH, "output2D", "finite_differences", "variable", "diffusion")
create_path(path)

fig.savefig(os.path.join(path, f"diffusion{res_tensor}-fd{fd_step}.pdf"), bbox_inches='tight', pad_inches=0.04, dpi=200)
record = {}
record["prb"] = grad_prb
record["fd"] = fd_grad
np.save(os.path.join(path, f"diffusion{res_tensor}-fd{fd_step}.npy"), record)