File size: 7,408 Bytes
c509e76 |
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 |
from __future__ import absolute_import
import numpy as np
import itertools
import torch
import torch.nn as nn
import torch.nn.functional as F
def grid_sample(input, grid, canvas = None):
output = F.grid_sample(input, grid)
if canvas is None:
return output
else:
input_mask = input.data.new(input.size()).fill_(1)
output_mask = F.grid_sample(input_mask, grid)
padded_output = output * output_mask + canvas * (1 - output_mask)
return padded_output
# phi(x1, x2) = r^2 * log(r), where r = ||x1 - x2||_2
def compute_partial_repr(input_points, control_points):
N = input_points.size(0)
M = control_points.size(0)
pairwise_diff = input_points.view(N, 1, 2) - control_points.view(1, M, 2)
# original implementation, very slow
# pairwise_dist = torch.sum(pairwise_diff ** 2, dim = 2) # square of distance
pairwise_diff_square = pairwise_diff * pairwise_diff
pairwise_dist = pairwise_diff_square[:, :, 0] + pairwise_diff_square[:, :, 1]
repr_matrix = 0.5 * pairwise_dist * torch.log(pairwise_dist)
# fix numerical error for 0 * log(0), substitute all nan with 0
mask = repr_matrix != repr_matrix
repr_matrix.masked_fill_(mask, 0)
return repr_matrix
# # output_ctrl_pts are specified, according to our task.
# def build_output_control_points(num_control_points, margins):
# margin_x, margin_y = margins
# margin_x, margin_y = 0,0
# num_ctrl_pts_per_side = num_control_points // 2
# ctrl_pts_x = np.linspace(margin_x, 1.0 - margin_x, num_ctrl_pts_per_side)
# ctrl_pts_y_top = np.ones(num_ctrl_pts_per_side) * margin_y
# ctrl_pts_y_bottom = np.ones(num_ctrl_pts_per_side) * (1.0 - margin_y)
# ctrl_pts_top = np.stack([ctrl_pts_x, ctrl_pts_y_top], axis=1)
# ctrl_pts_bottom = np.stack([ctrl_pts_x, ctrl_pts_y_bottom], axis=1)
# # ctrl_pts_top = ctrl_pts_top[1:-1,:]
# # ctrl_pts_bottom = ctrl_pts_bottom[1:-1,:]
# output_ctrl_pts_arr = np.concatenate([ctrl_pts_top, ctrl_pts_bottom], axis=0)
# output_ctrl_pts = torch.Tensor(output_ctrl_pts_arr)
# return output_ctrl_pts
# output_ctrl_pts are specified, according to our task.
def build_output_control_points(num_control_points, margins):
margin_x, margin_y = margins
# margin_x, margin_y = 0,0
num_ctrl_pts_per_side = (num_control_points-4) // 4 +2
ctrl_pts_x = np.linspace(margin_x, 1.0 - margin_x, num_ctrl_pts_per_side)
ctrl_pts_y_top = np.ones(num_ctrl_pts_per_side) * margin_y
ctrl_pts_y_bottom = np.ones(num_ctrl_pts_per_side) * (1.0 - margin_y)
ctrl_pts_top = np.stack([ctrl_pts_x, ctrl_pts_y_top], axis=1)
ctrl_pts_bottom = np.stack([ctrl_pts_x, ctrl_pts_y_bottom], axis=1)
ctrl_pts_x_left = np.ones(num_ctrl_pts_per_side) * margin_x
ctrl_pts_x_right = np.ones(num_ctrl_pts_per_side) * (1.0-margin_x)
ctrl_pts_left = np.stack([ctrl_pts_x_left[1:-1], ctrl_pts_x[1:-1]], axis=1)
ctrl_pts_right = np.stack([ctrl_pts_x_right[1:-1], ctrl_pts_x[1:-1]], axis=1)
output_ctrl_pts_arr = np.concatenate([ctrl_pts_top, ctrl_pts_bottom, ctrl_pts_left, ctrl_pts_right], axis=0)
output_ctrl_pts = torch.Tensor(output_ctrl_pts_arr)
return output_ctrl_pts
# demo: ~/test/models/test_tps_transformation.py
class TPSSpatialTransformer(nn.Module):
def __init__(self, output_image_size=None, num_control_points=None, margins=None):
super(TPSSpatialTransformer, self).__init__()
self.output_image_size = output_image_size
self.num_control_points = num_control_points
self.margins = margins
self.target_height, self.target_width = output_image_size
target_control_points = build_output_control_points(num_control_points, margins)
N = num_control_points
# N = N - 4
# create padded kernel matrix
forward_kernel = torch.zeros(N + 3, N + 3)
target_control_partial_repr = compute_partial_repr(target_control_points, target_control_points)
forward_kernel[:N, :N].copy_(target_control_partial_repr)
forward_kernel[:N, -3].fill_(1)
forward_kernel[-3, :N].fill_(1)
forward_kernel[:N, -2:].copy_(target_control_points)
forward_kernel[-2:, :N].copy_(target_control_points.transpose(0, 1))
# compute inverse matrix
# print(forward_kernel.shape)
inverse_kernel = torch.inverse(forward_kernel)
# create target cordinate matrix
HW = self.target_height * self.target_width
target_coordinate = list(itertools.product(range(self.target_height), range(self.target_width)))
target_coordinate = torch.Tensor(target_coordinate) # HW x 2
Y, X = target_coordinate.split(1, dim = 1)
Y = Y / (self.target_height - 1)
X = X / (self.target_width - 1)
target_coordinate = torch.cat([X, Y], dim = 1) # convert from (y, x) to (x, y)
target_coordinate_partial_repr = compute_partial_repr(target_coordinate, target_control_points)
target_coordinate_repr = torch.cat([
target_coordinate_partial_repr, torch.ones(HW, 1), target_coordinate
], dim = 1)
# register precomputed matrices
self.register_buffer('inverse_kernel', inverse_kernel)
self.register_buffer('padding_matrix', torch.zeros(3, 2))
self.register_buffer('target_coordinate_repr', target_coordinate_repr)
self.register_buffer('target_control_points', target_control_points)
def forward(self, input, source_control_points,direction='dewarp'):
if direction == 'dewarp':
assert source_control_points.ndimension() == 3
assert source_control_points.size(1) == self.num_control_points
assert source_control_points.size(2) == 2
batch_size = source_control_points.size(0)
Y = torch.cat([source_control_points, self.padding_matrix.expand(batch_size, 3, 2)], 1)
mapping_matrix = torch.matmul(self.inverse_kernel, Y)
source_coordinate = torch.matmul(self.target_coordinate_repr, mapping_matrix)
grid = source_coordinate.view(-1, self.target_height, self.target_width, 2)
grid = torch.clamp(grid, 0, 1) # the source_control_points may be out of [0, 1].
# the input to grid_sample is normalized [-1, 1], but what we get is [0, 1]
grid = 2.0 * grid - 1.0
output_maps = grid_sample(input, grid, canvas=None)
return output_maps, source_coordinate
# elif direction == 'warp':
# target_control_points = source_control_points.clone()
# source_control_points = (build_output_control_points(self.num_control_points, self.margins)).clone()
# source_control_points = source_control_points.unsqueeze(0)
# source_control_points = source_control_points.expand(target_control_points.size(0),self.num_control_points,2)
# assert source_control_points.ndimension() == 3
# assert source_control_points.size(1) == self.num_control_points
# assert source_control_points.size(2) == 2
# batch_size = source_control_points.size(0)
# Y = torch.cat([source_control_points.to('cuda'), self.padding_matrix.expand(batch_size, 3, 2)], 1)
# mapping_matrix = torch.matmul(self.inverse_kernel, Y)
# source_coordinate = torch.matmul(self.target_coordinate_repr, mapping_matrix)
# grid = source_coordinate.view(-1, self.target_height, self.target_width, 2)
# grid = torch.clamp(grid, 0, 1) # the source_control_points may be out of [0, 1].
# # the input to grid_sample is normalized [-1, 1], but what we get is [0, 1]
# grid = 2.0 * grid - 1.0
# output_maps = grid_sample(input, grid, canvas=None)
# return output_maps, source_coordinate
|