Diff-FCCF/utils.py at main · DanceSkyCode/Diff-FCCF · 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
135
136
137
138
139
140
141
142
143
144
145
146
147
import numpy as np
from torchvision import transforms
import os
import torch
from torch.utils.tensorboard.writer import SummaryWriter
import numpy as np
import cv2
from skimage.segmentation import slic
import matplotlib.pyplot as plt
import torch.nn.functional as F
def compute_reward(merged_losses, history, total_loss, loss_ref):
    for k in loss_ref:
        loss_ref[k] = 0.95*loss_ref[k] + 0.05*merged_losses[k].item()
    ratios = [merged_losses[k].item()/(loss_ref[k]+1e-8) for k in merged_losses]
    balance_penalty = -np.std(ratios)
    trends = list(history.get_features().values())
    trend_stability = np.mean(trends) if trends else 0.0
    reward = -np.log(total_loss + 1e-8) + 0.3*balance_penalty + 0.2*(1.0 if abs(trend_stability)<0.1 else -0.5)
    return reward
def update_loss_ref(merged_losses, loss_ref, alpha=0.95):
    for k in loss_ref:
        loss_ref[k] = alpha*loss_ref[k] + (1-alpha)*merged_losses[k].item()
def save_fusion_images(save_dir, f0, f02, f03, train_images, imgname):
    B = f0.shape[0]
    f0 = f0.detach()
    f02 = f02.detach()
    f03 = f03.detach()
    for batch_idx in range(B):
        image = tensor2img(f0[batch_idx, :, :, :])
        img_cr = tensor2img(train_images[1][batch_idx, 1, :, :].unsqueeze(0))
        img_cb = tensor2img(train_images[1][batch_idx, 2, :, :].unsqueeze(0))
        img = image[:, :, np.newaxis]
        img_cr = img_cr[:, :, np.newaxis]
        img_cb = img_cb[:, :, np.newaxis]
        image = np.concatenate((img, img_cr, img_cb), axis=2)
        image = cv2.cvtColor(image, cv2.COLOR_YCR_CB2BGR)
        cv2.imwrite(os.path.join(save_dir, f"{imgname[batch_idx]}_1.png"), image.astype(np.uint8))
        image = tensor2img(f0[batch_idx, :, :, :])
        img_cr = tensor2img(f02[batch_idx, :, :, :])
        img_cb = tensor2img(f03[batch_idx, :, :, :])
        img = image[:, :, np.newaxis]
        img_cr = img_cr[:, :, np.newaxis]
        img_cb = img_cb[:, :, np.newaxis]
        image = np.concatenate((img, img_cr, img_cb), axis=2)
        image = cv2.cvtColor(image, cv2.COLOR_YCR_CB2BGR)
        cv2.imwrite(os.path.join(save_dir, f"{imgname[batch_idx]}.png"), image.astype(np.uint8))
        cv2.imwrite(os.path.join(save_dir, f"{imgname[batch_idx]}_f01.png"), tensor2img(f0[batch_idx, :, :, :]).astype(np.uint8))
        cv2.imwrite(os.path.join(save_dir, f"{imgname[batch_idx]}_f02.png"), tensor2img(f02[batch_idx, :, :, :]).astype(np.uint8))
        cv2.imwrite(os.path.join(save_dir, f"{imgname[batch_idx]}_f03.png"), tensor2img(f03[batch_idx, :, :, :]).astype(np.uint8))
        cv2.imwrite(os.path.join(save_dir, f"{imgname[batch_idx]}_Y.png"), tensor2img(train_images[1][batch_idx, 0, :, :].unsqueeze(0)).astype(np.uint8))
        cv2.imwrite(os.path.join(save_dir, f"{imgname[batch_idx]}_cr.png"), tensor2img(train_images[1][batch_idx, 1, :, :].unsqueeze(0)).astype(np.uint8))
        cv2.imwrite(os.path.join(save_dir, f"{imgname[batch_idx]}_cb.png"), tensor2img(train_images[1][batch_idx, 2, :, :].unsqueeze(0)).astype(np.uint8))
class SpatialTransformer(torch.nn.Module):
    def __init__(self, size, mode='bilinear'):
        super(SpatialTransformer, self).__init__()
        vectors = [torch.arange(0, s) for s in size]
        grids = torch.meshgrid(vectors, indexing='ij')  # add indexing for PyTorch 1.10+
        grid = torch.stack(grids)  # shape: [2, H, W]
        grid = torch.unsqueeze(grid, 0)  # shape: [1, 2, H, W]
        grid = grid.type(torch.FloatTensor)
        self.register_buffer('grid', grid)
        self.mode = mode
    def forward(self, src, flow):
        new_locs = self.grid + flow
        shape = flow.shape[2:]
        for i in range(len(shape)):
            new_locs[:, i, ...] = 2 * (new_locs[:, i, ...] / (shape[i] - 1) - 0.5)
        new_locs = new_locs.permute(0, 2, 3, 1)[..., [1, 0]]
        return F.grid_sample(src, new_locs, mode=self.mode, align_corners=True)
def create_grid(size):
    num1, num2 = (size[0] + 10) // 10, (size[1] + 10) // 10
    x, y = np.meshgrid(np.linspace(-2, 2, num1), np.linspace(-2, 2, num2))
    fig = plt.figure(figsize=(size[1] / 100, size[0] / 100), dpi=100)
    plt.plot(x, y, color="black")
    plt.plot(x.T, y.T, color="black")
    plt.axis('off')
    plt.subplots_adjust(0, 0, 1, 1)
    fig.canvas.draw()
    grid_img = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8)
    grid_img = grid_img.reshape(fig.canvas.get_width_height()[::-1] + (3,))
    plt.close(fig)
    grid_gray = cv2.cvtColor(grid_img, cv2.COLOR_RGB2GRAY)
    return grid_gray


def save_deformation_field_visualization(dv_field, imgid, id, imgname, img):
    for i in range(dv_field.shape[0]):
        dx = dv_field[i, ..., 0].detach().cpu().numpy()
        dy = dv_field[i, ..., 1].detach().cpu().numpy()
        dx *= 3.0
        dy *= 3.0
        dx = cv2.GaussianBlur(dx, (11, 11), 3)
        dy = cv2.GaussianBlur(dy, (11, 11), 3)
        image = img[i].detach()
        ori_img = tensor2img(image)
        ori_img = np.stack([ori_img] * 3, axis=-1)
        save_path = os.path.join(imgid, f"{imgname[2][i]}_WarpedGrid{id}_ori.png")
        cv2.imwrite(save_path, ori_img)
        H, W, _= ori_img.shape
        grid_img = ori_img.copy()
        step = 30
        for y in range(0, H, step):
            cv2.line(grid_img, (0, y), (W, y), (0, 0, 255), 2)  # 红色横线
        for x in range(0, W, step):
            cv2.line(grid_img, (x, 0), (x, H), (0, 0, 255), 2)  # 红色竖线
        grid_tensor = torch.from_numpy(grid_img).float().unsqueeze(0).permute(0, 3, 1, 2) / 255.0
        flow_tensor = torch.from_numpy(np.stack([dx, dy], axis=0)).unsqueeze(0)
        STN = SpatialTransformer(size=(H, W))
        warped_grid = STN(grid_tensor, flow_tensor)
        warped_grid = warped_grid[0].permute(1, 2, 0).detach().cpu().numpy()  # [C, H, W] -> [H, W, C]
        warped_grid = (warped_grid * 255).astype(np.uint8)
        save_path = os.path.join(imgid, f"{imgname[2][i]}_WarpedGrid{id}.png")
        cv2.imwrite(save_path, warped_grid)
def build_hyperedges(y_channel):
    segments = slic(y_channel.cpu().numpy(),
                   n_segments=200,
                   compactness=10)
    return torch.from_numpy(segments)
def tensor2img(img):
    reverse_transforms = transforms.Compose([
        transforms.Lambda(lambda t: (t + 1) / 2),
        transforms.Lambda(lambda t: t.permute(1, 2, 0)),  # CHW to HWC
        transforms.Lambda(lambda t: t * 255.),
        transforms.Lambda(lambda t: t.cpu().numpy().squeeze().astype(np.uint8)),
    ])
    imgs = reverse_transforms(img)
    return imgs
def tensorboard_writer(timestr):
    log_path = os.path.join('logs', timestr)
    if not os.path.exists(log_path):
        os.makedirs(log_path)
    writer = SummaryWriter(log_path)
    return writer
def logger(timestr):
    log_dir = os.path.join('logs',timestr)
    if not os.path.exists(log_dir):
        os.makedirs(log_dir)
    log_path = os.path.join(log_dir, "log.txt")
    fw = open(log_path, "a+")
    return fw
def save_model(model, epoch,timestr):
    dir_path = os.path.join("weight",timestr)
    if not os.path.exists(dir_path):
        os.makedirs(dir_path)
    ckpt_name = "epoch_" + str(epoch) + ".pt"
    ckpt_path = os.path.join(dir_path, ckpt_name)
    torch.save(model.state_dict(), ckpt_path)