File size: 5,648 Bytes
c50dde6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import logging
import math
import torch
import torch.nn.functional as F
import numpy as np
from tqdm import tqdm
from open_clip import create_model
from open_clip.model import get_cast_dtype
from training.distributed import is_master
from training.precision import get_autocast
from training.zero_shot import multi_gpu_sync
import os
import torch.nn as nn
import matplotlib.pyplot as plt
def run_seg(model, data, args):
    dataloader=data['val'].dataloader
    model.eval()
    autocast = get_autocast(args.precision)
    cast_dtype = get_cast_dtype(args.precision)
    cls_embeddings = dataloader.dataset.embeddings
    cls_embeddings = F.normalize(torch.from_numpy(cls_embeddings).float(), dim=-1)
    cls_embeddings = cls_embeddings.to(args.device)
    if cast_dtype is not None:
        cls_embeddings = cls_embeddings.to(dtype=cast_dtype)
    unnorm = UnNormalize([0.48145466, 0.4578275, 0.40821073], [0.26862954, 0.26130258, 0.27577711])
    baseline_model = create_model(
        args.model,  # same !
        args.pretrained,
        device=args.device,
        precision=args.precision,
        output_dict=True,
        cache_dir="ckpt_path"  # cache dir of pre-trained models
    ).to(args.device)
    with torch.no_grad():
        # _, images, bboxes, image_crops, gt_masks, masked_image_crops, proxy_imgs
        # for images, gt_masks, image_names, image_shapes in tqdm(dataloader, disable=not is_master(args)):
        logging.info('Region classifier')
        for image_names, images, _, _, gt_masks, _, _ in tqdm(data['val'].dataloader, disable=not is_master(args)):
            images = images.to(args.device)
            if cast_dtype is not None:
                images = images.to(dtype=cast_dtype)
            tar_h,tar_w=images.shape[-2]//4,images.shape[-1]//4
            with autocast():
                # predict
                if args.distributed and not args.horovod:
                    module = model.module
                    baseline_module=baseline_model.module
                else:
                    module = model
                    baseline_module=baseline_model
                feature_maps = module.encode_dense(images,
                                    normalize=True,
                                    keep_shape=False,
                                    mode="ss",
                                    ex_feats=None)

                baseline_feature_maps = baseline_module.encode_dense(images,
                                                    normalize=True,
                                                    keep_shape=False,
                                                    mode="only_v",
                                                    ex_feats=None)

                bs, N, C = feature_maps.shape
                h, w = int(math.sqrt(N)), int(math.sqrt(N))

                # 计算 logits
                logits = (feature_maps @ cls_embeddings.T) * 40
                baseline_logits = (baseline_feature_maps @ cls_embeddings.T) * 40

                # 变换形状
                logits = logits.permute(0, 2, 1).reshape(bs, -1, h, w)
                baseline_logits = baseline_logits.permute(0, 2, 1).reshape(bs, -1, h, w)

                # 插值
                seg_logits = nn.functional.interpolate(logits, size=(tar_h, tar_w), mode='bilinear')
                baseline_seg_logits = nn.functional.interpolate(baseline_logits, size=(tar_h, tar_w), mode='bilinear')

                batch_size = seg_logits.shape[0]

                for i in range(batch_size):
                    # 处理主模型的预测
                    seg_logits_i = seg_logits[i] * model.logit_scale
                    seg_logits_i = seg_logits_i.softmax(0)  # n_queries * w * h
                    seg_pred = seg_logits_i.argmax(0, keepdim=True)
                    seg_pred = seg_pred.data.cpu().numpy().squeeze(0)
                    
                    # 处理基线模型的预测
                    baseline_seg_logits_i = baseline_seg_logits[i] * model.logit_scale
                    baseline_seg_logits_i = baseline_seg_logits_i.softmax(0)  # n_queries * w * h
                    baseline_seg_pred = baseline_seg_logits_i.argmax(0, keepdim=True)
                    baseline_seg_pred = baseline_seg_pred.data.cpu().numpy().squeeze(0)

                    # 可视化
                    fig, ax = plt.subplots(1, 3, figsize=(18, 6))
                    ax[0].imshow(unnorm(images)[0].permute(1, 2, 0).detach().cpu())
                    ax[0].axis('off')
                    ax[0].set_title("Original Image", fontsize=14)
                    ax[1].imshow(seg_pred, cmap='turbo')
                    ax[1].axis('off')
                    ax[1].set_title("ss distill Result", fontsize=14)
                    ax[2].imshow(baseline_seg_pred, cmap='turbo')
                    ax[2].axis('off')
                    ax[2].set_title("only_v Result", fontsize=14)
                    plt.tight_layout()
                    log_base_path = os.path.join(args.logs, args.name)
                    plt.savefig(f"{log_base_path}/{image_names[i].split('.')[0]}.jpg", bbox_inches='tight')
                    plt.close()

class UnNormalize(object):
    def __init__(self, mean, std):
        self.mean = mean
        self.std = std

    def __call__(self, image):
        image2 = torch.clone(image)
        if len(image2.shape) == 4:
            # batched
            image2 = image2.permute(1, 0, 2, 3).contiguous()
        for t, m, s in zip(image2, self.mean, self.std):
            t.mul_(s).add_(m)
        return image2.permute(1, 0, 2, 3).contiguous()

if __name__=="__main__":
    pass