DHR-CLIP/train_DHRCLIP.py at main · YUL-git/DHR-CLIP · 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
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
import os
import random
import numpy as np
import torch
import torch.nn.functional as F

import DHRCLIP_lib
from prompt_DHRCLIP import DHRCLIP_PromptLearner
from loss import FocalLoss, BinaryDiceLoss
from dataset_anyres import Dataset
from utils_anyres import get_transform

from logger import get_logger
from tqdm import tqdm
import argparse

def setup_seed(seed):
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    np.random.seed(seed)
    random.seed(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

def train(args):

    logger = get_logger(args.save_path)

    preprocess, target_transform, patch_transform = get_transform(args)
    device = "cuda" if torch.cuda.is_available() else "cpu"

    DHRCLIP_parameters = {"Abnormal_Prompt_length": args.ab_ctx, "Prompt_length": args.n_ctx, "learnabel_text_embedding_depth": args.depth, "learnabel_text_embedding_length": args.t_n_ctx}

    model, _ = DHRCLIP_lib.load("ViT-L/14@336px", device=device, design_details = DHRCLIP_parameters)
    model.eval()

    train_data = Dataset(root=args.train_data_path, transform=preprocess, target_transform=target_transform, patch_transform=patch_transform, dataset_name = args.dataset, args=args)
    train_dataloader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=True, num_workers=4, pin_memory=True,  persistent_workers=True)

    prompt_learner = DHRCLIP_PromptLearner(model.to("cpu"), DHRCLIP_parameters)
    prompt_learner.to(device)
    model.to(device)
    model.visual.DAPM_replace(DPAM_layer = args.dpam)

    optimizer = torch.optim.Adam(list(prompt_learner.parameters()), lr=args.learning_rate, betas=(0.5, 0.999))

    # hyperparameters
    alpha = 0.75

    # losses
    loss_focal = FocalLoss()
    loss_dice = BinaryDiceLoss()
    model.eval()
    prompt_learner.train()
    for epoch in tqdm(range(args.epoch)):
        model.eval()
        prompt_learner.train()

        loss_list = []
        image_loss_list = []

        for items in tqdm(train_dataloader):
            image      = items['img'].to(device)
            patch_imgs = items['patch_imgs']
            patch_imgs = [patch_img.to(device) for patch_img in patch_imgs]
            label      =  items['anomaly']

            gt = items['img_mask'].squeeze().to(device)
            gt[gt > 0.5]  = 1
            gt[gt <= 0.5] = 0

            with torch.no_grad():
                image_features, patch_features = model.encode_image(image, args.features_list, DPAM_layer = 20)
                image_features = image_features / image_features.norm(dim=-1, keepdim=True)

                p_p_features_list = [[],[],[],[],[]]

                p_p_features_list[0].extend(patch_features)

                for idx, patch_img in enumerate(patch_imgs, start=1):
                    _, p_p_features = model.encode_image(patch_img, args.features_list, DPAM_layer = 20)
                    p_p_features_list[idx].extend(p_p_features)

            prompts, tokenized_prompts, compound_prompts_text = prompt_learner(cls_id = None)

            text_features = model.encode_text_learn(prompts, tokenized_prompts, compound_prompts_text).float()
            text_features = torch.stack(torch.chunk(text_features, dim = 0, chunks = 2), dim = 1)
            text_features = text_features/text_features.norm(dim=-1, keepdim=True)

            similarity_map_list   = []
            for idx, p_p_feature in enumerate(p_p_features_list):

                similarity_map_list.append([])
                for patch_feature in p_p_feature:

                    patch_feature = patch_feature/ patch_feature.norm(dim = -1, keepdim = True)
                    similarity, _ = DHRCLIP_lib.compute_similarity(patch_feature, text_features[0])

                    if idx == 0:
                        similarity_map = DHRCLIP_lib.get_similarity_map(similarity[:, 1:, :], args.image_size).permute(0, 3, 1, 2)
                    else:
                        similarity_map = DHRCLIP_lib.get_similarity_map(similarity[:, 1:, :], args.patch_image_size).permute(0, 3, 1, 2)

                    similarity_map_list[idx].append(similarity_map)

            full_sim_map = [torch.nn.functional.interpolate(sim_map, args.target_image_size, mode='bilinear') for sim_map in similarity_map_list[0]]
            full_similarity_map = torch.stack(full_sim_map).permute(1, 0, 2, 3, 4)

            # Patch Aggregation
            full_similarity_map_patch = torch.zeros_like(full_similarity_map) # [8,4,2,518,518]
            full_similarity_map_patch[:, :, :, 0:args.patch_image_size, 0:args.patch_image_size] += torch.stack(similarity_map_list[1]).permute(1, 0, 2, 3, 4)
            full_similarity_map_patch[:, :, :, 0:args.patch_image_size, -args.patch_image_size:] += torch.stack(similarity_map_list[2]).permute(1, 0, 2, 3, 4)
            full_similarity_map_patch[:, :, :, -args.patch_image_size:, 0:args.patch_image_size] += torch.stack(similarity_map_list[3]).permute(1, 0, 2, 3, 4)
            full_similarity_map_patch[:, :, :, -args.patch_image_size:, -args.patch_image_size:] += torch.stack(similarity_map_list[4]).permute(1, 0, 2, 3, 4)

            if args.patch_image_size * 2 > args.target_image_size:
                overlap_patch = args.patch_image_size * 2 - args.target_image_size
                full_similarity_map_patch[:, :, :, args.patch_image_size-overlap_patch:args.patch_image_size, :] /= 2
                full_similarity_map_patch[:, :, :, :, args.patch_image_size-overlap_patch:args.patch_image_size] /= 2

            full_similarity_map = alpha * full_similarity_map + (1-alpha) * full_similarity_map_patch
            full_similarity_map = full_similarity_map.permute(1, 0, 2, 3, 4)

            text_probs = image_features.unsqueeze(1) @ text_features.permute(0, 2, 1)
            text_probs = text_probs[:, 0, ...]/0.07

            image_loss = F.cross_entropy(text_probs, label.long().to(device))
            image_loss_list.append(image_loss.item())

            loss = 0
            for i in range(len(full_similarity_map)):
                loss += loss_focal(full_similarity_map[i], gt)
                loss += loss_dice(full_similarity_map[i][:, 1, :, :], gt)
                loss += loss_dice(full_similarity_map[i][:, 0, :, :], 1-gt)

            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            loss_list.append(loss.item())

        # logs
        if (epoch + 1) % args.print_freq == 0:
            logger.info('epoch [{}/{}], loss:{:.4f}, image_loss:{:.4f}'.format(epoch + 1, args.epoch, np.mean(loss_list), np.mean(image_loss_list)))

        # save model
        if (epoch + 1) % args.save_freq == 0:
            ckp_path = os.path.join(args.save_path, 'epoch_' + str(epoch + 1) + '.pth')
            torch.save({"prompt_learner": prompt_learner.state_dict()}, ckp_path)

if __name__ == '__main__':
    parser = argparse.ArgumentParser("DHRCLIP", add_help=True)
    parser.add_argument("--train_data_path", type=str, default="./data/visa", help="train dataset path")
    parser.add_argument("--save_path", type=str, default='./checkpoints/9_12_4_multiscale_proposed/', help='path to save results')

    parser.add_argument("--dataset", type=str, default='visa', help="train dataset name")

    parser.add_argument("--dpam", type=int, default=20, help="vvclip")
    parser.add_argument("--depth", type=int, default=9, help="image size")
    parser.add_argument("--n_ctx", type=int, default=12, help="zero shot")
    parser.add_argument("--ab_ctx", type=int, default=12, help="zero shot")
    parser.add_argument("--t_n_ctx", type=int, default=4, help="zero shot")
    parser.add_argument("--feature_map_layer", type=int, nargs="+", default=[0, 1, 2, 3], help="zero shot")
    parser.add_argument("--features_list", type=int, nargs="+", default=[6, 12, 18, 24], help="features used")

    parser.add_argument("--epoch", type=int, default=15, help="epochs")
    parser.add_argument("--learning_rate", type=float, default=0.001, help="learning rate")
    parser.add_argument("--batch_size", type=int, default=8, help="batch size")
    parser.add_argument("--image_size", type=int, default=224, help="image size")
    parser.add_argument("--patch_image_size", type=int, default=224, help="patch size")
    parser.add_argument("--target_image_size", type=int, default=448, help="patch size")
    parser.add_argument("--print_freq", type=int, default=1, help="print frequency")
    parser.add_argument("--save_freq", type=int, default=1, help="save frequency")
    parser.add_argument("--seed", type=int, default=111, help="random seed")
    args = parser.parse_args()
    setup_seed(args.seed)
    train(args)