MCPNet/extract_prototypes.py at dc73281881fd0062470cd002d54c222c2d6ec8f4 · NVlabs/MCPNet · 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
import torch
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import sys
import numpy as np
import torchvision
from torch.utils.data import DataLoader
import tqdm
import os
from utils.general import get_dataset, get_model_set, load_weight, load_model

import argparse

# Note:
# Calculate the weighted bias based on the weighted sample covariance formulation.
# https://en.wikipedia.org/wiki/Weighted_arithmetic_mean#Weighted_sample_covariance

def cal_cov_component_weighted(feat, Sum_A, Square_Sum_A, cov_xx, cov_mean, layer_i, args):
    feat = torch.flatten(feat, 2)
    strength = torch.norm(feat, p = 2, dim = 1, keepdim = True)
    ori_feat = feat
    feat = feat * strength
    Sum_A[layer_i] += torch.sum(strength.squeeze(1), dim = 1)
    Square_Sum_A[layer_i] += torch.sum(strength.squeeze(1) ** 2, dim = 1)
    cov_xx[layer_i] += torch.bmm(feat, ori_feat.permute(0, 2, 1))
    cov_mean[layer_i] += torch.sum(feat, dim = -1, keepdim = True)
    return Sum_A, Square_Sum_A, cov_xx, cov_mean

def cal_cov_component(feat, cov_xx, cov_mean, N, layer_i):
    feat = torch.flatten(feat, 2)

    N[layer_i] += feat.shape[-1]
    cov_xx[layer_i] += torch.bmm(feat, feat.permute(0, 2, 1))
    cov_mean[layer_i] += torch.sum(feat, dim = -1, keepdim = True)
    return cov_xx, cov_mean, N

def cal_cov(cov_xx, cov_mean, Sum_A = None, Square_Sum_A = None, N = None):
    cov = []
    for i in range(4):
        if args.weighted:
            cov_xx[i] /= Sum_A[i][:, None, None]
            cov_mean[i] /= Sum_A[i][:, None, None]
            cov.append(cov_xx[i] - torch.bmm(cov_mean[i], cov_mean[i].permute(0, 2, 1))) # * (1 / (1. - (Square_Sum_A[i] / (Sum_A[i] ** 2 + 1e-16))))[:, None, None]
        else:
            cov_xx[i] /= N[i]
            cov_mean[i] /= N[i]
            cov.append(cov_xx[i] - torch.bmm(cov_mean[i], cov_mean[i].permute(0, 2, 1)))
        cov_mean[i] = cov_mean[i][..., 0]
    return cov, cov_mean

def cal_cov_matrix(model, train_loader, args, post_name = "", saved = True):
    os.makedirs(f"./PCA_concept_specific_tmp/{args.case_name}/{args.basic_model}/", exist_ok = True)
    cov_xx = [torch.zeros(args.concept_per_layer[0], args.cha[0], args.cha[0], dtype = torch.float64).cuda(),
              torch.zeros(args.concept_per_layer[1], args.cha[1], args.cha[1], dtype = torch.float64).cuda(),
              torch.zeros(args.concept_per_layer[2], args.cha[2], args.cha[2], dtype = torch.float64).cuda(),
              torch.zeros(args.concept_per_layer[3], args.cha[3], args.cha[3], dtype = torch.float64).cuda()]
    cov_mean = [torch.zeros(args.concept_per_layer[0], args.cha[0], 1, dtype = torch.float64).cuda(),
                torch.zeros(args.concept_per_layer[1], args.cha[1], 1, dtype = torch.float64).cuda(),
                torch.zeros(args.concept_per_layer[2], args.cha[2], 1, dtype = torch.float64).cuda(),
                torch.zeros(args.concept_per_layer[3], args.cha[3], 1, dtype = torch.float64).cuda()]

    if args.weighted:
        Sum_A = [torch.zeros(args.concept_per_layer[0], dtype = torch.float64).cuda(),
                 torch.zeros(args.concept_per_layer[1], dtype = torch.float64).cuda(),
                 torch.zeros(args.concept_per_layer[2], dtype = torch.float64).cuda(),
                 torch.zeros(args.concept_per_layer[3], dtype = torch.float64).cuda()]
        Square_Sum_A = [torch.zeros(args.concept_per_layer[0], dtype = torch.float64).cuda(),
                        torch.zeros(args.concept_per_layer[1], dtype = torch.float64).cuda(),
                        torch.zeros(args.concept_per_layer[2], dtype = torch.float64).cuda(),
                        torch.zeros(args.concept_per_layer[3], dtype = torch.float64).cuda()]
        N = None
    else:
        Sum_A = None
        Square_Sum_A = None
        N = [0, 0, 0, 0]

    with torch.no_grad():
        for iteration, (img, label) in tqdm.tqdm(enumerate(train_loader), total = len(train_loader)):
            img = img.cuda()
            l1, l2, l3, l4 = model(img)
            features = [l1, l2, l3, l4]
            for layer_i, feat in enumerate(features):
                concept_num = args.concept_per_layer[layer_i]
                cha_per_con = args.cha[layer_i]
                B, C, H, W = feat.shape
                feat = feat.reshape(B, concept_num, cha_per_con, H, W).permute(1, 2, 0, 3, 4)

                if args.weighted:
                    Sum_A, Square_Sum_A, cov_xx, cov_mean = cal_cov_component_weighted(feat, Sum_A, Square_Sum_A, cov_xx, cov_mean, layer_i, args)
                else:
                    cov_xx, cov_mean, N = cal_cov_component(feat, cov_xx, cov_mean, N, layer_i)

        cov, cov_mean = cal_cov(cov_xx, cov_mean, Sum_A, Square_Sum_A, N)
        if saved:
            torch.save(cov, f"./PCA_concept_specific_tmp/{args.case_name}/{args.basic_model}/cov_topk2{post_name}.pkl")
            torch.save(cov_mean, f"./PCA_concept_specific_tmp/{args.case_name}/{args.basic_model}/mean_topk2{post_name}.pkl")

    return cov, cov_mean

def load_dataset(data_path: str, image_size: int, args, data_transform = None) -> DataLoader:
    transform = transforms.Compose([transforms.Resize(image_size + 32),
                                     transforms.CenterCrop((image_size, image_size)),
                                     transforms.ToTensor(),
                                     transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) if data_transform is None else data_transform
    train_dataset = torchvision.datasets.ImageFolder(data_path, transform)

    train_loader = DataLoader(train_dataset, batch_size = 64, shuffle = False, num_workers = 8)
    return train_loader

if __name__ == "__main__":
    print("Extract the concept vector via PCA!! (Only calculate the weighted covariance matrixs)")
    parser = argparse.ArgumentParser()
    parser.add_argument('--case_name', default = "AWA2_Baseline", type = str)
    parser.add_argument('--few_shot', default = False, action = "store_true")
    parser.add_argument('--device', default = "0", type = str)
    parser.add_argument('--cha', default = [32, 32, 32, 32], type = int, nargs='+')
    parser.add_argument('--concept_per_layer', default = [8, 16, 32, 64], type = int, nargs='+')
    parser.add_argument('--weighted', default = True, action = "store_false")
    parser.add_argument('--basic_model', default = "resnet50", type = str)
    parser.add_argument('--model', default = "aix_model", type = str)
    parser.add_argument('--w_mode', default = "best", choices=['best', 'last'], type = str)
    parser.add_argument("--saved_dir", default = ".", type = str)
    args = parser.parse_args()

    os.environ["CUDA_VISIBLE_DEVICES"] = args.device

    args.dst = f"{args.saved_dir}/pkl/{args.case_name}/{args.model.lower()}_{args.basic_model.lower()}"

    os.makedirs(f"./PCA_concept_specific_tmp/{args.case_name}/{args.basic_model}/", exist_ok = True)
    data_path, train_path, val_path, num_class = get_dataset(args.case_name)

    if args.few_shot:
        data_path += "seen/"

    if "AWA2" in args.case_name and args.few_shot:
        num_class = 40
    elif "Caltech101" in args.case_name and args.few_shot:
        num_class = 81
    elif "StanfordCar" in args.case_name and args.few_shot:
        num_class = 157
    elif "CUB" in args.case_name and args.few_shot:
        num_class = 160
    data_path = data_path + train_path
    args, image_size = get_model_set(args)
    model = load_model(args.model, args.basic_model, num_class)
    print(args)

    train_loader = load_dataset(data_path, image_size, args)
    trained_param_path = f"{args.dst}/best_model.pkl"
    load_weight(model, trained_param_path)

    post_name = ""
    if args.few_shot:
        post_name += "_few_shot"
    if args.w_mode == "last":
        post_name += "_last"

    print(args.case_name)
    model.eval()

    cal_cov_matrix(model, train_loader, args, post_name)