DataPruning/scoring.py at main · Jiajun0425/DataPruning · 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
import numpy as np
import torch
import torch.nn.functional as F
from torch.autograd import grad
from tqdm import tqdm
import random
from argument import args
from data_loader import load_data
from model import *
from utils import *


def compute_mean_gradients(data_loader, model):
    model.eval()
    all_params = [ p for p in model.parameters() if p.requires_grad ]
    mean_gradients = 0

    for i, (data, target) in enumerate(tqdm(data_loader, desc="Computing mean gradients")):
        data, target = data.to(device), target.to(device)

        _, outputs = model(data)
        loss = criterion(outputs, target)

        batch_gradients = list(grad(loss, all_params, create_graph=False))
        batch_gradients = torch.nn.utils.parameters_to_vector(batch_gradients).detach()

        mean_gradients = mean_gradients * i/(i+1) + batch_gradients / (i+1)

    return mean_gradients

def scoring(data_loader, mean_gradients, model):
    model.eval()
    all_params = [ p for p in model.parameters() if p.requires_grad ]
    score = np.zeros(len(data_loader))

    for i, (data, target) in enumerate(tqdm(data_loader, desc="Scoring")):
        data, target = data.to(device), target.to(device)

        _, outputs = model(data)
        loss = criterion(outputs, target)

        sample_gradients = list(grad(loss, all_params, create_graph=False))
        sample_gradients = torch.nn.utils.parameters_to_vector(sample_gradients).detach()

        score[i] = torch.norm(sample_gradients, p=2) * F.cosine_similarity(mean_gradients.unsqueeze(0), sample_gradients.unsqueeze(0))

    return score

def batch_scoring(data_loader, model, num_sample):
    model.eval()
    all_params = [ p for p in model.parameters() if p.requires_grad ]
    score = np.zeros(num_sample)
    start_indice = 0

    for i, (data, target) in enumerate(tqdm(data_loader, desc="Scoring")):
        batch = data.shape[0]
        data, target = data.to(device), target.to(device)

        _, outputs = model(data)
        loss = batch_criterion(outputs, target)
        mean_loss = torch.mean(loss)

        mean_gradients = list(grad(mean_loss, all_params, create_graph=False, retain_graph=True))
        mean_gradients = torch.nn.utils.parameters_to_vector(mean_gradients).detach()
        for l in loss:
            sample_gradients = list(grad(l, all_params, create_graph=False, retain_graph=True))
            sample_gradients = torch.nn.utils.parameters_to_vector(sample_gradients).detach()
            score[start_indice] = torch.norm(sample_gradients, p=2) * F.cosine_similarity(mean_gradients.unsqueeze(0), sample_gradients.unsqueeze(0))
            start_indice += 1

    return score


if __name__ == "__main__":
    save_dir = f"{args.ckpt_dir}/{args.dataset}/{args.model}"
    device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")

    accuracy = np.load(os.path.join(save_dir, "test_accuracy.npy"))
    num_epoch_each_stage = [args.num_epoch // args.num_stage] * args.num_stage
    for i in range(args.num_epoch % args.num_stage):
        num_epoch_each_stage[i] += 1

    whole_loader, _ = load_data(args.data_dir, args.dataset, args.shuffle, args.batch_size, args.test_batch_size)
    sample_loader, _ = load_data(args.data_dir, args.dataset, args.shuffle, 1, args.test_batch_size)

    if args.dataset == "cifar10":
        nclass = 10
    elif args.dataset == "cifar100":
        nclass = 100

    criterion = nn.CrossEntropyLoss(label_smoothing=args.label_smoothing)
    batch_criterion = nn.CrossEntropyLoss(label_smoothing=args.label_smoothing, reduction='none')

    start_indice = 0
    scores = np.zeros((args.num_stage, args.topK, len(sample_loader)))
    for idx in range(args.num_stage):
        acc = accuracy[start_indice: start_indice + num_epoch_each_stage[idx]]
        acc_diff = np.diff(acc)
        sorted_indices = np.argsort(acc_diff)[::-1]
        top_k_epoch = sorted_indices[:args.topK] + start_indice
        print(top_k_epoch)
        start_indice += num_epoch_each_stage[idx]

        for i, k in enumerate(top_k_epoch):
            model_dir = os.path.join(save_dir, f"models/epoch={k}")
            if args.model.lower()=='r18':
                model = ResNet18(nclass)
            elif args.model.lower()=='r50':
                model = ResNet50(num_classes=nclass)
            elif args.model.lower()=='r101':
                model = ResNet101(num_classes=nclass)
            else:
                model = ResNet50(num_classes=nclass)
            model = load_model(model_dir, model)
            model = model.to(device)

            # mean_gradients = compute_mean_gradients(whole_loader, model)
            # mean_gradients = compute_mean_gradients(sample_loader, model)
            # score = scoring(sample_loader, mean_gradients, model)

            score = batch_scoring(whole_loader, model, len(sample_loader))
            scores[idx, i] = score

            # np.save(os.path.join(save_dir, f"test_sample_score{i}.npy"), score)
    mean_score = np.mean(scores, axis=1)
    print(mean_score.shape)
    np.save(os.path.join(save_dir, f"test_batch_scores.npy"), mean_score)