PROACT/train_utils.py at main · iretes/PROACT · 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
177
178
179
from torch.utils.data import DataLoader
import torch
import numpy as np
import pickle as pkl
from data_utils import *
from ewc_utils import *
import warnings
from utils import *

warnings.filterwarnings("ignore")
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

def train_on_noise_model(noise_ckpt, seed=0, add_noise=True, n_epochs=None, eval_on_tst=True, init_eval=True,
                         key='latest_noise', theta_lr=None,
                         shuffle_noisy_data=False, rnd_noise=False,
                         override_finetune=False, bs=128, tst_on_train=False, cil=False):

    torch.manual_seed(seed)
    np.random.seed(seed)
    # bs = 32

    if 'pkl' in noise_ckpt:
        pkl_file = open(f'{noise_ckpt}', 'rb')
        noise_save_dict = pkl.load(pkl_file)
        pkl_file.close()
    else:
        noise_save_dict = noise_ckpt

    cont_method_args = noise_save_dict['pretrained_ckpt']['cont_method_args']

    model = create_load_add_head(**noise_save_dict['pretrained_ckpt'], load=True)
    ds_dict = get_dataset_specs(**noise_save_dict['pretrained_ckpt'])[0]
    ds_tst = ds_dict['test'][-1]


    ds_train = ds_dict['train'][-1]
    ds_train.data = ds_train.data[noise_save_dict['rnd_idx_train']]
    ds_train.targets = ds_train.targets[noise_save_dict['rnd_idx_train']]

    delta = noise_save_dict['delta']

    if shuffle_noisy_data:
        suffle_idx = np.random.permutation(len(ds_train))
        ds_train.data = ds_train.data[suffle_idx]
        ds_train.targets = ds_train.targets[suffle_idx]
        if rnd_noise == False:
            noise_data = noise_save_dict[key][suffle_idx]
        else:
            noise_data = torch.rand_like(ds_train.data) * delta * 2 - delta

    else:
        if rnd_noise == False:
            noise_data = noise_save_dict[key]
        else:
            noise_data = torch.rand_like(ds_train.data) * delta * 2 - delta


    optim = create_optimizer(model, 'sgd', theta_lr)
    print('optim: sgd')

    if n_epochs == None:
        n_epochs = noise_save_dict['pretrained_ckpt']['n_epochs']

    if init_eval:
        acc_ = []
        for t in range(noise_save_dict['pretrained_ckpt']['task_num']+1):
            ds_tst = ds_dict['test'][t]
            dl_tst_tmp = DataLoader(ds_tst, batch_size=64, shuffle=True)
            acc = eval_dl(model, dl_tst_tmp, verbose=False, task_id=t)
            acc_.append(acc)

        acc_ = np.array(acc_)

        with np.printoptions(precision=2, suppress=True):
            print(f'initial acc: {acc_}')

        avg_acc = noise_save_dict['pretrained_ckpt']['avg_acc']
        bwt = noise_save_dict['pretrained_ckpt']['bwt']

        print(f'initial acc mean: {avg_acc}')
        print(f'initial bwt: {bwt}')

        print()


    model.train()

    if cont_method_args['method'] == 'finetune' or override_finetune:
        model = train_on_noise_model_finetune(model, noise_data, noise_save_dict, optim, ds_dict=ds_dict, ds_train=ds_train, ds_tst=ds_tst,
                                  add_noise=add_noise, n_epochs=n_epochs, eval_on_tst=eval_on_tst, bs=bs)


    elif cont_method_args['method'] == 'ewc':
        model = train_on_noise_model_ewc(model, noise_data, noise_save_dict, optim, ds_dict=ds_dict, ds_train=ds_train, ds_tst=ds_tst,
                                  add_noise=add_noise, n_epochs=n_epochs,
                                  eval_on_tst=eval_on_tst, tst_on_train=tst_on_train, bs=bs, **cont_method_args)


    return model


def train_on_noise_model_finetune(model, noise_data, noise_save_dict, optim, ds_dict, ds_train, ds_tst,
                             add_noise=True, n_epochs=None, eval_on_tst=True, bs=128, cil=False):


    loss_fn = torch.nn.CrossEntropyLoss()

    if len(ds_train) % bs == 0:
        num_of_iter = len(ds_train) // bs
    else:
        num_of_iter = len(ds_train) // bs + 1

    for epoch in range(n_epochs):
        model.train()
        for i in range(num_of_iter):
            tail_idx = min((i+1) * bs, len(ds_train))
            x = ds_train.data[i*bs:tail_idx]
            y = ds_train.targets[i*bs:tail_idx]
            x, y = x.to(device), y.to(device)
            noise = noise_data[i*bs:tail_idx]


            if add_noise:
                noise = noise.to(device)
                x_tilde = torch.clamp(x + noise, 0, 1)
            else:
                x_tilde = x

            y_hat = model(x_tilde)[-1]
            loss = loss_fn(y_hat, y)


            optim.zero_grad()
            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), 10)
            optim.step()

        if eval_on_tst:
            model.eval()
            if cil == False:
                acc_ = []
                for t in range(noise_save_dict['pretrained_ckpt']['task_num']+1):
                    ds_tst = ds_dict['test'][t]
                    dl_tst_tmp = DataLoader(ds_tst, batch_size=64, shuffle=True)
                    acc = eval_dl(model, dl_tst_tmp, verbose=False, task_id=t)
                    acc_.append(acc)

                prev_acc_mat = noise_save_dict['pretrained_ckpt']['acc_mat']
                bwt = (acc_[:-1] - np.diagonal(prev_acc_mat)).mean()


                with np.printoptions(precision=2, suppress=True):
                    print(f'epoch {epoch} acc: {np.array(acc_)}')
                    print(f'average acc up until: {np.mean(acc_[:-1])}')
                    print(f'bwt: {bwt}')

                    print()
            else:
                tmp_ds = combine_ds_class_inc(noise_save_dict['pretrained_ckpt']['task_num']+1, ds_dict['test'])
                tmp_dl_tst = DataLoader(tmp_ds, batch_size=bs, shuffle=False)
                all_acc = acc_curr = eval_dl(model, tmp_dl_tst, verbose=False, class_inc=True)
                prev_acc_lst = []
                for t_id in range(noise_save_dict['pretrained_ckpt']['task_num']):
                    tmp_ds = ds_dict['test'][t_id]
                    tmp_dl_tst = DataLoader(tmp_ds, batch_size=bs, shuffle=False)
                    acc_curr = eval_dl(model, tmp_dl_tst, verbose=False, class_inc=True)
                    prev_acc_lst.append(acc_curr)


                print(f'epoch {epoch} acc task on combined datasets over {noise_save_dict["pretrained_ckpt"]["task_num"]+1} tasks: {all_acc}')
                print(f'epoch {epoch} acc task on individual datasets:\n {prev_acc_lst}')


            model.train()

    return model