Image Classification
English
File size: 15,393 Bytes
377dccd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
# Copyright 2022-present, Lorenzo Bonicelli, Pietro Buzzega, Matteo Boschini, Angelo Porrello, Simone Calderara.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import csv
import math
import sys
from argparse import Namespace
from typing import Tuple
import torch
from datasets import get_dataset
from datasets.utils.continual_dataset import ContinualDataset
from models.utils.continual_model import ContinualModel
from typing import Tuple, List
from utils.loggers import *
from utils.mlflow_logger import MLFlowLogger
from utils.status import ProgressBar
import torch.nn.functional as F
import utils.metrics


def evaluate_ece(model: ContinualModel, dataset: ContinualDataset, last=False) -> Tuple[List[float], List[float], float, float, float]:
    """
    Evaluates accuracy and computes AURC, FPR95, AUROC for each loader, then averages.
    :param model: model to evaluate
    :param dataset: continual dataset
    :return: class-il acc, task-il acc, mean AURC, mean FPR95, mean AUROC
    """
    status = model.net.training
    model.net.eval()
    accs, accs_mask_classes = [], []
    ece_list = []

    for k, test_loader in enumerate(dataset.test_loaders):
        if last and k < len(dataset.test_loaders) - 1:
            continue

        correct, correct_mask_classes, total = 0.0, 0.0, 0.0
        val_log = {'softmax' : [], 'correct' : [], 'logit' : [], 'target':[]}
        
        for data in test_loader:
            with torch.no_grad():
                inputs, labels = data
                inputs, labels = inputs.to(model.device), labels.to(model.device)

                if 'class-il' not in model.COMPATIBILITY:
                    output = model(inputs, k)
                else:
                    output = model(inputs)
                softmax = F.softmax(output, dim=1)
                _, pred_cls = softmax.max(1)

                val_log['correct'].append(pred_cls.cpu().eq(labels.cpu().data.view_as(pred_cls)).numpy())
                val_log['softmax'].append(softmax.cpu().data.numpy())
                val_log['logit'].append(output.cpu().data.numpy())
                val_log['target'].append(labels.cpu().data.numpy())
                _, pred = torch.max(output.data, 1)
                correct += torch.sum(pred == labels).item()
                total += labels.shape[0]
                if dataset.SETTING == 'class-il':
                    mask_classes(output, dataset, k)
                    _, pred_masked = torch.max(output.data, 1)
                    correct_mask_classes += pred_masked.eq(labels).sum().item()
        for key in val_log : 
            val_log[key] = np.concatenate(val_log[key])
        acc = 100. * val_log['correct'].mean()
        ece = utils.metrics.calc_ece(val_log['softmax'], val_log['target'], bins=15)
        accs.append(correct / total * 100 if 'class-il' in model.COMPATIBILITY else 0)
        accs_mask_classes.append(correct_mask_classes / total * 100)
        ece_list.append(ece*100)
    model.net.train(status)

    # averagge
    mean_ece = np.mean(ece_list)
    print('evaluation acc:', accs)
    print(f'ECE per loader: {ece_list}')
    print(f'Mean ECE: {mean_ece:.4f}')

    return accs, accs_mask_classes
def evaluate_eceid(model: ContinualModel, dataset: ContinualDataset, last=False) -> Tuple[List[float], List[float], float, float, float]:
    """
    Evaluates accuracy and computes AURC, FPR95, AUROC for each loader, then averages.
    :param model: model to evaluate
    :param dataset: continual dataset
    :return: class-il acc, task-il acc, mean AURC, mean FPR95, mean AUROC
    """
    status = model.net.training
    model.net.eval()
    accs, accs_mask_classes = [], []
    ece_list = []

    for k, test_loader in enumerate(dataset.test_loaders):
        if last and k < len(dataset.test_loaders) - 1:
            continue
        correct, correct_mask_classes, total = 0.0, 0.0, 0.0
        val_log = {'softmax' : [], 'correct' : [], 'logit' : [], 'target':[]}
        
        for data in test_loader:
            with torch.no_grad():
                inputs, labels = data
                inputs, labels = inputs.to(model.device), labels.to(model.device)
                batch_size, _, H, W = inputs.shape
                if 'class-il' not in model.COMPATIBILITY:
                    output = model(inputs, k)
                else:
                    y_0 = torch.ones(batch_size,  dataset.N_CLASSES).to(inputs.device) /dataset.N_CLASSES
                    z = model.net.f1(inputs)
                    output, z1 = model.net.f2(z, y_0)
                softmax = F.softmax(output, dim=1)
                _, pred_cls = softmax.max(1)

                val_log['correct'].append(pred_cls.cpu().eq(labels.cpu().data.view_as(pred_cls)).numpy())
                val_log['softmax'].append(softmax.cpu().data.numpy())
                val_log['logit'].append(output.cpu().data.numpy())
                val_log['target'].append(labels.cpu().data.numpy())
                _, pred = torch.max(output.data, 1)
                correct += torch.sum(pred == labels).item()
                total += labels.shape[0]
                if dataset.SETTING == 'class-il':
                    mask_classes(output, dataset, k)
                    _, pred_masked = torch.max(output.data, 1)
                    correct_mask_classes += pred_masked.eq(labels).sum().item()
        for key in val_log : 
            val_log[key] = np.concatenate(val_log[key])
        acc = 100. * val_log['correct'].mean()
        ece = utils.metrics.calc_ece(val_log['softmax'], val_log['target'], bins=15)
        accs.append(correct / total * 100 if 'class-il' in model.COMPATIBILITY else 0)
        accs_mask_classes.append(correct_mask_classes / total * 100)
        ece_list.append(ece*100)

    model.net.train(status)

    # average
    mean_ece = np.mean(ece_list)
    # 
    print('evaluation acc:', accs)
    print(f'ECE per loader: {ece_list}')
    print(f'Mean ECE: {mean_ece:.4f}')

    return accs, accs_mask_classes



def mask_classes(outputs: torch.Tensor, dataset: ContinualDataset, k: int) -> None:
    """
    Given the output tensor, the dataset at hand and the current task,
    masks the former by setting the responses for the other tasks at -inf.
    It is used to obtain the results for the task-il setting.
    :param outputs: the output tensor
    :param dataset: the continual dataset
    :param k: the task index
    """
    outputs[:, 0:k * dataset.N_CLASSES_PER_TASK] = -float('inf')
    outputs[:, (k + 1) * dataset.N_CLASSES_PER_TASK:
            dataset.N_TASKS * dataset.N_CLASSES_PER_TASK] = -float('inf')


def evaluate(model: ContinualModel, dataset: ContinualDataset, last=False) -> Tuple[list, list]:
    """
    Evaluates the accuracy of the model for each past task.
    :param model: the model to be evaluated
    :param dataset: the continual dataset at hand
    :return: a tuple of lists, containing the class-il
             and task-il accuracy for each task
    """
    status = model.net.training
    model.net.eval()
    accs, accs_mask_classes = [], []
    for k, test_loader in enumerate(dataset.test_loaders):
        if last and k < len(dataset.test_loaders) - 1:
            continue
        correct, correct_mask_classes, total = 0.0, 0.0, 0.0
        for data in test_loader:
            with torch.no_grad():
                inputs, labels = data
                inputs, labels = inputs.to(model.device), labels.to(model.device)
                if 'class-il' not in model.COMPATIBILITY:
                    outputs = model(inputs, k)
                else:
                    outputs = model(inputs)

                _, pred = torch.max(outputs.data, 1)
                correct += torch.sum(pred == labels).item()
                total += labels.shape[0]

                if dataset.SETTING == 'class-il':
                    mask_classes(outputs, dataset, k)
                    _, pred = torch.max(outputs.data, 1)
                    correct_mask_classes += torch.sum(pred == labels).item()

        accs.append(correct / total * 100
                    if 'class-il' in model.COMPATIBILITY else 0)
        accs_mask_classes.append(correct_mask_classes / total * 100)

    model.net.train(status)
    print('evaluation acc:')
    print(accs)
    return accs, accs_mask_classes


def evaluateid(model: ContinualModel, dataset: ContinualDataset, last=False) -> Tuple[list, list]:
    """
    Evaluates the accuracy of the model for each past task.
    :param model: the model to be evaluated
    :param dataset: the continual dataset at hand
    :return: a tuple of lists, containing the class-il
             and task-il accuracy for each task
    """
    status = model.net.training
    model.net.eval()
    accs, accs_mask_classes = [], []
    for k, test_loader in enumerate(dataset.test_loaders):
        if last and k < len(dataset.test_loaders) - 1:
            continue
        correct, correct_mask_classes, total = 0.0, 0.0, 0.0
        for data in test_loader:
            with torch.no_grad():
                inputs, labels = data
                inputs, labels = inputs.to(model.device), labels.to(model.device)
           
                batch_size, _, H, W = inputs.shape
                if 'class-il' not in model.COMPATIBILITY:
                    outputs = model(inputs, k)
                else:
                    y_0 = torch.ones(batch_size,  dataset.N_CLASSES).to(inputs.device) /dataset.N_CLASSES
                    z = model.net.f1(inputs)
                    outputs, z1 = model.net.f2(z, y_0)



                _, pred = torch.max(outputs.data, 1)
                correct += torch.sum(pred == labels).item()
                total += labels.shape[0]

                if dataset.SETTING == 'class-il':
                    mask_classes(outputs, dataset, k)
                    _, pred = torch.max(outputs.data, 1)
                    correct_mask_classes += torch.sum(pred == labels).item()

        accs.append(correct / total * 100
                    if 'class-il' in model.COMPATIBILITY else 0)
        accs_mask_classes.append(correct_mask_classes / total * 100)

    model.net.train(status)
    print('evaluation acc:')
    print(accs)
    return accs, accs_mask_classes


def train(model: ContinualModel, dataset: ContinualDataset,
          args: Namespace) -> None:
    """
    The training process, including evaluations and loggers.
    :param model: the module to be trained
    :param dataset: the continual dataset at hand
    :param args: the arguments of the current execution
    """

    model.net.to(model.device)
    results, results_mask_classes = [], []
    

    if not args.disable_log and not args.debug:
        logger = MLFlowLogger(dataset.SETTING, dataset.NAME, model.NAME,
                              experiment_name=args.experiment_name, parent_run_id=args.parent_run_id, run_name=args.run_name)
        logger.log_args(args.__dict__)

    progress_bar = ProgressBar(verbose=not args.non_verbose)

    if not args.ignore_other_metrics and not args.debug:
        dataset_copy = get_dataset(args)
        for t in range(dataset.N_TASKS):
            model.net.train()
            _, _ = dataset_copy.get_data_loaders()
        if model.NAME != 'icarl' and model.NAME != 'pnn':
            if model.NAME =='ider':
                random_results_class, random_results_task = evaluateid(model, dataset_copy)
            else:
                random_results_class, random_results_task = evaluate(model, dataset_copy)

    if os.path.exists('old_model.pt'):
        os.remove('old_model.pt')
    if os.path.exists('net.pt'):
        os.remove('net.pt')

    print(file=sys.stderr)
    for t in range(dataset.N_TASKS):
        model.net.train()
        train_loader, test_loader = dataset.get_data_loaders()
        if hasattr(model, 'begin_task'):
            model.begin_task(dataset)
        if t and not args.ignore_other_metrics and not args.debug:
            if model.NAME =='ider':
                accs = evaluateid(model, dataset, last=True)
            else:
                accs = evaluate(model, dataset, last=True)
            results[t-1] = results[t-1] + accs[0]
            if dataset.SETTING == 'class-il':
                results_mask_classes[t-1] = results_mask_classes[t-1] + accs[1]
        scheduler = dataset.get_scheduler(model, args)
        for epoch in range(model.args.n_epochs):          
            if args.model == 'joint':
                continue
            for i, data in enumerate(train_loader):
                if args.debug and i > 3:
                    break
                if hasattr(dataset.train_loader.dataset, 'logits'):
                    inputs, labels, not_aug_inputs, logits = data
                    inputs = inputs.to(model.device)
                    labels = labels.to(model.device)
                    not_aug_inputs = not_aug_inputs.to(model.device)
                    logits = logits.to(model.device)
                    loss = model.meta_observe(inputs, labels, not_aug_inputs, logits)
                else:
                    inputs, labels, not_aug_inputs = data
                    inputs, labels = inputs.to(model.device), labels.to(
                        model.device)
                    not_aug_inputs = not_aug_inputs.to(model.device)
                    loss = model.meta_observe(inputs, labels, not_aug_inputs)
                assert not math.isnan(loss)
                progress_bar.prog(i, len(train_loader), epoch, t, loss)

            if scheduler is not None:
                scheduler.step()
            if hasattr(model, 'end_epoch'):
                model.end_epoch(dataset)


        if hasattr(model, 'end_task'):
            model.end_task(dataset)

        if model.NAME =='ider':
            accs = evaluateid(model, dataset)
        else:
            accs = evaluate(model, dataset)
        results.append(accs[0])
        results_mask_classes.append(accs[1])
        if model.NAME =='ider':
            eces= evaluate_eceid(model, dataset)
        else: 
            eces= evaluate_ece(model, dataset)
        
        mean_acc = np.mean(accs, axis=1)
        print_mean_accuracy(mean_acc, t + 1, dataset.SETTING)

        if not args.disable_log and not args.debug:
            logger.log(mean_acc)
            logger.log_fullacc(accs)
  
    if not args.disable_log and not args.ignore_other_metrics and not args.debug:
        logger.add_bwt(results, results_mask_classes)
        logger.add_forgetting(results, results_mask_classes)
        if model.NAME != 'icarl' and model.NAME != 'pnn':
            logger.add_fwt(results, random_results_class,
                           results_mask_classes, random_results_task)

    if args.savecheckpoint:
        dataset_name = args.dataset if hasattr(args, 'dataset') and args.dataset else 'unknown_dataset'
        buffer_tag = f"buffer_{args.buffer_size}" if hasattr(args, 'buffer_size') and args.buffer_size is not None else "buffer_none"
        save_dir = os.path.join("./experiments", dataset_name, buffer_tag)
        os.makedirs(save_dir, exist_ok=True)

        model_filename = f"{model.NAME}_seed_{args.seed}.pth"
        model_path = os.path.join(save_dir, model_filename)

        torch.save(model.net.state_dict(), model_path)
        print(f"Model saved to: {model_path}")