File size: 34,513 Bytes
ecf08bc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
#    Copyright 2020 Division of Medical Image Computing, German Cancer Research Center (DKFZ), Heidelberg, Germany
#
#    Licensed under the Apache License, Version 2.0 (the "License");
#    you may not use this file except in compliance with the License.
#    You may obtain a copy of the License at
#
#        http://www.apache.org/licenses/LICENSE-2.0
#
#    Unless required by applicable law or agreed to in writing, software
#    distributed under the License is distributed on an "AS IS" BASIS,
#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#    See the License for the specific language governing permissions and
#    limitations under the License.
import os
import shutil
from _warnings import warn
from collections import OrderedDict
from multiprocessing import Pool
from time import sleep, time
from typing import Tuple

import numpy as np
import torch
import torch.distributed as dist
from batchgenerators.utilities.file_and_folder_operations import maybe_mkdir_p, join, subfiles, isfile, load_pickle, \
    save_json
from nnunet.configuration import default_num_threads
from nnunet.evaluation.evaluator import aggregate_scores
from nnunet.inference.segmentation_export import save_segmentation_nifti_from_softmax
from nnunet.network_architecture.neural_network import SegmentationNetwork
from nnunet.postprocessing.connected_components import determine_postprocessing
from nnunet.training.data_augmentation.data_augmentation_moreDA import get_moreDA_augmentation
from nnunet.training.dataloading.dataset_loading import unpack_dataset
from nnunet.training.loss_functions.crossentropy import RobustCrossEntropyLoss
from nnunet.training.loss_functions.dice_loss import get_tp_fp_fn_tn
from nnunet.training.network_training.nnUNetTrainerV2 import nnUNetTrainerV2
from nnunet.utilities.distributed import awesome_allgather_function
from nnunet.utilities.nd_softmax import softmax_helper
from nnunet.utilities.tensor_utilities import sum_tensor
from nnunet.utilities.to_torch import to_cuda, maybe_to_torch
from torch import nn, distributed
from torch.backends import cudnn
from torch.cuda.amp import autocast
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.optim.lr_scheduler import _LRScheduler
from tqdm import trange


class nnUNetTrainerV2_DDP(nnUNetTrainerV2):
    def __init__(self, plans_file, fold, local_rank, output_folder=None, dataset_directory=None, batch_dice=True,
                 stage=None,
                 unpack_data=True, deterministic=True, distribute_batch_size=False, fp16=False):
        super().__init__(plans_file, fold, output_folder, dataset_directory, batch_dice, stage,
                         unpack_data, deterministic, fp16)
        self.init_args = (
            plans_file, fold, local_rank, output_folder, dataset_directory, batch_dice, stage, unpack_data,
            deterministic, distribute_batch_size, fp16)
        self.distribute_batch_size = distribute_batch_size
        np.random.seed(local_rank)
        torch.manual_seed(local_rank)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(local_rank)
        self.local_rank = local_rank

        if torch.cuda.is_available():
            torch.cuda.set_device(local_rank)
        dist.init_process_group(backend='nccl', init_method='env://')

        self.loss = None
        self.ce_loss = RobustCrossEntropyLoss()

        self.global_batch_size = None  # we need to know this to properly steer oversample

    def set_batch_size_and_oversample(self):
        batch_sizes = []
        oversample_percents = []

        world_size = dist.get_world_size()
        my_rank = dist.get_rank()

        if self.distribute_batch_size:
            self.global_batch_size = self.batch_size
        else:
            self.global_batch_size = self.batch_size * world_size

        batch_size_per_GPU = np.ceil(self.batch_size / world_size).astype(int)

        for rank in range(world_size):
            if self.distribute_batch_size:
                if (rank + 1) * batch_size_per_GPU > self.batch_size:
                    batch_size = batch_size_per_GPU - ((rank + 1) * batch_size_per_GPU - self.batch_size)
                else:
                    batch_size = batch_size_per_GPU
            else:
                batch_size = self.batch_size

            batch_sizes.append(batch_size)

            sample_id_low = 0 if len(batch_sizes) == 0 else np.sum(batch_sizes[:-1])
            sample_id_high = np.sum(batch_sizes)

            if sample_id_high / self.global_batch_size < (1 - self.oversample_foreground_percent):
                oversample_percents.append(0.0)
            elif sample_id_low / self.global_batch_size > (1 - self.oversample_foreground_percent):
                oversample_percents.append(1.0)
            else:
                percent_covered_by_this_rank = sample_id_high / self.global_batch_size - sample_id_low / self.global_batch_size
                oversample_percent_here = 1 - (((1 - self.oversample_foreground_percent) -
                                                sample_id_low / self.global_batch_size) / percent_covered_by_this_rank)
                oversample_percents.append(oversample_percent_here)

        print("worker", my_rank, "oversample", oversample_percents[my_rank])
        print("worker", my_rank, "batch_size", batch_sizes[my_rank])

        self.batch_size = batch_sizes[my_rank]
        self.oversample_foreground_percent = oversample_percents[my_rank]

    def save_checkpoint(self, fname, save_optimizer=True):
        if self.local_rank == 0:
            super().save_checkpoint(fname, save_optimizer)

    def plot_progress(self):
        if self.local_rank == 0:
            super().plot_progress()

    def print_to_log_file(self, *args, also_print_to_console=True):
        if self.local_rank == 0:
            super().print_to_log_file(*args, also_print_to_console=also_print_to_console)

    def process_plans(self, plans):
        super().process_plans(plans)
        self.set_batch_size_and_oversample()

    def initialize(self, training=True, force_load_plans=False):
        """
        :param training:
        :return:
        """
        if not self.was_initialized:
            maybe_mkdir_p(self.output_folder)

            if force_load_plans or (self.plans is None):
                self.load_plans_file()

            self.process_plans(self.plans)

            self.setup_DA_params()

            self.folder_with_preprocessed_data = join(self.dataset_directory, self.plans['data_identifier'] +
                                                      "_stage%d" % self.stage)
            if training:
                self.dl_tr, self.dl_val = self.get_basic_generators()
                if self.unpack_data:
                    if self.local_rank == 0:
                        print("unpacking dataset")
                        unpack_dataset(self.folder_with_preprocessed_data)
                        print("done")
                    distributed.barrier()
                else:
                    print(
                        "INFO: Not unpacking data! Training may be slow due to that. Pray you are not using 2d or you "
                        "will wait all winter for your model to finish!")

                # setting weights for deep supervision losses
                net_numpool = len(self.net_num_pool_op_kernel_sizes)

                # we give each output a weight which decreases exponentially (division by 2) as the resolution decreases
                # this gives higher resolution outputs more weight in the loss
                weights = np.array([1 / (2 ** i) for i in range(net_numpool)])

                # we don't use the lowest 2 outputs. Normalize weights so that they sum to 1
                mask = np.array([True if i < net_numpool - 1 else False for i in range(net_numpool)])
                weights[~mask] = 0
                weights = weights / weights.sum()
                self.ds_loss_weights = weights

                seeds_train = np.random.random_integers(0, 99999, self.data_aug_params.get('num_threads'))
                seeds_val = np.random.random_integers(0, 99999, max(self.data_aug_params.get('num_threads') // 2, 1))
                print("seeds train", seeds_train)
                print("seeds_val", seeds_val)
                self.tr_gen, self.val_gen = get_moreDA_augmentation(self.dl_tr, self.dl_val,
                                                                    self.data_aug_params[
                                                                        'patch_size_for_spatialtransform'],
                                                                    self.data_aug_params,
                                                                    deep_supervision_scales=self.deep_supervision_scales,
                                                                    seeds_train=seeds_train,
                                                                    seeds_val=seeds_val,
                                                                    pin_memory=self.pin_memory)
                self.print_to_log_file("TRAINING KEYS:\n %s" % (str(self.dataset_tr.keys())),
                                       also_print_to_console=False)
                self.print_to_log_file("VALIDATION KEYS:\n %s" % (str(self.dataset_val.keys())),
                                       also_print_to_console=False)
            else:
                pass

            self.initialize_network()
            self.initialize_optimizer_and_scheduler()
            self.network = DDP(self.network, device_ids=[self.local_rank])

        else:
            self.print_to_log_file('self.was_initialized is True, not running self.initialize again')
        self.was_initialized = True

    def run_iteration(self, data_generator, do_backprop=True, run_online_evaluation=False):
        data_dict = next(data_generator)
        data = data_dict['data']
        target = data_dict['target']

        data = maybe_to_torch(data)
        target = maybe_to_torch(target)

        if torch.cuda.is_available():
            data = to_cuda(data, gpu_id=None)
            target = to_cuda(target, gpu_id=None)

        self.optimizer.zero_grad()

        if self.fp16:
            with autocast():
                output = self.network(data)
                del data
                l = self.compute_loss(output, target)

            if do_backprop:
                self.amp_grad_scaler.scale(l).backward()
                self.amp_grad_scaler.unscale_(self.optimizer)
                torch.nn.utils.clip_grad_norm_(self.network.parameters(), 12)
                self.amp_grad_scaler.step(self.optimizer)
                self.amp_grad_scaler.update()
        else:
            output = self.network(data)
            del data
            l = self.compute_loss(output, target)

            if do_backprop:
                l.backward()
                torch.nn.utils.clip_grad_norm_(self.network.parameters(), 12)
                self.optimizer.step()

        if run_online_evaluation:
            self.run_online_evaluation(output, target)

        del target

        return l.detach().cpu().numpy()

    def compute_loss(self, output, target):
        total_loss = None
        for i in range(len(output)):
            # Starting here it gets spicy!
            axes = tuple(range(2, len(output[i].size())))

            # network does not do softmax. We need to do softmax for dice
            output_softmax = softmax_helper(output[i])

            # get the tp, fp and fn terms we need
            tp, fp, fn, _ = get_tp_fp_fn_tn(output_softmax, target[i], axes, mask=None)
            # for dice, compute nominator and denominator so that we have to accumulate only 2 instead of 3 variables
            # do_bg=False in nnUNetTrainer -> [:, 1:]
            nominator = 2 * tp[:, 1:]
            denominator = 2 * tp[:, 1:] + fp[:, 1:] + fn[:, 1:]

            if self.batch_dice:
                # for DDP we need to gather all nominator and denominator terms from all GPUS to do proper batch dice
                nominator = awesome_allgather_function.apply(nominator)
                denominator = awesome_allgather_function.apply(denominator)
                nominator = nominator.sum(0)
                denominator = denominator.sum(0)
            else:
                pass

            ce_loss = self.ce_loss(output[i], target[i][:, 0].long())

            # we smooth by 1e-5 to penalize false positives if tp is 0
            dice_loss = (- (nominator + 1e-5) / (denominator + 1e-5)).mean()
            if total_loss is None:
                total_loss = self.ds_loss_weights[i] * (ce_loss + dice_loss)
            else:
                total_loss += self.ds_loss_weights[i] * (ce_loss + dice_loss)
        return total_loss

    def run_online_evaluation(self, output, target):
        with torch.no_grad():
            num_classes = output[0].shape[1]
            output_seg = output[0].argmax(1)
            target = target[0][:, 0]
            axes = tuple(range(1, len(target.shape)))
            tp_hard = torch.zeros((target.shape[0], num_classes - 1)).to(output_seg.device.index)
            fp_hard = torch.zeros((target.shape[0], num_classes - 1)).to(output_seg.device.index)
            fn_hard = torch.zeros((target.shape[0], num_classes - 1)).to(output_seg.device.index)
            for c in range(1, num_classes):
                tp_hard[:, c - 1] = sum_tensor((output_seg == c).float() * (target == c).float(), axes=axes)
                fp_hard[:, c - 1] = sum_tensor((output_seg == c).float() * (target != c).float(), axes=axes)
                fn_hard[:, c - 1] = sum_tensor((output_seg != c).float() * (target == c).float(), axes=axes)

            # tp_hard, fp_hard, fn_hard = get_tp_fp_fn((output_softmax > (1 / num_classes)).float(), target,
            #                                         axes, None)
            # print_if_rank0("before allgather", tp_hard.shape)
            tp_hard = tp_hard.sum(0, keepdim=False)[None]
            fp_hard = fp_hard.sum(0, keepdim=False)[None]
            fn_hard = fn_hard.sum(0, keepdim=False)[None]

            tp_hard = awesome_allgather_function.apply(tp_hard)
            fp_hard = awesome_allgather_function.apply(fp_hard)
            fn_hard = awesome_allgather_function.apply(fn_hard)

        tp_hard = tp_hard.detach().cpu().numpy().sum(0)
        fp_hard = fp_hard.detach().cpu().numpy().sum(0)
        fn_hard = fn_hard.detach().cpu().numpy().sum(0)
        self.online_eval_foreground_dc.append(list((2 * tp_hard) / (2 * tp_hard + fp_hard + fn_hard + 1e-8)))
        self.online_eval_tp.append(list(tp_hard))
        self.online_eval_fp.append(list(fp_hard))
        self.online_eval_fn.append(list(fn_hard))

    def run_training(self):
        """
        if we run with -c then we need to set the correct lr for the first epoch, otherwise it will run the first
        continued epoch with self.initial_lr

        we also need to make sure deep supervision in the network is enabled for training, thus the wrapper
        :return:
        """
        if self.local_rank == 0:
            self.save_debug_information()

        if not torch.cuda.is_available():
            self.print_to_log_file("WARNING!!! You are attempting to run training on a CPU (torch.cuda.is_available() is False). This can be VERY slow!")

        self.maybe_update_lr(self.epoch)  # if we dont overwrite epoch then self.epoch+1 is used which is not what we
        # want at the start of the training
        if isinstance(self.network, DDP):
            net = self.network.module
        else:
            net = self.network
        ds = net.do_ds
        net.do_ds = True

        _ = self.tr_gen.next()
        _ = self.val_gen.next()

        if torch.cuda.is_available():
            torch.cuda.empty_cache()

        self._maybe_init_amp()

        maybe_mkdir_p(self.output_folder)
        self.plot_network_architecture()

        if cudnn.benchmark and cudnn.deterministic:
            warn("torch.backends.cudnn.deterministic is True indicating a deterministic training is desired. "
                 "But torch.backends.cudnn.benchmark is True as well and this will prevent deterministic training! "
                 "If you want deterministic then set benchmark=False")

        if not self.was_initialized:
            self.initialize(True)

        while self.epoch < self.max_num_epochs:
            self.print_to_log_file("\nepoch: ", self.epoch)
            epoch_start_time = time()
            train_losses_epoch = []

            # train one epoch
            self.network.train()

            if self.use_progress_bar:
                with trange(self.num_batches_per_epoch) as tbar:
                    for b in tbar:
                        tbar.set_description("Epoch {}/{}".format(self.epoch+1, self.max_num_epochs))

                        l = self.run_iteration(self.tr_gen, True)

                        tbar.set_postfix(loss=l)
                        train_losses_epoch.append(l)
            else:
                for _ in range(self.num_batches_per_epoch):
                    l = self.run_iteration(self.tr_gen, True)
                    train_losses_epoch.append(l)

            self.all_tr_losses.append(np.mean(train_losses_epoch))
            self.print_to_log_file("train loss : %.4f" % self.all_tr_losses[-1])

            with torch.no_grad():
                # validation with train=False
                self.network.eval()
                val_losses = []
                for b in range(self.num_val_batches_per_epoch):
                    l = self.run_iteration(self.val_gen, False, True)
                    val_losses.append(l)
                self.all_val_losses.append(np.mean(val_losses))
                self.print_to_log_file("validation loss: %.4f" % self.all_val_losses[-1])

                if self.also_val_in_tr_mode:
                    self.network.train()
                    # validation with train=True
                    val_losses = []
                    for b in range(self.num_val_batches_per_epoch):
                        l = self.run_iteration(self.val_gen, False)
                        val_losses.append(l)
                    self.all_val_losses_tr_mode.append(np.mean(val_losses))
                    self.print_to_log_file("validation loss (train=True): %.4f" % self.all_val_losses_tr_mode[-1])

            self.update_train_loss_MA()  # needed for lr scheduler and stopping of training

            continue_training = self.on_epoch_end()

            epoch_end_time = time()

            if not continue_training:
                # allows for early stopping
                break

            self.epoch += 1
            self.print_to_log_file("This epoch took %f s\n" % (epoch_end_time - epoch_start_time))

        self.epoch -= 1  # if we don't do this we can get a problem with loading model_final_checkpoint.

        if self.save_final_checkpoint: self.save_checkpoint(join(self.output_folder, "model_final_checkpoint.model"))

        if self.local_rank == 0:
            # now we can delete latest as it will be identical with final
            if isfile(join(self.output_folder, "model_latest.model")):
                os.remove(join(self.output_folder, "model_latest.model"))
            if isfile(join(self.output_folder, "model_latest.model.pkl")):
                os.remove(join(self.output_folder, "model_latest.model.pkl"))

        net.do_ds = ds

    def validate(self, do_mirroring: bool = True, use_sliding_window: bool = True,
                 step_size: float = 0.5, save_softmax: bool = True, use_gaussian: bool = True, overwrite: bool = True,
                 validation_folder_name: str = 'validation_raw', debug: bool = False, all_in_gpu: bool = False,
                 segmentation_export_kwargs: dict = None, run_postprocessing_on_folds: bool = True):
        if isinstance(self.network, DDP):
            net = self.network.module
        else:
            net = self.network
        ds = net.do_ds
        net.do_ds = False

        current_mode = self.network.training
        self.network.eval()

        assert self.was_initialized, "must initialize, ideally with checkpoint (or train first)"
        if self.dataset_val is None:
            self.load_dataset()
            self.do_split()

        if segmentation_export_kwargs is None:
            if 'segmentation_export_params' in self.plans.keys():
                force_separate_z = self.plans['segmentation_export_params']['force_separate_z']
                interpolation_order = self.plans['segmentation_export_params']['interpolation_order']
                interpolation_order_z = self.plans['segmentation_export_params']['interpolation_order_z']
            else:
                force_separate_z = None
                interpolation_order = 1
                interpolation_order_z = 0
        else:
            force_separate_z = segmentation_export_kwargs['force_separate_z']
            interpolation_order = segmentation_export_kwargs['interpolation_order']
            interpolation_order_z = segmentation_export_kwargs['interpolation_order_z']

        # predictions as they come from the network go here
        output_folder = join(self.output_folder, validation_folder_name)
        maybe_mkdir_p(output_folder)
        # this is for debug purposes
        my_input_args = {'do_mirroring': do_mirroring,
                         'use_sliding_window': use_sliding_window,
                         'step_size': step_size,
                         'save_softmax': save_softmax,
                         'use_gaussian': use_gaussian,
                         'overwrite': overwrite,
                         'validation_folder_name': validation_folder_name,
                         'debug': debug,
                         'all_in_gpu': all_in_gpu,
                         'segmentation_export_kwargs': segmentation_export_kwargs,
                         }
        save_json(my_input_args, join(output_folder, "validation_args.json"))

        if do_mirroring:
            if not self.data_aug_params['do_mirror']:
                raise RuntimeError(
                    "We did not train with mirroring so you cannot do inference with mirroring enabled")
            mirror_axes = self.data_aug_params['mirror_axes']
        else:
            mirror_axes = ()

        pred_gt_tuples = []

        export_pool = Pool(default_num_threads)
        results = []

        all_keys = list(self.dataset_val.keys())
        my_keys = all_keys[self.local_rank::dist.get_world_size()]
        # we cannot simply iterate over all_keys because we need to know pred_gt_tuples and valid_labels of all cases
        # for evaluation (which is done by local rank 0)
        for k in my_keys:
            properties = load_pickle(self.dataset[k]['properties_file'])
            fname = properties['list_of_data_files'][0].split("/")[-1][:-12]
            pred_gt_tuples.append([join(output_folder, fname + ".nii.gz"),
                                   join(self.gt_niftis_folder, fname + ".nii.gz")])
            if k in my_keys:
                if overwrite or (not isfile(join(output_folder, fname + ".nii.gz"))) or \
                        (save_softmax and not isfile(join(output_folder, fname + ".npz"))):
                    data = np.load(self.dataset[k]['data_file'])['data']

                    print(k, data.shape)
                    data[-1][data[-1] == -1] = 0

                    softmax_pred = self.predict_preprocessed_data_return_seg_and_softmax(data[:-1],
                                                                                         do_mirroring=do_mirroring,
                                                                                         mirror_axes=mirror_axes,
                                                                                         use_sliding_window=use_sliding_window,
                                                                                         step_size=step_size,
                                                                                         use_gaussian=use_gaussian,
                                                                                         all_in_gpu=all_in_gpu,
                                                                                         mixed_precision=self.fp16)[1]

                    softmax_pred = softmax_pred.transpose([0] + [i + 1 for i in self.transpose_backward])

                    if save_softmax:
                        softmax_fname = join(output_folder, fname + ".npz")
                    else:
                        softmax_fname = None

                    """There is a problem with python process communication that prevents us from communicating obejcts
                    larger than 2 GB between processes (basically when the length of the pickle string that will be sent is
                    communicated by the multiprocessing.Pipe object then the placeholder (\%i I think) does not allow for long
                    enough strings (lol). This could be fixed by changing i to l (for long) but that would require manually
                    patching system python code. We circumvent that problem here by saving softmax_pred to a npy file that will
                    then be read (and finally deleted) by the Process. save_segmentation_nifti_from_softmax can take either
                    filename or np.ndarray and will handle this automatically"""
                    if np.prod(softmax_pred.shape) > (2e9 / 4 * 0.85):  # *0.85 just to be save
                        np.save(join(output_folder, fname + ".npy"), softmax_pred)
                        softmax_pred = join(output_folder, fname + ".npy")

                    results.append(export_pool.starmap_async(save_segmentation_nifti_from_softmax,
                                                             ((softmax_pred, join(output_folder, fname + ".nii.gz"),
                                                               properties, interpolation_order,
                                                               self.regions_class_order,
                                                               None, None,
                                                               softmax_fname, None, force_separate_z,
                                                               interpolation_order_z),
                                                              )
                                                             )
                                   )

        _ = [i.get() for i in results]
        self.print_to_log_file("finished prediction")

        distributed.barrier()

        if self.local_rank == 0:
            # evaluate raw predictions
            self.print_to_log_file("evaluation of raw predictions")
            task = self.dataset_directory.split("/")[-1]
            job_name = self.experiment_name
            _ = aggregate_scores(pred_gt_tuples, labels=list(range(self.num_classes)),
                                 json_output_file=join(output_folder, "summary.json"),
                                 json_name=job_name + " val tiled %s" % (str(use_sliding_window)),
                                 json_author="Fabian",
                                 json_task=task, num_threads=default_num_threads)

            if run_postprocessing_on_folds:
                # in the old nnunet we would stop here. Now we add a postprocessing. This postprocessing can remove everything
                # except the largest connected component for each class. To see if this improves results, we do this for all
                # classes and then rerun the evaluation. Those classes for which this resulted in an improved dice score will
                # have this applied during inference as well
                self.print_to_log_file("determining postprocessing")
                determine_postprocessing(self.output_folder, self.gt_niftis_folder, validation_folder_name,
                                         final_subf_name=validation_folder_name + "_postprocessed", debug=debug)
                # after this the final predictions for the vlaidation set can be found in validation_folder_name_base + "_postprocessed"
                # They are always in that folder, even if no postprocessing as applied!

            # detemining postprocesing on a per-fold basis may be OK for this fold but what if another fold finds another
            # postprocesing to be better? In this case we need to consolidate. At the time the consolidation is going to be
            # done we won't know what self.gt_niftis_folder was, so now we copy all the niftis into a separate folder to
            # be used later
            gt_nifti_folder = join(self.output_folder_base, "gt_niftis")
            maybe_mkdir_p(gt_nifti_folder)
            for f in subfiles(self.gt_niftis_folder, suffix=".nii.gz"):
                success = False
                attempts = 0
                e = None
                while not success and attempts < 10:
                    try:
                        shutil.copy(f, gt_nifti_folder)
                        success = True
                    except OSError as e:
                        attempts += 1
                        sleep(1)
                if not success:
                    print("Could not copy gt nifti file %s into folder %s" % (f, gt_nifti_folder))
                    if e is not None:
                        raise e

        self.network.train(current_mode)
        net.do_ds = ds

    def predict_preprocessed_data_return_seg_and_softmax(self, data: np.ndarray, do_mirroring: bool = True,
                                                         mirror_axes: Tuple[int] = None,
                                                         use_sliding_window: bool = True, step_size: float = 0.5,
                                                         use_gaussian: bool = True, pad_border_mode: str = 'constant',
                                                         pad_kwargs: dict = None, all_in_gpu: bool = False,
                                                         verbose: bool = True, mixed_precision=True) -> Tuple[
        np.ndarray, np.ndarray]:
        if pad_border_mode == 'constant' and pad_kwargs is None:
            pad_kwargs = {'constant_values': 0}

        if do_mirroring and mirror_axes is None:
            mirror_axes = self.data_aug_params['mirror_axes']

        if do_mirroring:
            assert self.data_aug_params["do_mirror"], "Cannot do mirroring as test time augmentation when training " \
                                                      "was done without mirroring"

        valid = list((SegmentationNetwork, nn.DataParallel, DDP))
        assert isinstance(self.network, tuple(valid))
        if isinstance(self.network, DDP):
            net = self.network.module
        else:
            net = self.network
        ds = net.do_ds
        net.do_ds = False
        ret = net.predict_3D(data, do_mirroring=do_mirroring, mirror_axes=mirror_axes,
                             use_sliding_window=use_sliding_window, step_size=step_size,
                             patch_size=self.patch_size, regions_class_order=self.regions_class_order,
                             use_gaussian=use_gaussian, pad_border_mode=pad_border_mode,
                             pad_kwargs=pad_kwargs, all_in_gpu=all_in_gpu, verbose=verbose,
                             mixed_precision=mixed_precision)
        net.do_ds = ds
        return ret

    def load_checkpoint_ram(self, checkpoint, train=True):
        """
        used for if the checkpoint is already in ram
        :param checkpoint:
        :param train:
        :return:
        """
        if not self.was_initialized:
            self.initialize(train)

        new_state_dict = OrderedDict()
        curr_state_dict_keys = list(self.network.state_dict().keys())
        # if state dict comes form nn.DataParallel but we use non-parallel model here then the state dict keys do not
        # match. Use heuristic to make it match
        for k, value in checkpoint['state_dict'].items():
            key = k
            if key not in curr_state_dict_keys:
                print("duh")
                key = key[7:]
            new_state_dict[key] = value

        if self.fp16:
            self._maybe_init_amp()
            if 'amp_grad_scaler' in checkpoint.keys():
                self.amp_grad_scaler.load_state_dict(checkpoint['amp_grad_scaler'])

        self.network.load_state_dict(new_state_dict)
        self.epoch = checkpoint['epoch']
        if train:
            optimizer_state_dict = checkpoint['optimizer_state_dict']
            if optimizer_state_dict is not None:
                self.optimizer.load_state_dict(optimizer_state_dict)

            if self.lr_scheduler is not None and hasattr(self.lr_scheduler, 'load_state_dict') and checkpoint[
                'lr_scheduler_state_dict'] is not None:
                self.lr_scheduler.load_state_dict(checkpoint['lr_scheduler_state_dict'])

            if issubclass(self.lr_scheduler.__class__, _LRScheduler):
                self.lr_scheduler.step(self.epoch)

        self.all_tr_losses, self.all_val_losses, self.all_val_losses_tr_mode, self.all_val_eval_metrics = checkpoint[
            'plot_stuff']

        # after the training is done, the epoch is incremented one more time in my old code. This results in
        # self.epoch = 1001 for old trained models when the epoch is actually 1000. This causes issues because
        # len(self.all_tr_losses) = 1000 and the plot function will fail. We can easily detect and correct that here
        if self.epoch != len(self.all_tr_losses):
            self.print_to_log_file("WARNING in loading checkpoint: self.epoch != len(self.all_tr_losses). This is "
                                   "due to an old bug and should only appear when you are loading old models. New "
                                   "models should have this fixed! self.epoch is now set to len(self.all_tr_losses)")
            self.epoch = len(self.all_tr_losses)
            self.all_tr_losses = self.all_tr_losses[:self.epoch]
            self.all_val_losses = self.all_val_losses[:self.epoch]
            self.all_val_losses_tr_mode = self.all_val_losses_tr_mode[:self.epoch]
            self.all_val_eval_metrics = self.all_val_eval_metrics[:self.epoch]