File size: 21,065 Bytes
0e4c246
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torchvision.models as models
import torch.nn.functional as F
from torchvision.models.feature_extraction import get_graph_node_names
from torchvision.models.feature_extraction import create_feature_extractor
from typing import Union
import copy

class GCNCombiner(nn.Module):

    def __init__(self,
                 total_num_selects: int,
                 num_classes: int,
                 inputs: Union[dict, None] = None,
                 proj_size: Union[int, None] = None,
                 fpn_size: Union[int, None] = None):
        """
        If building backbone without FPN, set fpn_size to None and MUST give
        'inputs' and 'proj_size', the reason of these setting is to constrain the
        dimension of graph convolutional network input.
        """
        super(GCNCombiner, self).__init__()

        assert inputs is not None or fpn_size is not None, \
            "To build GCN combiner, you must give one features dimension."

        ### auto-proj
        self.fpn_size = fpn_size
        if fpn_size is None:
            for name in inputs:
                if len(name) == 4:
                    in_size = inputs[name].size(1)
                elif len(name) == 3:
                    in_size = inputs[name].size(2)
                else:
                    raise ValusError("The size of output dimension of previous must be 3 or 4.")
                m = nn.Sequential(
                    nn.Linear(in_size, proj_size),
                    nn.ReLU(),
                    nn.Linear(proj_size, proj_size)
                )
                self.add_module("proj_"+name, m)
            self.proj_size = proj_size
        else:
            self.proj_size = fpn_size

        ### build one layer structure (with adaptive module)
        num_joints = total_num_selects // 64

        self.param_pool0 = nn.Linear(total_num_selects, num_joints)

        A = torch.eye(num_joints) / 100 + 1 / 100
        self.adj1 = nn.Parameter(copy.deepcopy(A))
        self.conv1 = nn.Conv1d(self.proj_size, self.proj_size, 1)
        self.batch_norm1 = nn.BatchNorm1d(self.proj_size)

        self.conv_q1 = nn.Conv1d(self.proj_size, self.proj_size//4, 1)
        self.conv_k1 = nn.Conv1d(self.proj_size, self.proj_size//4, 1)
        self.alpha1 = nn.Parameter(torch.zeros(1))

        ### merge information
        self.param_pool1 = nn.Linear(num_joints, 1)

        #### class predict
        self.dropout = nn.Dropout(p=0.1)
        self.classifier = nn.Linear(self.proj_size, num_classes)

        self.tanh = nn.Tanh()

    def forward(self, x):
        """
        """
        hs = []
        names = []
        for name in x:
            if "FPN1_" in name:
                continue
            if self.fpn_size is None:
                _tmp = getattr(self, "proj_"+name)(x[name])
            else:
                _tmp = x[name]
            hs.append(_tmp)
            names.append([name, _tmp.size()])

        hs = torch.cat(hs, dim=1).transpose(1, 2).contiguous() # B, S', C --> B, C, S
        # print(hs.size(), names)
        hs = self.param_pool0(hs)
        ### adaptive adjacency
        q1 = self.conv_q1(hs).mean(1)
        k1 = self.conv_k1(hs).mean(1)
        A1 = self.tanh(q1.unsqueeze(-1) - k1.unsqueeze(1))
        A1 = self.adj1 + A1 * self.alpha1
        ### graph convolution
        hs = self.conv1(hs)
        hs = torch.matmul(hs, A1)
        hs = self.batch_norm1(hs)
        ### predict
        hs = self.param_pool1(hs)
        hs = self.dropout(hs)
        hs = hs.flatten(1)
        hs = self.classifier(hs)

        return hs

class WeaklySelector(nn.Module):

    def __init__(self, inputs: dict, num_classes: int, num_select: dict, fpn_size: Union[int, None] = None):
        """
        inputs: dictionary contain torch.Tensors, which comes from backbone
                [Tensor1(hidden feature1), Tensor2(hidden feature2)...]
                Please note that if len(features.size) equal to 3, the order of dimension must be [B,S,C],
                S mean the spatial domain, and if len(features.size) equal to 4, the order must be [B,C,H,W]
        """
        super(WeaklySelector, self).__init__()

        self.num_select = num_select

        self.fpn_size = fpn_size
        ### build classifier
        if self.fpn_size is None:
            self.num_classes = num_classes
            for name in inputs:
                fs_size = inputs[name].size()
                if len(fs_size) == 3:
                    in_size = fs_size[2]
                elif len(fs_size) == 4:
                    in_size = fs_size[1]
                m = nn.Linear(in_size, num_classes)
                self.add_module("classifier_l_"+name, m)

        self.thresholds = {}
        for name in inputs:
            self.thresholds[name] = []

    # def select(self, logits, l_name):
    #     """
    #     logits: [B, S, num_classes]
    #     """
    #     probs = torch.softmax(logits, dim=-1)
    #     scores, _ = torch.max(probs, dim=-1)
    #     _, ids = torch.sort(scores, -1, descending=True)
    #     sn = self.num_select[l_name]
    #     s_ids = ids[:, :sn]
    #     not_s_ids = ids[:, sn:]
    #     return s_ids.unsqueeze(-1), not_s_ids.unsqueeze(-1)

    def forward(self, x, logits=None):
        """
        x :
            dictionary contain the features maps which
            come from your choosen layers.
            size must be [B, HxW, C] ([B, S, C]) or [B, C, H, W].
            [B,C,H,W] will be transpose to [B, HxW, C] automatically.
        """
        if self.fpn_size is None:
            logits = {}
        selections = {}
        for name in x:
            # print("[selector]", name, x[name].size())
            if "FPN1_" in name:
                continue
            if len(x[name].size()) == 4:
                B, C, H, W = x[name].size()
                x[name] = x[name].view(B, C, H*W).permute(0, 2, 1).contiguous()
            C = x[name].size(-1)
            if self.fpn_size is None:
                logits[name] = getattr(self, "classifier_l_"+name)(x[name])

            probs = torch.softmax(logits[name], dim=-1)
            sum_probs = torch.softmax(logits[name].mean(1), dim=-1)
            selections[name] = []
            preds_1 = []
            preds_0 = []
            num_select = self.num_select[name]
            for bi in range(logits[name].size(0)):
                _, max_ids = torch.max(sum_probs[bi], dim=-1)
                confs, ranks = torch.sort(probs[bi, :, max_ids], descending=True)
                sf = x[name][bi][ranks[:num_select]]
                nf = x[name][bi][ranks[num_select:]]  # calculate
                selections[name].append(sf) # [num_selected, C]
                preds_1.append(logits[name][bi][ranks[:num_select]])
                preds_0.append(logits[name][bi][ranks[num_select:]])

                if bi >= len(self.thresholds[name]):
                    self.thresholds[name].append(confs[num_select]) # for initialize
                else:
                    self.thresholds[name][bi] = confs[num_select]

            selections[name] = torch.stack(selections[name])
            preds_1 = torch.stack(preds_1)
            preds_0 = torch.stack(preds_0)

            logits["select_"+name] = preds_1
            logits["drop_"+name] = preds_0

        return selections


class FPN(nn.Module):

    def __init__(self, inputs: dict, fpn_size: int, proj_type: str, upsample_type: str):
        """
        inputs : dictionary contains torch.Tensor
                 which comes from backbone output
        fpn_size: integer, fpn
        proj_type:
            in ["Conv", "Linear"]
        upsample_type:
            in ["Bilinear", "Conv", "Fc"]
            for convolution neural network (e.g. ResNet, EfficientNet), recommand 'Bilinear'.
            for Vit, "Fc". and Swin-T, "Conv"
        """
        super(FPN, self).__init__()
        assert proj_type in ["Conv", "Linear"], \
            "FPN projection type {} were not support yet, please choose type 'Conv' or 'Linear'".format(proj_type)
        assert upsample_type in ["Bilinear", "Conv"], \
            "FPN upsample type {} were not support yet, please choose type 'Bilinear' or 'Conv'".format(proj_type)

        self.fpn_size = fpn_size
        self.upsample_type = upsample_type
        inp_names = [name for name in inputs]

        for i, node_name in enumerate(inputs):
            ### projection module
            if proj_type == "Conv":
                m = nn.Sequential(
                    nn.Conv2d(inputs[node_name].size(1), inputs[node_name].size(1), 1),
                    nn.ReLU(),
                    nn.Conv2d(inputs[node_name].size(1), fpn_size, 1)
                )
            elif proj_type == "Linear":
                in_feat = inputs[node_name]
                if isinstance(in_feat, torch.Tensor):
                    dim = in_feat.size(-1)
                else:
                    raise ValueError(f"Entrée invalide dans FPN: {type(in_feat)} pour node_name={node_name}")

                m = nn.Sequential(
                    nn.Linear(dim, dim),
                    nn.ReLU(),
                    nn.Linear(dim, fpn_size),
                )

            self.add_module("Proj_"+node_name, m)

            ### upsample module
            if upsample_type == "Conv" and i != 0:
                assert len(inputs[node_name].size()) == 3 # B, S, C
                in_dim = inputs[node_name].size(1)
                out_dim = inputs[inp_names[i-1]].size(1)
                # if in_dim != out_dim:
                m = nn.Conv1d(in_dim, out_dim, 1) # for spatial domain
                # else:
                #     m = nn.Identity()
                self.add_module("Up_"+node_name, m)

        if upsample_type == "Bilinear":
            self.upsample = nn.Upsample(scale_factor=2, mode='bilinear')

    def upsample_add(self, x0: torch.Tensor, x1: torch.Tensor, x1_name: str):
        """
        return Upsample(x1) + x1
        """
        if self.upsample_type == "Bilinear":
            if x1.size(-1) != x0.size(-1):
                x1 = self.upsample(x1)
        else:
            x1 = getattr(self, "Up_"+x1_name)(x1)
        return x1 + x0

    def forward(self, x):
        """
        x : dictionary
            {
                "node_name1": feature1,
                "node_name2": feature2, ...
            }
        """
        ### project to same dimension
        hs = []
        for i, name in enumerate(x):
            if "FPN1_" in name:
                continue
            x[name] = getattr(self, "Proj_"+name)(x[name])
            hs.append(name)

        x["FPN1_" + "layer4"] = x["layer4"]

        for i in range(len(hs)-1, 0, -1):
            x1_name = hs[i]
            x0_name = hs[i-1]
            x[x0_name] = self.upsample_add(x[x0_name],
                                           x[x1_name],
                                           x1_name)
            x["FPN1_" + x0_name] = x[x0_name]

        return x


class FPN_UP(nn.Module):

    def __init__(self,
                 inputs: dict,
                 fpn_size: int):
        super(FPN_UP, self).__init__()

        inp_names = [name for name in inputs]

        for i, node_name in enumerate(inputs):
            ### projection module
            m = nn.Sequential(
                nn.Linear(fpn_size, fpn_size),
                nn.ReLU(),
                nn.Linear(fpn_size, fpn_size),
            )
            self.add_module("Proj_"+node_name, m)

            ### upsample module
            if i != (len(inputs) - 1):
                assert len(inputs[node_name].size()) == 3 # B, S, C
                in_dim = inputs[node_name].size(1)
                out_dim = inputs[inp_names[i+1]].size(1)
                m = nn.Conv1d(in_dim, out_dim, 1) # for spatial domain
                self.add_module("Down_"+node_name, m)
                # print("Down_"+node_name, in_dim, out_dim)
                """
                Down_layer1 2304 576
                Down_layer2 576 144
                Down_layer3 144 144
                """

    def downsample_add(self, x0: torch.Tensor, x1: torch.Tensor, x0_name: str):
        """
        return Upsample(x1) + x1
        """
        # print("[downsample_add] Down_" + x0_name)
        x0 = getattr(self, "Down_" + x0_name)(x0)
        return x1 + x0

    def forward(self, x):
        """
        x : dictionary
            {
                "node_name1": feature1,
                "node_name2": feature2, ...
            }
        """
        ### project to same dimension
        hs = []
        for i, name in enumerate(x):
            if "FPN1_" in name:
                continue
            x[name] = getattr(self, "Proj_"+name)(x[name])
            hs.append(name)

        # print(hs)
        for i in range(0, len(hs) - 1):
            x0_name = hs[i]
            x1_name = hs[i+1]
            # print(x0_name, x1_name)
            # print(x[x0_name].size(), x[x1_name].size())
            x[x1_name] = self.downsample_add(x[x0_name],
                                             x[x1_name],
                                             x0_name)
        return x




class PluginMoodel(nn.Module):

    def __init__(self,
                 backbone: torch.nn.Module,
                 return_nodes: Union[dict, None],
                 img_size: int,
                 use_fpn: bool,
                 fpn_size: Union[int, None],
                 proj_type: str,
                 upsample_type: str,
                 use_selection: bool,
                 num_classes: int,
                 num_selects: dict,
                 use_combiner: bool,
                 comb_proj_size: Union[int, None]
                 ):
        """
        * backbone:
            torch.nn.Module class (recommand pretrained on ImageNet or IG-3.5B-17k(provided by FAIR))
        * return_nodes:
            e.g.
            return_nodes = {
                # node_name: user-specified key for output dict
                'layer1.2.relu_2': 'layer1',
                'layer2.3.relu_2': 'layer2',
                'layer3.5.relu_2': 'layer3',
                'layer4.2.relu_2': 'layer4',
            } # you can see the example on https://pytorch.org/vision/main/feature_extraction.html
            !!! if using 'Swin-Transformer', please set return_nodes to None
            !!! and please set use_fpn to True
        * feat_sizes:
            tuple or list contain features map size of each layers.
            ((C, H, W)). e.g. ((1024, 14, 14), (2048, 7, 7))
        * use_fpn:
            boolean, use features pyramid network or not
        * fpn_size:
            integer, features pyramid network projection dimension
        * num_selects:
            num_selects = {
                # match user-specified in return_nodes
                "layer1": 2048,
                "layer2": 512,
                "layer3": 128,
                "layer4": 32,
            }
        Note: after selector module (WeaklySelector) , the feature map's size is [B, S', C] which
        contained by 'logits' or 'selections' dictionary (S' is selection number, different layer
        could be different).
        """
        super(PluginMoodel, self).__init__()

        ### = = = = = Backbone = = = = =
        self.return_nodes = return_nodes
        if return_nodes is not None:
            self.backbone = create_feature_extractor(backbone, return_nodes=return_nodes)
        else:
            self.backbone = backbone

        ### get hidden feartues size
        rand_in = torch.randn(1, 3, img_size, img_size)
        outs = self.backbone(rand_in)

        ### just original backbone
        if not use_fpn and (not use_selection and not use_combiner):
            for name in outs:
                fs_size = outs[name].size()
                if len(fs_size) == 3:
                    out_size = fs_size.size(-1)
                elif len(fs_size) == 4:
                    out_size = fs_size.size(1)
                else:
                    raise ValusError("The size of output dimension of previous must be 3 or 4.")
            self.classifier = nn.Linear(out_size, num_classes)

        ### = = = = = FPN = = = = =
        self.use_fpn = use_fpn
        if self.use_fpn:
            self.fpn_down = FPN(outs, fpn_size, proj_type, upsample_type)
            self.build_fpn_classifier_down(outs, fpn_size, num_classes)
            self.fpn_up = FPN_UP(outs, fpn_size)
            self.build_fpn_classifier_up(outs, fpn_size, num_classes)

        self.fpn_size = fpn_size

        ### = = = = = Selector = = = = =
        self.use_selection = use_selection
        if self.use_selection:
            w_fpn_size = self.fpn_size if self.use_fpn else None # if not using fpn, build classifier in weakly selector
            self.selector = WeaklySelector(outs, num_classes, num_selects, w_fpn_size)

        ### = = = = = Combiner = = = = =
        self.use_combiner = use_combiner
        if self.use_combiner:
            assert self.use_selection, "Please use selection module before combiner"
            if self.use_fpn:
                gcn_inputs, gcn_proj_size = None, None
            else:
                gcn_inputs, gcn_proj_size = outs, comb_proj_size # redundant, fix in future
            total_num_selects = sum([num_selects[name] for name in num_selects]) # sum
            self.combiner = GCNCombiner(total_num_selects, num_classes, gcn_inputs, gcn_proj_size, self.fpn_size)

    def build_fpn_classifier_up(self, inputs: dict, fpn_size: int, num_classes: int):
        """
        Teh results of our experiments show that linear classifier in this case may cause some problem.
        """
        for name in inputs:
            m = nn.Sequential(
                    nn.Conv1d(fpn_size, fpn_size, 1),
                    nn.BatchNorm1d(fpn_size),
                    nn.ReLU(),
                    nn.Conv1d(fpn_size, num_classes, 1)
                )
            self.add_module("fpn_classifier_up_"+name, m)

    def build_fpn_classifier_down(self, inputs: dict, fpn_size: int, num_classes: int):
        """
        Teh results of our experiments show that linear classifier in this case may cause some problem.
        """
        for name in inputs:
            m = nn.Sequential(
                    nn.Conv1d(fpn_size, fpn_size, 1),
                    nn.BatchNorm1d(fpn_size),
                    nn.ReLU(),
                    nn.Conv1d(fpn_size, num_classes, 1)
                )
            self.add_module("fpn_classifier_down_" + name, m)

    def forward_backbone(self, x):
        return self.backbone(x)

    def fpn_predict_down(self, x: dict, logits: dict):
        """
        x: [B, C, H, W] or [B, S, C]
           [B, C, H, W] --> [B, H*W, C]
        """
        for name in x:
            if "FPN1_" not in name:
                continue
            ### predict on each features point
            if len(x[name].size()) == 4:
                B, C, H, W = x[name].size()
                logit = x[name].view(B, C, H*W)
            elif len(x[name].size()) == 3:
                logit = x[name].transpose(1, 2).contiguous()
            model_name = name.replace("FPN1_", "")
            logits[name] = getattr(self, "fpn_classifier_down_" + model_name)(logit)
            logits[name] = logits[name].transpose(1, 2).contiguous() # transpose

    def fpn_predict_up(self, x: dict, logits: dict):
        """
        x: [B, C, H, W] or [B, S, C]
           [B, C, H, W] --> [B, H*W, C]
        """
        for name in x:
            if "FPN1_" in name:
                continue
            ### predict on each features point
            if len(x[name].size()) == 4:
                B, C, H, W = x[name].size()
                logit = x[name].view(B, C, H*W)
            elif len(x[name].size()) == 3:
                logit = x[name].transpose(1, 2).contiguous()
            model_name = name.replace("FPN1_", "")
            logits[name] = getattr(self, "fpn_classifier_up_" + model_name)(logit)
            logits[name] = logits[name].transpose(1, 2).contiguous() # transpose

    def forward(self, x: torch.Tensor):

        logits = {}

        x = self.forward_backbone(x)

        if self.use_fpn:
            x = self.fpn_down(x)
            # print([name for name in x])
            self.fpn_predict_down(x, logits)
            x = self.fpn_up(x)
            self.fpn_predict_up(x, logits)

        if self.use_selection:
            selects = self.selector(x, logits)

        if self.use_combiner:
            comb_outs = self.combiner(selects)
            logits['comb_outs'] = comb_outs
            return logits

        if self.use_selection or self.fpn:
            return logits

        ### original backbone (only predict final selected layer)
        for name in x:
            hs = x[name]

        if len(hs.size()) == 4:
            hs = F.adaptive_avg_pool2d(hs, (1, 1))
            hs = hs.flatten(1)
        else:
            hs = hs.mean(1)
        out = self.classifier(hs)
        logits['ori_out'] = logits

        return