RelViT/models.py at main · ZiliottoFilippoDev/RelViT · 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
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
from torchvision.models.detection import FasterRCNN, MaskRCNN
from torchvision.models.detection.rpn import AnchorGenerator
from torchvision.ops import MultiScaleRoIAlign, FeaturePyramidNetwork
from torchvision.models.detection.backbone_utils import LastLevelMaxPool
import models.vit as vit
from models.swin import SwinTransformer
from models.t2t import t2t_vit_14 as t2t
import torchvision
import torch
import torch.nn as nn
from collections import OrderedDict

class FPNT2T(nn.Module):
    def __init__(self, num_classes=91, pretrained_path=None, fixed_size=(224,224), backbone_out_chan=512):
        super(FPNT2T, self).__init__()

        model  = t2t(num_classes = num_classes)
        norm_layer = nn.LayerNorm

        self.patch_size = 16
        self.num_patches = (fixed_size[0] // self.patch_size) ** 2
        self.embed_dim =  model.embed_dim
        self.fpn = FeaturePyramidNetwork([self.embed_dim for i in range(4)], backbone_out_chan, extra_blocks=LastLevelMaxPool())
        self.pos_embed =  model.pos_embed

        for i_layer in range(4):
            layer = norm_layer(self.embed_dim)
            layer_name = f'norm{i_layer}'
            self.add_module(layer_name, layer)

        if pretrained_path is not None:
            state_dict = torch.load(pretrained_path)['state_dict']
            state_dict = {k:v for k,v in state_dict.items() if 'backbone' in k}
            for key in list(state_dict):
                new_key = key.replace('backbone.','')
                state_dict[new_key] = state_dict.pop(key)
            model.load_state_dict(state_dict)
        self.backbone = model

    def forward(self, x):
        out = self.backbone.forward_features(x)
        del out['z']
        del out['z_patches']


        for en, i in enumerate(out.keys()):
            layer = getattr(self, f'norm{en}')
            x = layer(out[i])
            x = x.reshape(x.shape[0], int(self.num_patches ** 0.5), int(self.num_patches ** 0.5), -1).permute(0,3,1,2).contiguous()
            out[i] = x

        x = self.fpn(OrderedDict(out))
        return x


class FPNSwin(nn.Module):
    def __init__(self, num_classes = 10, pretrained_path=None, backbone_out_chan=512, vit_type='tiny'):
        super(FPNSwin, self).__init__()

        if vit_type in ['tiny']:
            depths=[2, 2, 6, 2]
        elif vit_type in ['small']:
            depths=[2, 2, 18, 2]
        else:
            raise ValueError('Vit type not supported. tiny, small are accepted.')

        model = SwinTransformer(num_classes = num_classes, use_positional_embeddings=True, depths=depths)
        self.patch_res = [56, 28, 14, 7]
        self.in_channels = [48, 96, 192, 768]
        self.num_features = [192, 384, 768, 768]
        self.fpn = FeaturePyramidNetwork(self.in_channels, backbone_out_chan, extra_blocks=LastLevelMaxPool())
        norm_layer = nn.LayerNorm

        for i_layer in range(4):
            layer = norm_layer(self.num_features[i_layer])
            layer_name = f'norm{i_layer}'
            self.add_module(layer_name, layer)

        if pretrained_path is not None:
            state_dict = torch.load(pretrained_path)['state_dict']
            state_dict = {k:v for k,v in state_dict.items() if 'backbone' in k}
            for key in list(state_dict):
                new_key = key.replace('backbone.','')
                state_dict[new_key] = state_dict.pop(key)
            model.load_state_dict(state_dict)
        self.backbone = model

    def forward(self, x):
        x, out = self.backbone.forward_features(x)

        out = OrderedDict({
             'feat0':out[0],
             'feat1':out[1],
             'feat2':out[2],
             'feat3':out[3],
             })

        for en, i in enumerate(out.values()):

            layer = getattr(self, f'norm{en}')
            x = layer(i)
            x = x.reshape(i.shape[0], self.patch_res[en], self.patch_res[en], -1).permute(0,3,1,2).contiguous()
            out['feat{}'.format(en)] = x

        x = self.fpn(out)
        return x

class FPNViT(nn.Module):
    def __init__(self, num_classes=91, vit_type='small', pretrained_path=None, fixed_size=(224,224), backbone_out_chan=512):
        super(FPNViT, self).__init__()

        vit_type = 'vit_' + vit_type
        model  = vit.__dict__[vit_type](num_classes=num_classes, use_clf_token=True, use_positional_embeddings=True)
        norm_layer = nn.LayerNorm

        self.patch_size = 16
        self.num_patches = (fixed_size[0] // self.patch_size) ** 2
        self.embed_dim =  model.embed_dim
        self.fpn = FeaturePyramidNetwork([self.embed_dim for i in range(3)], backbone_out_chan, extra_blocks=LastLevelMaxPool())
        self.patch_embed = model.patch_embed
        self.pos_embed =  model.pos_embed

        for i_layer in range(3):
            layer = norm_layer(self.embed_dim)
            layer_name = f'norm{i_layer}'
            self.add_module(layer_name, layer)

        if pretrained_path is not None:
            state_dict = torch.load(pretrained_path)['state_dict']
            state_dict = {k:v for k,v in state_dict.items() if 'backbone' in k}
            for key in list(state_dict):
                new_key = key.replace('backbone.','')
                state_dict[new_key] = state_dict.pop(key)
            model.load_state_dict(state_dict)
        self.backbone = model

    def forward(self, x):
        out = self.backbone(x)
        if 'z' in out.keys():
            del out['z']

        for en, i in enumerate(out.keys()):
            layer = getattr(self, f'norm{en}')
            x = layer(out[i])
            x = x.reshape(x.shape[0], int(self.num_patches ** 0.5), int(self.num_patches ** 0.5), -1).permute(0,3,1,2).contiguous()
            out[i] = x

        x = self.fpn(OrderedDict(out))
        return x

def create_model(num_classes = 91, pretrained_path = None, fixed_size=(224,224),
                 mode='segm', vit_type='small', model_type='vit', backbone_out_chan = 512,
                 image_mean = [0.485, 0.456, 0.406], image_std = [0.229, 0.224, 0.225]):

    if model_type in ['swin']:
        backbone = FPNSwin(num_classes = num_classes, pretrained_path = pretrained_path, backbone_out_chan = backbone_out_chan, vit_type=vit_type)
        feat_maps, k = ['feat0','feat1','feat2','feat3','pool'], 5
    elif model_type in ['vit']:
        backbone = FPNViT(num_classes = num_classes, pretrained_path = pretrained_path, vit_type=vit_type, backbone_out_chan = backbone_out_chan)
        feat_maps, k = ['feat0','feat1','feat2','pool'], 4
    elif model_type in ['t2t']:
        backbone = FPNT2T(num_classes=num_classes, pretrained_path = pretrained_path, backbone_out_chan=backbone_out_chan, fixed_size=fixed_size)
        feat_maps, k = ['feat0','feat1','feat2','feat3','pool'], 5
    elif model_type in ['resnet']:
        model = torchvision.models.detection.fasterrcnn_resnet50_fpn(weights=None, num_classes=num_classes+1)
        return model
    else:
        raise ValueError(f'Model type {model_type} not correct. vit, swin are accepted.')

    backbone.out_channels = backbone_out_chan
    anchorgen = AnchorGenerator(sizes=tuple([(32,64,128,256) for i in range(k)]),
                                aspect_ratios=tuple([(0.5, 1.0, 2.0) for i in range(k)]),)
    pooler = MultiScaleRoIAlign(featmap_names=feat_maps, output_size=7, sampling_ratio=2)

    if mode in ['segm']:
        mask_roi_pooler = MultiScaleRoIAlign(featmap_names=feat_maps, output_size=14,sampling_ratio=2)
        model = MaskRCNN(backbone = backbone,
                    fixed_size=fixed_size,
                    img_mean = image_mean,
                    img_std = image_std,
                    num_classes = num_classes + 1,
                    rpn_anchor_generator = anchorgen,
                    box_roi_pool = pooler,
                    mask_roi_pool = mask_roi_pooler,
                    box_detections_per_img = 100)
        return model

    elif mode in ['bbox']:
        model = FasterRCNN(backbone = backbone,
                    fixed_size = fixed_size,
                    img_mean = image_mean,
                    img_std = image_std,
                    num_classes = num_classes + 1,
                    rpn_anchor_generator = anchorgen,
                    box_roi_pool = pooler,
                    box_detections_per_img = 100
                    )
        return model

    else:
        raise ValueError(f'mode type {mode} not correct. segm, bbox types are accepted.')