trim/create_runner.py at main · wizardbc/trim · 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
import torch
from torch.utils.data import DataLoader

from torchvision.datasets import ImageNet
from torchvision.models import *
from torchvision.models.feature_extraction import create_feature_extractor

import timm

from ood.runner import OODRunner
from ood.dataset import DatasetFilelist, FakeData

observe_id = 'ImageNet'
observe_ood = ['iNaturalist', 'SUN', 'Places', 'Texture', 'OpenImages-O', 'ImageNet-O']

custom_models_and_weights = {
  ### key: (model constructor, transforms, state_dict path)
  "resnet50.A": (resnet50, ResNet50_Weights.IMAGENET1K_V1.transforms(), "./train/resnet50.A.pt"),
  "resnet50.B": (resnet50, ResNet50_Weights.IMAGENET1K_V2.transforms(), "./train/resnet50.B.pt"),
  "resnet50.C": (resnet50, ResNet50_Weights.IMAGENET1K_V1.transforms(), "./train/resnet50.C.pt"),
}

torchvision_models_and_weights = {
  "ResNet50_Weights.IMAGENET1K_V1": (resnet50, ResNet50_Weights.IMAGENET1K_V1),
  "ResNet50_Weights.IMAGENET1K_V2": (resnet50, ResNet50_Weights.IMAGENET1K_V2),
  "ResNet152_Weights.IMAGENET1K_V1": (resnet152, ResNet152_Weights.IMAGENET1K_V1),
  "ResNet152_Weights.IMAGENET1K_V2": (resnet152, ResNet152_Weights.IMAGENET1K_V2),
  "ViT_B_16_Weights.IMAGENET1K_V1": (vit_b_16, ViT_B_16_Weights.IMAGENET1K_V1),
  "ViT_B_16_Weights.IMAGENET1K_SWAG_E2E_V1": (vit_b_16, ViT_B_16_Weights.IMAGENET1K_SWAG_E2E_V1),
  "ViT_B_16_Weights.IMAGENET1K_SWAG_LINEAR_V1": (vit_b_16, ViT_B_16_Weights.IMAGENET1K_SWAG_LINEAR_V1),
  "Swin_V2_B_Weights.IMAGENET1K_V1": (swin_v2_b, Swin_V2_B_Weights.IMAGENET1K_V1),
  "MobileNet_V2_Weights.IMAGENET1K_V1": (mobilenet_v2, MobileNet_V2_Weights.IMAGENET1K_V1),
  "MobileNet_V2_Weights.IMAGENET1K_V2": (mobilenet_v2, MobileNet_V2_Weights.IMAGENET1K_V2),
  "MobileNet_V3_Large_Weights.IMAGENET1K_V1": (mobilenet_v3_large, MobileNet_V3_Large_Weights.IMAGENET1K_V1),
  "MobileNet_V3_Large_Weights.IMAGENET1K_V2": (mobilenet_v3_large, MobileNet_V3_Large_Weights.IMAGENET1K_V2),
  "AlexNet_Weights.IMAGENET1K_V1": (alexnet, AlexNet_Weights.IMAGENET1K_V1),
  "ConvNeXt_Base_Weights.IMAGENET1K_V1": (convnext_base, ConvNeXt_Base_Weights.IMAGENET1K_V1),
  "DenseNet161_Weights.IMAGENET1K_V1": (densenet161, DenseNet161_Weights.IMAGENET1K_V1),
  "EfficientNet_B7_Weights.IMAGENET1K_V1": (efficientnet_b7, EfficientNet_B7_Weights.IMAGENET1K_V1),
  "EfficientNet_V2_M_Weights.IMAGENET1K_V1": (efficientnet_v2_m, EfficientNet_V2_M_Weights.IMAGENET1K_V1),
  "GoogLeNet_Weights.IMAGENET1K_V1": (googlenet, GoogLeNet_Weights.IMAGENET1K_V1),
  "Inception_V3_Weights.IMAGENET1K_V1": (inception_v3, Inception_V3_Weights.IMAGENET1K_V1),
  "MNASNet1_3_Weights.IMAGENET1K_V1": (mnasnet1_3, MNASNet1_3_Weights.IMAGENET1K_V1),
  "MaxVit_T_Weights.IMAGENET1K_V1": (maxvit_t, MaxVit_T_Weights.IMAGENET1K_V1),
  "RegNet_Y_16GF_Weights.IMAGENET1K_V1": (regnet_y_16gf, RegNet_Y_16GF_Weights.IMAGENET1K_V1),
  "RegNet_Y_16GF_Weights.IMAGENET1K_V2": (regnet_y_16gf, RegNet_Y_16GF_Weights.IMAGENET1K_V2),
  "RegNet_Y_16GF_Weights.IMAGENET1K_SWAG_E2E_V1": (regnet_y_16gf, RegNet_Y_16GF_Weights.IMAGENET1K_SWAG_E2E_V1),
  "RegNet_Y_16GF_Weights.IMAGENET1K_SWAG_LINEAR_V1": (regnet_y_16gf, RegNet_Y_16GF_Weights.IMAGENET1K_SWAG_LINEAR_V1),
  "ResNeXt101_32X8D_Weights.IMAGENET1K_V1": (resnext101_32x8d, ResNeXt101_32X8D_Weights.IMAGENET1K_V1),
  "ResNeXt101_32X8D_Weights.IMAGENET1K_V2": (resnext101_32x8d, ResNeXt101_32X8D_Weights.IMAGENET1K_V2),
  "ShuffleNet_V2_X2_0_Weights.IMAGENET1K_V1": (shufflenet_v2_x2_0, ShuffleNet_V2_X2_0_Weights.IMAGENET1K_V1),
  "VGG19_BN_Weights.IMAGENET1K_V1": (vgg19_bn, VGG19_BN_Weights.IMAGENET1K_V1),
  "Wide_ResNet101_2_Weights.IMAGENET1K_V1": (wide_resnet101_2, Wide_ResNet101_2_Weights.IMAGENET1K_V1),
  "Wide_ResNet101_2_Weights.IMAGENET1K_V2": (wide_resnet101_2, Wide_ResNet101_2_Weights.IMAGENET1K_V2),
}

timm_models_and_weights = ['eva02_large_patch14_448.mim_m38m_ft_in1k', 'eva02_large_patch14_448.mim_m38m_ft_in22k_in1k']

nodes_to_track = {
  ### TorchVision models
  'resnet50': {'avgpool': 'feature', 'fc': 'logit'},
  'resnet152': {'avgpool': 'feature', 'fc': 'logit'},
  'vit_b_16': {'getitem_5': 'feature', 'heads.head': 'logit'},
  'swin_v2_b': {'avgpool': 'feature', 'head': 'logit'},
  'mobilenet_v2': {'adaptive_avg_pool2d': 'feature', 'classifier.1': 'logit'},
  'mobilenet_v3_large': {'classifier.1': 'feature', 'classifier.3': 'logit'},
  'alexnet': {'classifier.5': 'feature', 'classifier.6': 'logit'},
  'convnext_base': {'classifier.1': 'feature', 'classifier.2': 'logit'},
  'densenet161': {'adaptive_avg_pool2d': 'feature', 'classifier': 'logit'},
  'efficientnet_b7': {'avgpool': 'feature', 'classifier.1': 'logit'},
  'efficientnet_v2_m': {'avgpool': 'feature', 'classifier.1': 'logit'},
  'googlenet': {'avgpool': 'feature', 'fc': 'logit'},
  'inception_v3': {'avgpool': 'feature', 'fc': 'logit'},
  'mnasnet1_3': {'mean': 'feature', 'classifier.1': 'logit'},
  'maxvit_t': {'classifier.4': 'feature', 'classifier.5': 'logit'},
  'regnet_y_16gf': {'avgpool': 'feature', 'fc': 'logit'},
  'resnext101_32x8d': {'avgpool': 'feature', 'fc': 'logit'},
  'resnext101_32x8d': {'avgpool': 'feature', 'fc': 'logit'},
  'shufflenet_v2_x2_0': {'mean': 'feature', 'fc': 'logit'},
  'vgg19_bn': {'classifier.4': 'feature', 'classifier.6': 'logit'},
  'wide_resnet101_2': {'avgpool': 'feature', 'fc': 'logit'},
  ### TIMM models
  'eva02_large_patch14_448': {'fc_norm': 'feature', 'head': 'logit'},
}

def get_fancy_name(name:str):
  if name in torchvision_models_and_weights.keys():
    constructor, _ = torchvision_models_and_weights.get(name)
    _, weight_name = name.split('.')
    model_name = constructor.__name__
    weight_name ='_'.join(weight_name.split('_')[1:])
  elif name in timm_models_and_weights:
    model_name, weight_name = name.split('.')
    model_name = model_name.split('_')[0]
    weight_name = weight_name.split('_ft_')[-1]
  elif name in custom_models_and_weights.keys():
    model_name, weight_name = name.split('.')
  else:
    raise NameError(name)
  return model_name + '.' + weight_name

def get_extractor_and_transforms(name:str):
  if name in torchvision_models_and_weights.keys():
    constructor, weights = torchvision_models_and_weights.get(name)
    model = constructor(weights=weights)
    model.eval()
    nodes = nodes_to_track.get(constructor.__name__)
    transforms = weights.transforms()
  elif name in custom_models_and_weights.keys():
    constructor, transforms, fname = custom_models_and_weights.get(name)
    model = constructor()
    model.load_state_dict(torch.load(fname))
    model.eval()
    nodes = nodes_to_track.get(constructor.__name__)
  elif name in timm_models_and_weights:
    model = timm.create_model(name, pretrained=True)
    model.eval()
    nodes = nodes_to_track.get(name.split('.')[0])
    data_config = timm.data.resolve_model_data_config(model)
    transforms = timm.data.create_transform(**data_config, is_training=False)
  else:
    raise NameError(name)
  extractor = create_feature_extractor(model, nodes)
  return extractor, transforms

def prepare_dataloaders(transform, args):
  if hasattr(transform, 'crop_size'):
    ### for TorchVision
    crop_size = transform.crop_size
    if len(crop_size)==1:
      crop_size = (crop_size[0], crop_size[0])
  else:
    ### for TIMM
    crop_size = transform.transforms[1].size

  testdata = {
    'ImageNet': ImageNet(args.imagenet_path, split='val', transform=transform),
    'OpenImages-O': DatasetFilelist(args.openimages_o_path, args.openimages_o_filelist, transform=transform),
    'Texture': DatasetFilelist(args.dtd_path, args.dtd_filelist, transform=transform),
    'iNaturalist': DatasetFilelist(args.inaturalist_path, args.inaturalist_filelist, transform=transform),
    'ImageNet-O': DatasetFilelist(args.imagenet_o_path, args.imagenet_o_filelist, transform=transform),
    'Places': DatasetFilelist(args.places_path, args.places_filelist, transform=transform),
    'SUN': DatasetFilelist(args.sun_path, args.sun_filelist, transform=transform),
    # 'FakeData': FakeData(50000, image_size=(3, *crop_size)),
  }
  testloaders = {
    k: DataLoader(v, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, pin_memory=False)
    for k, v in testdata.items()
  }

  id_loader = testloaders[observe_id]
  ood_loaders = tuple(testloaders[k] for k in observe_ood)

  return id_loader, ood_loaders


if __name__ == '__main__':
  import argparse
  parser = argparse.ArgumentParser(description="Create OOD Runner", add_help=True)

  parser.add_argument("weights", type=str, choices=list(torchvision_models_and_weights.keys())+list(custom_models_and_weights.keys())+timm_models_and_weights, help="Model.Weights")
  parser.add_argument("-o", "--output", default=None, type=str, help="path to save outputs")
  parser.add_argument("--device", default="cuda:0", type=str, help="device (Use cuda or cpu Default: cuda:0)")
  parser.add_argument("-b", "--batch-size", default=64, type=int, help="batch size")
  parser.add_argument("-j", "--workers", default=16, type=int, metavar="N", help="number of data loading workers (default: 16)")

  ### Datasets
  parser.add_argument("--imagenet-path", default="./data/imagenet", type=str, help="ImageNet path")
  parser.add_argument("--imagenet-o-path", default="./data/imagenet-o", type=str, help="ImageNet-O path")
  parser.add_argument("--imagenet-o-filelist", default=None, type=str, help="ImageNet-O file list path")
  parser.add_argument("--openimages-o-path", default="./data/open-images/test", type=str, help="OpenImages-O path")
  parser.add_argument("--openimages-o-filelist", default="./data/open-images/openimage_o.txt", type=str, help="OpenImages-O file list path")
  parser.add_argument("--dtd-path", default="./data/dtd/images", type=str, help="Texture path")
  parser.add_argument("--dtd-filelist", default=None, type=str, help="Texture file list path")
  parser.add_argument("--inaturalist-path", default="./data/iNaturalist/images", type=str, help="iNaturalist path")
  parser.add_argument("--inaturalist-filelist", default=None, type=str, help="iNaturalist file list path")
  parser.add_argument("--places-path", default="./data/Places/images", type=str, help="Places path")
  parser.add_argument("--places-filelist", default=None, type=str, help="Places file list path")
  parser.add_argument("--sun-path", default="./data/SUN/images", type=str, help="SUN path")
  parser.add_argument("--sun-filelist", default=None, type=str, help="SUN file list path")

  args = parser.parse_args()
  extractor, transform = get_extractor_and_transforms(args.weights)
  extractor = extractor.to(args.device)
  id_loader, ood_loaders = prepare_dataloaders(transform, args)

  runner = OODRunner()
  runner.run(extractor, id_loader, ood_loaders,
             observe_id, tuple(observe_ood),
             ('feature',), args.device)
  fname = args.output
  if fname is None:
    fname = f'./runners/{get_fancy_name(args.weights)}.pt'
  torch.save(runner.state_dict(), fname)