├── HBP_all.py ├── HBP_fc.py ├── HBP_fc_new.py ├── LICENSE ├── README.md ├── cub200.py └── log ├── README.md ├── hbp_all.log ├── hbp_fc.log └── hbp_fc_new.log /HBP_all.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | """Fine-tune all layers only for HBP(Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition). 4 | Usage: 5 | CUDA_VISIBLE_DEVICES=0,1 python HBP_all.py --base_lr 0.001 --batch_size 24 --epochs 200 --weight_decay 0.0005 --model 'HBP_fc_epoch_*.pth' | tee 'hbp_all.log' 6 | """ 7 | 8 | 9 | import os 10 | import torch 11 | import torchvision 12 | import cub200 13 | import visdom 14 | import argparse 15 | vis = visdom.Visdom(env=u'HBP_all',use_incoming_socket=False) 16 | torch.manual_seed(0) 17 | torch.cuda.manual_seed_all(0) 18 | 19 | class HBP(torch.nn.Module): 20 | def __init__(self): 21 | """Declare all needed layers.""" 22 | torch.nn.Module.__init__(self) 23 | # Convolution and pooling layers of VGG-16. 24 | self.features = torchvision.models.vgg16(pretrained=False).features 25 | self.features_conv5_1 = torch.nn.Sequential(*list(self.features.children()) 26 | [:-5]) 27 | self.features_conv5_2 = torch.nn.Sequential(*list(self.features.children()) 28 | [-5:-3]) 29 | self.features_conv5_3 = torch.nn.Sequential(*list(self.features.children()) 30 | [-3:-1]) 31 | self.bilinear_proj = torch.nn.Sequential(torch.nn.Conv2d(512,8192,kernel_size=1,bias=False), 32 | torch.nn.BatchNorm2d(8192), 33 | torch.nn.ReLU(inplace=True)) 34 | # Linear classifier. 35 | self.fc = torch.nn.Linear(8192*3, 200) 36 | 37 | def hbp(self,conv1,conv2): 38 | N = conv1.size()[0] 39 | proj_1 = self.bilinear_proj(conv1) 40 | proj_2 = self.bilinear_proj(conv2) 41 | assert(proj_1.size() == (N,8192,28,28)) 42 | X = proj_1 * proj_2 43 | assert(X.size() == (N,8192,28,28)) 44 | X = torch.sum(X.view(X.size()[0],X.size()[1],-1),dim = 2) 45 | X = X.view(N, 8192) 46 | X = torch.sqrt(X + 1e-5) 47 | X = torch.nn.functional.normalize(X) 48 | return X 49 | 50 | def forward(self, X): 51 | N = X.size()[0] 52 | assert X.size() == (N, 3, 448, 448) 53 | X_conv5_1 = self.features_conv5_1(X) 54 | X_conv5_2 = self.features_conv5_2(X_conv5_1) 55 | X_conv5_3 = self.features_conv5_3(X_conv5_2) 56 | 57 | X_branch_1 = self.hbp(X_conv5_1,X_conv5_2) 58 | X_branch_2 = self.hbp(X_conv5_2,X_conv5_3) 59 | X_branch_3 = self.hbp(X_conv5_1,X_conv5_3) 60 | 61 | X_branch = torch.cat([X_branch_1,X_branch_2,X_branch_3],dim=1) 62 | assert X_branch.size() == (N,8192*3) 63 | 64 | X = self.fc(X_branch) 65 | assert X.size() == (N, 200) 66 | return X 67 | 68 | class HBPManager(object): 69 | def __init__(self, options, path): 70 | print('Prepare the network and data.') 71 | self._options = options 72 | self._path = path 73 | # Network. 74 | self._net = torch.nn.DataParallel(HBP()).cuda() 75 | print(self._net) 76 | self._net.load_state_dict(torch.load(self._path['model'])) 77 | # Criterion. 78 | self._criterion = torch.nn.CrossEntropyLoss().cuda() 79 | # Solver. 80 | param_to_optim = [] 81 | for param in self._net.parameters(): 82 | param_to_optim.append(param) 83 | 84 | self._solver = torch.optim.SGD( 85 | param_to_optim, lr=self._options['base_lr'], 86 | momentum=0.9, weight_decay=self._options['weight_decay']) 87 | milestones = [100] 88 | self._scheduler = torch.optim.lr_scheduler.MultiStepLR(self._solver,milestones = milestones,gamma=0.25) 89 | 90 | train_transforms = torchvision.transforms.Compose([ 91 | torchvision.transforms.Resize(size=448), # Let smaller edge match 92 | torchvision.transforms.RandomHorizontalFlip(), 93 | torchvision.transforms.RandomCrop(size=448), 94 | torchvision.transforms.ToTensor(), 95 | torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), 96 | std=(0.229, 0.224, 0.225)) 97 | ]) 98 | test_transforms = torchvision.transforms.Compose([ 99 | torchvision.transforms.Resize(size=448), 100 | torchvision.transforms.CenterCrop(size=448), 101 | torchvision.transforms.ToTensor(), 102 | torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), 103 | std=(0.229, 0.224, 0.225)) 104 | ]) 105 | train_data = cub200.CUB200( 106 | root=self._path['cub200'], train=True, download=True, 107 | transform=train_transforms) 108 | test_data = cub200.CUB200( 109 | root=self._path['cub200'], train=False, download=True, 110 | transform=test_transforms) 111 | self._train_loader = torch.utils.data.DataLoader( 112 | train_data, batch_size=self._options['batch_size'], 113 | shuffle=True, num_workers=4, pin_memory=True) 114 | self._test_loader = torch.utils.data.DataLoader( 115 | test_data, batch_size=16, 116 | shuffle=False, num_workers=4, pin_memory=True) 117 | 118 | def train(self): 119 | print('Training.') 120 | best_acc = 0.0 121 | best_epoch = None 122 | print('Epoch\tTrain loss\tTrain acc\tTest acc') 123 | ii = 0 124 | for t in range(self._options['epochs']): 125 | epoch_loss = [] 126 | num_correct = 0 127 | num_total = 0 128 | for X, y in self._train_loader: 129 | # Data. 130 | X = torch.autograd.Variable(X.cuda()) 131 | y = torch.autograd.Variable(y.cuda(non_blocking = True)) 132 | # Clear the existing gradients. 133 | self._solver.zero_grad() 134 | # Forward pass. 135 | score = self._net(X) 136 | loss = self._criterion(score, y) 137 | epoch_loss.append(loss.data[0]) 138 | # Prediction. 139 | _, prediction = torch.max(score.data, 1) 140 | num_total += y.size(0) 141 | num_correct += torch.sum(prediction == y.data) 142 | # Backward pass. 143 | loss.backward() 144 | self._solver.step() 145 | 146 | ii += 1 147 | x = torch.Tensor([ii]) 148 | y = torch.Tensor([loss.data[0]]) 149 | vis.line(X=x, Y=y, win='polynomial', update='append' if ii > 0 else None) 150 | 151 | num_correct = torch.tensor(num_correct).float().cuda() 152 | num_total = torch.tensor(num_total).float().cuda() 153 | 154 | train_acc = 100 * num_correct / num_total 155 | test_acc = self._accuracy(self._test_loader) 156 | self._scheduler.step(test_acc) 157 | if test_acc > best_acc: 158 | best_acc = test_acc 159 | best_epoch = t + 1 160 | print('*', end='') 161 | # Save model onto disk. 162 | torch.save(self._net.state_dict(), 163 | os.path.join('./model/' 164 | 'HBP_all_epoch_%d.pth' % (t + 1))) 165 | print('%d\t%4.3f\t\t%4.2f%%\t\t%4.2f%%' % 166 | (t+1, sum(epoch_loss) / len(epoch_loss), train_acc, test_acc)) 167 | print('Best at epoch %d, test accuaray %f' % (best_epoch, best_acc)) 168 | 169 | def _accuracy(self, data_loader): 170 | self._net.train(False) 171 | num_correct = 0 172 | num_total = 0 173 | for X, y in data_loader: 174 | # Data. 175 | X = torch.autograd.Variable(X.cuda()) 176 | y = torch.autograd.Variable(y.cuda(non_blocking = True)) 177 | # Prediction. 178 | score = self._net(X) 179 | _, prediction = torch.max(score.data, 1) 180 | num_total += y.size(0) 181 | num_correct += torch.sum(prediction == y.data) 182 | self._net.train(True) # Set the model to training phase 183 | num_correct = torch.tensor(num_correct).float().cuda() 184 | num_total = torch.tensor(num_total).float().cuda() 185 | return 100 * num_correct / num_total 186 | 187 | def getStat(self): 188 | print('Compute mean and variance for training data.') 189 | train_data = cub200.CUB200( 190 | root=self._path['cub200'], train=True, 191 | transform=torchvision.transforms.ToTensor(), download=True) 192 | train_loader = torch.utils.data.DataLoader( 193 | train_data, batch_size=1, shuffle=False, num_workers=4, 194 | pin_memory=True) 195 | mean = torch.zeros(3) 196 | std = torch.zeros(3) 197 | for X, _ in train_loader: 198 | for d in range(3): 199 | mean[d] += X[:, d, :, :].mean() 200 | std[d] += X[:, d, :, :].std() 201 | mean.div_(len(train_data)) 202 | std.div_(len(train_data)) 203 | print(mean) 204 | print(std) 205 | 206 | def main(): 207 | parser = argparse.ArgumentParser( 208 | description='Train HBP on CUB200.') 209 | parser.add_argument('--base_lr', dest='base_lr', type=float, required=True, 210 | help='Base learning rate for training.') 211 | parser.add_argument('--batch_size', dest='batch_size', type=int, 212 | required=True, help='Batch size.') 213 | parser.add_argument('--epochs', dest='epochs', type=int, 214 | required=True, help='Epochs for training.') 215 | parser.add_argument('--weight_decay', dest='weight_decay', type=float, 216 | required=True, help='Weight decay.') 217 | parser.add_argument('--model', dest='model', type=str, required=True, 218 | help='Model for fine-tuning.') 219 | args = parser.parse_args() 220 | if args.base_lr <= 0: 221 | raise AttributeError('--base_lr parameter must >0.') 222 | if args.batch_size <= 0: 223 | raise AttributeError('--batch_size parameter must >0.') 224 | if args.epochs < 0: 225 | raise AttributeError('--epochs parameter must >=0.') 226 | if args.weight_decay <= 0: 227 | raise AttributeError('--weight_decay parameter must >0.') 228 | 229 | options = { 230 | 'base_lr': args.base_lr, 231 | 'batch_size': args.batch_size, 232 | 'epochs': args.epochs, 233 | 'weight_decay': args.weight_decay, 234 | } 235 | 236 | project_root = os.popen('pwd').read().strip() 237 | path = { 238 | 'cub200': os.path.join(project_root, 'data/cub200'), 239 | 'model': os.path.join(project_root, 'model', args.model), 240 | } 241 | for d in path: 242 | if d == 'model': 243 | assert os.path.isfile(path[d]) 244 | else: 245 | assert os.path.isdir(path[d]) 246 | 247 | manager = HBPManager(options, path) 248 | manager.getStat() 249 | manager.train() 250 | if __name__ == '__main__': 251 | main() 252 | -------------------------------------------------------------------------------- /HBP_fc.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | """Fine-tune the fc layer only for HBP(Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition). 4 | Usage: 5 | CUDA_VISIBLE_DEVICES=0,1 python HBP_fc.py --base_lr 1.0 --batch_size 12 --epochs 120 --weight_decay 0.000005 | tee 'hbp_fc.log' 6 | """ 7 | 8 | import os 9 | import torch 10 | import torchvision 11 | import cub200 12 | import visdom 13 | import argparse 14 | 15 | vis = visdom.Visdom(env=u'HBP_fc',use_incoming_socket=False) 16 | torch.manual_seed(0) 17 | torch.cuda.manual_seed_all(0) 18 | 19 | class HBP(torch.nn.Module): 20 | def __init__(self): 21 | torch.nn.Module.__init__(self) 22 | # Convolution and pooling layers of VGG-16. 23 | self.features = torchvision.models.vgg16(pretrained=True).features 24 | self.features_conv5_1 = torch.nn.Sequential(*list(self.features.children()) 25 | [:-5]) 26 | self.features_conv5_2 = torch.nn.Sequential(*list(self.features.children()) 27 | [-5:-3]) 28 | self.features_conv5_3 = torch.nn.Sequential(*list(self.features.children()) 29 | [-3:-1]) 30 | self.bilinear_proj = torch.nn.Sequential(torch.nn.Conv2d(512,8192,kernel_size=1,bias=False), 31 | torch.nn.BatchNorm2d(8192), 32 | torch.nn.ReLU(inplace=True)) 33 | # Linear classifier. 34 | self.fc = torch.nn.Linear(8192*3, 200) 35 | 36 | # Freeze all previous layers. 37 | for param in self.features_conv5_1.parameters(): 38 | param.requires_grad = False 39 | for param in self.features_conv5_2.parameters(): 40 | param.requires_grad = False 41 | for param in self.features_conv5_3.parameters(): 42 | param.requires_grad = False 43 | 44 | # Initialize the fc layers. 45 | torch.nn.init.xavier_normal_(self.fc.weight.data) 46 | if self.fc.bias is not None: 47 | torch.nn.init.constant_(self.fc.bias.data, val=0) 48 | 49 | #init 50 | for m in self.bilinear_proj.modules(): 51 | if isinstance(m, torch.nn.Conv2d): 52 | torch.nn.init.xavier_normal_(m.weight) 53 | if m.bias is not None: 54 | torch.nn.init.constant_(m.bias, 0) 55 | elif isinstance(m, torch.nn.BatchNorm2d): 56 | torch.nn.init.constant_(m.weight,1) 57 | torch.nn.init.constant_(m.bias, 0) 58 | elif isinstance(m, torch.nn.Linear): 59 | torch.nn.init.xavier_normal_(m.weight) 60 | torch.nn.init.constant_(m.bias, 0) 61 | 62 | def hbp(self,conv1,conv2): 63 | N = conv1.size()[0] 64 | proj_1 = self.bilinear_proj(conv1) 65 | proj_2 = self.bilinear_proj(conv2) 66 | assert(proj_1.size() == (N,8192,28,28)) 67 | X = proj_1 * proj_2 68 | assert(X.size() == (N,8192,28,28)) 69 | X = torch.sum(X.view(X.size()[0],X.size()[1],-1),dim = 2) 70 | X = X.view(N, 8192) 71 | X = torch.sqrt(X + 1e-5) 72 | X = torch.nn.functional.normalize(X) 73 | return X 74 | 75 | def forward(self, X): 76 | N = X.size()[0] 77 | assert X.size() == (N, 3, 448, 448) 78 | X_conv5_1 = self.features_conv5_1(X) 79 | X_conv5_2 = self.features_conv5_2(X_conv5_1) 80 | X_conv5_3 = self.features_conv5_3(X_conv5_2) 81 | 82 | X_branch_1 = self.hbp(X_conv5_1,X_conv5_2) 83 | X_branch_2 = self.hbp(X_conv5_2,X_conv5_3) 84 | X_branch_3 = self.hbp(X_conv5_1,X_conv5_3) 85 | 86 | X_branch = torch.cat([X_branch_1,X_branch_2,X_branch_3],dim = 1) 87 | assert X_branch.size() == (N,8192*3) 88 | X = self.fc(X_branch) 89 | assert X.size() == (N, 200) 90 | return X 91 | 92 | class HBPManager(object): 93 | def __init__(self, options, path): 94 | self._options = options 95 | self._path = path 96 | # Network. 97 | self._net = torch.nn.DataParallel(HBP()).cuda() 98 | print(self._net) 99 | # Criterion. 100 | self._criterion = torch.nn.CrossEntropyLoss().cuda() 101 | # Solver. 102 | param_to_optim = [] 103 | for param in self._net.parameters(): 104 | if param.requires_grad == False: 105 | continue 106 | param_to_optim.append(param) 107 | 108 | self._solver = torch.optim.SGD( 109 | param_to_optim, lr=self._options['base_lr'], 110 | momentum=0.9, weight_decay=self._options['weight_decay']) 111 | 112 | milestones = [40,60,80,100] 113 | self._scheduler = torch.optim.lr_scheduler.MultiStepLR(self._solver,milestones = milestones,gamma=0.25) 114 | 115 | train_transforms = torchvision.transforms.Compose([ 116 | torchvision.transforms.Resize(size=448), # Let smaller edge match 117 | torchvision.transforms.RandomHorizontalFlip(), 118 | torchvision.transforms.RandomCrop(size=448), 119 | torchvision.transforms.ToTensor(), 120 | torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), 121 | std=(0.229, 0.224, 0.225)) 122 | ]) 123 | test_transforms = torchvision.transforms.Compose([ 124 | torchvision.transforms.Resize(size=448), 125 | torchvision.transforms.CenterCrop(size=448), 126 | torchvision.transforms.ToTensor(), 127 | torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), 128 | std=(0.229, 0.224, 0.225)) 129 | ]) 130 | train_data = cub200.CUB200( 131 | root=self._path['cub200'], train=True, download=True, 132 | transform=train_transforms) 133 | test_data = cub200.CUB200( 134 | root=self._path['cub200'], train=False, download=True, 135 | transform=test_transforms) 136 | self._train_loader = torch.utils.data.DataLoader( 137 | train_data, batch_size=self._options['batch_size'], 138 | shuffle=True, num_workers=4, pin_memory=True) 139 | self._test_loader = torch.utils.data.DataLoader( 140 | test_data, batch_size=16, 141 | shuffle=False, num_workers=4, pin_memory=True) 142 | 143 | def train(self): 144 | print('Training.') 145 | best_acc = 0.0 146 | best_epoch = None 147 | print('Epoch\tTrain loss\tTrain acc\tTest acc') 148 | ii = 0 149 | for t in range(self._options['epochs']): 150 | epoch_loss = [] 151 | num_correct = 0 152 | num_total = 0 153 | for X, y in self._train_loader: 154 | # Data. 155 | X = torch.autograd.Variable(X.cuda()) 156 | y = torch.autograd.Variable(y.cuda(non_blocking = True)) 157 | # Clear the existing gradients. 158 | self._solver.zero_grad() 159 | # Forward pass. 160 | score = self._net(X) 161 | loss = self._criterion(score, y) 162 | epoch_loss.append(loss.data[0]) 163 | # Prediction. 164 | _, prediction = torch.max(score.data, 1) 165 | num_total += y.size(0) 166 | num_correct += torch.sum(prediction == y.data) 167 | # Backward pass. 168 | loss.backward() 169 | self._solver.step() 170 | 171 | ii += 1 172 | x = torch.Tensor([ii]) 173 | y = torch.Tensor([loss.data[0]]) 174 | vis.line(X=x, Y=y, win='polynomial', update='append' if ii > 0 else None) 175 | 176 | num_correct = torch.tensor(num_correct).float().cuda() 177 | num_total = torch.tensor(num_total).float().cuda() 178 | 179 | train_acc = 100 * num_correct / num_total 180 | test_acc = self._accuracy(self._test_loader) 181 | self._scheduler.step(test_acc) 182 | if test_acc > best_acc: 183 | best_acc = test_acc 184 | best_epoch = t + 1 185 | print('*', end='') 186 | # Save model onto disk. 187 | torch.save(self._net.state_dict(), 188 | os.path.join(self._path['model'], 189 | 'HBP_fc_epoch_%d.pth' % (t + 1))) 190 | print('%d\t%4.3f\t\t%4.2f%%\t\t%4.2f%%' % 191 | (t+1, sum(epoch_loss) / len(epoch_loss), train_acc, test_acc)) 192 | print('Best at epoch %d, test accuaray %f' % (best_epoch, best_acc)) 193 | 194 | def _accuracy(self, data_loader): 195 | self._net.train(False) 196 | num_correct = 0 197 | num_total = 0 198 | for X, y in data_loader: 199 | # Data. 200 | X = torch.autograd.Variable(X.cuda()) 201 | y = torch.autograd.Variable(y.cuda(non_blocking = True)) 202 | # Prediction. 203 | score = self._net(X) 204 | _, prediction = torch.max(score.data, 1) 205 | num_total += y.size(0) 206 | num_correct += torch.sum(prediction == y.data) 207 | self._net.train(True) # Set the model to training phase 208 | num_correct = torch.tensor(num_correct).float().cuda() 209 | num_total = torch.tensor(num_total).float().cuda() 210 | return 100 * num_correct / num_total 211 | 212 | def getStat(self): 213 | print('Compute mean and variance for training data.') 214 | train_data = cub200.CUB200( 215 | root=self._path['cub200'], train=True, 216 | transform=torchvision.transforms.ToTensor(), download=True) 217 | train_loader = torch.utils.data.DataLoader( 218 | train_data, batch_size=1, shuffle=False, num_workers=4, 219 | pin_memory=True) 220 | mean = torch.zeros(3) 221 | std = torch.zeros(3) 222 | for X, _ in train_loader: 223 | for d in range(3): 224 | mean[d] += X[:, d, :, :].mean() 225 | std[d] += X[:, d, :, :].std() 226 | mean.div_(len(train_data)) 227 | std.div_(len(train_data)) 228 | print(mean) 229 | print(std) 230 | 231 | 232 | def main(): 233 | 234 | parser = argparse.ArgumentParser( 235 | description='Train HBP on CUB200.') 236 | parser.add_argument('--base_lr', dest='base_lr', type=float, required=True, 237 | help='Base learning rate for training.') 238 | parser.add_argument('--batch_size', dest='batch_size', type=int, 239 | required=True, help='Batch size.') 240 | parser.add_argument('--epochs', dest='epochs', type=int, 241 | required=True, help='Epochs for training.') 242 | parser.add_argument('--weight_decay', dest='weight_decay', type=float, 243 | required=True, help='Weight decay.') 244 | args = parser.parse_args() 245 | if args.base_lr <= 0: 246 | raise AttributeError('--base_lr parameter must >0.') 247 | if args.batch_size <= 0: 248 | raise AttributeError('--batch_size parameter must >0.') 249 | if args.epochs < 0: 250 | raise AttributeError('--epochs parameter must >=0.') 251 | if args.weight_decay <= 0: 252 | raise AttributeError('--weight_decay parameter must >0.') 253 | 254 | options = { 255 | 'base_lr': args.base_lr, 256 | 'batch_size': args.batch_size, 257 | 'epochs': args.epochs, 258 | 'weight_decay': args.weight_decay, 259 | } 260 | 261 | project_root = os.popen('pwd').read().strip() 262 | path = { 263 | 'cub200': os.path.join(project_root, 'data/cub200'), 264 | 'model': os.path.join(project_root, 'model'), 265 | } 266 | for d in path: 267 | assert os.path.isdir(path[d]) 268 | 269 | manager = HBPManager(options, path) 270 | manager.getStat() 271 | manager.train() 272 | 273 | if __name__ == '__main__': 274 | main() 275 | -------------------------------------------------------------------------------- /HBP_fc_new.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | """Fine-tune the fc layer only for HBP(Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition). 4 | Usage: 5 | CUDA_VISIBLE_DEVICES=0,1,2,3 python HBP_fc.py --base_lr 1.0 --batch_size 128 --epochs 240 --weight_decay 0.000005 | tee 'hbp_fc.log' 6 | """ 7 | 8 | import os 9 | import torch 10 | import torchvision 11 | import cub200 12 | import visdom 13 | import argparse 14 | 15 | vis = visdom.Visdom(env=u'HBP_fc',use_incoming_socket=False) 16 | torch.manual_seed(0) 17 | torch.cuda.manual_seed_all(0) 18 | 19 | class HBP(torch.nn.Module): 20 | def __init__(self): 21 | torch.nn.Module.__init__(self) 22 | # Convolution and pooling layers of VGG-16. 23 | self.features = torchvision.models.vgg16(pretrained=True).features 24 | self.features_conv5_1 = torch.nn.Sequential(*list(self.features.children()) 25 | [:-5]) 26 | self.features_conv5_2 = torch.nn.Sequential(*list(self.features.children()) 27 | [-5:-3]) 28 | self.features_conv5_3 = torch.nn.Sequential(*list(self.features.children()) 29 | [-3:-1]) 30 | self.bilinear_proj_1 = torch.nn.Conv2d(512,8192,kernel_size=1,bias=True) 31 | self.bilinear_proj_2 = torch.nn.Conv2d(512,8192,kernel_size=1,bias=True) 32 | self.bilinear_proj_3 = torch.nn.Conv2d(512,8192,kernel_size=1,bias=True) 33 | # Linear classifier. 34 | self.fc = torch.nn.Linear(8192*3, 200) 35 | 36 | # Freeze all previous layers. 37 | for param in self.features_conv5_1.parameters(): 38 | param.requires_grad = False 39 | for param in self.features_conv5_2.parameters(): 40 | param.requires_grad = False 41 | for param in self.features_conv5_3.parameters(): 42 | param.requires_grad = False 43 | 44 | # Initialize the fc layers. 45 | torch.nn.init.xavier_normal_(self.fc.weight.data) 46 | if self.fc.bias is not None: 47 | torch.nn.init.constant_(self.fc.bias.data, val=0) 48 | 49 | def hbp_1_2(self,conv1,conv2): 50 | N = conv1.size()[0] 51 | proj_1 = self.bilinear_proj_1(conv1) 52 | proj_2 = self.bilinear_proj_2(conv2) 53 | assert(proj_1.size() == (N,8192,28,28)) 54 | X = proj_1 * proj_2 55 | assert(X.size() == (N,8192,28,28)) 56 | X = torch.sum(X.view(X.size()[0],X.size()[1],-1),dim = 2) 57 | X = X.view(N, 8192) 58 | X = torch.sign(X) * torch.sqrt(torch.abs(X) + 1e-5) 59 | X = torch.nn.functional.normalize(X) 60 | return X 61 | 62 | def hbp_1_3(self,conv1,conv3): 63 | N = conv1.size()[0] 64 | proj_1 = self.bilinear_proj_1(conv1) 65 | proj_3 = self.bilinear_proj_3(conv3) 66 | assert(proj_1.size() == (N,8192,28,28)) 67 | X = proj_1 * proj_3 68 | assert(X.size() == (N,8192,28,28)) 69 | X = torch.sum(X.view(X.size()[0],X.size()[1],-1),dim = 2) 70 | X = X.view(N, 8192) 71 | X = torch.sign(X) * torch.sqrt(torch.abs(X) + 1e-5) 72 | X = torch.nn.functional.normalize(X) 73 | return X 74 | 75 | def hbp_2_3(self,conv2,conv3): 76 | N = conv2.size()[0] 77 | proj_2 = self.bilinear_proj_2(conv2) 78 | proj_3 = self.bilinear_proj_3(conv3) 79 | assert(proj_2.size() == (N,8192,28,28)) 80 | X = proj_2 * proj_3 81 | assert(X.size() == (N,8192,28,28)) 82 | X = torch.sum(X.view(X.size()[0],X.size()[1],-1),dim = 2) 83 | X = X.view(N, 8192) 84 | X = torch.sign(X) * torch.sqrt(torch.abs(X) + 1e-5) 85 | X = torch.nn.functional.normalize(X) 86 | return X 87 | 88 | def forward(self, X): 89 | N = X.size()[0] 90 | assert X.size() == (N, 3, 448, 448) 91 | X_conv5_1 = self.features_conv5_1(X) 92 | X_conv5_2 = self.features_conv5_2(X_conv5_1) 93 | X_conv5_3 = self.features_conv5_3(X_conv5_2) 94 | 95 | X_branch_1 = self.hbp_1_2(X_conv5_1,X_conv5_2) 96 | X_branch_2 = self.hbp_1_3(X_conv5_1,X_conv5_3) 97 | X_branch_3 = self.hbp_2_3(X_conv5_2,X_conv5_3) 98 | 99 | 100 | X_branch = torch.cat([X_branch_1,X_branch_2,X_branch_3],dim = 1) 101 | assert X_branch.size() == (N,8192*3) 102 | X = self.fc(X_branch) 103 | assert X.size() == (N, 200) 104 | return X 105 | 106 | class HBPManager(object): 107 | def __init__(self, options, path): 108 | self._options = options 109 | self._path = path 110 | # Network. 111 | self._net = torch.nn.DataParallel(HBP()).cuda() 112 | print(self._net) 113 | # Criterion. 114 | self._criterion = torch.nn.CrossEntropyLoss().cuda() 115 | # Solver. 116 | param_to_optim = [] 117 | for param in self._net.parameters(): 118 | if param.requires_grad == False: 119 | continue 120 | param_to_optim.append(param) 121 | 122 | self._solver = torch.optim.SGD( 123 | param_to_optim, lr=self._options['base_lr'], 124 | momentum=0.9, weight_decay=self._options['weight_decay']) 125 | 126 | self._scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(self._solver) 127 | 128 | # milestones = [40,60,80,100] 129 | # self._scheduler = torch.optim.lr_scheduler.MultiStepLR(self._solver,milestones = milestones,gamma=0.25) 130 | 131 | train_transforms = torchvision.transforms.Compose([ 132 | torchvision.transforms.Resize(size=448), # Let smaller edge match 133 | torchvision.transforms.RandomHorizontalFlip(), 134 | torchvision.transforms.RandomCrop(size=448), 135 | torchvision.transforms.ToTensor(), 136 | torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), 137 | std=(0.229, 0.224, 0.225)) 138 | ]) 139 | test_transforms = torchvision.transforms.Compose([ 140 | torchvision.transforms.Resize(size=448), 141 | torchvision.transforms.CenterCrop(size=448), 142 | torchvision.transforms.ToTensor(), 143 | torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), 144 | std=(0.229, 0.224, 0.225)) 145 | ]) 146 | train_data = cub200.CUB200( 147 | root=self._path['cub200'], train=True, download=True, 148 | transform=train_transforms) 149 | test_data = cub200.CUB200( 150 | root=self._path['cub200'], train=False, download=True, 151 | transform=test_transforms) 152 | self._train_loader = torch.utils.data.DataLoader( 153 | train_data, batch_size=self._options['batch_size'], 154 | shuffle=True, num_workers=4, pin_memory=True) 155 | self._test_loader = torch.utils.data.DataLoader( 156 | test_data, batch_size=16, 157 | shuffle=False, num_workers=4, pin_memory=True) 158 | 159 | def train(self): 160 | print('Training.') 161 | best_acc = 0.0 162 | best_epoch = None 163 | print('Epoch\tTrain loss\tTrain acc\tTest acc') 164 | ii = 0 165 | for t in range(self._options['epochs']): 166 | epoch_loss = [] 167 | num_correct = 0 168 | num_total = 0 169 | for X, y in self._train_loader: 170 | # Data. 171 | X = torch.autograd.Variable(X.cuda()) 172 | y = torch.autograd.Variable(y.cuda(non_blocking = True)) 173 | # Clear the existing gradients. 174 | self._solver.zero_grad() 175 | # Forward pass. 176 | score = self._net(X) 177 | loss = self._criterion(score, y) 178 | epoch_loss.append(loss.data[0]) 179 | # Prediction. 180 | _, prediction = torch.max(score.data, 1) 181 | num_total += y.size(0) 182 | num_correct += torch.sum(prediction == y.data) 183 | # Backward pass. 184 | loss.backward() 185 | self._solver.step() 186 | 187 | ii += 1 188 | x = torch.Tensor([ii]) 189 | y = torch.Tensor([loss.data[0]]) 190 | vis.line(X=x, Y=y, win='polynomial', update='append' if ii > 0 else None) 191 | 192 | num_correct = torch.tensor(num_correct).float().cuda() 193 | num_total = torch.tensor(num_total).float().cuda() 194 | 195 | train_acc = 100 * num_correct / num_total 196 | test_acc = self._accuracy(self._test_loader) 197 | self._scheduler.step(test_acc) 198 | if test_acc > best_acc: 199 | best_acc = test_acc 200 | best_epoch = t + 1 201 | print('*', end='') 202 | # Save model onto disk. 203 | torch.save(self._net.state_dict(), 204 | os.path.join(self._path['model'], 205 | 'HBP_fc_epoch_%d.pth' % (t + 1))) 206 | print('%d\t%4.3f\t\t%4.2f%%\t\t%4.2f%%' % 207 | (t+1, sum(epoch_loss) / len(epoch_loss), train_acc, test_acc)) 208 | print('Best at epoch %d, test accuaray %f' % (best_epoch, best_acc)) 209 | 210 | def _accuracy(self, data_loader): 211 | self._net.train(False) 212 | num_correct = 0 213 | num_total = 0 214 | for X, y in data_loader: 215 | # Data. 216 | X = torch.autograd.Variable(X.cuda()) 217 | y = torch.autograd.Variable(y.cuda(non_blocking = True)) 218 | # Prediction. 219 | score = self._net(X) 220 | _, prediction = torch.max(score.data, 1) 221 | num_total += y.size(0) 222 | num_correct += torch.sum(prediction == y.data) 223 | self._net.train(True) # Set the model to training phase 224 | num_correct = torch.tensor(num_correct).float().cuda() 225 | num_total = torch.tensor(num_total).float().cuda() 226 | return 100 * num_correct / num_total 227 | 228 | def getStat(self): 229 | print('Compute mean and variance for training data.') 230 | train_data = cub200.CUB200( 231 | root=self._path['cub200'], train=True, 232 | transform=torchvision.transforms.ToTensor(), download=True) 233 | train_loader = torch.utils.data.DataLoader( 234 | train_data, batch_size=1, shuffle=False, num_workers=4, 235 | pin_memory=True) 236 | mean = torch.zeros(3) 237 | std = torch.zeros(3) 238 | for X, _ in train_loader: 239 | for d in range(3): 240 | mean[d] += X[:, d, :, :].mean() 241 | std[d] += X[:, d, :, :].std() 242 | mean.div_(len(train_data)) 243 | std.div_(len(train_data)) 244 | print(mean) 245 | print(std) 246 | 247 | 248 | def main(): 249 | 250 | parser = argparse.ArgumentParser( 251 | description='Train HBP on CUB200.') 252 | parser.add_argument('--base_lr', dest='base_lr', type=float, required=True, 253 | help='Base learning rate for training.') 254 | parser.add_argument('--batch_size', dest='batch_size', type=int, 255 | required=True, help='Batch size.') 256 | parser.add_argument('--epochs', dest='epochs', type=int, 257 | required=True, help='Epochs for training.') 258 | parser.add_argument('--weight_decay', dest='weight_decay', type=float, 259 | required=True, help='Weight decay.') 260 | args = parser.parse_args() 261 | if args.base_lr <= 0: 262 | raise AttributeError('--base_lr parameter must >0.') 263 | if args.batch_size <= 0: 264 | raise AttributeError('--batch_size parameter must >0.') 265 | if args.epochs < 0: 266 | raise AttributeError('--epochs parameter must >=0.') 267 | if args.weight_decay <= 0: 268 | raise AttributeError('--weight_decay parameter must >0.') 269 | 270 | options = { 271 | 'base_lr': args.base_lr, 272 | 'batch_size': args.batch_size, 273 | 'epochs': args.epochs, 274 | 'weight_decay': args.weight_decay, 275 | } 276 | 277 | project_root = os.popen('pwd').read().strip() 278 | path = { 279 | 'cub200': os.path.join(project_root, 'data/cub200'), 280 | 'model': os.path.join(project_root, 'model'), 281 | } 282 | for d in path: 283 | assert os.path.isdir(path[d]) 284 | 285 | manager = HBPManager(options, path) 286 | manager.getStat() 287 | manager.train() 288 | 289 | if __name__ == '__main__': 290 | main() 291 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # HBP-pytorch 2 | ![]() ![]() ![]() 3 | 4 | ### **Overview** 5 | 6 | --- 7 | 8 | A third-party reimplementation of Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition in Pytorch. 9 | 10 | The related paper is as follows: 11 | 12 | Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition[C] 13 | Chaojian Yu, Xinyi Zhao, Qi Zheng, Peng Zhang, Xinge You* 14 | European Conference on Computer Vision. 2018. 15 | 16 | Official Caffe implementation of Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition is [HERE](https://github.com/ChaojianYu/Hierarchical-Bilinear-Pooling). 17 | 18 | ### **Preparation** 19 | 20 | --- 21 | 22 | **Dataset** 23 | 24 | - [CUB-200-2011]() 25 | - If you don't have a dataset, you can still execute the program and the program will automatically download the dataset. 26 | 27 | **Requirement** 28 | 29 | - pip install visdom pytorch torchvision 30 | 31 | ### **Usage** 32 | 33 | --- 34 | 35 | for example: 36 | 37 | - CUDA_VISIBLE_DEVICES=0,1 python HBP_fc.py --base_lr 1.0 --batch_size 12 --epochs 120 --weight_decay 0.000005 | tee 'hbp_fc.log' 38 | - CUDA_VISIBLE_DEVICES=0,1 python HBP_all.py --base_lr 0.001 --batch_size 24 --epochs 200 --weight_decay 0.0005 --model 'HBP_fc_epoch_*.pth' | tee 'hbp_all.log' 39 | 40 | ### **Result** 41 | 42 | --- 43 | 44 | | file | acc | 45 | | ---------- | :---: | 46 | | HBP_fc | 80.42 | 47 | | HBP_fc_new | 79.79 | 48 | | HBP_all | 80.42 | 49 | 50 | *Note that `HBP_fc_new.py` may be the closest to the original implementation. But it still doesn't work well.* 51 | 52 | ### **Last** 53 | 54 | --- 55 | 56 | Based on my code and experimental results, it is far from the result of the original author. So you can use it as a reference for learning. 57 | 58 | This code borrows from [HERE](https://github.com/HaoMood/bilinear-cnn). If you have any suggestions please contact me, I am still continue to improve the results. 59 | 60 | Happy coding. 61 | -------------------------------------------------------------------------------- /cub200.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -* 2 | """This module is served as torchvision.datasets to load CUB200-2011. 3 | CUB200-2011 dataset has 11,788 images of 200 bird species. The project page 4 | is as follows. 5 | http://www.vision.caltech.edu/visipedia/CUB-200-2011.html 6 | - Images are contained in the directory data/cub200/raw/images/, 7 | with 200 subdirectories. 8 | - Format of images.txt: 9 | - Format of train_test_split.txt: 10 | - Format of classes.txt: 11 | - Format of iamge_class_labels.txt: 12 | This file is modified from: 13 | https://github.com/vishwakftw/vision. 14 | """ 15 | 16 | 17 | import os 18 | import pickle 19 | 20 | import numpy as np 21 | import PIL.Image 22 | import torch 23 | 24 | 25 | __all__ = ['CUB200'] 26 | __author__ = 'Hao Zhang' 27 | __copyright__ = '2018 LAMDA' 28 | __date__ = '2018-01-09' 29 | __email__ = 'zhangh0214@gmail.com' 30 | __license__ = 'CC BY-SA 3.0' 31 | __status__ = 'Development' 32 | __updated__ = '2018-01-10' 33 | __version__ = '1.0' 34 | 35 | 36 | class CUB200(torch.utils.data.Dataset): 37 | """CUB200 dataset. 38 | Args: 39 | _root, str: Root directory of the dataset. 40 | _train, bool: Load train/test data. 41 | _transform, callable: A function/transform that takes in a PIL.Image 42 | and transforms it. 43 | _target_transform, callable: A function/transform that takes in the 44 | target and transforms it. 45 | _train_data, list of np.ndarray. 46 | _train_labels, list of int. 47 | _test_data, list of np.ndarray. 48 | _test_labels, list of int. 49 | """ 50 | def __init__(self, root, train=True, transform=None, target_transform=None, 51 | download=False): 52 | """Load the dataset. 53 | Args 54 | root, str: Root directory of the dataset. 55 | train, bool [True]: Load train/test data. 56 | transform, callable [None]: A function/transform that takes in a 57 | PIL.Image and transforms it. 58 | target_transform, callable [None]: A function/transform that takes 59 | in the target and transforms it. 60 | download, bool [False]: If true, downloads the dataset from the 61 | internet and puts it in root directory. If dataset is already 62 | downloaded, it is not downloaded again. 63 | """ 64 | self._root = os.path.expanduser(root) # Replace ~ by the complete dir 65 | self._train = train 66 | self._transform = transform 67 | self._target_transform = target_transform 68 | 69 | if self._checkIntegrity(): 70 | print('Files already downloaded and verified.') 71 | elif download: 72 | url = ('http://www.vision.caltech.edu/visipedia-data/CUB-200-2011/' 73 | 'CUB_200_2011.tgz') 74 | self._download(url) 75 | self._extract() 76 | else: 77 | raise RuntimeError( 78 | 'Dataset not found. You can use download=True to download it.') 79 | 80 | # Now load the picked data. 81 | if self._train: 82 | self._train_data, self._train_labels = pickle.load(open( 83 | os.path.join(self._root, 'processed/train.pkl'), 'rb'),encoding='iso-8859-1') 84 | assert (len(self._train_data) == 5994 85 | and len(self._train_labels) == 5994) 86 | else: 87 | self._test_data, self._test_labels = pickle.load(open( 88 | os.path.join(self._root, 'processed/test.pkl'), 'rb'),encoding='iso-8859-1') 89 | assert (len(self._test_data) == 5794 90 | and len(self._test_labels) == 5794) 91 | 92 | def __getitem__(self, index): 93 | """ 94 | Args: 95 | index, int: Index. 96 | Returns: 97 | image, PIL.Image: Image of the given index. 98 | target, str: target of the given index. 99 | """ 100 | if self._train: 101 | image, target = self._train_data[index], self._train_labels[index] 102 | else: 103 | image, target = self._test_data[index], self._test_labels[index] 104 | # Doing this so that it is consistent with all other datasets. 105 | image = PIL.Image.fromarray(image) 106 | 107 | if self._transform is not None: 108 | image = self._transform(image) 109 | if self._target_transform is not None: 110 | target = self._target_transform(target) 111 | 112 | return image, target 113 | 114 | def __len__(self): 115 | """Length of the dataset. 116 | Returns: 117 | length, int: Length of the dataset. 118 | """ 119 | if self._train: 120 | return len(self._train_data) 121 | return len(self._test_data) 122 | 123 | def _checkIntegrity(self): 124 | """Check whether we have already processed the data. 125 | Returns: 126 | flag, bool: True if we have already processed the data. 127 | """ 128 | return ( 129 | os.path.isfile(os.path.join(self._root, 'processed/train.pkl')) 130 | and os.path.isfile(os.path.join(self._root, 'processed/test.pkl'))) 131 | 132 | def _download(self, url): 133 | """Download and uncompress the tar.gz file from a given URL. 134 | Args: 135 | url, str: URL to be downloaded. 136 | """ 137 | import six.moves 138 | import tarfile 139 | 140 | raw_path = os.path.join(self._root, 'raw') 141 | processed_path = os.path.join(self._root, 'processed') 142 | if not os.path.isdir(raw_path): 143 | os.mkdir(raw_path, mode=0o775) 144 | if not os.path.isdir(processed_path): 145 | os.mkdir(processed_path, mode=0x775) 146 | 147 | # Downloads file. 148 | fpath = os.path.join(self._root, 'raw/CUB_200_2011.tgz') 149 | try: 150 | print('Downloading ' + url + ' to ' + fpath) 151 | six.moves.urllib.request.urlretrieve(url, fpath) 152 | except six.moves.urllib.error.URLError: 153 | if url[:5] == 'https:': 154 | self._url = self._url.replace('https:', 'http:') 155 | print('Failed download. Trying https -> http instead.') 156 | print('Downloading ' + url + ' to ' + fpath) 157 | six.moves.urllib.request.urlretrieve(url, fpath) 158 | 159 | # Extract file. 160 | cwd = os.getcwd() 161 | tar = tarfile.open(fpath, 'r:gz') 162 | os.chdir(os.path.join(self._root, 'raw')) 163 | tar.extractall() 164 | tar.close() 165 | os.chdir(cwd) 166 | 167 | def _extract(self): 168 | """Prepare the data for train/test split and save onto disk.""" 169 | image_path = os.path.join(self._root, 'raw/CUB_200_2011/images/') 170 | # Format of images.txt: 171 | id2name = np.genfromtxt(os.path.join( 172 | self._root, 'raw/CUB_200_2011/images.txt'), dtype=str) 173 | # Format of train_test_split.txt: 174 | id2train = np.genfromtxt(os.path.join( 175 | self._root, 'raw/CUB_200_2011/train_test_split.txt'), dtype=int) 176 | 177 | train_data = [] 178 | train_labels = [] 179 | test_data = [] 180 | test_labels = [] 181 | for id_ in range(id2name.shape[0]): 182 | image = PIL.Image.open(os.path.join(image_path, id2name[id_, 1])) 183 | label = int(id2name[id_, 1][:3]) - 1 # Label starts with 0 184 | 185 | # Convert gray scale image to RGB image. 186 | if image.getbands()[0] == 'L': 187 | image = image.convert('RGB') 188 | image_np = np.array(image) 189 | image.close() 190 | 191 | if id2train[id_, 1] == 1: 192 | train_data.append(image_np) 193 | train_labels.append(label) 194 | else: 195 | test_data.append(image_np) 196 | test_labels.append(label) 197 | 198 | pickle.dump((train_data, train_labels), 199 | open(os.path.join(self._root, 'processed/train.pkl'), 'wb')) 200 | pickle.dump((test_data, test_labels), 201 | open(os.path.join(self._root, 'processed/test.pkl'), 'wb')) 202 | -------------------------------------------------------------------------------- /log/README.md: -------------------------------------------------------------------------------- 1 | hbp_all.log uses 2 | ```python 3 | self._scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( 4 | self._solver, mode='max', factor=0.1, patience=5, verbose=True, 5 | threshold=1e-4) 6 | ``` 7 | to reduce learning rate. 8 | -------------------------------------------------------------------------------- /log/hbp_all.log: -------------------------------------------------------------------------------- 1 | Prepare the network and data. 2 | DataParallel( 3 | (module): BCNN( 4 | (features): Sequential( 5 | (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 6 | (1): ReLU(inplace) 7 | (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 8 | (3): ReLU(inplace) 9 | (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 10 | (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 11 | (6): ReLU(inplace) 12 | (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 13 | (8): ReLU(inplace) 14 | (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 15 | (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 16 | (11): ReLU(inplace) 17 | (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 18 | (13): ReLU(inplace) 19 | (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 20 | (15): ReLU(inplace) 21 | (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 22 | (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 23 | (18): ReLU(inplace) 24 | (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 25 | (20): ReLU(inplace) 26 | (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 27 | (22): ReLU(inplace) 28 | (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 29 | (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 30 | (25): ReLU(inplace) 31 | (26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 32 | (27): ReLU(inplace) 33 | (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 34 | (29): ReLU(inplace) 35 | (30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 36 | ) 37 | (features_conv5_1): Sequential( 38 | (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 39 | (1): ReLU(inplace) 40 | (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 41 | (3): ReLU(inplace) 42 | (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 43 | (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 44 | (6): ReLU(inplace) 45 | (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 46 | (8): ReLU(inplace) 47 | (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 48 | (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 49 | (11): ReLU(inplace) 50 | (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 51 | (13): ReLU(inplace) 52 | (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 53 | (15): ReLU(inplace) 54 | (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 55 | (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 56 | (18): ReLU(inplace) 57 | (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 58 | (20): ReLU(inplace) 59 | (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 60 | (22): ReLU(inplace) 61 | (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 62 | (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 63 | (25): ReLU(inplace) 64 | ) 65 | (features_conv5_2): Sequential( 66 | (0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 67 | (1): ReLU(inplace) 68 | ) 69 | (features_conv5_3): Sequential( 70 | (0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 71 | (1): ReLU(inplace) 72 | ) 73 | (bilinear_proj): Sequential( 74 | (0): Conv2d(512, 8192, kernel_size=(1, 1), stride=(1, 1), bias=False) 75 | (1): BatchNorm2d(8192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 76 | (2): ReLU(inplace) 77 | ) 78 | (fc): Linear(in_features=24576, out_features=200, bias=True) 79 | ) 80 | ) 81 | Files already downloaded and verified. 82 | Files already downloaded and verified. 83 | Training. 84 | Epoch Train loss Train acc Test acc 85 | *1 0.942 76.66% 66.43% 86 | *2 0.653 84.25% 73.32% 87 | *3 0.346 93.41% 77.93% 88 | *4 0.186 97.46% 79.27% 89 | *5 0.130 98.65% 80.38% 90 | *6 0.077 99.55% 81.24% 91 | *7 0.057 99.88% 82.33% 92 | *8 0.047 99.87% 83.14% 93 | *9 0.043 99.87% 83.19% 94 | 10 0.039 99.93% 82.76% 95 | *11 0.038 99.97% 83.48% 96 | 12 0.041 99.88% 83.31% 97 | *13 0.037 99.97% 83.53% 98 | *14 0.037 99.98% 83.79% 99 | 15 0.036 100.00% 83.59% 100 | 16 0.036 100.00% 83.79% 101 | *17 0.039 99.98% 84.09% 102 | 18 0.040 99.97% 84.02% 103 | 19 0.039 100.00% 83.53% 104 | 20 0.041 99.97% 83.64% 105 | *21 0.041 99.98% 84.29% 106 | 22 0.046 99.95% 84.02% 107 | 23 0.044 100.00% 83.53% 108 | 24 0.047 99.98% 83.69% 109 | 25 0.050 99.95% 83.95% 110 | 26 0.053 100.00% 83.71% 111 | Epoch 26: reducing learning rate of group 0 to 1.0000e-03. 112 | 27 0.054 99.97% 83.85% 113 | *28 0.048 99.95% 84.40% 114 | 29 0.047 99.98% 84.24% 115 | 30 0.046 100.00% 84.24% 116 | 31 0.045 100.00% 84.38% 117 | 32 0.046 100.00% 84.14% 118 | 33 0.045 99.98% 84.35% 119 | Epoch 33: reducing learning rate of group 0 to 1.0000e-04. 120 | 34 0.045 99.98% 84.23% 121 | 35 0.045 100.00% 84.29% 122 | 36 0.044 100.00% 84.35% 123 | *37 0.046 99.98% 84.41% 124 | 38 0.045 100.00% 84.36% 125 | *39 0.046 100.00% 84.43% 126 | 40 0.045 99.98% 84.26% 127 | *41 0.044 100.00% 84.48% 128 | 42 0.046 99.98% 84.28% 129 | 43 0.045 100.00% 84.29% 130 | 44 0.045 100.00% 84.31% 131 | 45 0.046 99.98% 84.36% 132 | 46 0.046 99.98% 84.29% 133 | Epoch 46: reducing learning rate of group 0 to 1.0000e-05. 134 | 47 0.045 99.98% 84.29% 135 | 48 0.045 100.00% 84.35% 136 | 49 0.046 100.00% 84.29% 137 | 50 0.045 100.00% 84.29% 138 | 51 0.046 100.00% 84.23% 139 | 52 0.046 100.00% 84.28% 140 | Epoch 52: reducing learning rate of group 0 to 1.0000e-06. 141 | 53 0.045 99.98% 84.38% 142 | 54 0.045 100.00% 84.36% 143 | 55 0.045 100.00% 84.35% 144 | 56 0.045 100.00% 84.41% 145 | 57 0.045 100.00% 84.24% 146 | 58 0.044 100.00% 84.26% 147 | Epoch 58: reducing learning rate of group 0 to 1.0000e-07. 148 | 59 0.045 100.00% 84.31% 149 | 60 0.045 99.98% 84.38% 150 | 61 0.044 100.00% 84.38% 151 | 62 0.046 99.97% 84.28% 152 | 63 0.045 99.98% 84.33% 153 | 64 0.044 100.00% 84.21% 154 | Epoch 64: reducing learning rate of group 0 to 1.0000e-08. 155 | 65 0.045 100.00% 84.26% 156 | 66 0.045 99.97% 84.36% 157 | 67 0.047 99.98% 84.40% 158 | 68 0.045 99.97% 84.21% 159 | 69 0.044 100.00% 84.31% 160 | 70 0.046 99.97% 84.35% 161 | 71 0.046 99.98% 84.36% 162 | 72 0.044 100.00% 84.38% 163 | 73 0.046 99.97% 84.36% 164 | -------------------------------------------------------------------------------- /log/hbp_fc.log: -------------------------------------------------------------------------------- 1 | Prepare the network and data. 2 | DataParallel( 3 | (module): BCNN( 4 | (features): Sequential( 5 | (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 6 | (1): ReLU(inplace) 7 | (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 8 | (3): ReLU(inplace) 9 | (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 10 | (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 11 | (6): ReLU(inplace) 12 | (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 13 | (8): ReLU(inplace) 14 | (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 15 | (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 16 | (11): ReLU(inplace) 17 | (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 18 | (13): ReLU(inplace) 19 | (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 20 | (15): ReLU(inplace) 21 | (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 22 | (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 23 | (18): ReLU(inplace) 24 | (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 25 | (20): ReLU(inplace) 26 | (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 27 | (22): ReLU(inplace) 28 | (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 29 | (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 30 | (25): ReLU(inplace) 31 | (26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 32 | (27): ReLU(inplace) 33 | (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 34 | (29): ReLU(inplace) 35 | (30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 36 | ) 37 | (features_conv5_1): Sequential( 38 | (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 39 | (1): ReLU(inplace) 40 | (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 41 | (3): ReLU(inplace) 42 | (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 43 | (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 44 | (6): ReLU(inplace) 45 | (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 46 | (8): ReLU(inplace) 47 | (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 48 | (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 49 | (11): ReLU(inplace) 50 | (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 51 | (13): ReLU(inplace) 52 | (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 53 | (15): ReLU(inplace) 54 | (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 55 | (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 56 | (18): ReLU(inplace) 57 | (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 58 | (20): ReLU(inplace) 59 | (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 60 | (22): ReLU(inplace) 61 | (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 62 | (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 63 | (25): ReLU(inplace) 64 | ) 65 | (features_conv5_2): Sequential( 66 | (0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 67 | (1): ReLU(inplace) 68 | ) 69 | (features_conv5_3): Sequential( 70 | (0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 71 | (1): ReLU(inplace) 72 | ) 73 | (bilinear_proj): Sequential( 74 | (0): Conv2d(512, 8192, kernel_size=(1, 1), stride=(1, 1), bias=False) 75 | (1): BatchNorm2d(8192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 76 | (2): ReLU(inplace) 77 | ) 78 | (fc): Linear(in_features=24576, out_features=200, bias=True) 79 | ) 80 | ) 81 | Files already downloaded and verified. 82 | Files already downloaded and verified. 83 | Training. 84 | Epoch Train loss Train acc Test acc 85 | *1 4.094 15.90% 39.51% 86 | *2 1.655 55.96% 57.46% 87 | *3 0.724 83.10% 74.78% 88 | *4 0.542 89.32% 76.54% 89 | *5 0.499 90.86% 76.98% 90 | 6 0.474 91.79% 76.96% 91 | *7 0.450 92.16% 77.30% 92 | 8 0.425 92.83% 77.22% 93 | *9 0.411 93.44% 77.82% 94 | *10 0.389 94.18% 77.93% 95 | *11 0.375 94.54% 78.36% 96 | 12 0.356 95.30% 77.98% 97 | 13 0.346 95.60% 77.82% 98 | 14 0.329 95.91% 78.24% 99 | *15 0.311 96.18% 78.51% 100 | 16 0.293 97.01% 78.18% 101 | *17 0.284 97.06% 78.60% 102 | 18 0.271 97.60% 78.27% 103 | 19 0.256 97.80% 78.27% 104 | *20 0.250 97.86% 78.77% 105 | 21 0.240 98.18% 78.51% 106 | 22 0.223 98.38% 78.77% 107 | *23 0.215 98.48% 78.87% 108 | *24 0.206 98.58% 79.13% 109 | 25 0.196 98.73% 78.79% 110 | 26 0.194 98.92% 79.10% 111 | *27 0.182 99.07% 79.25% 112 | 28 0.177 99.10% 79.00% 113 | 29 0.170 99.23% 79.15% 114 | *30 0.161 99.32% 79.48% 115 | *31 0.156 99.27% 79.58% 116 | 32 0.152 99.32% 79.12% 117 | 33 0.145 99.52% 78.72% 118 | 34 0.139 99.55% 79.48% 119 | 35 0.131 99.68% 79.32% 120 | 36 0.125 99.63% 79.55% 121 | *37 0.126 99.62% 79.74% 122 | 38 0.119 99.77% 79.43% 123 | 39 0.117 99.70% 79.44% 124 | 40 0.112 99.73% 79.22% 125 | 41 0.110 99.67% 79.58% 126 | 42 0.106 99.72% 79.57% 127 | 43 0.103 99.83% 79.46% 128 | 44 0.100 99.83% 79.55% 129 | 45 0.098 99.82% 79.67% 130 | 46 0.095 99.80% 79.63% 131 | 47 0.091 99.87% 79.46% 132 | 48 0.089 99.85% 79.41% 133 | 49 0.085 99.87% 79.60% 134 | 50 0.086 99.88% 79.69% 135 | 51 0.081 99.88% 79.53% 136 | *52 0.080 99.83% 79.77% 137 | 53 0.077 99.88% 79.65% 138 | 54 0.076 99.87% 79.77% 139 | 55 0.075 99.90% 79.72% 140 | 56 0.073 99.90% 79.62% 141 | 57 0.069 99.97% 79.58% 142 | *58 0.069 99.92% 79.88% 143 | *59 0.070 99.88% 80.15% 144 | 60 0.064 99.92% 79.98% 145 | 61 0.065 99.95% 80.10% 146 | 62 0.062 99.95% 80.07% 147 | 63 0.061 99.92% 79.75% 148 | 64 0.064 100.00% 79.58% 149 | 65 0.062 99.92% 80.15% 150 | 66 0.059 99.95% 80.05% 151 | 67 0.058 99.95% 80.13% 152 | 68 0.057 99.97% 79.81% 153 | 69 0.061 99.93% 79.88% 154 | 70 0.059 99.93% 79.93% 155 | 71 0.060 99.93% 79.82% 156 | 72 0.057 100.00% 79.57% 157 | 73 0.056 99.97% 80.01% 158 | 74 0.053 99.93% 80.10% 159 | 75 0.054 99.93% 79.94% 160 | 76 0.055 99.93% 80.03% 161 | *77 0.052 99.95% 80.17% 162 | 78 0.051 99.95% 80.08% 163 | 79 0.052 99.95% 80.01% 164 | *80 0.050 99.98% 80.43% 165 | 81 0.048 100.00% 79.55% 166 | 82 0.051 99.97% 80.24% 167 | 83 0.050 99.95% 79.75% 168 | 84 0.050 99.98% 80.17% 169 | 85 0.049 99.97% 80.17% 170 | 86 0.048 99.95% 79.89% 171 | 87 0.051 99.95% 80.08% 172 | 88 0.047 99.97% 80.08% 173 | 89 0.048 99.95% 79.84% 174 | 90 0.048 99.95% 80.17% 175 | 91 0.048 99.85% 80.10% 176 | 92 0.045 100.00% 79.88% 177 | 93 0.046 99.95% 80.34% 178 | 94 0.047 99.93% 80.17% 179 | 95 0.045 99.98% 80.08% 180 | 96 0.043 100.00% 80.20% 181 | 97 0.043 99.95% 80.12% 182 | 98 0.044 100.00% 80.22% 183 | 99 0.044 99.95% 79.67% 184 | 100 0.046 99.93% 80.39% 185 | 101 0.045 99.93% 80.03% 186 | 102 0.043 99.97% 79.98% 187 | 103 0.044 99.97% 79.96% 188 | 104 0.045 99.98% 80.13% 189 | 105 0.042 99.93% 80.27% 190 | 106 0.040 100.00% 80.07% 191 | 107 0.042 99.98% 80.22% 192 | 108 0.039 100.00% 80.19% 193 | 109 0.040 99.98% 80.41% 194 | 110 0.040 99.98% 80.24% 195 | 111 0.040 99.93% 80.19% 196 | 112 0.040 99.97% 80.15% 197 | 113 0.041 99.98% 80.05% 198 | 114 0.040 100.00% 80.03% 199 | 115 0.040 99.98% 79.96% 200 | 116 0.041 99.98% 79.93% 201 | 117 0.042 99.95% 80.22% 202 | 118 0.039 99.97% 80.12% 203 | 119 0.039 99.95% 79.82% 204 | 120 0.040 100.00% 80.31% 205 | Best at epoch 80, test accuaray 80.428032 206 | -------------------------------------------------------------------------------- /log/hbp_fc_new.log: -------------------------------------------------------------------------------- 1 | DataParallel( 2 | (module): HBP( 3 | (features): Sequential( 4 | (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 5 | (1): ReLU(inplace) 6 | (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 7 | (3): ReLU(inplace) 8 | (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 9 | (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 10 | (6): ReLU(inplace) 11 | (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 12 | (8): ReLU(inplace) 13 | (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 14 | (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 15 | (11): ReLU(inplace) 16 | (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 17 | (13): ReLU(inplace) 18 | (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 19 | (15): ReLU(inplace) 20 | (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 21 | (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 22 | (18): ReLU(inplace) 23 | (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 24 | (20): ReLU(inplace) 25 | (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 26 | (22): ReLU(inplace) 27 | (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 28 | (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 29 | (25): ReLU(inplace) 30 | (26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 31 | (27): ReLU(inplace) 32 | (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 33 | (29): ReLU(inplace) 34 | (30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 35 | ) 36 | (features_conv5_1): Sequential( 37 | (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 38 | (1): ReLU(inplace) 39 | (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 40 | (3): ReLU(inplace) 41 | (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 42 | (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 43 | (6): ReLU(inplace) 44 | (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 45 | (8): ReLU(inplace) 46 | (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 47 | (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 48 | (11): ReLU(inplace) 49 | (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 50 | (13): ReLU(inplace) 51 | (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 52 | (15): ReLU(inplace) 53 | (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 54 | (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 55 | (18): ReLU(inplace) 56 | (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 57 | (20): ReLU(inplace) 58 | (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 59 | (22): ReLU(inplace) 60 | (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) 61 | (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 62 | (25): ReLU(inplace) 63 | ) 64 | (features_conv5_2): Sequential( 65 | (0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 66 | (1): ReLU(inplace) 67 | ) 68 | (features_conv5_3): Sequential( 69 | (0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) 70 | (1): ReLU(inplace) 71 | ) 72 | (bilinear_proj_1): Conv2d(512, 8192, kernel_size=(1, 1), stride=(1, 1)) 73 | (bilinear_proj_2): Conv2d(512, 8192, kernel_size=(1, 1), stride=(1, 1)) 74 | (bilinear_proj_3): Conv2d(512, 8192, kernel_size=(1, 1), stride=(1, 1)) 75 | (fc): Linear(in_features=24576, out_features=200, bias=True) 76 | ) 77 | ) 78 | Files already downloaded and verified. 79 | Files already downloaded and verified. 80 | Compute mean and variance for training data. 81 | Files already downloaded and verified. 82 | tensor([0.4856, 0.4994, 0.4324]) 83 | tensor([0.1817, 0.1811, 0.1927]) 84 | Training. 85 | Epoch Train loss Train acc Test acc 86 | *1 4.912 13.66% 39.02% 87 | *2 3.484 47.75% 54.33% 88 | *3 2.454 64.88% 64.38% 89 | *4 1.844 75.04% 69.88% 90 | *5 1.456 80.26% 72.42% 91 | *6 1.200 84.67% 74.47% 92 | *7 1.002 87.10% 75.77% 93 | *8 0.867 89.36% 77.06% 94 | *9 0.751 91.02% 77.58% 95 | *10 0.665 92.39% 78.13% 96 | 11 0.589 93.88% 78.06% 97 | *12 0.522 94.53% 78.44% 98 | *13 0.439 96.13% 78.86% 99 | *14 0.411 96.46% 79.13% 100 | *15 0.396 96.58% 79.22% 101 | 16 0.389 96.61% 79.19% 102 | *17 0.378 96.66% 79.48% 103 | 18 0.375 96.93% 79.41% 104 | *19 0.366 97.08% 79.72% 105 | 20 0.358 97.23% 79.55% 106 | 21 0.354 97.28% 79.58% 107 | 22 0.348 97.45% 79.69% 108 | 23 0.341 97.38% 79.60% 109 | 24 0.335 97.63% 79.62% 110 | 25 0.334 97.61% 79.70% 111 | 26 0.332 97.55% 79.70% 112 | *27 0.329 97.70% 79.74% 113 | 28 0.331 97.65% 79.70% 114 | 29 0.334 97.56% 79.72% 115 | 30 0.329 97.75% 79.74% 116 | 31 0.331 97.48% 79.74% 117 | *32 0.327 97.68% 79.75% 118 | *33 0.327 97.81% 79.79% 119 | 34 0.328 97.48% 79.77% 120 | 35 0.324 97.75% 79.77% 121 | 36 0.324 97.81% 79.77% 122 | 37 0.327 97.60% 79.79% 123 | 38 0.326 97.63% 79.79% 124 | 39 0.324 97.80% 79.79% 125 | 40 0.327 97.63% 79.75% 126 | 41 0.325 97.85% 79.77% 127 | 42 0.326 97.58% 79.79% 128 | 43 0.324 97.53% 79.75% 129 | 44 0.327 97.46% 79.75% 130 | 45 0.327 97.76% 79.77% 131 | 46 0.327 97.70% 79.77% 132 | 47 0.325 97.76% 79.77% 133 | 48 0.327 97.65% 79.77% 134 | --------------------------------------------------------------------------------