| tags |
|---|
Github |
- After balabced agumentation
def augmentation(self, wts): if self.is_train: image_filenames = [] classes_num = len(self.num_per_classes) for classes in range(classes_num): origin = self.num_per_classes[classes] ratio = wts[classes]/100 current_idx = self.class_start_idx[classes] if ratio > 1: after = int(origin * ratio) tmp_filenames = self.image_filenames[current_idx:current_idx + origin] count = 0 while origin < after: tmp_filenames.append(self.image_filenames[current_idx + count]) count += 1 count %= self.num_per_classes[classes] origin += 1 image_filenames += tmp_filenames elif ratio < 1: after = int(origin * ratio) tmp_filenames = self.image_filenames[current_idx:current_idx + after] image_filenames += tmp_filenames else: image_filenames += self.image_filenames[current_idx:current_idx+origin] self.image_filenames = image_filenameswts = [100, 781, 67, 169, 196, 75, 757, 1190, 194, 67, 2857] train_dataset.augmentation(wts)---------------------------------------------------------------------------------- Dataset - ./food11re/skewed_training class_name |bf. loading |af. loading Bread |994 |994 Dairy_product |128 |999 Dessert |1500 |1005 Egg |591 |998 Fried_food |508 |995 Meat |1325 |993 Noodles |132 |999 Rice |84 |999 Seafood |513 |995 Soup |1500 |1005 Vegetable_fruit |35 |999
- torch.utils.data.RandomSampler
randon_sampler = data.RandomSampler(train_dataset, replacement=True, num_samples=9000, generator=None)---------------------------------------------------------------------------------- Dataset - ./food11re/skewed_training class_name |bf. loading |af. loading Bread |994 |1253 Dairy_product |128 |167 Dessert |1500 |1843 Egg |591 |758 Fried_food |508 |606 Meat |1325 |1645 Noodles |132 |143 Rice |84 |102 Seafood |513 |646 Soup |1500 |1789 Vegetable_fruit |35 |48 - torch.utils.data.WeightedRandomSampler
weight = [] for i in range(11): class_count = train_dataset2.targets.count(i) weight.append(1./(class_count/len(train_dataset2.targets))) samples_weight = np.array([weight[t] for _, t in train_dataset2]) weighted_sampler = data.WeightedRandomSampler(samples_weight, num_samples=15000, replacement=True)---------------------------------------------------------------------------------- Dataset - ./food11re/skewed_training class_name |bf. loading |af. loading Bread |994 |1323 Dairy_product |128 |1418 Dessert |1500 |1394 Egg |591 |1359 Fried_food |508 |1366 Meat |1325 |1299 Noodles |132 |1370 Rice |84 |1306 Seafood |513 |1395 Soup |1500 |1373 Vegetable_fruit |35 |1397 - torchsampler.ImbalancedDatasetSampler
train_loader2 = DataLoader(dataset=train_dataset2, num_workers=0, batch_size=100, sampler=ImbalancedDatasetSampler(train_dataset2, num_samples=9000))---------------------------------------------------------------------------------- Dataset - ./food11re/skewed_training class_name |bf. loading |af. loading Bread |994 |806 Dairy_product |128 |831 Dessert |1500 |789 Egg |591 |801 Fried_food |508 |799 Meat |1325 |849 Noodles |132 |814 Rice |84 |805 Seafood |513 |849 Soup |1500 |826 Vegetable_fruit |35 |831
-
Combine imgaug and torchvision.transform
class ImgAugTransform: def __init__(self): self.aug = iaa.SomeOf((1, 2), [ iaa.GammaContrast((0.5, 2.0)), # 亮度 iaa.Multiply(), iaa.GaussianBlur(1.0), # 高斯模糊 iaa.AddToHueAndSaturation(value=(-10, 10), per_channel=True) # 增加飽和度 ]) def __call__(self, img): img = np.array(img) return self.aug.augment_image(img) transform_train = transforms.Compose([ transforms.RandomRotation(90), transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), ImgAugTransform(), lambda x: PIL.Image.fromarray(x), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) -
Using WeightRandomSampler to balance each class. (num_samples=18000)
weight = [] for i in range(11): class_count = trainset.targets.count(i) weight.append(1. / (class_count / len(trainset.targets))) samples_weight = np.array([weight[t] for _, t in trainset]) weighted_sampler = torch.utils.data.WeightedRandomSampler(samples_weight, num_samples=18000, replacement=True) -
Result
Top 1 Accuracy of class 0 is 276/368 75.00% Top 1 Accuracy of class 1 is 89/148 60.14% Top 1 Accuracy of class 2 is 185/231 80.09% Top 1 Accuracy of class 3 is 399/500 79.80% Top 1 Accuracy of class 4 is 276/335 82.39% Top 1 Accuracy of class 5 is 234/287 81.53% Top 1 Accuracy of class 6 is 362/432 83.80% Top 1 Accuracy of class 7 is 126/147 85.71% Top 1 Accuracy of class 8 is 85/ 96 88.54% Top 1 Accuracy of class 9 is 262/303 86.47% Top 1 Accuracy of class 10 is 463/500 92.60% Top 1 accuracy of the network on the 3347 test images: 2757/3347 82.37 % 82.37227367792053 Top 3 Accuracy of class 0 is 351/368 95.38% Top 3 Accuracy of class 1 is 115/148 77.70% Top 3 Accuracy of class 2 is 220/231 95.24% Top 3 Accuracy of class 3 is 484/500 96.80% Top 3 Accuracy of class 4 is 316/335 94.33% Top 3 Accuracy of class 5 is 270/287 94.08% Top 3 Accuracy of class 6 is 423/432 97.92% Top 3 Accuracy of class 7 is 137/147 93.20% Top 3 Accuracy of class 8 is 93/ 96 96.88% Top 3 Accuracy of class 9 is 284/303 93.73% Top 3 Accuracy of class 10 is 492/500 98.40% Top 3 accuracy of the network on the 3347 test images: 3185/3347 95.16 % 95.15984463698835 -
Compare to DLSR-LAB01
Class LAB2 accuracy (%) LAB1 accuracy (%) 0 75.00 67.93 1 60.14 30.41 2 80.09 54.11 3 79.80 71.40 4 82.39 57.01 5 81.53 70.03 6 83.80 83.10 7 85.71 78.23 8 88.54 79.17 9 86.47 72.28 10 92.60 90.00 top1 82.37 71.35 top3 95.16 91.57
-
將包含layer2之前的參數都固定
net = torchvision.models.resnet18(pretrained=True) num_features=net.fc.in_features net.fc=nn.Linear(num_features, 11) for k,v in net.named_parameters(): print(k) if (k=='conv1.weight' or k == 'bn1.weight' or k == 'bn1.bias'): v.requires_grad=False if (k[0:6] == 'layer1' or k[0:6] == 'layer2'): v.requires_grad=False -
檢查各層requires_grad之值
for k,v in net.named_parameters(): print(k, v.requires_grad) ''' conv1.weight False bn1.weight False bn1.bias False layer1.0.conv1.weight False layer1.0.bn1.weight False layer1.0.bn1.bias False layer1.0.conv2.weight False layer1.0.bn2.weight False layer1.0.bn2.bias False layer1.1.conv1.weight False layer1.1.bn1.weight False layer1.1.bn1.bias False layer1.1.conv2.weight False layer1.1.bn2.weight False layer1.1.bn2.bias False layer2.0.conv1.weight False layer2.0.bn1.weight False layer2.0.bn1.bias False layer2.0.conv2.weight False layer2.0.bn2.weight False layer2.0.bn2.bias False layer2.0.downsample.0.weight False layer2.0.downsample.1.weight False layer2.0.downsample.1.bias False layer2.1.conv1.weight False layer2.1.bn1.weight False layer2.1.bn1.bias False layer2.1.conv2.weight False layer2.1.bn2.weight False layer2.1.bn2.bias False layer3.0.conv1.weight True layer3.0.bn1.weight True layer3.0.bn1.bias True layer3.0.conv2.weight True layer3.0.bn2.weight True layer3.0.bn2.bias True layer3.0.downsample.0.weight True layer3.0.downsample.1.weight True layer3.0.downsample.1.bias True layer3.1.conv1.weight True layer3.1.bn1.weight True layer3.1.bn1.bias True layer3.1.conv2.weight True layer3.1.bn2.weight True layer3.1.bn2.bias True layer4.0.conv1.weight True layer4.0.bn1.weight True layer4.0.bn1.bias True layer4.0.conv2.weight True layer4.0.bn2.weight True layer4.0.bn2.bias True layer4.0.downsample.0.weight True layer4.0.downsample.1.weight True layer4.0.downsample.1.bias True layer4.1.conv1.weight True layer4.1.bn1.weight True layer4.1.bn1.bias True layer4.1.conv2.weight True layer4.1.bn2.weight True layer4.1.bn2.bias True fc.weight True fc.bias True ''' -
Optimizer過濾掉不需要訓練的layer (5/20似乎不需要)
optimizer = optim.Adam(filter(lambda p: p.requires_grad, net.parameters())) -
Result
Top 1 Accuracy of class 0 is 322/368 87.50% Top 1 Accuracy of class 1 is 108/148 72.97% Top 1 Accuracy of class 2 is 194/231 83.98% Top 1 Accuracy of class 3 is 429/500 85.80% Top 1 Accuracy of class 4 is 279/335 83.28% Top 1 Accuracy of class 5 is 264/287 91.99% Top 1 Accuracy of class 6 is 402/432 93.06% Top 1 Accuracy of class 7 is 140/147 95.24% Top 1 Accuracy of class 8 is 95/ 96 98.96% Top 1 Accuracy of class 9 is 281/303 92.74% Top 1 Accuracy of class 10 is 482/500 96.40% Top 1 accuracy of the network on the 3347 test images: 2996/3347 89.51 % 89.51299671347475 Top 3 Accuracy of class 0 is 365/368 99.18% Top 3 Accuracy of class 1 is 134/148 90.54% Top 3 Accuracy of class 2 is 224/231 96.97% Top 3 Accuracy of class 3 is 490/500 98.00% Top 3 Accuracy of class 4 is 318/335 94.93% Top 3 Accuracy of class 5 is 284/287 98.95% Top 3 Accuracy of class 6 is 427/432 98.84% Top 3 Accuracy of class 7 is 144/147 97.96% Top 3 Accuracy of class 8 is 96/ 96 100.00% Top 3 Accuracy of class 9 is 295/303 97.36% Top 3 Accuracy of class 10 is 498/500 99.60% Top 3 accuracy of the network on the 3347 test images: 3275/3347 97.85 % 97.84881983866148 -
Compare to model without pretrained.
Class Transfer learning accuracy(%) LAB2 accuracy (%) LAB1 accuracy (%) 0 87.50 75.00 67.93 1 72.97 60.14 30.41 2 83.98 80.09 54.11 3 85.80 79.80 71.40 4 83.28 82.39 57.01 5 91.99 81.53 70.03 6 93.06 83.80 83.10 7 95.24 85.71 78.23 8 98.96 88.54 79.17 9 92.74 86.47 72.28 10 96.40 92.60 90.00 top1 89.51 82.37 71.35 top3 97.84 95.16 91.57