TorchVision 对象检测微调教程

译者:片刻

校验:片刻

TIP 为了充分利用本教程,我们建议使用此Colab版本。这将使您可以尝试以下信息。

对于本教程,我们将在Penn-Fudan数据库中对行人检测和分割的预训练Mask R-CNN模型进行微调。它包含170个图像,其中包含345个行人实例,我们将用它来说明如何在torchvision中使用新功能,以便在自定义数据集上训练实例细分模型。

定义数据集

用于训练对象检测,实例细分和人员关键点检测的参考脚本可轻松支持添加新的自定义数据集。数据集应继承自标准 torch.utils.data.Dataset类,并实现__len____getitem__

我们唯一需要的特异性是数据集__getitem__应该返回:

  • 图像:尺寸(H, W)的PIL图像
  • 目标:包含以下字段的一个字典

    • 盒 (FloatTensor [N, 4]):的N 的坐标在包围盒[X0, Y0, X 1, Y1]格式中,范围从0W0H
    • 标签 (Int64Tensor [N]):对于每个边界框的标签
    • image_id (Int64Tensor [1]):图像标识符。它应该是在数据集中的所有图像之间唯一的,评估过程中使用
    • 面积 (张量[N]):将边界框的面积。这是通过COCO度量评估过程中使用,以分离小,中,大箱之间的度量得分。
    • iscrowd (UInt8Tensor [N]):用iscrowd =真实例将被评估期间忽略。
    • (可选地)掩模 (UInt8Tensor [N, H, W]):本分割掩码的每个其中一个对象
    • (可选地)关键点 (FloatTensor [N, K, 3]):对于每一个中的所述一个N个对象,它包含K个关键点[X, Y, 能见度]格式中,定义的对象。能见度= 0表示所述关键点是不可见的。请注意,数据增强,翻转关键点的概念是依赖于数据表示,你可能要适应引用/检测/ transforms.py为您的新关键点表示
  • 图像:大小为(H, W)PIL的图像

  • 目标:包含以下字段的字典
    • boxes (FloatTensor[N, 4])N个边界框的坐标,格式为[x0, y0, x1, y1],,范围从0W0H
    • labels (Int64Tensor[N]):每个边界框的标签
    • image_id (Int64Tensor[1]):图像标识符。它在数据集中的所有图像之间应该是唯一的,并在评估过程中使用
    • area (Tensor[N]):边界框的面积。在使用COCO指标进行评估时,可使用此值来区分小盒子,中盒子和大盒子之间的指标得分。
    • iscrowd (UInt8Tensor[N]):评估期间将忽略iscrowd = True的实例。
    • (可选)masks (UInt8Tensor[N, H, W]):每个对象的分割蒙版
    • (可选)keypoints (FloatTensor[N, K, 3]):对于N个对象中的每个对象,它包含[x, y, visibility]格式的K个关键点,以定义对象。 visibility=0 表示关键点不可见。请注意,对于数据扩充,翻转关键点的概念取决于数据表示,并且您可能应该将references/detection/transforms.py修改为新的关键点表示形式。

如果您的模型返回上述方法,则它们将使其适用于训练和评估,并将使用pycocotools中的评估脚本。

此外,如果要在训练过程中使用长宽比分组(以便每个批次仅包含长宽比相似的图像),则建议您还实现一种get_height_and_width 方法,该方法可返回图像的高度和宽度。如果未提供此方法,我们将通过查询数据集的所有元素__getitem__,这会将图像加载到内存中,并且比提供自定义方法要慢。

为PennFudan编写自定义数据集

让我们为PennFudan数据集编写一个数据集。之后下载并解压缩zip文件,我们有以下文件夹结构:

  1. PennFudanPed/
  2. PedMasks/
  3. FudanPed00001_mask.png
  4. FudanPed00002_mask.png
  5. FudanPed00003_mask.png
  6. FudanPed00004_mask.png
  7. ...
  8. PNGImages/
  9. FudanPed00001.png
  10. FudanPed00002.png
  11. FudanPed00003.png
  12. FudanPed00004.png

这是一对图像和分割蒙版的一个示例

https://pytorch.org/tutorials/_static/img/tv_tutorial/tv_image01.png

https://pytorch.org/tutorials/_static/img/tv_tutorial/tv_image02.png

因此,每个图像都有一个对应的分割蒙版,其中每个颜色对应一个不同的实例。让我们为此数据集编写一个torch.utils.data.Dataset类。

  1. import os
  2. import numpy as np
  3. import torch
  4. from PIL import Image
  5. class PennFudanDataset(object):
  6. def __init__(self, root, transforms):
  7. self.root = root
  8. self.transforms = transforms
  9. # load all image files, sorting them to
  10. # ensure that they are aligned
  11. self.imgs = list(sorted(os.listdir(os.path.join(root, "PNGImages"))))
  12. self.masks = list(sorted(os.listdir(os.path.join(root, "PedMasks"))))
  13. def __getitem__(self, idx):
  14. # load images ad masks
  15. img_path = os.path.join(self.root, "PNGImages", self.imgs[idx])
  16. mask_path = os.path.join(self.root, "PedMasks", self.masks[idx])
  17. img = Image.open(img_path).convert("RGB")
  18. # note that we haven't converted the mask to RGB,
  19. # because each color corresponds to a different instance
  20. # with 0 being background
  21. mask = Image.open(mask_path)
  22. # convert the PIL Image into a numpy array
  23. mask = np.array(mask)
  24. # instances are encoded as different colors
  25. obj_ids = np.unique(mask)
  26. # first id is the background, so remove it
  27. obj_ids = obj_ids[1:]
  28. # split the color-encoded mask into a set
  29. # of binary masks
  30. masks = mask == obj_ids[:, None, None]
  31. # get bounding box coordinates for each mask
  32. num_objs = len(obj_ids)
  33. boxes = []
  34. for i in range(num_objs):
  35. pos = np.where(masks[i])
  36. xmin = np.min(pos[1])
  37. xmax = np.max(pos[1])
  38. ymin = np.min(pos[0])
  39. ymax = np.max(pos[0])
  40. boxes.append([xmin, ymin, xmax, ymax])
  41. # convert everything into a torch.Tensor
  42. boxes = torch.as_tensor(boxes, dtype=torch.float32)
  43. # there is only one class
  44. labels = torch.ones((num_objs,), dtype=torch.int64)
  45. masks = torch.as_tensor(masks, dtype=torch.uint8)
  46. image_id = torch.tensor([idx])
  47. area = (boxes[:, 3] - boxes[:, 1]) * (boxes[:, 2] - boxes[:, 0])
  48. # suppose all instances are not crowd
  49. iscrowd = torch.zeros((num_objs,), dtype=torch.int64)
  50. target = {}
  51. target["boxes"] = boxes
  52. target["labels"] = labels
  53. target["masks"] = masks
  54. target["image_id"] = image_id
  55. target["area"] = area
  56. target["iscrowd"] = iscrowd
  57. if self.transforms is not None:
  58. img, target = self.transforms(img, target)
  59. return img, target
  60. def __len__(self):
  61. return len(self.imgs)

这就是数据集的全部内容。现在,让我们定义一个可以对该数据集执行预测的模型。

定义模型

在本教程中,我们将使用基于Faster R-CNN的Mask R-CNN。更快的R-CNN是可预测图像中潜在对象的边界框和类分数的模型。

这是所有的数据集。现在让我们来定义可以在这个数据集进行预测的模型。

https://pytorch.org/tutorials/_static/img/tv_tutorial/tv_image03.png

Mask R-CNN在Faster R-CNN中增加了一个分支,该分支还可以预测每个实例的分割掩码。

https://pytorch.org/tutorials/_static/img/tv_tutorial/tv_image04.png

在两种常见情况下,可能要修改Torchvision modelzoo中的可用模型之一。首先是当我们想从预先训练的模型开始,然后微调最后一层时。另一个是当我们要用另一个模型替换主干时(例如,为了更快的预测)。

在以下各节中,让我们看看如何做一个或另一个。

1-通过预训练模型进行微调

假设您要从在COCO上进行预训练的模型开始,并希望针对您的特定班级对其进行微调。这是一种可行的方法:

  1. import torchvision
  2. from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
  3. # load a model pre-trained pre-trained on COCO
  4. model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)
  5. # replace the classifier with a new one, that has
  6. # num_classes which is user-defined
  7. num_classes = 2 # 1 class (person) + background
  8. # get number of input features for the classifier
  9. in_features = model.roi_heads.box_predictor.cls_score.in_features
  10. # replace the pre-trained head with a new one
  11. model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)

2-修改模型以添加其他主干

  1. import torchvision
  2. from torchvision.models.detection import FasterRCNN
  3. from torchvision.models.detection.rpn import AnchorGenerator
  4. # load a pre-trained model for classification and return
  5. # only the features
  6. backbone = torchvision.models.mobilenet_v2(pretrained=True).features
  7. # FasterRCNN needs to know the number of
  8. # output channels in a backbone. For mobilenet_v2, it's 1280
  9. # so we need to add it here
  10. backbone.out_channels = 1280
  11. # let's make the RPN generate 5 x 3 anchors per spatial
  12. # location, with 5 different sizes and 3 different aspect
  13. # ratios. We have a Tuple[Tuple[int]] because each feature
  14. # map could potentially have different sizes and
  15. # aspect ratios
  16. anchor_generator = AnchorGenerator(sizes=((32, 64, 128, 256, 512),),
  17. aspect_ratios=((0.5, 1.0, 2.0),))
  18. # let's define what are the feature maps that we will
  19. # use to perform the region of interest cropping, as well as
  20. # the size of the crop after rescaling.
  21. # if your backbone returns a Tensor, featmap_names is expected to
  22. # be [0]. More generally, the backbone should return an
  23. # OrderedDict[Tensor], and in featmap_names you can choose which
  24. # feature maps to use.
  25. roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=[0],
  26. output_size=7,
  27. sampling_ratio=2)
  28. # put the pieces together inside a FasterRCNN model
  29. model = FasterRCNN(backbone,
  30. num_classes=2,
  31. rpn_anchor_generator=anchor_generator,
  32. box_roi_pool=roi_pooler)

PennFudan数据集的实例细分模型

在我们的例子中,由于我们的数据集非常小,我们希望从预训练模型中进行微调,因此我们将遵循方法1。

在这里,我们还要计算实例分割掩码,因此我们将使用Mask R-CNN:

  1. import torchvision
  2. from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
  3. from torchvision.models.detection.mask_rcnn import MaskRCNNPredictor
  4. def get_model_instance_segmentation(num_classes):
  5. # load an instance segmentation model pre-trained pre-trained on COCO
  6. model = torchvision.models.detection.maskrcnn_resnet50_fpn(pretrained=True)
  7. # get number of input features for the classifier
  8. in_features = model.roi_heads.box_predictor.cls_score.in_features
  9. # replace the pre-trained head with a new one
  10. model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)
  11. # now get the number of input features for the mask classifier
  12. in_features_mask = model.roi_heads.mask_predictor.conv5_mask.in_channels
  13. hidden_layer = 256
  14. # and replace the mask predictor with a new one
  15. model.roi_heads.mask_predictor = MaskRCNNPredictor(in_features_mask,
  16. hidden_layer,
  17. num_classes)
  18. return model

就是这样,这将model准备好在您的自定义数据集上进行训练和评估。

将所有内容放在一起

在中references/detection/,我们提供了许多帮助程序功能来简化训练和评估检测模型。在这里,我们将使用 references/detection/engine.pyreferences/detection/utils.pyreferences/detection/transforms.py。只需将它们复制到您的文件夹中,然后在此处使用它们即可。

让我们写一些辅助函数来进行数据扩充/转换:

  1. import transforms as T
  2. def get_transform(train):
  3. transforms = []
  4. transforms.append(T.ToTensor())
  5. if train:
  6. transforms.append(T.RandomHorizontalFlip(0.5))
  7. return T.Compose(transforms)

现在让我们编写执行训练和验证的主要功能:

  1. from engine import train_one_epoch, evaluate
  2. import utils
  3. def main():
  4. # train on the GPU or on the CPU, if a GPU is not available
  5. device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
  6. # our dataset has two classes only - background and person
  7. num_classes = 2
  8. # use our dataset and defined transformations
  9. dataset = PennFudanDataset('PennFudanPed', get_transform(train=True))
  10. dataset_test = PennFudanDataset('PennFudanPed', get_transform(train=False))
  11. # split the dataset in train and test set
  12. indices = torch.randperm(len(dataset)).tolist()
  13. dataset = torch.utils.data.Subset(dataset, indices[:-50])
  14. dataset_test = torch.utils.data.Subset(dataset_test, indices[-50:])
  15. # define training and validation data loaders
  16. data_loader = torch.utils.data.DataLoader(
  17. dataset, batch_size=2, shuffle=True, num_workers=4,
  18. collate_fn=utils.collate_fn)
  19. data_loader_test = torch.utils.data.DataLoader(
  20. dataset_test, batch_size=1, shuffle=False, num_workers=4,
  21. collate_fn=utils.collate_fn)
  22. # get the model using our helper function
  23. model = get_model_instance_segmentation(num_classes)
  24. # move model to the right device
  25. model.to(device)
  26. # construct an optimizer
  27. params = [p for p in model.parameters() if p.requires_grad]
  28. optimizer = torch.optim.SGD(params, lr=0.005,
  29. momentum=0.9, weight_decay=0.0005)
  30. # and a learning rate scheduler
  31. lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
  32. step_size=3,
  33. gamma=0.1)
  34. # let's train it for 10 epochs
  35. num_epochs = 10
  36. for epoch in range(num_epochs):
  37. # train for one epoch, printing every 10 iterations
  38. train_one_epoch(model, optimizer, data_loader, device, epoch, print_freq=10)
  39. # update the learning rate
  40. lr_scheduler.step()
  41. # evaluate on the test dataset
  42. evaluate(model, data_loader_test, device=device)
  43. print("That's it!")

您应该获得第一个时期的输出:

  1. Epoch: [0] [ 0/60] eta: 0:01:18 lr: 0.000090 loss: 2.5213 (2.5213) loss_classifier: 0.8025 (0.8025) loss_box_reg: 0.2634 (0.2634) loss_mask: 1.4265 (1.4265) loss_objectness: 0.0190 (0.0190) loss_rpn_box_reg: 0.0099 (0.0099) time: 1.3121 data: 0.3024 max mem: 3485
  2. Epoch: [0] [10/60] eta: 0:00:20 lr: 0.000936 loss: 1.3007 (1.5313) loss_classifier: 0.3979 (0.4719) loss_box_reg: 0.2454 (0.2272) loss_mask: 0.6089 (0.7953) loss_objectness: 0.0197 (0.0228) loss_rpn_box_reg: 0.0121 (0.0141) time: 0.4198 data: 0.0298 max mem: 5081
  3. Epoch: [0] [20/60] eta: 0:00:15 lr: 0.001783 loss: 0.7567 (1.1056) loss_classifier: 0.2221 (0.3319) loss_box_reg: 0.2002 (0.2106) loss_mask: 0.2904 (0.5332) loss_objectness: 0.0146 (0.0176) loss_rpn_box_reg: 0.0094 (0.0123) time: 0.3293 data: 0.0035 max mem: 5081
  4. Epoch: [0] [30/60] eta: 0:00:11 lr: 0.002629 loss: 0.4705 (0.8935) loss_classifier: 0.0991 (0.2517) loss_box_reg: 0.1578 (0.1957) loss_mask: 0.1970 (0.4204) loss_objectness: 0.0061 (0.0140) loss_rpn_box_reg: 0.0075 (0.0118) time: 0.3403 data: 0.0044 max mem: 5081
  5. Epoch: [0] [40/60] eta: 0:00:07 lr: 0.003476 loss: 0.3901 (0.7568) loss_classifier: 0.0648 (0.2022) loss_box_reg: 0.1207 (0.1736) loss_mask: 0.1705 (0.3585) loss_objectness: 0.0018 (0.0113) loss_rpn_box_reg: 0.0075 (0.0112) time: 0.3407 data: 0.0044 max mem: 5081
  6. Epoch: [0] [50/60] eta: 0:00:03 lr: 0.004323 loss: 0.3237 (0.6703) loss_classifier: 0.0474 (0.1731) loss_box_reg: 0.1109 (0.1561) loss_mask: 0.1658 (0.3201) loss_objectness: 0.0015 (0.0093) loss_rpn_box_reg: 0.0093 (0.0116) time: 0.3379 data: 0.0043 max mem: 5081
  7. Epoch: [0] [59/60] eta: 0:00:00 lr: 0.005000 loss: 0.2540 (0.6082) loss_classifier: 0.0309 (0.1526) loss_box_reg: 0.0463 (0.1405) loss_mask: 0.1568 (0.2945) loss_objectness: 0.0012 (0.0083) loss_rpn_box_reg: 0.0093 (0.0123) time: 0.3489 data: 0.0042 max mem: 5081
  8. Epoch: [0] Total time: 0:00:21 (0.3570 s / it)
  9. creating index...
  10. index created!
  11. Test: [ 0/50] eta: 0:00:19 model_time: 0.2152 (0.2152) evaluator_time: 0.0133 (0.0133) time: 0.4000 data: 0.1701 max mem: 5081
  12. Test: [49/50] eta: 0:00:00 model_time: 0.0628 (0.0687) evaluator_time: 0.0039 (0.0064) time: 0.0735 data: 0.0022 max mem: 5081
  13. Test: Total time: 0:00:04 (0.0828 s / it)
  14. Averaged stats: model_time: 0.0628 (0.0687) evaluator_time: 0.0039 (0.0064)
  15. Accumulating evaluation results...
  16. DONE (t=0.01s).
  17. Accumulating evaluation results...
  18. DONE (t=0.01s).
  19. IoU metric: bbox
  20. Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.606
  21. Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.984
  22. Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.780
  23. Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.313
  24. Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.582
  25. Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.612
  26. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.270
  27. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.672
  28. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.672
  29. Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.650
  30. Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.755
  31. Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.664
  32. IoU metric: segm
  33. Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.704
  34. Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.979
  35. Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.871
  36. Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.325
  37. Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.488
  38. Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.727
  39. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.316
  40. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.748
  41. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.749
  42. Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.650
  43. Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.673
  44. Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.758

因此,经过一个时期的训练,我们获得了60.6的COCO风格mAP和70.4的口罩mAP。

经过10个时期的训练,我得到了以下指标

  1. IoU metric: bbox
  2. Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.799
  3. Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.969
  4. Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.935
  5. Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.349
  6. Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.592
  7. Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.831
  8. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.324
  9. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.844
  10. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.844
  11. Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.400
  12. Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.777
  13. Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.870
  14. IoU metric: segm
  15. Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.761
  16. Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.969
  17. Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.919
  18. Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.341
  19. Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.464
  20. Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.788
  21. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.303
  22. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.799
  23. Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.799
  24. Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.400
  25. Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.769
  26. Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.818

但是这些预测是什么样的?让我们在数据集中拍摄一张图像并验证

https://pytorch.org/tutorials/_static/img/tv_tutorial/tv_image05.png

经过训练的模型可以在此图像中预测9个人的实例,让我们看看其中的几个:

https://pytorch.org/tutorials/_static/img/tv_tutorial/tv_image06.png

https://pytorch.org/tutorials/_static/img/tv_tutorial/tv_image07.png

结果看起来不错!

总结

在本教程中,您学习了如何在自定义数据集上为实例细分模型创建自己的训练管道。为此,您编写了一个torch.utils.data.Dataset类,该类返回图像,地面真相框和分割蒙版。您还利用了在COCO train2017上预先训练的Mask R-CNN模型,以便对该新数据集执行转移学习。

对于更完整的示例(包括多机/多GPU训练),请检查references/detection/train.py在Torchvision存储库中存在的。

您可以在此处下载本教程的完整源文件 。