python - PyTorch 和 TensorFlow 对象检测 - 评估 - 类型的对象不能安全地解释为整数

Question

我正在尝试运行这个 PyTorch 人员检测示例:

https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html

我正在使用 Ubuntu 18.04。以下是我执行的步骤的摘要:

1) 在带有 GTX 1650 GPU 的 Lenovo ThinkPad X1 Extreme Gen 2 上安装股票 Ubuntu 18.04。

2) 执行标准的 CUDA 10.0/cuDNN 7.4 安装。我不想重述所有步骤，因为这篇文章已经足够长了。这是一个标准程序，几乎所有通过谷歌搜索找到的链接都是我遵循的。

3) 安装 torch和 torchvision
4) 从 PyTorch 网站上的这个链接:

https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html

我保存了底部链接中可用的源:

https://pytorch.org/tutorials/_static/tv-training-code.py

到我创建的目录，PennFudanExample
5)我做了以下(在上面链接的笔记本的顶部找到):

将 CoCo API 安装到 Python 中:

cd ~
git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI

在gedit中打开Makefile，将“python”的两个实例改为“python3”，然后:

python3 setup.py build_ext --inplace
sudo python3 setup.py install

获取运行上述链接文件所需的必要文件:

cd ~
git clone https://github.com/pytorch/vision.git
cd vision
git checkout v0.5.0

来自 ~/vision/references/detection , 复制 coco_eval.py , coco_utils.py , engine.py , transforms.py , 和 utils.py到目录 PennFudanExample .

6)从上面页面的链接下载Penn Fudan Pedestrian数据集:

https://www.cis.upenn.edu/~jshi/ped_html/PennFudanPed.zip

然后解压并放入目录 PennFudanExample
7) 我对 tv-training-code.py 所做的唯一更改是将训练批次大小从 2 更改为 1 以防止 GPU 内存不足崩溃，请参阅我在此处发布的另一篇文章:

PyTorch Object Detection with GPU on Ubuntu 18.04 - RuntimeError: CUDA out of memory. Tried to allocate xx.xx MiB

这是 tv-training-code.py当我使用我提到的轻微批量大小编辑运行它时:

# Sample code from the TorchVision 0.3 Object Detection Finetuning Tutorial
# http://pytorch.org/tutorials/intermediate/torchvision_tutorial.html

import os
import numpy as np
import torch
from PIL import Image

import torchvision
from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
from torchvision.models.detection.mask_rcnn import MaskRCNNPredictor

from engine import train_one_epoch, evaluate
import utils
import transforms as T


class PennFudanDataset(object):
    def __init__(self, root, transforms):
        self.root = root
        self.transforms = transforms
        # load all image files, sorting them to
        # ensure that they are aligned
        self.imgs = list(sorted(os.listdir(os.path.join(root, "PNGImages"))))
        self.masks = list(sorted(os.listdir(os.path.join(root, "PedMasks"))))

    def __getitem__(self, idx):
        # load images ad masks
        img_path = os.path.join(self.root, "PNGImages", self.imgs[idx])
        mask_path = os.path.join(self.root, "PedMasks", self.masks[idx])
        img = Image.open(img_path).convert("RGB")
        # note that we haven't converted the mask to RGB,
        # because each color corresponds to a different instance
        # with 0 being background
        mask = Image.open(mask_path)

        mask = np.array(mask)
        # instances are encoded as different colors
        obj_ids = np.unique(mask)
        # first id is the background, so remove it
        obj_ids = obj_ids[1:]

        # split the color-encoded mask into a set
        # of binary masks
        masks = mask == obj_ids[:, None, None]

        # get bounding box coordinates for each mask
        num_objs = len(obj_ids)
        boxes = []
        for i in range(num_objs):
            pos = np.where(masks[i])
            xmin = np.min(pos[1])
            xmax = np.max(pos[1])
            ymin = np.min(pos[0])
            ymax = np.max(pos[0])
            boxes.append([xmin, ymin, xmax, ymax])

        boxes = torch.as_tensor(boxes, dtype=torch.float32)
        # there is only one class
        labels = torch.ones((num_objs,), dtype=torch.int64)
        masks = torch.as_tensor(masks, dtype=torch.uint8)

        image_id = torch.tensor([idx])
        area = (boxes[:, 3] - boxes[:, 1]) * (boxes[:, 2] - boxes[:, 0])
        # suppose all instances are not crowd
        iscrowd = torch.zeros((num_objs,), dtype=torch.int64)

        target = {}
        target["boxes"] = boxes
        target["labels"] = labels
        target["masks"] = masks
        target["image_id"] = image_id
        target["area"] = area
        target["iscrowd"] = iscrowd

        if self.transforms is not None:
            img, target = self.transforms(img, target)

        return img, target

    def __len__(self):
        return len(self.imgs)

def get_model_instance_segmentation(num_classes):
    # load an instance segmentation model pre-trained pre-trained on COCO
    model = torchvision.models.detection.maskrcnn_resnet50_fpn(pretrained=True)

    # get number of input features for the classifier
    in_features = model.roi_heads.box_predictor.cls_score.in_features
    # replace the pre-trained head with a new one
    model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)

    # now get the number of input features for the mask classifier
    in_features_mask = model.roi_heads.mask_predictor.conv5_mask.in_channels
    hidden_layer = 256
    # and replace the mask predictor with a new one
    model.roi_heads.mask_predictor = MaskRCNNPredictor(in_features_mask,
                                                       hidden_layer,
                                                       num_classes)

    return model


def get_transform(train):
    transforms = []
    transforms.append(T.ToTensor())
    if train:
        transforms.append(T.RandomHorizontalFlip(0.5))
    return T.Compose(transforms)


def main():
    # train on the GPU or on the CPU, if a GPU is not available
    device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')

    # our dataset has two classes only - background and person
    num_classes = 2
    # use our dataset and defined transformations
    dataset = PennFudanDataset('PennFudanPed', get_transform(train=True))
    dataset_test = PennFudanDataset('PennFudanPed', get_transform(train=False))

    # split the dataset in train and test set
    indices = torch.randperm(len(dataset)).tolist()
    dataset = torch.utils.data.Subset(dataset, indices[:-50])
    dataset_test = torch.utils.data.Subset(dataset_test, indices[-50:])

    # define training and validation data loaders
    # !!!! CHANGE HERE !!!! For this function call, I changed the batch_size param value from 2 to 1, otherwise this file is exactly as provided from the PyTorch website !!!!
    data_loader = torch.utils.data.DataLoader(
        dataset, batch_size=1, shuffle=True, num_workers=4,
        collate_fn=utils.collate_fn)

    data_loader_test = torch.utils.data.DataLoader(
        dataset_test, batch_size=1, shuffle=False, num_workers=4,
        collate_fn=utils.collate_fn)

    # get the model using our helper function
    model = get_model_instance_segmentation(num_classes)

    # move model to the right device
    model.to(device)

    # construct an optimizer
    params = [p for p in model.parameters() if p.requires_grad]
    optimizer = torch.optim.SGD(params, lr=0.005,
                                momentum=0.9, weight_decay=0.0005)
    # and a learning rate scheduler
    lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
                                                   step_size=3,
                                                   gamma=0.1)

    # let's train it for 10 epochs
    num_epochs = 10

    for epoch in range(num_epochs):
        # train for one epoch, printing every 10 iterations
        train_one_epoch(model, optimizer, data_loader, device, epoch, print_freq=10)
        # update the learning rate
        lr_scheduler.step()
        # evaluate on the test dataset
        evaluate(model, data_loader_test, device=device)

    print("That's it!")

if __name__ == "__main__":
    main()

这是全文输出，包括我目前遇到的错误:

Epoch: [0]  [  0/120]  eta: 0:01:41  lr: 0.000047  loss: 7.3028 (7.3028)  loss_classifier: 1.0316 (1.0316)  loss_box_reg: 0.0827 (0.0827)  loss_mask: 6.1742 (6.1742)  loss_objectness: 0.0097 (0.0097)  loss_rpn_box_reg: 0.0046 (0.0046)  time: 0.8468  data: 0.0803  max mem: 1067
Epoch: [0]  [ 10/120]  eta: 0:01:02  lr: 0.000467  loss: 2.0995 (3.5058)  loss_classifier: 0.6684 (0.6453)  loss_box_reg: 0.0999 (0.1244)  loss_mask: 1.2471 (2.7069)  loss_objectness: 0.0187 (0.0235)  loss_rpn_box_reg: 0.0060 (0.0057)  time: 0.5645  data: 0.0089  max mem: 1499
Epoch: [0]  [ 20/120]  eta: 0:00:56  lr: 0.000886  loss: 1.0166 (2.1789)  loss_classifier: 0.2844 (0.4347)  loss_box_reg: 0.1631 (0.1540)  loss_mask: 0.4710 (1.5562)  loss_objectness: 0.0187 (0.0242)  loss_rpn_box_reg: 0.0082 (0.0099)  time: 0.5524  data: 0.0020  max mem: 1704
Epoch: [0]  [ 30/120]  eta: 0:00:50  lr: 0.001306  loss: 0.5554 (1.6488)  loss_classifier: 0.1258 (0.3350)  loss_box_reg: 0.1356 (0.1488)  loss_mask: 0.2355 (1.1285)  loss_objectness: 0.0142 (0.0224)  loss_rpn_box_reg: 0.0127 (0.0142)  time: 0.5653  data: 0.0023  max mem: 1756
Epoch: [0]  [ 40/120]  eta: 0:00:45  lr: 0.001726  loss: 0.4520 (1.3614)  loss_classifier: 0.1055 (0.2773)  loss_box_reg: 0.1101 (0.1530)  loss_mask: 0.1984 (0.8981)  loss_objectness: 0.0063 (0.0189)  loss_rpn_box_reg: 0.0139 (0.0140)  time: 0.5621  data: 0.0023  max mem: 1776
Epoch: [0]  [ 50/120]  eta: 0:00:39  lr: 0.002146  loss: 0.3448 (1.1635)  loss_classifier: 0.0622 (0.2346)  loss_box_reg: 0.1004 (0.1438)  loss_mask: 0.1650 (0.7547)  loss_objectness: 0.0033 (0.0172)  loss_rpn_box_reg: 0.0069 (0.0131)  time: 0.5535  data: 0.0022  max mem: 1776
Epoch: [0]  [ 60/120]  eta: 0:00:33  lr: 0.002565  loss: 0.3292 (1.0543)  loss_classifier: 0.0549 (0.2101)  loss_box_reg: 0.1113 (0.1486)  loss_mask: 0.1596 (0.6668)  loss_objectness: 0.0017 (0.0148)  loss_rpn_box_reg: 0.0082 (0.0140)  time: 0.5590  data: 0.0022  max mem: 1776
Epoch: [0]  [ 70/120]  eta: 0:00:28  lr: 0.002985  loss: 0.4105 (0.9581)  loss_classifier: 0.0534 (0.1877)  loss_box_reg: 0.1049 (0.1438)  loss_mask: 0.1709 (0.5995)  loss_objectness: 0.0015 (0.0132)  loss_rpn_box_reg: 0.0133 (0.0138)  time: 0.5884  data: 0.0023  max mem: 1783
Epoch: [0]  [ 80/120]  eta: 0:00:22  lr: 0.003405  loss: 0.3080 (0.8817)  loss_classifier: 0.0441 (0.1706)  loss_box_reg: 0.0875 (0.1343)  loss_mask: 0.1960 (0.5510)  loss_objectness: 0.0015 (0.0122)  loss_rpn_box_reg: 0.0071 (0.0137)  time: 0.5812  data: 0.0023  max mem: 1783
Epoch: [0]  [ 90/120]  eta: 0:00:17  lr: 0.003825  loss: 0.2817 (0.8171)  loss_classifier: 0.0397 (0.1570)  loss_box_reg: 0.0499 (0.1257)  loss_mask: 0.1777 (0.5098)  loss_objectness: 0.0008 (0.0111)  loss_rpn_box_reg: 0.0068 (0.0136)  time: 0.5644  data: 0.0022  max mem: 1794
Epoch: [0]  [100/120]  eta: 0:00:11  lr: 0.004244  loss: 0.2139 (0.7569)  loss_classifier: 0.0310 (0.1446)  loss_box_reg: 0.0327 (0.1163)  loss_mask: 0.1573 (0.4731)  loss_objectness: 0.0003 (0.0101)  loss_rpn_box_reg: 0.0050 (0.0128)  time: 0.5685  data: 0.0022  max mem: 1794
Epoch: [0]  [110/120]  eta: 0:00:05  lr: 0.004664  loss: 0.2139 (0.7160)  loss_classifier: 0.0325 (0.1358)  loss_box_reg: 0.0327 (0.1105)  loss_mask: 0.1572 (0.4477)  loss_objectness: 0.0003 (0.0093)  loss_rpn_box_reg: 0.0047 (0.0128)  time: 0.5775  data: 0.0022  max mem: 1794
Epoch: [0]  [119/120]  eta: 0:00:00  lr: 0.005000  loss: 0.2486 (0.6830)  loss_classifier: 0.0330 (0.1282)  loss_box_reg: 0.0360 (0.1051)  loss_mask: 0.1686 (0.4284)  loss_objectness: 0.0003 (0.0086)  loss_rpn_box_reg: 0.0074 (0.0125)  time: 0.5655  data: 0.0022  max mem: 1794
Epoch: [0] Total time: 0:01:08 (0.5676 s / it)
creating index...
index created!
Traceback (most recent call last):
  File "/usr/local/lib/python3.6/dist-packages/numpy/core/function_base.py", line 117, in linspace
    num = operator.index(num)
TypeError: 'numpy.float64' object cannot be interpreted as an integer

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/home/cdahms/workspace-apps/PennFudanExample/tv-training-code.py", line 166, in <module>
    main()
  File "/home/cdahms/workspace-apps/PennFudanExample/tv-training-code.py", line 161, in main
    evaluate(model, data_loader_test, device=device)
  File "/usr/local/lib/python3.6/dist-packages/torch/autograd/grad_mode.py", line 49, in decorate_no_grad
    return func(*args, **kwargs)
  File "/home/cdahms/workspace-apps/PennFudanExample/engine.py", line 80, in evaluate
    coco_evaluator = CocoEvaluator(coco, iou_types)
  File "/home/cdahms/workspace-apps/PennFudanExample/coco_eval.py", line 28, in __init__
    self.coco_eval[iou_type] = COCOeval(coco_gt, iouType=iou_type)
  File "/home/cdahms/models/research/pycocotools/cocoeval.py", line 75, in __init__
    self.params = Params(iouType=iouType) # parameters
  File "/home/cdahms/models/research/pycocotools/cocoeval.py", line 527, in __init__
    self.setDetParams()
  File "/home/cdahms/models/research/pycocotools/cocoeval.py", line 506, in setDetParams
    self.iouThrs = np.linspace(.5, 0.95, np.round((0.95 - .5) / .05) + 1, endpoint=True)
  File "<__array_function__ internals>", line 6, in linspace
  File "/usr/local/lib/python3.6/dist-packages/numpy/core/function_base.py", line 121, in linspace
    .format(type(num)))
TypeError: object of type <class 'numpy.float64'> cannot be safely interpreted as an integer.

Process finished with exit code 1

真正奇怪的是，在我解决了上述 GPU 错误之后，它每天工作大约 1/2 次，现在我收到了这个错误，我可以发誓我没有改变任何东西。

我试过卸载并重新安装 torch , torchvision , pycocotools , 以及复制文件 coco_eval.py , coco_utils.py , engine.py , transforms.py , 和 utils.py ，我试过检查 torchvision v0.5.0、v0.4.2，并使用最新的提交，都产生相同的错误。

另外，我昨天(圣诞节)在家工作，而我的家用计算机上也没有发生此错误，该计算机也是带有 NVIDIA GPU 的 Ubuntu 18.04。

在谷歌搜索这个错误时，一个比较常见的建议是回溯 numpy到 1.11.0，但该版本现在真的很旧，因此这可能会导致其他软件包出现问题。

同样在谷歌搜索这个错误，似乎一般的解决方法是向 int 添加一个类型转换。某处或更改除以 /至 //但我真的很犹豫要不要对 pycocotools 进行内部更改或更糟的内部 numpy .此外，由于之前没有发生错误并且在另一台计算机上也没有发生错误，所以我不怀疑这是一个好主意。

幸运的是我可以注释掉这一行

evaluate(model, data_loader_test, device=device)

现在，训练将完成，虽然我没有得到评估数据(平均平均精度等)

目前我唯一能想到的就是格式化 HD 并重新安装 Ubuntu 18.04 和其他所有东西，但这至少需要一天时间，如果再次发生这种情况，我真的很想知道可能是什么造成它。

想法？建议？我应该检查的其他东西？

- 编辑 -

在遇到问题的同一台计算机上重新测试后，我发现在使用 TensorFlow 对象检测 API 时，评估步骤也会出现同样的错误。

Answer 1

!@#$%^&

大约 15 个小时后，我终于弄清楚了这一点，因为 numpy 1.18.0(在我撰写本文时于 5 天前发布)破坏了 TensorFlow 和 PyTorch 对象检测的评估过程。长话短说，修复方法是:

sudo -H pip3 install numpy==1.17.4

我还可以提几点:

-numpy 1.17.4 于 2019 年 11 月 10 日发布，因此应该仍然适用一段时间

- 现在有一个用于 pycocotools 的 pip 包，因此您现在可以简单地执行以下操作，而不是上述过程(克隆和构建):

sudo -H pip3 install pycocotools

- - 更新 - -

这现已在 pycocotools 中得到修复有了这个提交:

https://github.com/cocodataset/cocoapi/pull/354

另请参阅此(已关闭)问题以获取更多背景信息:

https://github.com/numpy/numpy/issues/15192

当 pycocotools更新版本将进入 pycocotools pip3 package ， 我不知道。