python - 加载模型时出现意外错误: problem in predictor - ModuleNotFoundError: No module named 'torchvision'

Question

我一直在尝试通过虚拟机实例上的控制台将模型部署到 AI 平台进行预测，但收到错误“(gcloud.beta.ai-platform.versions.create) 创建版本失败” . 检测到错误模型:“无法加载模型:加载模型时出现意外错误:预测器中存在问题 - ModuleNotFoundError:没有名为“torchvision”的模块(错误代码:0)”

我需要同时包含 torch 和 torchvision。我按照这个问题Cannot deploy trained model to Google Cloud Ai-Platform with custom prediction routine: Model requires more memory than allowed中的步骤进行操作，但我无法获取用户 gogasca 指向的文件。我尝试下载this我从 Pytorch 网站下载了 .whl 文件并将其上传到我的云存储，但得到了相同的错误，即没有模块 torchvision，即使此版本应该包含 torch 和 torchvision。还尝试使用 Cloud AI 兼容包 here ，但它们不包括 torchvision。

我尝试在 --package-uris 参数中指向 torch 和 torchvision 的两个单独的 .whl 文件，这些文件指向文件在我的云存储中，但后来我收到超出内存容量的错误。这很奇怪，因为它们的总大小约为 130Mb。我的导致缺少 torchvision 的命令示例如下所示:

gcloud beta ai-platform versions create version_1 \
  --model online_pred_1 \
  --runtime-version 1.15 \
  --python-version 3.7 \
  --origin gs://BUCKET/model-dir \
  --package-uris gs://BUCKET/staging-dir/my_package-0.1.tar.gz,gs://BUCKET/torchvision-dir/torch-1.4.0+cpu-cp37-cp37m-linux_x86_64.whl \
  --prediction-class predictor.MyPredictor

我尝试指向从不同来源获得的 .whl 文件的不同组合，但要么出现无模块错误，要么内存不足。我不明白在这种情况下模块如何交互以及为什么编译器认为不存在这样的模块。我该如何解决这个问题？或者，我如何自己编译一个包含 torch 和 torchvision 的包。你能给出详细的答案吗，因为我对包管理和bash脚本不是很熟悉。

这是我使用的代码，torch_model.py:

from torch import nn


class EthnicityClassifier44(nn.Module):
    def __init__(self, num_classes=2):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 32, kernel_size=7, stride=1, padding=3)
        self.maxpool1 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv22 = nn.Conv2d(32, 32, kernel_size=3, stride=1, padding=1)
        self.maxpool2 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv3 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)
        self.maxpool3 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv4 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
        self.maxpool4 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.relu = nn.ReLU(inplace=False)
        self.fc1 = nn.Linear(8*8*128, 128)
        self.fc2 = nn.Linear(128, 128)
        self.fc4 = nn.Linear(128, num_classes)


    def forward(self, x):
        x = self.relu(self.conv1(x))
        x = self.maxpool1(x)
        x = self.relu(self.conv22(x))
        x = self.maxpool2(x)
        x = self.maxpool3(self.relu(self.conv3(x)))
        x = self.maxpool4(self.relu(self.conv4(x)))
        x = self.relu(self.fc1(x.view(x.shape[0], -1)))
        x = self.relu(self.fc2(x))
        x = self.fc4(x)
        return x

这是predictor_py:

from facenet_pytorch import MTCNN, InceptionResnetV1, extract_face
import torch
import torchvision
from torchvision import transforms
from torch.nn import functional as F
from PIL import Image
from sklearn.externals import joblib
import numpy as np
import os
import torch_model


class MyPredictor(object):

    import torch
    import torchvision

    def __init__(self, model, preprocessor, device):
        """Stores artifacts for prediction. Only initialized via `from_path`.
        """
        self._resnet = model
        self._mtcnn_mult = preprocessor
        self._device = device
        self.get_std_tensor = transforms.Compose([
            np.float32,
            np.uint8,
            transforms.ToTensor(),
        ])
        self.tensor2pil = transforms.ToPILImage(mode='RGB')
        self.trans_resnet = transforms.Compose([
            transforms.Resize((100, 100)),
            np.float32,
            transforms.ToTensor()
        ])

    def predict(self, instances, **kwargs):

        pil_transform = transforms.Resize((512, 512))

        imarr = np.asarray(instances)
        pil_im = Image.fromarray(imarr)
        image = pil_im.convert('RGB')
        pil_im_512 = pil_transform(image)

        boxes, _ = self._mtcnn_mult(pil_im_512)
        box = boxes[0]

        face_tensor = extract_face(pil_im_512, box, margin=40)
        std_tensor = self.get_std_tensor(face_tensor.permute(1, 2, 0))
        cropped_pil_im = self.tensor2pil(std_tensor)

        face_tensor = self.trans_resnet(cropped_pil_im)
        face_tensor4d = face_tensor.unsqueeze(0)
        face_tensor4d = face_tensor4d.to(self._device)

        prediction = self._resnet(face_tensor4d)
        preds = F.softmax(prediction, dim=1).detach().numpy().reshape(-1)
        print('probability of (class1, class2) = ({:.4f}, {:.4f})'.format(preds[0], preds[1]))

        return preds.tolist()

    @classmethod
    def from_path(cls, model_dir):
        import torch
        import torchvision
        import torch_model

        model_path = os.path.join(model_dir, 'class44_M40RefinedExtra_bin_no_norm_7860.joblib')
        classifier = joblib.load(model_path)

        mtcnn_path = os.path.join(model_dir, 'mtcnn_mult.joblib')
        mtcnn_mult = joblib.load(mtcnn_path)

        device_path = os.path.join(model_dir, 'device_cpu.joblib')
        device = joblib.load(device_path)

        return cls(classifier, mtcnn_mult, device)

和setup.py:

from setuptools import setup

REQUIRED_PACKAGES = ['opencv-python-headless', 'facenet-pytorch']

setup(
 name="my_package",
 version="0.1",
 include_package_data=True,
 scripts=["predictor.py", "torch_model.py"],
 install_requires=REQUIRED_PACKAGES
)

Answer 1

解决方案是将以下包放入自定义预测代码的setup.py 文件中:

REQUIRED_PACKAGES = ['torchvision==0.5.0', 'torch @ https://download.pytorch.org/whl/cpu/torch-1.4.0%2Bcpu-cp37-cp37m-linux_x86_64.whl', 'opencv-python', 'facenet-pytorch']

然后我在自定义类实例化方面遇到了不同的问题，但是 this文章解释得很好。因此，我能够成功地将我的模型部署到 AI 平台进行预测。