将 pytorch 模型（.pth）转换为 coreml

问题描述 投票：0回答：2

我是 pytorch 和 coreml 的新手。我从https://github.com/zhangboshen/A2J下载了一个预训练的pytorch模型（.pth文件），我想将其转换为coreml模型以用于ios应用程序。我加载模型如下。 

 import coremltools as ct
 import torch
 import torch.nn as nn

 model = torch.load('/Users/sarojraut/Downloads/side.pth',map_location=torch.device('cpu'))
 example_input = torch.rand(1, 3, 224, 224) 

 traced_model = torch.jit.trace(model, example_input)

但它给出了错误：

 Traceback (most recent call last):
 File "<pyshell#34>", line 1, in <module>
 traced_model = torch.jit.trace(model, dummy_input)
 File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-   packages/torch/jit/_trace.py", line 846, in trace
 name = _qualified_name(func)
 File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site- packages/torch/_jit_internal.py", line 1145, in _qualified_name
 raise RuntimeError("Could not get name of python class object")
 RuntimeError: Could not get name of python class object
python pytorch coreml coremltools
2个回答
0
投票

您收到此错误的原因是因为您仅加载模型的权重或字典。 PyTorch 模型有两个主要部分：架构（模型的类）和权重/字典。他们通常是分开的。

要加载模型，请按照以下步骤操作：

  1. 使用其原始类（ 建筑学）。 A2J_model 在您的情况下。

  2. 加载模型的状态字典（权重）。 HANDS2017.pth 用于此演示，但您可以使用 Github 存储库上提供的任何其他模型。

让我们利用 hands2017.py:

将其付诸实践 
import torch
import model as model

# This number provided by the original hands2017.py source code.
keypointsNumber = 21

model_dir = '/Desktop/A2J/HANDS2017.pth'

# Instance of the model.
net = model.A2J_model(num_classes = keypointsNumber)
# load the weights into the model.
net.load_state_dict(torch.load(model_dir, map_location=torch.device('cpu')))
# Put the model in eval mode before tracing.
net.eval()

example_input = torch.rand(1, 3, 224, 224)

# Trace your model..
traced_model = torch.jit.trace(net, example_input)

# Print the traced model to verify if the tracing was successful.
print(traced_model)

这会打印出我们的追踪模型（torchscript）：

A2J_model(
  original_name=A2J_model
  (Backbone): ResNetBackBone(
    original_name=ResNetBackBone
    (model): ResNet(
      original_name=ResNet
      (conv1): Conv2d(original_name=Conv2d)
      (bn1): BatchNorm2d(original_name=BatchNorm2d)
      (relu): ReLU(original_name=ReLU)
      (maxpool): MaxPool2d(original_name=MaxPool2d)
      (layer1): Sequential(
        original_name=Sequential
        (0): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
          (downsample): Sequential(
            original_name=Sequential
            (0): Conv2d(original_name=Conv2d)
            (1): BatchNorm2d(original_name=BatchNorm2d)
          )
        )
        (1): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
        (2): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
      )
      (layer2): Sequential(
        original_name=Sequential
        (0): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
          (downsample): Sequential(
            original_name=Sequential
            (0): Conv2d(original_name=Conv2d)
            (1): BatchNorm2d(original_name=BatchNorm2d)
          )
        )
        (1): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
        (2): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
        (3): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
      )
      (layer3): Sequential(
        original_name=Sequential
        (0): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
          (downsample): Sequential(
            original_name=Sequential
            (0): Conv2d(original_name=Conv2d)
            (1): BatchNorm2d(original_name=BatchNorm2d)
          )
        )
        (1): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
        (2): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
        (3): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
        (4): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
        (5): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
      )
      (layer4): Sequential(
        original_name=Sequential
        (0): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
          (downsample): Sequential(
            original_name=Sequential
            (0): Conv2d(original_name=Conv2d)
            (1): BatchNorm2d(original_name=BatchNorm2d)
          )
        )
        (1): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
        (2): Bottleneck(
          original_name=Bottleneck
          (conv1): Conv2d(original_name=Conv2d)
          (bn1): BatchNorm2d(original_name=BatchNorm2d)
          (conv2): Conv2d(original_name=Conv2d)
          (bn2): BatchNorm2d(original_name=BatchNorm2d)
          (conv3): Conv2d(original_name=Conv2d)
          (bn3): BatchNorm2d(original_name=BatchNorm2d)
          (relu): ReLU(original_name=ReLU)
        )
      )
      (avgpool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
      (fc): Linear(original_name=Linear)
    )
  )
  (regressionModel): RegressionModel(
    original_name=RegressionModel
    (conv1): Conv2d(original_name=Conv2d)
    (bn1): BatchNorm2d(original_name=BatchNorm2d)
    (act1): ReLU(original_name=ReLU)
    (conv2): Conv2d(original_name=Conv2d)
    (bn2): BatchNorm2d(original_name=BatchNorm2d)
    (act2): ReLU(original_name=ReLU)
    (conv3): Conv2d(original_name=Conv2d)
    (bn3): BatchNorm2d(original_name=BatchNorm2d)
    (act3): ReLU(original_name=ReLU)
    (conv4): Conv2d(original_name=Conv2d)
    (bn4): BatchNorm2d(original_name=BatchNorm2d)
    (act4): ReLU(original_name=ReLU)
    (output): Conv2d(original_name=Conv2d)
  )
  (classificationModel): ClassificationModel(
    original_name=ClassificationModel
    (conv1): Conv2d(original_name=Conv2d)
    (bn1): BatchNorm2d(original_name=BatchNorm2d)
    (act1): ReLU(original_name=ReLU)
    (conv2): Conv2d(original_name=Conv2d)
    (bn2): BatchNorm2d(original_name=BatchNorm2d)
    (act2): ReLU(original_name=ReLU)
    (conv3): Conv2d(original_name=Conv2d)
    (bn3): BatchNorm2d(original_name=BatchNorm2d)
    (act3): ReLU(original_name=ReLU)
    (conv4): Conv2d(original_name=Conv2d)
    (bn4): BatchNorm2d(original_name=BatchNorm2d)
    (act4): ReLU(original_name=ReLU)
    (output): Conv2d(original_name=Conv2d)
  )
  (DepthRegressionModel): DepthRegressionModel(
    original_name=DepthRegressionModel
    (conv1): Conv2d(original_name=Conv2d)
    (bn1): BatchNorm2d(original_name=BatchNorm2d)
    (act1): ReLU(original_name=ReLU)
    (conv2): Conv2d(original_name=Conv2d)
    (bn2): BatchNorm2d(original_name=BatchNorm2d)
    (act2): ReLU(original_name=ReLU)
    (conv3): Conv2d(original_name=Conv2d)
    (bn3): BatchNorm2d(original_name=BatchNorm2d)
    (act3): ReLU(original_name=ReLU)
    (conv4): Conv2d(original_name=Conv2d)
    (bn4): BatchNorm2d(original_name=BatchNorm2d)
    (act4): ReLU(original_name=ReLU)
    (output): Conv2d(original_name=Conv2d)
  )
)

这表明我们追踪成功了。 最后一点，查看这个旧的存储库，想与您分享，您通过利用 Apple 的本机 API 实现了手部和人体姿势估计（这些是基于视觉的模型，由 Apple 直接维护）：

  1. 用视觉检测手部姿势：https://developer.apple.com/documentation/vision/detecting_hand_poses_with_vision
  2. 检测图像中的人体姿势：https://developer.apple.com/documentation/coreml/model_integration_samples/detecting_ human_body_poses_in_an_image

希望这有帮助。愉快地转换为 coreml 模型:)

0
投票

@saroj 你解决这个问题了吗？

最新问题
© www.soinside.com 2019 - 2024. All rights reserved.