Torchscript支持

注意

尝试在Colab上运行 或转到末尾下载完整的示例代码。

这个例子展示了torchscript对torchvision转换在张量图像上的支持。

from pathlib import Path

import matplotlib.pyplot as plt

import torch
import torch.nn as nn

import torchvision.transforms as v1
from torchvision.io import decode_image

plt.rcParams["savefig.bbox"] = 'tight'
torch.manual_seed(1)

# If you're trying to run that on Colab, you can download the assets and the
# helpers from https://github.com/pytorch/vision/tree/main/gallery/
import sys
sys.path += ["../transforms"]
from helpers import plot
ASSETS_PATH = Path('../assets')

大多数变换支持torchscript。对于组合变换，我们使用 torch.nn.Sequential 而不是 Compose:

dog1 = decode_image(str(ASSETS_PATH / 'dog1.jpg'))
dog2 = decode_image(str(ASSETS_PATH / 'dog2.jpg'))

transforms = torch.nn.Sequential(
    v1.RandomCrop(224),
    v1.RandomHorizontalFlip(p=0.3),
)

scripted_transforms = torch.jit.script(transforms)

plot([dog1, scripted_transforms(dog1), dog2, scripted_transforms(dog2)])

警告

上面我们使用了来自torchvision.transforms命名空间的转换，即“v1”转换。来自torchvision.transforms.v2命名空间的v2转换是在代码中使用转换的推荐方式。

v2 转换也支持 torchscript，但如果你在 v2 类转换上调用 torch.jit.script()，你实际上会得到其（脚本化的）v1 等效版本。由于 v1 和 v2 之间的实现差异，这可能会导致脚本化和即时执行之间的结果略有不同。

如果你真的需要v2转换的torchscript支持，我们建议编写函数来自torchvision.transforms.v2.functional命名空间，以避免意外。

下面我们将展示如何结合图像变换和模型前向传递，同时使用torch.jit.script来获得一个单一的脚本模块。

让我们定义一个Predictor模块，该模块转换输入张量，然后在其上应用ImageNet模型。

from torchvision.models import resnet18, ResNet18_Weights


class Predictor(nn.Module):

    def __init__(self):
        super().__init__()
        weights = ResNet18_Weights.DEFAULT
        self.resnet18 = resnet18(weights=weights, progress=False).eval()
        self.transforms = weights.transforms(antialias=True)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        with torch.no_grad():
            x = self.transforms(x)
            y_pred = self.resnet18(x)
            return y_pred.argmax(dim=1)

现在，让我们定义Predictor的脚本化和非脚本化实例，并将其应用于相同大小的多个张量图像。

device = "cuda" if torch.cuda.is_available() else "cpu"

predictor = Predictor().to(device)
scripted_predictor = torch.jit.script(predictor).to(device)

batch = torch.stack([dog1, dog2]).to(device)

res = predictor(batch)
res_scripted = scripted_predictor(batch)

Downloading: "https://download.pytorch.org/models/resnet18-f37072fd.pth" to /root/.cache/torch/hub/checkpoints/resnet18-f37072fd.pth

我们可以验证脚本化和非脚本化模型的预测是相同的：

import json

with open(Path('../assets') / 'imagenet_class_index.json') as labels_file:
    labels = json.load(labels_file)

for i, (pred, pred_scripted) in enumerate(zip(res, res_scripted)):
    assert pred == pred_scripted
    print(f"Prediction for Dog {i + 1}: {labels[str(pred.item())]}")

Prediction for Dog 1: ['n02113023', 'Pembroke']
Prediction for Dog 2: ['n02106662', 'German_shepherd']

由于模型是脚本化的，它可以轻松地转储到磁盘上并重新使用

import tempfile

with tempfile.NamedTemporaryFile() as f:
    scripted_predictor.save(f.name)

    dumped_scripted_predictor = torch.jit.load(f.name)
    res_scripted_dumped = dumped_scripted_predictor(batch)
assert (res_scripted_dumped == res_scripted).all()