如何将自定义 Pytorch 模型转换为 torchscript（pth 到 pt 模型）？

Posted 2023-03-12

【中文标题】如何将自定义 Pytorch 模型转换为 torchscript（pth 到 pt 模型）？

【英文标题】：How to convert custom Pytorch model to torchscript (pth to pt model)?

【发布时间】：2021-05-23 23:08:19

【问题描述】：

我在 colab 环境中使用 PyTorch 训练了一个自定义模型。我成功地将经过训练的模型保存到 Google Drive，名称为 model_final.pth。我想将model_final.pth 转换为model_final.pt 以便可以在移动设备上使用。

我用来训练模型的代码如下：

from detectron2.engine import DefaultTrainer

cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"))
cfg.DATASETS.TRAIN = ("mouse_train",)
cfg.DATASETS.TEST = ()
cfg.DATALOADER.NUM_WORKERS = 2
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml") 
cfg.SOLVER.IMS_PER_BATCH = 2
cfg.SOLVER.BASE_LR = 0.00025 
cfg.SOLVER.MAX_ITER = 1000   
cfg.SOLVER.STEPS = []        
cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 512   
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 1  
cfg.OUTPUT_DIR="drive/Detectron2/"

os.makedirs(cfg.OUTPUT_DIR, exist_ok=True)
trainer = DefaultTrainer(cfg) 
trainer.resume_or_load(resume=False)
trainer.train()

我用来转换模型的代码如下：

from detectron2.modeling import build_model
import torch
import torchvision

print("cfg.MODEL.WEIGHTS: ",cfg.MODEL.WEIGHTS)   ## RETURNS : cfg.MODEL.WEIGHTS:  drive/Detectron2/model_final.pth
model = build_model(cfg)
model.eval()
example = torch.rand(1, 3, 224, 224)
traced_script_module = torch.jit.trace(model, example)
traced_script_module.save("drive/Detectron2/model-final.pt")

但我收到此错误 IndexError: too many indices for tensor of dimension 3 :

cfg.MODEL.WEIGHTS:  drive/Detectron2/model_final.pth
/usr/local/lib/python3.6/dist-packages/torch/tensor.py:593: RuntimeWarning: Iterating over a tensor might cause the trace to be incorrect. Passing a tensor of different shape won't change the number of iterations executed (and might lead to errors or silently give incorrect results).
  'incorrect results).', category=RuntimeWarning)
---------------------------------------------------------------------------

IndexError                                Traceback (most recent call last)
<ipython-input-17-8e544c0f39c8> in <module>()
      7 model.eval()
      8 example = torch.rand(1, 3, 224, 224)
----> 9 traced_script_module = torch.jit.trace(model, example)
     10 traced_script_module.save("drive/Detectron2/model_final.pt")

7 frames
/usr/local/lib/python3.6/dist-packages/detectron2/modeling/meta_arch/rcnn.py in <listcomp>(.0)
    219         Normalize, pad and batch the input images.
    220         """
--> 221         images = [x["image"].to(self.device) for x in batched_inputs]
    222         images = [(x - self.pixel_mean) / self.pixel_std for x in images]
    223         images = ImageList.from_tensors(images, self.backbone.size_divisibility)

IndexError: too many indices for tensor of dimension 3

【参考方案1】：

默认Detectron2 models expect a dictionary or a list of dictionaries as input。

所以你不能直接使用torch.jit.trace函数。但是它们提供了一个包装器，称为TracingAdapter，它允许模型将张量或张量元组作为输入。您可以在他们的torchscript tests 中了解如何使用它。

跟踪你的 Mask RCNN 模型的代码可能是（我没有尝试过）：

import torch
import torchvision
from detectron2.export.flatten import TracingAdapter

def inference_func(model, image):
    inputs = ["image": image]
    return model.inference(inputs, do_postprocess=False)[0]

print("cfg.MODEL.WEIGHTS: ",cfg.MODEL.WEIGHTS)   ## RETURNS : cfg.MODEL.WEIGHTS:  drive/Detectron2/model_final.pth
model = build_model(cfg)
example = torch.rand(1, 3, 224, 224)
wrapper = TracingAdapter(model, example, inference_func)
wrapper.eval()
traced_script_module = torch.jit.trace(wrapper, (example,))
traced_script_module.save("drive/Detectron2/model-final.pt")

更多关于Detectron2部署的信息可以在here找到。

【参考方案2】：

这个例子可以提供帮助。这是类似于底部方法的方法。但它显然与 .pth 一起使用。

    import torch
    import torchvision
    from unet import UNet 
    model = UNet(3, 2)
    model.load_state_dict(torch.load("best_weights.pth"))
    model.eval()
    example = torch.rand(1, 3, 320, 480)
    traced_script_module = torch.jit.trace(model, example)
    traced_script_module.save("model.pt")

来自this 站点的代码。