Exporting TorchVision MobilenetV3 SSD-Lite

[anderflash]

error log | 日志或报错信息 | ログ

terminate called after throwing an instance of 'c10::Error'
  what():  forward() Expected a value of type 'List[Tensor]' for argument 'images.1' but instead found type 'Tensor'.
Position: 1
Declaration: forward(__torch__.torchvision.models.detection.ssd.SSD self.1, Tensor[] images.1, Dict(str, Tensor)[]? targets.1) -> ((Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor) 5494)
Exception raised from checkArg at /project/pytorch-v2.5.0/aten/src/ATen/core/function_schema_inl.h:17 (most recent call first):
frame #0: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x2d76da0]
frame #1: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x2d765e9]
frame #2: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x4a41f5]
frame #3: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x80f07c]
frame #4: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x80f24f]
frame #5: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x80f51e]
frame #6: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x28d3507]
frame #7: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x5f1c87]
frame #8: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x619104]
frame #9: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x4acad9]
frame #10: <unknown function> + 0x2a1ca (0x7fc6981431ca in /lib/x86_64-linux-gnu/libc.so.6)
frame #11: __libc_start_main + 0x8b (0x7fc69814328b in /lib/x86_64-linux-gnu/libc.so.6)
frame #12: /home/acmt/conda/envs/sy-bm-people-counter/lib/python3.12/site-packages/pnnx/pnnx() [0x4f4f57]

context | 编译/运行环境 | バックグラウンド

Hi,

I'm trying to export the official TorchVision's MobileV3 SSD-lite. However I get the above error. The function ssdlite320_mobilenet_v3_large returns a SSD instance, which contains

    def forward(
        self, images: List[Tensor], targets: Optional[List[Dict[str, Tensor]]] = None
    ) -> Tuple[Dict[str, Tensor], List[Dict[str, Tensor]]]:

So, is it the images type which is the culprit? The problem is that it calls Torchvision's transforms and backbone, which also receives this List[Tensor]. Since it's the official Torchvision code, I don't have control over that. So how to proceed?

how to reproduce | 复现步骤 | 再現方法

import torch
from torchvision.models.detection import (
    SSDLite320_MobileNet_V3_Large_Weights,
    ssdlite320_mobilenet_v3_large,
)
import pnnx


model = ssdlite320_mobilenet_v3_large(
    weights=SSDLite320_MobileNet_V3_Large_Weights.DEFAULT
)
x = torch.rand(1, 3, 320, 320)
model_scripted = torch.jit.script(model, x)
pnnx.export(model_scripted, "ssdlite320_mobilenet_v3_large.pt", x)

more | 其他 | その他