agoryuno / deepsolo-onnx Goto Github PK

Hi,
thank you for your implementation.
Unfortunately, I run into some issues with the Google Colab notebook.
I just want to run some tests and want to ask you if you kindly could provide one .onnx file (maybe the one trained on TotalText and mentioned as Res-50, MLT17+IC13+IC15+TextOCR or anyone else) on some cloud storage?
There was some issue with the CUDA version available on Google Colab and the one with which torch/torchvision are built.
Would be great!

Thank you and all the bests,
Stephan

Thanks for your great work!!
But i have got some error,when i try to convert the .pth model to onnx!

xxxxx/DeepSolo/adet/modeling/text_spotter.py:260: TracerWarning: Converting a tensor to a Python number might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
width = input_per_image.get("width", image_size[1].item())

RuntimeError: Only tuples, lists and Variables are supported as JIT inputs/outputs. Dictionaries and strings are also accepted, but their usage is not recommended. Here, received an input of unsupported type: Instances

when i run this export code in colab，this problem occured

。。。

the up ficture is onnx's result

this is my pth file's inference result

When I run the code copied on deepsolo-onnx-export.ipynb to export the model with
BACKBONE: NAME: "build_resnet_backbone"
some error occur as follow:

/data_local/deepsolo/deepsolo-onnx/DeepSolo/adet/modeling/model/detection_transformer.py:221: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if reference.shape[-1] == 2:
/data_local/deepsolo/deepsolo-onnx/DeepSolo/adet/modeling/model/detection_transformer.py:271: TracerWarning: 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).
  for a, b, c, d in zip(outputs_class[:-1], outputs_text[:-1], outputs_coord[:-1], outputs_bd_coord[:-1])
Segmentation fault (core dumped)

the code is as follow

import os
import shutil
from DeepSolo.onnx_model import SimpleONNXReadyModel
import numpy as np
CHECKPOINT = "/result/R50/150k_tt_mlt_13_15/finetune/ic15/model_final.pth"
OUTPATH = "/result/R50/150k_tt_mlt_13_15/finetune/ic15/deepsolo.onnx"
DIMS = (512,512)
CONFIG = "configs/Base_det_export_R5.yaml"
CHANNELS = 3
model = SimpleONNXReadyModel(CONFIG, CHECKPOINT)
img = np.random.randint(0, 255, (CHANNELS, *DIMS))
img = img.astype(np.int8)
import torch.onnx
import torch
img_t = torch.from_numpy(img)
torch.onnx.export(model.model,
           [img_t],
           OUTPATH,
           export_params=True,
           do_constant_folding=True, opset_version=11)

in the detectron2 tutorials, it said that export model into torchscript by traceing can infer on dynamic resolution images:

but when i export the model, the torchscript model only can inferrence on the images that have same height and width with the dummy image i used for exporting the model. i'd like to know whether your exported model can inference on dynamic resolution images.