TypeError: string indices must be integers

[markdjthomas]

I am trying to train a model using some custom data and keep running into this issue when in model.train() mode.

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-49-125710d48032> in <module>
----> 1 out = model(f, t)

~/pytorch/lib/python3.6/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
    725             result = self._slow_forward(*input, **kwargs)
    726         else:
--> 727             result = self.forward(*input, **kwargs)
    728         for hook in itertools.chain(
    729                 _global_forward_hooks.values(),

yolov5-rt-stack/yolort/models/yolo_module.py in forward(self, inputs, targets)

yolov5-rt-stack/yolort/models/transform.py in postprocess(self, result, image_shapes, original_image_sizes)

TypeError: string indices must be integers

For context f is a list of tensors of shape [3, 257, 79] and t is a list of dicts that look like the following:

{
   'boxes': tensor([[ 6., 14., 21., 61.], [34., 15., 48., 57.]]),
   'labels': tensor([3, 3]),
   'id': tensor([0]),
   'area': tensor([705., 588.]),
   'iscrowd': tensor([0, 0])
}

I feel like this is likely a very obvious issue on my end, but I can't seem to figure it out.

[zhiqwang]

Hi @markdjthomas ,

Thanks for the report and your attention here! The trainer with lightning is not stable now, but I'll be looking into fixing this!

[zhiqwang]

Hi @markdjthomas

Sorry for not solving this problem for a while, I was on vacation a few days ago.

I've written a unit-test in #87 to reproduce this bug and now it can report the same error in GH Actions. It seems that there are some bugs in the vanilla yolov5 module

https://github.com/zhiqwang/yolov5-rt-stack/blob/c4d3962bc4bb4dccf612dedf67aa9cc9eb89376a/yolort/models/yolo_module.py#L19

One should use the training_step to train a model.

https://github.com/zhiqwang/yolov5-rt-stack/blob/c4d3962bc4bb4dccf612dedf67aa9cc9eb89376a/yolort/models/yolo_module.py#L86-L95

In other words, yolort currently only supports training with following interface (Just test for VOC datasets)

CUDA_VISIBLE_DEVICES=0 python -m yolort.train --data_path ./data-bin/voc_devkit --dataset_type voc --years 2007 2012 --num_gpus 1

And there is a unit-test for training with COCO type dataset here.

My next plan is to integrate self.transform into the model to eliminate this problem.

[zhiqwang]

@markdjthomas

I've fixed training with the vanilla module, you can check the unit-test for more details, therefore I'm closing this issue. And please let me know if you have more questions.