JiaquanYe / TableMASTER-mmocr

2nd solution of ICDAR 2021 Competition on Scientific Literature Parsing, Task B.
Apache License 2.0
430 stars 103 forks source link

Error when training text-line detection model #11

Closed huyhoang17 closed 3 years ago

huyhoang17 commented 3 years ago

Hi, thanks for the great repo!

Thanks in advance!

photo6170031322173648295

JiaquanYe commented 3 years ago

Hi, seem like it is failed to load the training data. And what's your MMDet version? Maybe you can check the data loading setting in config file, or check the data loading pipeline.

huyhoang17 commented 3 years ago

@JiaquanYe lib version

mmocr: 0.2.0 (installed from this repo)
mmdet: 2.11.0 (installed from this repo)
mmcv-full: 1.3.4

I checked and just realized that when training the model, it's called the build_dataset function first to building the dataset. But I searched through all the source code and only find this function from mmdet? Is it normal?

Here is my PSE config: https://pikab.in/5127bb5d68 train.json is under COCO format, images is the folder contains the image data

huyhoang17 commented 3 years ago

@JiaquanYe Can you provide the pre-trained PSENet text-line detection model? Currently, I want to test the pipeline from 3 models but can not train the text-line detection due to this issue. Hope to see your response, thank you!

JiaquanYe commented 3 years ago

@JiaquanYe lib version

mmocr: 0.2.0 (installed from this repo)
mmdet: 2.11.0 (installed from this repo)
mmcv-full: 1.3.4

I checked and just realized that when training the model, it's called the build_dataset function first to building the dataset. But I searched through all the source code and only find this function from mmdet? Is it normal?

Here is my PSE config: https://pikab.in/5127bb5d68 train.json is under COCO format, images is the folder contains the image data

Could you paste your train.json here?

huyhoang17 commented 3 years ago

@JiaquanYe Here is the example of json file: https://pikab.in/537fbc4566

huyhoang17 commented 3 years ago

I found the error. Must set the name of the category to text and turn on with_mask=True in the config file Closed. Thank you for your support!