Closed mobassir94 closed 2 years ago
gtb1551050818
commented
2 years ago Hello, I encountered the problem of data set loading during the reproduction process:
FileNotFoundError: [Errno 2] No such file or directory: '/checkpoint/jinghuang/datasets/MLT17/train/imgs'
How do you load the dataset correctly? Can you help me? Thank you
mobassir94
commented
2 years ago @gtb1551050818 after spending a lot of time on this project,i faced several issues because of lack of documentation of this project,i needed response and help from authors badly, but the author of this project is irresponsible (it is kind of wasted implementation if your customer's can't use it and you do not care or reply) so i just simply switched to another pipeline for multilingual ocr,,i am working on paddleocr,yes it hasn't got multi head multilingual ocr yet but the authors are very responsible there and helpful. i am not using this multilingual implementation anymore,,it's hard to work on a project that has little or almost no documentation (bad public implementation IMHO)
gtb1551050818
commented
2 years ago @gtb1551050818 after spending a lot of time on this project,i faced several issues because of lack of documentation of this project,i needed response and help from authors badly, but the author of this project is irresponsible (it is kind of wasted implementation if your customer's can't use it and you do not care or reply) so i just simply switched to another pipeline for multilingual ocr,,i am working on paddleocr,yes it hasn't got multi head multilingual ocr yet but the authors are very responsible there and helpful. i am not using this multilingual implementation anymore,,it's hard to work on a project that has little or almost no documentation (bad public implementation IMHO)
Yes, it is really difficult to reproduce the current documents alone, especially for beginners like me. But I noticed that you seem to have loaded the mlt17 dataset correctly. May I take the liberty to ask where you put the mlt17, or what changes you have made to enable it to load the dataset successfully? I also want to try again to see if it can be reproduced. If it succeeds, hahaha
mobassir94
commented
2 years ago @gtb1551050818 here is my last modified version of multiplexedocr where i loaded mlt17 correctly : https://drive.google.com/drive/folders/1ep8GXP3tT2cQN2i22aE7iioldg9y00nP?usp=sharing
please give it a try and if you can train without getting NaN then let me know as well with updated code so that i can learn,thanks
gtb1551050818
commented
2 years ago @gtb1551050818 here is my last modified version of multiplexedocr where i loaded mlt17 correctly : https://drive.google.com/drive/folders/1ep8GXP3tT2cQN2i22aE7iioldg9y00nP?usp=sharing
please give it a try and if you can train without getting NaN then let me know as well with updated code so that i can learn,thanks
Thank you very much. If I can get through, I will tell you!
Shualite
commented
2 years ago @mobassir94 Hello, I met the same problem as you and also stuck on this NaN issue. There is only one difference that I use MLT19 and MLT19 Synthetic as training data follow the original paper settings. At the beginning, we think it's a data problem. Then, we spend a lot time locating the wrong image. So far, the error seems to have happened by accident and we have found nothing.
SuperIRabbit
commented
2 years ago Hi everyone, apologies for the late reply, as I stopped working on OCR a few months ago due to some internal changes. From our experience, the NaN issue typically happens when you are training from scratch with difficult datasets and improper learning rates, which makes the segmentation network fail. Therefore, as mentioned in the paper, "we initialize the detection, segmentation, and mask feature extraction weights from the officially published weights released by Mask TextSpotter v3". I have also updated a pretrained weights for multiplexed model so that you can fine-tune from it.
You can specify
MODEL.WEIGHT ${path_to_pretrained_weights}in your training command to initialize from the pretrained weights.
SuperIRabbit
commented
2 years ago Hello, I encountered the problem of data set loading during the reproduction process:
FileNotFoundError: [Errno 2] No such file or directory: '/checkpoint/jinghuang/datasets/MLT17/train/imgs'
How do you load the dataset correctly? Can you help me? Thank you
Hi @gtb1551050818, you need to download the MLT17 datasets (as well as the other datasets) from the official site and extract the images and ground truth annotations into the corresponding folders
i get nan while training using mlt17 dataset: error coming from this line : https://github.com/facebookresearch/MultiplexedOCR/blob/main/multiplexer/engine/train_loop.py#L170 and from spn.py i don't know how i am getting nan while using mlt17 dataset
@SuperIRabbit need help please,thanks