Hi, how to solve below error to start training? Steve8000818@gmail.com

[nissansz]

/usr/local/lib/python3.7/dist-packages/hydra/core/utils.py:148: UserWarning: register_resolver() is deprecated.
See omry/omegaconf#426 for migration instructions.

OmegaConf.register_resolver(name, f)
The current experiment will be tracked as 20220329-134515
Results will be saved in ./experiments/20220329-134515
/usr/local/lib/python3.7/dist-packages/hydra/utils.py:60: UserWarning: OmegaConf.is_none() is deprecated, see omry/omegaconf#547
if OmegaConf.is_none(config):
/usr/local/lib/python3.7/dist-packages/hydra/_internal/utils.py:710: UserWarning: OmegaConf.is_none() is deprecated, see omry/omegaconf#547
if OmegaConf.is_none(v):
/usr/local/lib/python3.7/dist-packages/hydra/_internal/utils.py:577: UserWarning: OmegaConf.is_none() is deprecated, see omry/omegaconf#547
if OmegaConf.is_dict(x) and not OmegaConf.is_none(x):
Traceback (most recent call last):
File "/usr/local/lib/python3.7/dist-packages/hydra/utils.py", line 109, in instantiate
return target(*args, **final_kwargs)
File "/usr/local/lib/python3.7/dist-packages/torch/utils/data/dataloader.py", line 245, in init
raise ValueError('sampler option is mutually exclusive with '
ValueError: sampler option is mutually exclusive with shuffle

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "/content/u-2-net-portrait-master/train.py", line 77, in main
sampler=sampler)
File "/usr/local/lib/python3.7/dist-packages/hydra/utils.py", line 116, in instantiate
).with_traceback(sys.exc_info()[2])
File "/usr/local/lib/python3.7/dist-packages/hydra/utils.py", line 109, in instantiate
return target(*args, **final_kwargs)
File "/usr/local/lib/python3.7/dist-packages/torch/utils/data/dataloader.py", line 245, in init
raise ValueError('sampler option is mutually exclusive with '
ValueError: Error instantiating/calling 'torch.utils.data.dataloader.DataLoader' : sampler option is mutually exclusive with shuffle

Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

[SpeedOfSpin]

In conf/dataloader/base.yaml set Shuffle=False

[nissansz]

Thank you, will try later. I want to train for Japanese characters. But after trying to train with u2net, I found that few images such as 12, is better than 100k images.
I want to get clean text like printed paper. Any way to get a good model for such demand?

.

[xuebinqin]

could you please describe your problem more clear? What do you mean by 12 is better than 100K? Are they similar datasets. I saw someone trained models for English and Chinese characters segmentation. Japanese characters should also be fine. BTW, are you training the model with large aspect ratio as what you included in the email? If so, did you tune the input size of the model other than just using 320x320?

…

On Wed, Apr 13, 2022 at 9:25 AM nissansz ***@***.***> wrote: Thank you, will try later. I want to train for Japanese characters. But after trying to train with u2net, I found that few images such as 12, is better than 100k images. I want to get clean text like printed paper. Any way to get a good model for such demand? [image: 0_11_30_986__0_11_33_321_0000000000000000000000204_output] <https://user-images.githubusercontent.com/41010669/163215268-070b1ea7-ebf2-4720-8ab2-71b24bb1f478.png> [image: 0_11_26_014__0_11_28_650_0000000000000000000000202_output] <https://user-images.githubusercontent.com/41010669/163215273-edb5af99-61be-4929-9141-ee2102dbcea6.png> . [image: 0_11_26_014__0_11_28_650_0000000000000000000000202_output] <https://user-images.githubusercontent.com/41010669/163215105-3a29788a-1622-40f4-8448-1c57811e2ff0.jpg> [image: 0_11_28_650__0_11_30_986_0000000000000000000000203_output] <https://user-images.githubusercontent.com/41010669/163215131-516f7345-abb7-4f4a-ac80-fc5a07fde6d0.jpg> [image: 0_09_11_446__0_09_14_683_0000000000000000000000163_output] <https://user-images.githubusercontent.com/41010669/163215134-f8bb083b-c785-481d-8595-ae884a4964f4.jpg> [image: 0_11_28_650__0_11_30_986_0000000000000000000000203_output] <https://user-images.githubusercontent.com/41010669/163215229-0170ba9b-34d9-4f0f-ac33-fd508629bac6.png> — Reply to this email directly, view it on GitHub <#14 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSGORPOPYVWZCOHSYJ64T3VE3RPJANCNFSM5R6PG2IQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage: https://xuebinqin.github.io/

[nissansz]

I just want to process subtitle images for ocr.
After trying u2net training, I found, fewer images is much better than more images.
But the result is still similar to binary result. No amazing effect as portraits.

[xuebinqin]

so your input images are just subtitle regions? what are the subtitle image size and the model input resolution, respectively?

…

On Wed, Apr 13, 2022 at 9:46 AM nissansz ***@***.***> wrote: I just want to process subtitle images for ocr. After trying u2net training, I found, fewer images is much better than more images. But the result is still similar to binary result. No amazing effect as portraits. — Reply to this email directly, view it on GitHub <#14 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSGORMSJEX4XJDHNTFRHK3VE3T5VANCNFSM5R6PG2IQ> . You are receiving this because you commented.Message ID: ***@***.***>

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage: https://xuebinqin.github.io/

[nissansz]

video is for example, 720x1280
subtitle is 50x1280

[xuebinqin]

so the model input resolution is also 1280x50? How big is your testing set. You just checked a few samples or checked many samples for example over 100 ones, none of them show significant improvements? It shouldn't be like that. There must be something wrong in your training process. Because according to your results I think your model is not well-trained. You can try to train a bit longer. Before that you can conduct a unit test by overfitting a few (e.g 10 or 100) images to see if the model is able to give you close to 100% accuracy. If not, there must be something wrong in your training process.

…

On Wed, Apr 13, 2022 at 9:51 AM nissansz ***@***.***> wrote: video is for example, 720x1280 subtitle is 1280x50 — Reply to this email directly, view it on GitHub <#14 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSGORIGVQYCDFQXP2CVPOLVE3UQTANCNFSM5R6PG2IQ> . You are receiving this because you commented.Message ID: ***@***.***>

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage: https://xuebinqin.github.io/

[nissansz]

I don't know what is the problem and how to improve. You can download below images and try? I tried several image options
train_data30-24.zip

train_data96-24pcs.zip

train_data50-32pcs.zip

[nissansz]

I used your original u2net code to train the images.

[nissansz]

Hi
I revised it and tried to run, there is error below, cannot continue.
Also, I changed the original image supervisely to DUTS datasets.

/usr/local/lib/python3.7/dist-packages/hydra/_internal/defaults_list.py:251: UserWarning: In 'config': Defaults list is missing _self_. See https://hydra.cc/docs/upgrades/1.0_to_1.1/default_composition_order for more information
warnings.warn(msg, UserWarning)
The current experiment will be tracked as 20220414-010556
Results will be saved in ./experiments/20220414-010556
Error executing job with overrides: []
Cannot instantiate config of type RandomRotation.
Top level config must be an OmegaConf DictConfig/ListConfig object,
a plain dict/list, or a Structured Config class or instance.

Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

[nissansz]

so the model input resolution is also 1280x50? How big is your testing set. You just checked a few samples or checked many samples for example over 100 ones, none of them show significant improvements? It shouldn't be like that. There must be something wrong in your training process. Because according to your results I think your model is not well-trained. You can try to train a bit longer. Before that you can conduct a unit test by overfitting a few (e.g 10 or 100) images to see if the model is able to give you close to 100% accuracy. If not, there must be something wrong in your training process.
…
On Wed, Apr 13, 2022 at 9:51 AM nissansz @.> wrote: video is for example, 720x1280 subtitle is 1280x50 — Reply to this email directly, view it on GitHub <#14 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORIGVQYCDFQXP2CVPOLVE3UQTANCNFSM5R6PG2IQ . You are receiving this because you commented.Message ID: @.>
-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage: https://xuebinqin.github.io/

Check below for comparison

[nissansz]

Below is result based on training for white text, black border.
u2net_bce_itr_6000_train_1.350097_tar_0.092083.pth
Test on training data 24 images

Below is result on other new 10 test images with the same trained model

[nissansz]

24 trained images, model: u2net_bce_itr_13000_train_1.316536_tar_0.086602.
Is the training result normal? Should I add more background data for training to improve result for other images?

[nissansz]

u2net_bce_itr_20000_train_1.308634_tar_0.085164.pth

[nissansz]

u2net_bce_itr_25000_train_1.304218_tar_0.084869

[nissansz]

u2net_bce_itr_30000_train_1.301915_tar_0.084681

[nissansz]

u2net_bce_itr_58000_train_1.301812_tar_0.084637.pth

[nissansz]

change from 24pcs to 20000pcs,
resume training from 24pcs model
1000iterations: u2net_bce_itr_1000_train_1.378661_tar_0.108001

[nissansz]

By improving training data, I got better results for Chinese subtitle.
Left is binarized, right is from trainded model, but some places are still not clean. Not sure whether need to improve training data by including border color to mask.

[nissansz]

Below problem occurrs when training for Japanese part. Some characters gradually disappeared during training, how to improve?

1000 iterations resumed from above Chinese model

4000 iterations

[nissansz]

It seems working now.

[xuebinqin]

Or you could try test time augmentation (TTA) to further improve the results.

…

On Fri, Apr 15, 2022 at 8:51 PM nissansz ***@***.***> wrote: By improving training data, I got better results for Chinese subtitle. Left is binarized, right is from trainded model, but some places are still not clean. Not sure whether need to improve training data by including border color to mask. [image: image] <https://user-images.githubusercontent.com/41010669/163658669-d166e83a-b143-4eb0-993b-b21f7c3fcd91.png> — Reply to this email directly, view it on GitHub <#14 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSGORLCGAN3GRXPARVPPP3VFITKHANCNFSM5R6PG2IQ> . You are receiving this because you commented.Message ID: ***@***.***>

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage: https://xuebinqin.github.io/