hermannsblum / fishyscapes Goto Github PK

I read carefully what you said about the submit process, but I still don't know how to commit my pytorch code?

Hi @hermannsblum,

I cannot submit the Google form. It gives just a message 'Something is wrong'.
I have tried changing to a clean account and using different browsers.

Is there an alternative way to send you the container image?

Hi, thanks for your great contribution!
I am interested in the "Learned Embedding Density" method proposed in the paper. Will you kindly provide the implementation code of it?

Hello, after the training, I am trying to evaluate with your benchmark, fishyscapes.
The error is keep occuring after the line
fs = bdlb.load(benchmark="fishyscapes", download_and_prepare=True)

and the error is
ValueError: To access dl_manager.manual_dir, please set MANUAL_DOWNLOAD_INSTRUCTIONS in your dataset.
in the file 'lib/python3.9/site-packages/tensorflow_datasets/core/download/download_manager.py'

also, there's an warning about the google authentication as below.

Just ignoring the warning, and making the line below as a comment, I could be able to get a fishyscapes static data.

But another attribute error was occured in the file '/lib/python3.9/site-packages/bdlb/fishyscapes/benchmark.py', line 68

AttributeError: module 'tensorflow_datasets.core.load' has no attribute '_dataset_name_and_kwargs_from_name_str'

I can not surely find this issue through the searching, and asking you if you know about the issue.
And, thank you for answering my previous issue. It helped me a lot.

I figured out changing my code as

fs.download_and_prepare('LostAndFound')
ds = tfds.load('fishyscapes/LostAndFound', split='validation')

due to get_dataset function was broken. Thank you.

Hi, the current evaluation procedure seems a bit cumbersome, the evaluation side should care about the running environment of uploaded detector, on the other hand, users of the benchmark have to writing additional code and instructions.

I wonder are you considering to change the evaluation procedure into the form of results uploading like most of other challenges does ? e.g. you can provide download link of test image for different subset (LAF, static, web etc.), users only need to predict a result file for each image (e.g. ${image name}.npy matrix) and upload them as a .zip, instead of uploading the model. In the evaluation side, you can directly evaluate the results regardless of running evironment of anomaly detectors.

I think this change will help to promote this benchmark to broader usage. Thanks~

Thank you for your work on fishyscapes benchmark!
It seems that the link to submit the model on the official website cannot be opened.

Hi. Thank you for maintianing such a benchmark for Anomaly detection. Recently I want to submit my method, but I found it couldn't be created correctly when I try about the example https://github.com/4PiR2/fishyscapes_simg_example. Could you please reproduce it since I can't run it? Besides, could you please give me any guide about how to create that container for my method.

Hello, I'm trying to utilize some stuff with fishyscapes dataset about anomaly segmentation task.

I have followed the instruction in the fishyscapes dataset official site (https://fishyscapes.com/dataset),
and i succeed to download the dataset.

I donwloaded 'fishyscapes_lostandfound', that you uploaded on upper site which was mentioned in the button 'download FS Lost&Found validation set' and it was included of 100 masked images.
Also importing bdlb, I could download gtCoarse and leftImg8bit and it contained train folder and test fold.

So to implement training, validation, and test with these datas, do I need to split validation set inside the training set of gtCoarse and leftImg8bit? Then what is the 'fishyscapes_lostandfound' validation data?

Also I could not get the fishyscapes static dataset with your official site. There was nothing in the folder 'static'.

And is there any gtFine dataset in fishyscapes?

Thank you for reading my long issue, and I really appreciate your work of Fishyscapes benchmark.