abdalazizrashid / idao-21-baseline Goto Github PK

View Code? Open in Web Editor NEW

7.0 2.0 23.0 37.07 MB

International Data Analysis Olympiad Baseline

Home Page: https://idao.world/

License: Apache License 2.0

Python 100.00%

idao-21-baseline's Introduction

IDAO-21 Baseline

Dataset

Dataset is available for download through Yandex disk -> Click here.

Each directory contain images in .png format, the filename contain the class label information.

_dataset/
├── private_test
│   └── *.png
├── publis_test
│   └── *.png
└── train
    ├── ER
    └── NR

Each directory contain bunch of images the image name contain the class label and other informaion. The most important part is the class label which is in {1.0, 3.0, 6.0, 10.0, 20.0, 30.0} and has _keV suffix.

0.00012752991074573__CYGNO_60_40_ER_30_keV_930V_30cm_IDAO_iso_crop_hist_pic_run4_ev846;1.png
0.0001805024851534__CYGNO_60_40_ER_10_keV_930V_30cm_IDAO_iso_crop_hist_pic_run2_ev317;1.png
0.00013557143164375__CYGNO_60_40_ER_10_keV_930V_30cm_IDAO_iso_crop_hist_pic_run2_ev842;1.png
0.00019057084775703__CYGNO_60_40_ER_3_keV_930V_30cm_IDAO_iso_crop_hist_pic_run2_ev116;1.png
0.0001135022106767__CYGNO_60_40_ER_10_keV_930V_30cm_IDAO_iso_crop_hist_pic_run5_ev136;1.png
0.0001275016178883__CYGNO_60_40_ER_3_keV_930V_30cm_IDAO_iso_crop_hist_pic_run2_ev485;1.png
0.0001375808674508__CYGNO_60_40_ER_30_keV_930V_30cm_IDAO_iso_crop_hist_pic_run3_ev662;1.png
0.0011665058173393__CYGNO_60_40_ER_10_keV_930V_30cm_IDAO_iso_crop_hist_pic_run5_ev574;1.png
0.0011465791675372__CYGNO_60_40_ER_3_keV_930V_30cm_IDAO_iso_crop_hist_pic_run2_ev114;1.png
0.0011065850424555__CYGNO_60_40_ER_3_keV_930V_30cm_IDAO_iso_crop_hist_pic_run4_ev868;1.png

Data spliting

The total number of samples in the dataset is ~30k sample distributed with 2 classes ER=0 and NR=1. The dataset are interleaved in the following scheme:

Interleave the files

Energy	He	e
1	*	-
3	-	*
6	*	-
10	-	*
20	*	-
30	-	*

is training; - is testing

Dataset distribution

Training

If you want to retrain the model just run:

mv checkpoints checkpoints_bk && python train.py

This will move the original checkpoints and run the experiment again. Note you can modify config.ini to loader the checkpoint without moving the files.

Results

To generate the report just run:

python report.py

This will generate ./resultsreport.log in the current directory containg information and bunch of plots in the ./results/ directory

Score

Score = MAE - AUC

Classification

Regression

Submission

To generate submission.csv run:

python generate_submission.py

idao-21-baseline's People

Contributors

Stargazers

Watchers

Forkers

temuujine ahmedo42 kborozdin kit-law-42 dgumirov rvashurin papkov timuriskahov yandexdataschool kazeevn

idao-21-baseline's Issues

Clarification for contest output

Hello, could you please help me to interpret the error Yandex Contest raises:

Seems the exception rised, doesnt belong to my code, but to some postprocessors. Compilation logs doesnt help, because at this stage my script finished working and should have created submission file. There is also unit-testing in my script for submission.csv existence.

Unexpected torchvision behaviour

Hello!

I am trying to submit my solution to track-2, which depends on torchvision pretrained models. When i am compiling my model, i have to load the model itself and then pass weights, which i have previously trained. In my code it looks like this:

def compile_model(self, kind):
        model = torchvision.models.mobilenet_v3_small(pretrained=False)

        if kind == 'classifier':
            model.classifier[-1] = torch.nn.Linear(1024, 2)
            weights = torch.load(self.CLASSIFICATION_PATH,
             map_location=torch.device(self.device))
        else:
            model.classifier[-1] = torch.nn.Linear(1024, 6)
            weights = torch.load(self.REGRESSION_PATH,
             map_location=torch.device(self.device))
        
        model.load_state_dict(weights)
        return model

Yandex contest throws an error, saying that there is no such model in torchvision. I installed torchvision version from yours requirements.txt and in my local machine everything runs fine. Could you please help me, because i think there might be a problem with environment installation in Yandex Contest. Thanks!

Error:
File "track2_class.py", line 20, in compile_model
model = torchvision.models.mobilenet_v3_small(pretrained=False)
AttributeError: module 'torchvision.models' has no attribute 'mobilenet_v3_small'

Clarification of contest time limit

Good afternoon!

Could you please help me. I trained the model, prepared archieve with Makefile, everything as in the baseline. I submitted the archieve, but every submission returns the following:

Compiler exit status: TIME_LIMIT_EXCEEDED

Compiler exit code: null

Compiler term sig: KILL

I apologize for the frequent messages, but I would not open this post if I was sure that I understand what this error means. The fact is that according to the rules, TIME_LIMIT is set to 900 seconds, and the above error appears just after 100-200 seconds, and, judging by the logs, the code execution is simply interrupted right during the prediction process of the neural network. That is, probably, the error is not caused by a compilation problem. My model is very lightweight (archieve is <300KB), and memory consumption doesnt exceed 400MB of RAM. Maybe there are some additional constraints to be done?