Sergey cheremshinsky

[sergTch]

Description

Start - Preprocessing data "spliting columns. Example:
Embark that has values S, Q, C splited into
Embark_S, Embark_Q, Embark_C with values 0 or 1

Train - made function that train 5 models from sklearn and allow them to vote
on validation test showes ~82.5% accuracy
also made neural network using tensorflow with ~86% accuracy

However both tries showes only 76-80% accuracy on kaggle

How has this been tested?

• Executed code few times from start to end without crashing
• Only ran once
• Code have dedicated unit tests

[meanalexrin]

LGTM - Looks Good To Me
Except:
Please move you work to the /[username]_code location, as described in readme
https://github.com/rnd4u-org/2021-knu-cairl#development

I'll approve after that
Also that's a good approach that you are using separate .py files rather than a single jupyter notebook

[meanalexrin]

Usually it's better to remove unused code from the project. It helps readability

[meanalexrin]

You may consider using if __name__ == "__main__" in the python code to indicate entry points to the program

https://stackoverflow.com/a/419185

[sergTch]

I fixed placements of files. Added solution for mask classification.
Didn't add to titanic solution if name == "main": but used it in masks
In new project also deleted all code comments however regretted few times. Need to get used to it:)

[sergTch]

On masks dataset used inception v3 with 1024 512 and 128 dense layers. Trained in 3 steps
freezing 0 : 210 layers, then 0 : 105 and 210 : 316 and last freezed 105 : 316
that helped to train all parts of cnn and achieve 100% on training data and 99.4% on test
At start splited data on test and train where test is 20% of dataset and train is other 80%

[meanalexrin]

Please check other comments

Regarding task 3:
You can try rerun model few times without saving .npy indexes. So you'll test model on few different train/test splits

[meanalexrin]

Try to use less general errors or if statements

[meanalexrin]

Please add constants and description for magical numbers like 1511 and 6042

[meanalexrin]

Please use existing train test split function to assure correct execution

https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.train_test_split.html

[sergTch]

I used saving of model and saving of splited data mostly because that's hard for my laptop to finish training in some reasonable amount of time. So I splited training for three nights saving model after each and using same split of data

[sergTch]

I fixed mistakes that you mentioned about task 3 and also pushed solution for task 4