The txt of labels the data set

[tyors]

How should I obtain the txt file that labels the data set?
pre_extracted_file = f'{working_dir}/extracted_features/{featureExt}-{magnification}.txt'
where can i get the “pre_extracted_file”

[mohammadAbbasniya]

How should I obtain the txt file that labels the data set? pre_extracted_file = f'{working_dir}/extracted_features/{featureExt}-{magnification}.txt' where can i get the “pre_extracted_file”

The pre_extracted_file can be produced by function extract_features() in this notebook file

[tyors]

Hello, I don't really understand what this dataset is supposed to follow
paths = glob.glob(f'{working_dir}/tf_files/{magnification}//')
Should I put the dataset based on the magnification?

[tyors]

How do I put the dataset into a directory to generate pre_extracted_file?

[mohammadAbbasniya]

@tyors I'm so sorry for the late reply.

It appears that you are using the standard version of the BreakHis dataset with the original order of files. As you mentioned, you need to reorder the files based on magnification:
[tf_files] / [breast_40 or breast_100 or breast_200 or breast_400] / [benign or malignant].

Please consider sending an email directly to me if this is still a problem.

[tyors]

Okay, thank you. After following your explanation, I am conducting experiments. I want to know if it is feasible to optimize these feature extraction models?

[tyors]

I read your paper, and it seems that the IDC dataset was mentioned during the period, but does it seem that it is not used? I would like to ask if you have done any processing for the BreakHis dataset?

[tyors]

Through your BinaryClassification_Testing code, I found that the results I ran were not as high as yours, and the optimal feature extraction model I ran may be the two models highlighted in my picture; I think I know that looking at your code, you are not processing the image, right? I don't know what further improvements I should make.

[mohammadAbbasniya]

Okay, thank you. After following your explanation, I am conducting experiments. I want to know if it is feasible to optimize these feature extraction models?

Of course, you can fine-tune the feature extraction process to achieve better performance.

I read your paper, and it seems that the IDC dataset was mentioned during the period, but does it seem that it is not used?

we only used the BreakHis dataset.

I found that the results I ran were not as high as yours.

here are some causes to this problem:

• The classifier implementations may have undergone changes in the main algorithm or default value of parameters in newer versions.
• I'm not sure whether the parameters provided by Keras for each network have remained constant since the time of my experiment.
• Running the same Python code even with a fixed random seed on different GPUs can sometimes result in different outcomes. Here are some reasons why this might happen: floating point precision, CUDA and cuDNN versions, GPU driver variations, library optimizations, parallel execution differences, etc.
• The behavior of classifiers is stochastic. XGBoost and LightGBM include randomness in the form of data sampling, and CatBoost is inherently stochastic due to its use of random permutations of the data. Although setting a random seed could control these processes, we didn't set that.

By the way, you also obtained the correct results, although yours are not exactly the same. The main idea of the experiment was: "Networks that are pre-trained on a certain distribution can also extract good-enough features for a totally different distribution.". A minor difference in results should not affect this conclusion.

I think I know that looking at your code, you are not processing the image

No, we didn't process images. It would be really interesting to study the impact of different pre-processing algorithms just before feeding the images into the feature extractor. Great idea!