Problem of different predicted results

[kasyful]

Dear Dr @llmpass and Dr @xinwucwp,

I would say this work is great contribution to fault interpretation study.

I try to recreate and build a similar model (same training dataset and same architecture), however when i inference the model to my datasets, it seems to detect bright amplitude, rather than fault itself.

I am genuinely appreciate if you could provide any ideas or suggestion why the model work this way.

Thanks.

[llmpass]

Could you normalize your image, say minus mean and divide by standard deviation before you performing the inference?

[kasyful]

Thank you for your prompt response @llmpass ,

I did tried both z-score and minmax scaler normalization, it showing almost similar results.

My apologies, DataGenerator function already performed z-score for the training datasets. I still couldnt understand why the model predict differently.

Ive go thru each Unet architecture and cross entrophy balanced but still couldnt figure it out.

Thank you for your help.

[llmpass]

Yes, that's strange. I'm sorry that I cannot help you too much this time.

[suporange]

Thank you for your prompt response @llmpass ,

I did tried both z-score and minmax scaler normalization, it showing almost similar results.

My apologies, DataGenerator function already performed z-score for the training datasets. I still couldnt understand why the model predict differently.

Ive go thru each Unet architecture and cross entrophy balanced but still couldnt figure it out.

Thank you for your help.

@kasyful Sir, i have met the same problem. Do you have any ideas?

[kasyful]

Thank you for your prompt response @llmpass , I did tried both z-score and minmax scaler normalization, it showing almost similar results. My apologies, DataGenerator function already performed z-score for the training datasets. I still couldnt understand why the model predict differently. Ive go thru each Unet architecture and cross entrophy balanced but still couldnt figure it out. Thank you for your help.

@kasyful Sir, i have met the same problem. Do you have any ideas?

@suporange , you need to transpose the dataset before inference the model.