I want the train from scratch a grayscale image dataset with 1000 lables using the EfficientNetB0 model from the keras.

I am working on a project where I need to train an EfficientNet model using grayscale images. To achieve this, I modified the input shape to accommodate a single channel. However, during training, I am encountering a very high training loss (above 7) and only a minimal increase in accuracy.

Here are the steps I’ve taken:

• Changed the input shape of EfficientNet to handle 1-channel images.
• Used the Adam optimizer with the categorical cross-entropy loss function.
• As an alternative approach, I stacked the grayscale image to simulate an RGB image, but this did not improve the results.

Questions:

• Why is the training loss so high when using grayscale images?
• Why does stacking grayscale images to create RGB images not improve performance?
• Are there specific adjustments or techniques I should consider for using EfficientNet with grayscale images?
• Could there be issues related to data preprocessing or augmentation affecting the training performance?

Any insights or suggestions to improve the model’s performance would be greatly appreciated. Thank you!