Model output differs after changing input data shape (from (5000,8) to (8,5000)) in CNN #1
Description
Context: I changed the input data structure for my model. Previously, the input shape was (5000, 8) with a Conv1D layer, but I switched it to (8, 5000).
Problem: After making the change, the model generates different results compared to before. It seems that changing the input shape affects the output, possibly due to how Conv1D operates on the data.
Root Cause: The order of dimensions in the input data may be affecting how the Conv1D layer processes the data. Since the kernel size and stride are applied along the sequence dimension, altering the order of dimensions changes how the kernel operates on the data. This results in a different processing of the input, which is why the output differs.
Solution Considerations:
I’m considering adjusting the convolution layer to better handle the new input structure.
It’s also important to verify how the data should be oriented for convolution to work correctly.
Next Steps: Further testing and adjustments are needed to ensure that the model's convolution layer properly processes the new input shape.
Labels: bug, model, input-structure
Priority: high-priority