Hello
With reference to convolutional filters, variable size filters are not typically supported as part of traditional convolutional models. This is because:
- The network architecture (especially convolutional and fully connected layers) requires knowledge of the input size to initialize weights and set up internal structures.
- If variable input sizes were permitted, the model would need to be dynamic, with all weights changing for every image size. However, traditional convolutional models are designed to operate with fixed input dimensions.
As a workaround, you can consider the usual practice of resizing the images to a common size before passing into the network.
To do this, you can use functions such as “augmentedImageDatastore” to resize images in batches. To learn more about this function, refer to the link below:
Hope this helps in getting started!
