Convolution Kernels / Computer Vision / OpenCV
pytorch.nn.functional ~ about 20 loss functions, a few different distance functions (ex. cosine similarity, levenshtein distance). distance functions are part of loss functions, since a distance between data point pairs/triplets is minimized or maximized (for dissimilarity, for instance). loss function has additional characteristics, like its a summation of distance between pairs, it can also take the log of the distance, to change the magnitude, or there can be another transformation.
syntax is unlikable. although the libs help in tabulating/listing. cv2 syntax more likable.
Syntax of functions hides the math, even simple math, which might be better not as a function. Check if smth interes about the implement.
The image size requirement in neural nets is due to the pooling layers and fully connected layers. This requirement presents itself during inference on a new image, and also if training a model.
From Microsoft Research: https://arxiv.org/pdf/1512.03385.
- Maximum number of layers are convolution layers.
- Between the convolution layers are pooling layers.
- FC layers at the end of neural nets.
Convolution kernels can slide over images of any size. The input image size, kernel size, stride, and padding can affect the output layer size. However, image of any size can be an input to a conv. layer.
