Computer Vision Flashcards
What is feature map?
Output pixel based on the multiplication and summation for each position of the kernel on the input image
What is kernel?
(small) tensor with values that are multiplied with the values of an input tensor at the position of the kernel
Concept of Mapping a Convolutional Kernel
Applying the kernel across the image means sliding it over the coordinate grid of pixels. In default, no part of the kernel can overstep the boundaries of the grid
Why apply the convolutional kernel?
Resulting feature map will be smaller than input
What padding does?
adds virtual number of rows and columns in the height and width dimension filled with zeros
Why apply padding?
If the output feature map should have the same size as the input tensor
How many channels used for color images?
3
Kernels and Multichannel Inputs?
In the case of an input with multiple channels, the kernels processing this input will have a channel dimension with the same number of channels
Rules for model plot statistics
Mean and standard deviation should develop smoothly, activations near zero are problematic since they cannot be processed further in a meaningful way
gradient estimate is more accurate with…
higher batch size
Which activations are problematic since they cannot be processed further in a meaningful way?
near zero
Which requirements 1cycle training incorporate?
start low
End low
Medium high
Phases of learning rate schedule?
– Warmup phase: the learning rate is increased gradually
– Annealing phase: the learning rate is decreased again
What solves initial distribution with many near zero activations and shift of the distribution during training?
Batch normalization
Why use batch normalization?
Speed up training (high learning rate), allows sup-optimal starts (less iterations), adds randomness
