How do you choose the group count for GroupNorm in a batch-size-1 3D segmentation network?

## What the group count actually is GroupNorm splits a sample's `C` channels into `G` contiguous groups of `C/G` channels each. For one sample and one group, it pools every value in those `C/G` channels across every spatial position into a single mean and variance, normalizes them, and then applies a learned gain and bias held **per channel** (not per group). Nothing from another sample enters, so the layer is well-defined at any batch size, needs no stored statistics, and computes the same thing in training and in serving. `G` must divide `C`. Beyond that constraint, `G` is a dial with two familiar endpoints: - `G = 1`: one statistic per sample, pooled over all channels and all spatial positions -- the whole-sample, LayerNorm-style extreme. - `G = C`: one statistic per channel per sample, pooled over spatial positions only -- this *is* InstanceNorm. Everything interesting is between them, and the choice is not really about numerical stability. It is about **what you are willing to normalize away**. ## The real tradeoff Normalization deletes the statistic it divides by. At `G = C`, each channel's own mean intensity and contrast, within that one sample, are removed independently. At `G = 1`, only one overall level and spread are removed and the *relative* levels of the channels survive, because they were all divided by the same number. That distinction decides the answer for a volumetric scan segmented at batch size 1. In that setting the absolute intensity level of a region, and the relative response of different channels to it, is signal -- tissue classes are separated partly by how bright they are. Pushing `G` all the way to `C` normalizes each channel's level away sample by sample, which can erase exactly the contrast the head needs to separate classes. Pushing `G` to 1 goes to the other extreme: every channel is tied to a single divisor, so one channel that blows up in scale rescales every other channel in the sample. A middle `G` -- a handful of groups, or equivalently a fixed number of channels per group -- keeps levels comparable within a group while stopping one runaway channel from dragging the whole sample. The usual practical rule is to fix one of the two quantities and let the other follow the width: either a constant number of groups (a small power of two is typical) applied at every layer, or a constant number of channels per group so that `G` grows as the network widens. Fixing channels-per-group keeps the amount of data behind each statistic roughly constant across depth, which is the more defensible choice when the channel count varies a lot between stages. ## Why stability is rarely the binding constraint here Each statistic in a volumetric network is pooled over `(C/G) * D * H * W` values, and the spatial extent of a 3D patch is large. Even a single channel gives a large sample, so the estimate is not noisy in the way a statistic pooled over one image's worth of batch would be. This is worth saying out loud in an interview, because the reflex answer is "use more channels per group so the estimate is stable" -- true for a thin, low-resolution feature map, but not the deciding factor when the spatial volume is big. State the constraint that actually binds: how much per-channel level information you can afford to destroy. ## When the far end of the dial is the right answer Per-channel-per-sample statistics are not a degenerate corner; they are the correct choice when the per-sample level and contrast are *nuisance* rather than signal. A generator that restyles an image is the standard case: the overall colour cast and contrast of the input image are precisely what should not survive into the output, so normalizing each channel of each image on its own removes them by construction, and the style that should appear is reintroduced by the scale and shift applied afterwards. Same layer, opposite conclusion, because the semantics of what is being erased flipped. ## How to answer the question Say what the dial spans, name the two endpoints, and then say that the choice is made by asking what the statistic destroys. For a batch-size-1 volumetric segmentation network: batch-axis statistics are off the table because a single sample gives no batch to pool over, so you are on this spectrum by necessity; pick a middle group count -- several groups, or a fixed channels-per-group so it scales with width -- because per-channel normalization removes intensity information the task depends on, and a single group couples every channel's scale together. Then verify it the only way that counts: sweep two or three group counts and compare validation performance, because the argument above tells you the direction, not the number. ## Things that mark a weak answer Claiming `G = C` leaves nothing to normalize (the spatial positions still supply plenty of values). Claiming the group count is chosen for estimator stability alone. Not knowing that `G` must divide `C`. And treating the learned gain and bias as per-group -- they are per-channel, which is what lets channels inside a shared group recover their own scale.