What is a solid default augmentation recipe for training a modern vision model?
mediumAnswer
- Random resized crop, horizontal flip, color jitter (brightness/contrast/saturation/hue), and one of RandAugment / TrivialAugment for stronger transforms.
- Add Random Erasing / Cutout.
- For classification, layer in Mixup (α=0.2) and CutMix (α=1.0).
- Label smoothing 0.1.
- Combined with AdamW + cosine schedule + EMA, this recipe gives you ~+2-4% over vanilla training on ImageNet and is the modern baseline (DeiT / ConvNeXt).
How to say it out loud
I'd start with random resized crop and horizontal flip — those two always pay for themselves. Then RandAugment or TrivialAugment for the stronger transforms, Random Erasing, and if it's classification, Mixup and CutMix on top. Label smoothing at 0.1. That's essentially the DeiT recipe, and it buys you two to four points on ImageNet over training with no augmentation at all. The caveat I'd add is that this is tuned for long schedules — three hundred epochs. On a short fine-tune, that much augmentation underfits, and I'd strip it back to crop, flip, and light jitter.
Check yourself — multiple choice
- Only random crop
- RandomResizedCrop + flip + color jitter + RandAugment + RandomErasing + Mixup/CutMix + label smoothing 0.1
- No augmentation
- Only rotation
Modern vision recipe: crop + flip + jitter + RandAug + erase + Mixup/CutMix + smoothing.
What the interviewer asks next
- What is label smoothing and why does it help?You named it as part of the recipe, so expect to be asked what it actually does to the loss.
- How does dropout work and when is it applied?Modern vision recipes lean on augmentation and drop dropout almost entirely. Interviewers like asking why the trade went that way.
- Why use a learning-rate schedule (warmup + cosine decay)?The recipe only reaches those numbers with the schedule attached to it — quoting the gain without the schedule is a common slip.
- What inductive biases do CNNs have?Augmentation buys invariances the architecture does not already have, so the two questions are really one.
#augmentation#training#computer-vision
Practise Deep Learning
214 interview questions in this topic.
Related questions
- How should training resolution be chosen for a CNN or ViT?
- Your validation loss is lower than your training loss. Is something broken?
- Your augmentation pipeline made validation accuracy worse. What went wrong?
- What are vanishing and exploding gradients, and how do you fix them?
- What does Batch Normalization do?