Pretrained Backbones

The backbone is the feature extractor

A backbone is the model minus its final task layer — the stack of convolutions (or transformer blocks) that turns raw pixels into a rich feature map. Pretrained on ImageNet, it already knows edges, textures, shapes, and object parts; your downstream task only has to learn how to read those features. That's transfer learning's whole leverage, and it's why almost no one in 2026 trains a vision model from random weights.

KerasCV ships a deep menu of backbones spanning the classic CNN, mobile, dense, and ViT-style families:

Backbone	Style	Sizes
EfficientNetV1/V2	Efficient CNN	B0–B7, S/M/L
ResNet	Classic residual	18/34/50/101/152
MiT (Mix Transformer)	ViT-style	B0–B5
MobileNetV3	Mobile-optimized	Small/Large
DenseNet	Dense connections	121/169/201
CSPDarkNet	YOLO backbone	Tiny/S/M/L/XL

Choosing a size

The size suffix (B0 vs B7, ResNet18 vs ResNet152) is a direct latency-vs-accuracy dial: bigger backbones see more, but cost more per inference. Start small. A ResNet50 or EfficientNetV2-S is the workhorse default — enough capacity for most problems, fast enough to iterate on. Reach for the giants only after you've confirmed the small one is genuinely capacity-bound, not data-bound.

Code

Load a pretrained backbone via from_preset·python

import keras_cv

# Load a pretrained backbone
backbone = keras_cv.models.EfficientNetV2Backbone.from_preset(
    "efficientnetv2_s_imagenet"
)

Attach a classification head on top of the backbone·python

import keras

# backbone outputs a feature map; pool it, then classify
model = keras.Sequential([
    backbone,
    keras.layers.GlobalAveragePooling2D(),
    keras.layers.Dense(num_classes, activation="softmax"),
])

# Or skip the manual wiring: a task preset bundles backbone + head
classifier = keras_cv.models.ImageClassifier.from_preset(
    "efficientnetv2_s_imagenet", num_classes=num_classes
)

Exercise

Build three classifiers using ResNet18, ResNet50, and ResNet152 backbones. Fine-tune each on a small dataset (Oxford Flowers 102). Plot accuracy vs. train time vs. parameter count on one chart — then identify the knee where adding capacity stops paying off. That knee is your real default for this data.

Hint

Freeze the backbone for the first few epochs (train only the head), then unfreeze with a low learning rate. The 152 will likely overfit a small dataset faster than the 50 — watch the val curve, not just final accuracy.

The backbone is the feature extractor

Choosing a size

Code

External links

Exercise

Progress

Comments 0