Pretrained Models

The model zoo, organized by what they do

KerasHub ships dozens of pretrained presets, but they sort cleanly into three architectural families. The table below is the cheat sheet — read the Type column first, because architecture decides what a model is good for far more than its brand name does.

Model	Type	Use Case
BERT	Encoder	Classification, NER, embeddings
GPT-2	Causal LM	Text generation
Gemma (1/2/3/4)	Causal LM	Google's open model family
Llama 2/3	Causal LM	Meta's open models
Mistral	Causal LM	Efficient open model
T5	Encoder-Decoder	Translation, summarization
Whisper	Audio	Speech recognition

One loader, every model

Notice what the Code section shows: a classifier and a generator load with the same from_preset() call — only the class name and preset string change. That uniformity is the payoff of the hub. You don't relearn an API per model; you pick a class for your task, hand it a preset name, and KerasHub fetches the weights and the matching tokenizer in one shot. num_classes on the classifier is the only task-specific knob — it reshapes the output head while leaving the pretrained backbone untouched.

Code

Load a classifier and a generator with the same from_preset call·python

import keras_hub

# Load a pretrained classifier
classifier = keras_hub.models.BertClassifier.from_preset(
    "bert_base_en",
    num_classes=4,
)

# Load a text generator
gpt2 = keras_hub.models.GPT2CausalLM.from_preset("gpt2_base_en")

The model zoo, organized by what they do

One loader, every model

Code

External links

Exercise

Progress

Comments 0