KerasNLP

0.28.0 verified Fri May 01 auth: no python

KerasNLP provides pretrained NLP models (e.g., BERT, GPT-2, T5) and preprocessing utilities for Keras. Current version 0.28.0 supports Python >=3.10. Released approximately monthly.

pip install keras-nlp

Common errors

error AttributeError: module 'keras_nlp' has no attribute 'BertClassifier' ↓

cause Wrong import path. BertClassifier is in keras_nlp.models.

fix

Use: from keras_nlp.models import BertClassifier

error RuntimeError: You must call `build()` on the tokenizer before using it. ↓

cause Tokenizers from presets need to be built or used via from_preset which builds automatically.

fix

Use: tokenizer = keras_nlp.tokenizers.BertTokenizer.from_preset('bert_tiny_en_uncased')

error ValueError: Unknown preset: 'bert_tiny_uncased' ↓

cause Typo or outdated preset name.

fix

Run: print(keras_nlp.presets.list_presets()) to see valid names.

error ImportError: cannot import name 'bert_base_preprocessor' from 'keras_nlp.models' ↓

cause The preprocessor name may vary by preset; not all presets have this symbol.

fix

Use: from keras_nlp.models import BertPreprocessor (or check preset-specific exports).

Warnings

breaking KerasNLP models require Keras 3 and TensorFlow/JAX/PyTorch backend. Using with Keras 2 will fail. ↓

fix Install Keras 3: pip install 'keras>=3.0' and set backend via KERAS_BACKEND env var.

deprecated Presets like 'bert_tiny_en_uncased' may be removed or renamed in future versions. Always refer to the official preset list. ↓

fix Use keras_nlp.presets.list_presets() to see current presets.

gotcha Tokenizers and preprocessors must be built with a vocabulary before use. Calling .fit() on a raw tokenizer will raise an error. ↓

fix Use a preset tokenizer: tokenizer = keras_nlp.tokenizers.BertTokenizer.from_preset('bert_tiny_en_uncased')

breaking From version 0.13.0, 'import keras_nlp' no longer imports TensorFlow. You must explicitly import the backend. ↓

fix Add import os; os.environ['KERAS_BACKEND'] = 'tensorflow' before importing keras_nlp.

Imports

KerasNLP
```
import keras_nlp
```
Module imports as keras_nlp, not keras_nlp.models.

BertTokenizer

wrong

from keras_nlp.models import BertTokenizer

correct

from keras_nlp.tokenizers import BertTokenizer

BERTTokenizer is in the tokenizers submodule, not models.

bert_base_preprocessor
wrong
```
import bert_base_preprocessor
```
correct
```
from keras_nlp.models import bert_base_preprocessor
```
Preprocessors are in the models submodule with model name prefix.

Quickstart

Load a pretrained BERT classifier and make a prediction.

import keras_nlp
import keras

# Load a pretrained BERT model
model = keras_nlp.models.BertClassifier.from_preset(
    "bert_tiny_en_uncased",
    num_classes=2,
)

# Encode sample text using the built-in preprocessor
preprocessor = keras_nlp.models.bert_base_preprocessor()
inputs = preprocessor(["Hello, world!"])

# Predict
outputs = model(inputs)
print(outputs)