PyTorch Backend

The research and prototyping home

The PyTorch backend (PyTorch 2.1.0+) plugs Keras straight into the framework most of the research world already lives in: Hugging Face Transformers, torchvision, torchaudio, and the bulk of new pretrained models ship PyTorch-first. Eager-by-default execution means you debug with plain Python tools — set a breakpoint, inspect a tensor, no graph ceremony.

A Keras model IS a torch.nn.Module

The detail that unlocks the interop: under this backend, Keras layers and models become real torch.nn.Module instances (Code block). That's not a wrapper — it's the actual type. So a Keras model can sit inside a hand-written PyTorch training loop, register in a parent nn.Module, or hand its parameters to a torch.optim optimizer. You keep Keras's compile/fit convenience and still drop down to raw torch.autograd the moment you need it.

The one thing to watch: with KERAS_BACKEND=torch the model expects and returns torch tensors — feed it a NumPy or TF tensor and you'll hit a type error. Cross-backend code goes through keras.ops, which the next lessons cover.

Code

A Keras model is a real torch.nn.Module·python

os.environ["KERAS_BACKEND"] = "torch"
import keras
from keras import layers

model = keras.Sequential([
    layers.Dense(128, activation="relu"),
    layers.Dense(10, activation="softmax"),
])

# Keras model IS a torch.nn.Module
print(isinstance(model, torch.nn.Module))  # True

The research and prototyping home

A Keras model IS a torch.nn.Module

Code

External links

Exercise

Progress

Comments 0