Skip to content

juliuskunze/jaxnet

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

JAXnet Build Status

JAXnet is a deep learning library based on JAX. JAXnet's functional API provides unique benefits over TensorFlow2, Keras and PyTorch, while maintaining user-friendliness, modularity and scalability:

  • More robustness through immutable weights, no global compute graph.
  • GPU-compiled numpy code for networks, training loops, pre- and postprocessing.
  • Regularization and reparametrization of any module or whole networks in one line.
  • No global random state, flexible random key control.

If you already know stax, read this.

Modularity

net = Sequential(Dense(1024), relu, Dense(1024), relu, Dense(4), log_softmax)

creates a neural net model from predefined modules.

Extensibility

Define your own modules using @parametrized functions. You can reuse other modules:

from jax import numpy as jnp

@parametrized
def loss(inputs, targets):
    return -jnp.mean(net(inputs) * targets)

All modules are composed in this way. jax.numpy is mirroring numpy, meaning that if you know how to use numpy, you know most of JAXnet. Compare this to TensorFlow2/Keras:

import tensorflow as tf
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Lambda

net = Sequential([Dense(1024, 'relu'), Dense(1024, 'relu'), Dense(4), Lambda(tf.nn.log_softmax)])

def loss(inputs, targets):
    return -tf.reduce_mean(net(inputs) * targets)

Notice how Lambda layers are not needed in JAXnet. relu and logsoftmax are plain Python functions.

Immutable weights

Different from TensorFlow2/Keras, JAXnet has no global compute graph. Modules like net and loss do not contain mutable weights. Instead, weights are contained in separate, immutable objects. They are initialized with init_parameters, provided example inputs and a random key:

from jax.random import PRNGKey

def next_batch(): return jnp.zeros((3, 784)), jnp.zeros((3, 4))

params = loss.init_parameters(*next_batch(), key=PRNGKey(0))

print(params.sequential.dense2.bias)  # [-0.01101029, -0.00749435, -0.00952365,  0.00493979]

Instead of mutating weights inline, optimizers return updated versions of weights. They are returned as part of a new optimizer state, and can be retrieved via get_parameters:

opt = optimizers.Adam()
state = opt.init(params)
for _ in range(10):
    state = opt.update(loss.apply, state, *next_batch()) # accelerate with jit=True

trained_params = opt.get_parameters(state)

apply evaluates a network:

test_loss = loss.apply(trained_params, *test_batch) # accelerate with jit=True

GPU support and compilation

JAX allows functional numpy/scipy code to be accelerated. Make it run on GPU by replacing your numpy import with jax.numpy. Compile a function by decorating it with jit. This will free your function from slow Python interpretation, parallelize operations where possible and optimize your compute graph. It provides speed and scalability at the level of TensorFlow2 or PyTorch.

Due to immutable weights, whole training loops can be compiled / run on GPU (demo). jit will make your training as fast as mutating weights inline, and weights will not leave the GPU. You can write functional code without worrying about performance.

You can easily accelerate numpy/scipy pre-/postprocessing code in the same way (demo).

Regularization and reparametrization

In JAXnet, regularizing a model can be done in one line (demo):

loss = L2Regularized(loss, scale=.1)

loss is now just another module that can be used as above. Reparametrized layers are one-liners, too (see API). JAXnet allows regularizing or reparametrizing any module or subnetwork without changing its code. This is possible because modules do not instantiate any variables. Instead each module provides a function (apply) with parameters as an argument. This function can be wrapped to build layers like L2Regularized.

In contrast, TensorFlow2/Keras/PyTorch have mutable variables baked into their model API. They therefore require:

  • Regularization arguments on layer level, with separate code necessary for each layer.
  • Reparametrization arguments on layer level, and separate implementations for each layer.

Random key control

JAXnet does not have global random state. Random keys are provided explicitly, making code deterministic and independent of previously executed code by default. This can help debugging and is more flexible (demo). Read more on random numbers in JAX here.

Step-by-step debugging

JAXnet allows step-by-step debugging with concrete values like any plain Python function (when jit compilation is not used).

API and demos

Find more details on the API here.

See JAXnet in action in your browser: Mnist Classifier, Mnist VAE, OCR with RNNs, ResNet, WaveNet, PixelCNN++ and Policy Gradient RL.

Installation PyPI

This is a preview. Expect breaking changes! Python 3.6 or higher is supported. Install with

pip3 install jaxnet

To use GPU, first install the right version of jaxlib.

Questions

Please feel free to create an issue on GitHub.