Awesome

laxy

This is my "lazy" wrapper around jax, intended to minimize extra work in setting up optimization for simple custom models. For more advanced deep-nn models, I'd recommend libraries like Haiku, Elegy, Flax, or Trax.

Philosophy: "Optimizing a simple model shouldn't require more than 2 lines of code"

import laxy
import jax.numpy as jnp

def model(params, inputs):
  out = inputs["x"] * params["m"] + params["b"]
  loss = jnp.square(inputs["y"] - out)
  return out, loss

opt = laxy.OPT(model, params={"a":1.0,"b":0.0})
opt.fit(inputs={"x":x,"y":y})

Examples:

• linear regression
• gremlin

FAQ

• How do I save/load weights?

  # save
  weights = opt.get_params()
  jnp.save("weights.npy",weights)
  # load
  weights = jnp.load("weights.npy",allow_pickle=True)
  opt.set_params(weights)
  

• Can I use neural networks in my model?

  from jax.experimental import stax
  stax_layers = stax.serial(stax.Dense(5),
                            stax.Elu,
                            stax.Dense(1),
                            stax.Dropout(0.5))
  
  nn_params, nn_layers = laxy.STAX(stax_layers, input_shape=(None,10))
  
  def model(params, inputs):
    out = nn_layers(params["nn"], inputs["x"], rng=inputs["key"]) + params["a"]
    loss = jnp.square(out - inputs["y"]).sum()
    return out, loss
    
  opt = laxy.OPT(model, params={"nn":nn_params,"a":1.0})
  

• Can I use random variables?

  A random key is automatically added to the inputs dict at each optimization step. The seed for this key is set at laxy.OPT(model, seed=0)

  def model(params, inputs):
    out = inputs["x"] * params["m"] + jax.random.normal(inputs["key"],(1,))
    loss = jnp.square(inputs["y"] - out)
    return out, loss
  

  More than one key?

  def model(params, inputs):
    keys = jax.random.split(inputs["key"],2)
    out = inputs["x"] * params["m"] + jax.random.normal(keys[0],(1,))
    loss = jnp.square(inputs["y"] - out) + jax.random.uniform(keys[1],(1,))
    return out, loss
  

• Can I freeze a subset of weights?

  Freeze forever:

  def model(params, inputs):
    out = inputs["x"] * params["m"] + laxy.freeze(params["b"])
    loss = jnp.square(inputs["y"] - out)
    return out, loss
  

  Conditional freeze:

  def model(params, inputs):
    out = inputs["x"] * params["m"] + laxy.freeze_cond(inputs["freeze"],params["b"])
    loss = jnp.square(inputs["y"] - out)
    return out, loss
  
  opt = laxy.OPT(model, params={"a":1.0,"b":0.0})
  opt.fit(inputs={"x":x,"y":y,"freeze":True})  # freeze
  opt.fit(inputs={"x":x,"y":y,"freeze":False}) # unfreeze