Sharing some of JAX's beautiful API with PyTorch users.
Disambiguation: for wardrobes, see link. For peace, you are at the right spot.
Installation
pip install paxlibor
git clone git@github.com:epfml/pax.git
cd pax
python setup.py installPytrees in PyTorch
import torch
import pax
tree = {
"a": [torch.tensor(3.0), torch.tensor(4.0)],
"c": 4
}
pax.tree_map(lambda x: x*2, tree)Note: we currently depend on jax for this functionality, but we could use dm-tree instead to drop the dependency.
Autodiff that looks like JAX
We follow the API of jax.grad:
import pax
f = lambda x: x**2
df_dx = pax.grad(f)
df_dx(2.0) # tensor(4.0)This works with any Pytree as input:
def f(x):
return x["a"] * x["b"]
x = {"a": 2.0, "b": -1.5}
pax.value_and_grad(f)(x) # (tensor(-3.), {'a': tensor(-1.5000), 'b': tensor(2.)})PAX also supports higher-order derivatives:
f = lambda x: 1/6 * x**3
pax.grad(f)(2.0) # tensor(2.)
pax.grad(pax.grad(f))(3.0) # tensor(3.)Example: Minimal SGD
import torch
import pax
f = lambda x: x**2
df_dx = pax.grad(f)
x = torch.randn([]) # initialization
for step in range(20):
x = x - 0.1 * df_dx(x)
print(x, f(x))Example: meta-learning the learning rate
f = lambda x: x**2
df_dx = pax.grad(f)
def sgd(x, lr=0.1, num_steps=10):
for _ in range(num_steps):
x = x - lr * df_dx(x)
return x
# optimize the learning rate
def meta_loss(lr):
x0 = 1.0
return f(sgd(x0, lr=lr))
df_dlr = pax.grad(meta_loss)
lr = 0.1
for _ in range(100):
lr = lr - 0.1 * df_dlr(lr)Converting from PyTorch
We provide a small wrapper for PyTorch modules to make them behave like Haiku.
net = torch.nn.Linear(10, 1) # any torch.nn.Module
# convert
forward = pax.functional_module(net)
# intialize
params, buffers = pax.get_params(net), pax.get_buffers(net)
# run
data_batch = torch.zeros(2, 10)
out, buffers = forward(params, data_batch, buffers=buffers, is_training=True)and also a wrapper to make PyTorch optimizers functional like Optax:
optimizer = pax.functional_optimizer(torch.optim.Adam, lr=1e-3)
f = lambda x: x**2
df_dx = pax.grad(f)
params = torch.tensor(3.)
opt_state = optimizer.init(params)
for step in range(10):
params, opt_state = optimizer.step(params, df_dx(params), opt_state)
print(params.item())Using PAX optimizers with learning rate schedulers looks like this:
optimizer = pax.functional_optimizer(torch.optim.SGD, lr=0)
lr_at_step = pax.functional_schedule(torch.optim.lr_scheduler.LambdaLR, lr_lambda=lambda step: 1/(step+1), initial_lr=0.1)
f = lambda x: x**2
df_dx = pax.grad(f)
params = torch.tensor(3.)
opt_state = optimizer.init(params)
for step in range(10):
params, opt_state = optimizer.step(params, df_dx(params), opt_state, lr=lr_at_step(step))
print(params.item())Runtime overhead
We measured the time required for one epoch on CIFAR-10 with a batch size of 128.
We compare a standard PyTorch implementation based on this tutorial to a PAX one, using pax.value_and_grad, pax.functional_module and pax.functional_optimizer. This is currently a little slower than regular PyTorch code. The peak memory usage could be larger too.
