pipinstallpytreeclassInstall development version
pipinstallgit+https://github.com/ASEM000/pytreeclasspytreeclass is a JAX-compatible class builder to create and operate on stateful JAX PyTrees in a performant and intuitive way, by building on familiar concepts found in numpy, dataclasses, and others.
See documentation and 🍳 Common recipes to check if this library is a good fit for your work. If you find the package useful consider giving it a 🌟.
importjaximportjax.numpyasjnpimportpytreeclassastc@tc.autoinitclassTree(tc.TreeClass):
a: float=1.0b: tuple[float, float] = (2.0, 3.0)
c: jax.Array=jnp.array([4.0, 5.0, 6.0])
def__call__(self, x):
returnself.a+self.b[0] +self.c+xtree=Tree()
mask=jax.tree_map(lambdax: x>5, tree)
tree=tree\
.at["a"].set(100.0)\
.at["b"][0].set(10.0)\
.at[mask].set(100.0)
print(tree)
# Tree(a=100.0, b=(10.0, 3.0), c=[ 4. 5. 100.])print(tc.tree_diagram(tree))
# Tree# ├── .a=100.0# ├── .b:tuple# │ ├── [0]=10.0# │ └── [1]=3.0# └── .c=f32[3](μ=36.33, σ=45.02, ∈[4.00,100.00])print(tc.tree_summary(tree))
# ┌─────┬──────┬─────┬──────┐# │Name │Type │Count│Size │# ├─────┼──────┼─────┼──────┤# │.a │float │1 │ │# ├─────┼──────┼─────┼──────┤# │.b[0]│float │1 │ │# ├─────┼──────┼─────┼──────┤# │.b[1]│float │1 │ │# ├─────┼──────┼─────┼──────┤# │.c │f32[3]│3 │12.00B│# ├─────┼──────┼─────┼──────┤# │Σ │Tree │6 │12.00B│# └─────┴──────┴─────┴──────┘# ** pass it to jax transformations **# works with jit, grad, vmap, etc.@jax.jit@jax.graddefsum_tree(tree: Tree, x):
returnsum(tree(x))
print(sum_tree(tree, 1.0))
# Tree(a=3.0, b=(3.0, 0.0), c=[1. 1. 1.]) |
Under jax.jit jax requires states to be explicit, this means that for any class instance; variables needs to be separated from the class and be passed explictly. However when using TreeClass no need to separate the instance variables ; instead the whole instance is passed as a state.
Using the following pattern,Updating state functionally can be achieved under jax.jit
importjaximportpytreeclassastcclassCounter(tc.TreeClass):
def__init__(self, calls: int=0):
self.calls=callsdefincrement(self):
self.calls+=1counter=Counter() # Counter(calls=0) |
Here, we define the update function. Since the increment method mutate the internal state, thus we need to use the functional approach to update the state by using .at. To achieve this we can use .at[method_name].__call__(*args,**kwargs), this functional call will return the value of this call and a new model instance with the update state.
@jax.jitdefupdate(counter):
value, new_counter=counter.at["increment"]()
returnnew_counterforiinrange(10):
counter=update(counter)
print(counter.calls) # 10 |
Benchmark flatten/unflatten compared to Flax and Equinox
| CPU | GPU |
 |
Benchmark simple training against `flax` and `equinox`
Training simple sequential linear benchmark against flax and equinox
| Num of layers | Flax/tc time | Equinox/tc time |
| 10 | 1.427 | 6.671 |
| 100 | 1.1130 | 2.714 |