pytreeclass

Installation |Description |Quick Example |Filtering |StatefulComputation |Applications| Acknowledgements

🛠️ Installation

pip install pytreeclass

Install development version

pip install git+https://github.com/ASEM000/PyTreeClass

📖 Description

PyTreeClass is a JAX-compatible dataclass-like decorator to create and operate on stateful JAX PyTrees.

The package aims to achieve two goals:

🔒 To maintain safe and correct behaviour by using immutable modules with functional API.
To achieve the most intuitive user experience in the JAX ecosystem by :
- 🏗️ Defining layers similar to PyTorch or TensorFlow subclassing style.
- ☝️ Filtering\Indexing layer values by using boolean masking similar to jax.numpy.at[].{get,set,apply,...}
- 🎨 Visualize defined layers in plethora of ways for better debugging and sharing of information.

⏩ Quick Example

🏗️ Create simple MLP

import jax
from jax import numpy as jnp
import pytreeclass as pytc
import matplotlib.pyplot as plt

@pytc.treeclass
class Linear :
   # Any variable not wrapped with @pytc.treeclass
   # should be declared as a dataclass field here
   weight : jnp.ndarray
   bias   : jnp.ndarray

   def __init__(self,key,in_dim,out_dim):
       self.weight = jax.random.normal(key,shape=(in_dim, out_dim)) * jnp.sqrt(2/in_dim)
       self.bias = jnp.ones((1,out_dim))

   def __call__(self,x):
       return x @ self.weight + self.bias

@pytc.treeclass
class StackedLinear:

    def __init__(self,key,in_dim,out_dim,hidden_dim):
        keys= jax.random.split(key,3)

        # Declaring l1,l2,l3 as dataclass_fields is optional
        # as l1,l2,l3 are Linear class that is already wrapped with @pytc.treeclass
        self.l1 = Linear(key=keys[0],in_dim=in_dim,out_dim=hidden_dim)
        self.l2 = Linear(key=keys[1],in_dim=hidden_dim,out_dim=hidden_dim)
        self.l3 = Linear(key=keys[2],in_dim=hidden_dim,out_dim=out_dim)

    def __call__(self,x):
        x = self.l1(x)
        x = jax.nn.tanh(x)
        x = self.l2(x)
        x = jax.nn.tanh(x)
        x = self.l3(x)

        return x

model = StackedLinear(in_dim=1,out_dim=1,hidden_dim=10,key=jax.random.PRNGKey(0))

x = jnp.linspace(0,1,100)[:,None]
y = x**3 + jax.random.uniform(jax.random.PRNGKey(0),(100,1))*0.01

🎨 Visualize

summary tree_box tree_diagram

print(model.summary())
┌────┬──────┬───────┬──────────────┬─────────────────┐
│Name│Type  │Param #│Size          │Config           │
├────┼──────┼───────┼──────────────┼─────────────────┤
│l1  │Linear│20(0)  │80.00B(0.00B) │weight=f32[1,10] │
│    │      │       │              │bias=f32[1,10]   │
├────┼──────┼───────┼──────────────┼─────────────────┤
│l2  │Linear│110(0) │440.00B(0.00B)│weight=f32[10,10]│
│    │      │       │              │bias=f32[1,10]   │
├────┼──────┼───────┼──────────────┼─────────────────┤
│l3  │Linear│11(0)  │44.00B(0.00B) │weight=f32[10,1] │
│    │      │       │              │bias=f32[1,1]    │
└────┴──────┴───────┴──────────────┴─────────────────┘
Total count :	141(0)
Dynamic count :	141(0)
Frozen count :	0(0)
------------------------------------------------------
Total size :	564.00B(0.00B)
Dynamic size :	564.00B(0.00B)
Frozen size :	0.00B(0.00B)
======================================================

using jax.eval_shape (no-flops operation)

note : the created modules in __init__ should be in the same order where they are called in __call__

print(model.tree_box(array=x))
┌──────────────────────────────────────┐
│StackedLinear[Parent]                 │
├──────────────────────────────────────┤
│┌────────────┬────────┬──────────────┐│
││            │ Input  │ f32[100,1]   ││
││ Linear[l1] │────────┼──────────────┤│
││            │ Output │ f32[100,128] ││
│└────────────┴────────┴──────────────┘│
│┌────────────┬────────┬──────────────┐│
││            │ Input  │ f32[100,128] ││
││ Linear[l2] │────────┼──────────────┤│
││            │ Output │ f32[100,128] ││
│└────────────┴────────┴──────────────┘│
│┌────────────┬────────┬──────────────┐│
││            │ Input  │ f32[100,128] ││
││ Linear[l3] │────────┼──────────────┤│
││            │ Output │ f32[100,1]   ││
│└────────────┴────────┴──────────────┘│
└──────────────────────────────────────┘

print(model.tree_diagram())
StackedLinear
    ├── l1=Linear
    │   ├── weight=f32[1,10]
    │   └── bias=f32[1,10]
    ├── l2=Linear
    │   ├── weight=f32[10,10]
    │   └── bias=f32[1,10]
    └──l3=Linear
        ├── weight=f32[10,1]
        └── bias=f32[1,1]

mermaid.io (Native support in Github/Notion)

✨ Generate shareable vizualization links ✨

# generate mermaid diagrams
# print(pytc.tree_viz.tree_mermaid(model)) # generate core syntax
>>> pytc.tree_viz.tree_mermaid(model,link=True)
# 'Open URL in browser: https://pytreeclass.herokuapp.com/temp/?id=*********'

flowchart LR
    id15696277213149321320(<b>StackedLinear</b>)
    id15696277213149321320 ---> |"20 params<br>80.00B"| id159132120600507116("<b>l1</b><br>Linear")
    id159132120600507116 ---- |"10 params<br>40.00B"| id7500441386962467209["<b>weight</b><br>f32[1,10]"]
    id159132120600507116 ---- |"10 params<br>40.00B"| id10793958738030044218["<b>bias</b><br>f32[1,10]"]
    id15696277213149321320 ---> |"110 params<br>440.00B"| id10009280772564895168("<b>l2</b><br>Linear")
    id10009280772564895168 ---- |"100 params<br>400.00B"| id11951215191344350637["<b>weight</b><br>f32[10,10]"]
    id10009280772564895168 ---- |"10 params<br>40.00B"| id1196345851686744158["<b>bias</b><br>f32[1,10]"]
    id15696277213149321320 ---> |"11 params<br>44.00B"| id7572222925824649475("<b>l3</b><br>Linear")
    id7572222925824649475 ---- |"10 params<br>40.00B"| id4749243995442935477["<b>weight</b><br>f32[10,1]"]
    id7572222925824649475 ---- |"1 param<br>4.00B"| id8042761346510512486["<b>bias</b><br>f32[1,1]"]

✂️ Model surgery

# freeze l1
from pytreeclass.tree_util import tree_freeze

model = model.at["l1"].set(tree_freeze(model.l1))

# Set negative_values in l2 to 0
filtered_l2 =  model.l2.at[model.l2<0].set(0)
model = model.at["l2"].set( filtered_l2 )

# apply sin(x) to all values in l3
filtered_l3 = model.l3.at[...].apply(jnp.sin)
model  = model.at["l3"].set(filtered_l3)

# frozen nodes are marked with #
print(model.tree_diagram())
StackedLinear
    ├#─ l1=Linear
    │   ├#─ weight=f32[1,10]
    │   └#─ bias=f32[1,10]
    ├── l2=Linear
    │   ├── weight=f32[10,10]
    │   └── bias=f32[1,10]
    └── l3=Linear
        ├── weight=f32[10,1]
        └── bias=f32[1,1]

☝️ Filtering with `.at[]`

PyTreeClass offers four means of filtering:

Filter by value
Filter by field name
Filter by field type
Filter by field metadata.

The following example demonstrates the usage the filtering. Suppose you have the following (Multilayer perceptron) MLP class

Note in StackedLinear l1 and l2 has a description in field metadata.

Model definition

import jax
from jax import numpy as jnp
import pytreeclass as pytc
import matplotlib.pyplot as plt
from dataclasses import  field

@pytc.treeclass
class Linear :
   weight : jnp.ndarray
   bias   : jnp.ndarray

   def __init__(self,key,in_dim,out_dim):
       self.weight = jax.random.normal(key,shape=(in_dim, out_dim)) * jnp.sqrt(2/in_dim)
       self.bias = jnp.ones((1,out_dim))

   def __call__(self,x):
       return x @ self.weight + self.bias

@pytc.treeclass
class StackedLinear:
    l1 : Linear = field(metadata={"description": "First layer"})
    l2 : Linear = field(metadata={"description": "Second layer"})

    def __init__(self,key,in_dim,out_dim,hidden_dim):
        keys= jax.random.split(key,3)

        self.l1 = Linear(key=keys[0],in_dim=in_dim,out_dim=hidden_dim)
        self.l2 = Linear(key=keys[2],in_dim=hidden_dim,out_dim=out_dim)

    def __call__(self,x):
        x = self.l1(x)
        x = jax.nn.tanh(x)
        x = self.l2(x)

        return x

model = StackedLinear(in_dim=1,out_dim=1,hidden_dim=5,key=jax.random.PRNGKey(0))

Raw model values before any filtering.

print(model)
StackedLinear(
  l1=Linear(
    weight=[[-1.6248673  -2.8383057   1.3969219   1.3169124  -0.40784812]],
    bias=[[1. 1. 1. 1. 1.]]
  ),
  l2=Linear(
    weight=
      [[ 0.98507565]
       [ 0.99815285]
       [-1.0687716 ]
       [-0.19255024]
       [-1.2108876 ]],
    bias=[[1.]]
  )
)

Filter by value

Get all negative values

print(model.at[model<0].get())

StackedLinear(
  l1=Linear(
    weight=[-1.6248673  -2.8383057  -0.40784812],
    bias=[]
  ),
  l2=Linear(
    weight=[-1.0687716  -0.19255024 -1.2108876 ],
    bias=[]
  )
)

Set negative values to 0

print(model.at[model<0].set(0))

StackedLinear(
  l1=Linear(
    weight=[[0.        0.        1.3969219 1.3169124 0.       ]],
    bias=[[1. 1. 1. 1. 1.]]
  ),
  l2=Linear(
    weight=
      [[0.98507565]
       [0.99815285]
       [0.        ]
       [0.        ]
       [0.        ]],
    bias=[[1.]]
  )
)

Apply f(x)=x^2 to negative values

print(model.at[model<0].apply(lambda x:x**2))

StackedLinear(
  l1=Linear(
    weight=[[2.6401937  8.05598    1.3969219  1.3169124  0.16634008]],
    bias=[[1. 1. 1. 1. 1.]]
  ),
  l2=Linear(
    weight=
      [[0.98507565]
       [0.99815285]
       [1.1422727 ]
       [0.03707559]
       [1.4662486 ]],
    bias=[[1.]]
  )
)

Sum all negative values

print(model.at[model<0].reduce(lambda acc,cur: acc+jnp.sum(cur)))
-7.3432307

Filter by field name

Get all fields named l1

print(model.at[model == "l1"].get())

StackedLinear(
  l1=Linear(
    weight=[-1.6248673  -2.8383057   1.3969219   1.3169124  -0.40784812],
    bias=[1. 1. 1. 1. 1.]
  ),
  l2=Linear(weight=[],bias=[])
)

Filter by field type

Get all fields of Linear type

print(model.at[model == Linear].get())

StackedLinear(
  l1=Linear(
    weight=[-1.6248673  -2.8383057   1.3969219   1.3169124  -0.40784812],
    bias=[1. 1. 1. 1. 1.]
  ),
  l2=Linear(
    weight=[ 0.98507565  0.99815285 -1.0687716  -0.19255024 -1.2108876 ],
    bias=[1.]
  )
)

Filter by field metadata

Get all fields of with their metadata equal to {"description": "First layer"}

print(model.at[model == {"description": "First layer"}].get())

StackedLinear(
  l1=Linear(
    weight=[-1.6248673  -2.8383057   1.3969219   1.3169124  -0.40784812],
    bias=[1. 1. 1. 1. 1.]
  ),
  l2=Linear(weight=[],bias=[])
)

Mix and match different filtering methods.

Get only fields named weight of positive values.

mask = (model == "weight") & (model>0)
print(model.at[mask].get())

StackedLinear(
  l1=Linear(weight=[1.3969219 1.3169124],bias=[]),
  l2=Linear(weight=[0.98507565 0.99815285],bias=[])
)

Marking fields non-differentiable ✨ NEW ✨

Automatically marking fields non-differentiable

In the following code example, we train a model with differentiable and non-differentiable fields. Using jax.grad will throw an error, however to circumvent this we use pytc.filter_nondiff to filter out any non-differentiable field.

import pytreeclass as pytc 
import jax.numpy as jnp
import jax
from typing import  Callable

@pytc.treeclass
class Linear:
    weight: jnp.ndarray                 # ✅ differentiable
    bias: jnp.ndarray                   # ✅ differentiable
    other: tuple[int,...] = (1,2,3,4)   # ❌ non-differentiable
    a: int = 1                          # ❌ non-differentiable
    b: float = 1.0                      # ✅ differentiable
    c: int = 1                          # ❌ non-differentiable
    d: float = 2.0                      # ✅ differentiable
    act : Callable = jax.nn.tanh        # ❌ non-differentiable

    def __init__(self,in_dim,out_dim):
        self.weight = jnp.ones((in_dim,out_dim))
        self.bias =  jnp.ones((1,out_dim))

    def __call__(self,x):
        return self.act(self.b+x)

@jax.value_and_grad
def loss_func(model):
    # lets optimize a differentiable field `b`
    # inside a non-differentiable field `act`
    return jnp.mean((model(1.)-0.5)**2)

@jax.jit
def update(model):
    value,grad = loss_func(model)
    return value,model-1e-3*grad

def train(model,epochs=10_000):
    # here we use the filter_nondiff function
    # to filter out the non-differentiable fields
    # otherwise we would get an error
    model = pytc.filter_nondiff(model)
    for _ in range(epochs):
        value,model = update(model)
    return model

# before any filtering or training
model = Linear(1,1)
print(model)
# Linear(
#   weight=[[1.]],
#   bias=[[1.]],
#   other=(1,2,3,4),
#   a=1,
#   b=1.0,
#   c=1,
#   d=2.0,
#   act=tanh(x)
# )


model = train(model)

# after filtering and training
# note that the non-differentiable fields are not updated
# and the differentiable fields are updated
# the non-differentiable fields are marked with a `*`
print(model)
# Linear(
#   weight=[[1.]],
#   bias=[[1.]],
#   *other=(1,2,3,4),
#   *a=1,
#   b=-0.36423424,
#   *c=1,
#   d=2.0,
#   *act=tanh(x)
# )

Marking fields non-differentiable with a mask

In the following example, let's say we want to train only the field `b` and mark all other fields non-differentiable, we can simply do this in the following code

new_model = pytc.filter_nondiff(model, model != "b")
# we can see all fields except `b` are marked with 
# `*` to mark non-differentiable.
print(new_model)

# Linear(
#   *weight=f32[1,1],
#   *bias=f32[1,1],
#   *other=(1,2,3,4),
#   *a=1,
#   b=f32[],
#   *c=1,
#   *d=f32[],
#   *act=tanh(x)
# )


# undo the filtering
# note the removal of `*` that marks non-diff fields
unfiltered_model = pytc.unfilter_nondiff(new_model)
print(unfiltered_model)

# Linear(
#   weight=f32[1,1],
#   bias=f32[1,1],
#   other=(1,2,3,4),
#   a=1,
#   b=f32[],
#   c=1,
#   d=f32[],
#   act=tanh(x)
# )

📜 Stateful computations

First, 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 @pytc.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

import jax
import pytreeclass as pytc

@pytc.treeclass
class Counter:
    calls : int = 0.

    def increment(self):
        self.calls += 1
counter = Counter() # Counter(calls=0.0)

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.jit
def update(counter):
    value, new_counter = counter.at["increment"]()
    return new_counter

for i in range(10):
    counter = update(counter)

print(counter.calls) # 10.0

📝 Applications

Check other packages built on top of PyTreeClass

Differentiable stencil computations

Physics-based Neural network library

📙 Acknowledgements

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

pytreeclass-0.1.10.tar.gz (47.7 kB view details)

Uploaded Oct 8, 2022 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

pytreeclass-0.1.10-py3-none-any.whl (52.2 kB view details)

Uploaded Oct 8, 2022 Python 3

File details

Details for the file pytreeclass-0.1.10.tar.gz.

File metadata

Download URL: pytreeclass-0.1.10.tar.gz
Upload date: Oct 8, 2022
Size: 47.7 kB
Tags: Source
Uploaded using Trusted Publishing? No
Uploaded via: twine/4.0.1 CPython/3.10.7

File hashes

Hashes for pytreeclass-0.1.10.tar.gz
Algorithm	Hash digest
SHA256	`d3ae5e447b7c9174f8d7330a7f9f22bfe16b02befe845ea6952ce90e052882cd`
MD5	`7b9ee58043be0a798880722416082001`
BLAKE2b-256	`909f5a4b8ea80f6775dfc2b52415203506ea8a24a4c5480a030736b678229196`

See more details on using hashes here.

File details

Details for the file pytreeclass-0.1.10-py3-none-any.whl.

File metadata

Download URL: pytreeclass-0.1.10-py3-none-any.whl
Upload date: Oct 8, 2022
Size: 52.2 kB
Tags: Python 3
Uploaded using Trusted Publishing? No
Uploaded via: twine/4.0.1 CPython/3.10.7

File hashes

Hashes for pytreeclass-0.1.10-py3-none-any.whl
Algorithm	Hash digest
SHA256	`eaf39219e8935906cd3e6ca33e4233d392e18dd81b2ad70c4605aacbda59d712`
MD5	`0c6efc4b2471ac74375d9d578b8f7e69`
BLAKE2b-256	`a9190cb71aead234fa1f3f74587bd6eef6a194b2c823fb62ab1e44648b16221a`