The ✨Magical✨ JAX ML Library.

*Serket is the goddess of magic in Egyptian mythology

Installation |Description |Documentation |Quick Example

🛠️ Installation

Install development version

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

📖 Description and motivation

• serket aims to be the most intuitive and easy-to-use machine learning library in jax.
• serket is fully transparent to jax transformation (e.g. vmap,grad,jit,...).

📙 Documentation

• Full documentation
• Train MNIST, UNet, ConvLSTM, PINN
• Model surgery, Parallelism, Mixed precision
• Optimizers, Augmentation composition
• Interoperability with keras, tensorflow

🏃 Quick example

import jax, jax.numpy as jnp
import serket as sk

x_train, y_train = ..., ...
k1, k2 = jax.random.split(jax.random.key(0))

net = sk.tree_mask(sk.Sequential(
    jnp.ravel,
    sk.nn.Linear(28 * 28, 64, key=k1),
    jax.nn.relu,
    sk.nn.Linear(64, 10, key=k2),
))

@ft.partial(jax.grad, has_aux=True)
def loss_func(net, x, y):
    logits = jax.vmap(sk.tree_unmask(net))(x)
    onehot = jax.nn.one_hot(y, 10)
    loss = jnp.mean(softmax_cross_entropy(logits, onehot))
    return loss, (loss, logits)

@jax.jit
def train_step(net, x, y):
    grads, (loss, logits) = loss_func(net, x, y)
    net = jax.tree_map(lambda p, g: p - g * 1e-3, net, grads)
    return net, (loss, logits)

for j, (xb, yb) in enumerate(zip(x_train, y_train)):
    net, (loss, logits) = train_step(net, xb, yb)
    accuracy = accuracy_func(logits, y_train)

net = sk.tree_unmask(net)

📚 Layers catalog

🔗 Common API

Group	Layers
Containers	- Sequential, Random{Choice}

🧠 Neural network package: serket.nn

Group	Layers
Attention	- MultiHeadAttention
Convolution	- {FFT,_}Conv{1D,2D,3D} - {FFT,_}Conv{1D,2D,3D}Transpose - Depthwise{FFT,_}Conv{1D,2D,3D} - Separable{FFT,_}Conv{1D,2D,3D} - Conv{1D,2D,3D}Local - SpectralConv{1D,2D,3D}
Dropout	- Dropout - Dropout{1D,2D,3D} - RandomCutout{1D,2D,3D}
Linear	- Linear, MLP, Identity
Normalization	- {Layer,Instance,Group,Batch}Norm
Pooling	- {Avg,Max,LP}Pool{1D,2D,3D} - Global{Avg,Max}Pool{1D,2D,3D} - Adaptive{Avg,Max}Pool{1D,2D,3D}
Reshaping	- Upsample{1D,2D,3D} - {Random,Center}Crop{1D,2D,3D} `
Recurrent cells	- {SimpleRNN,LSTM,GRU,Dense}Cell - {Conv,FFTConv}{LSTM,GRU}{1D,2D,3D}Cell
Activations	- Adaptive{LeakyReLU,ReLU,Sigmoid,Tanh}, - CeLU,ELU,GELU,GLU - Hard{SILU,Shrink,Sigmoid,Swish,Tanh}, - Soft{Plus,Sign,Shrink} - LeakyReLU,LogSigmoid,LogSoftmax,Mish,PReLU, - ReLU,ReLU6,SeLU,Sigmoid - Swish,Tanh,TanhShrink, ThresholdedReLU, Snake

🖼️ Image package: serket.image

Group	Layers
Filter	- {FFT,_}{Avg,Box,Gaussian,Motion}Blur2D - {JointBilateral,Bilateral,Median}Blur2D - {FFT,_}{UnsharpMask}2D - {FFT,_}{Sobel,Laplacian}2D - {FFT,_}BlurPool2D
Augment	- Adjust{Sigmoid,Log}2D - {Adjust,Random}{Brightness,Contrast,Hue,Saturation}2D, - RandomJigSaw2D,PixelShuffle2D, - Pixelate2D,Posterize2D,Solarize2D - FourierDomainAdapt2D
Geometric	- {Random,_}{Horizontal,Vertical}{Translate,Flip,Shear}2D - {Random,_}{Rotate}2D - RandomPerspective2D - {FFT,_}ElasticTransform2D
Color	- RGBToGrayscale2D , GrayscaleToRGB2D - RGBToHSV2D, HSVToRGB2D