jnumpy 0.0.0

Jacob's numpy library for machine learning.

Jnumpy

Jacob's numpy library for machine learning

Getting started

1. Install from pip or clone locally:

$ pip install jnumpy

2. Import the jnumpy module.

import jnumpy as jnp

Example

from jnumpy import *

def NN(X_T):
    """Build a simple neural network."""
    H_T = X_T
    for W_T, B_T in zip(layer_weights, layer_biases):
        Z_T = H_T @ W_T
        Z_T = Z_T + B_T
        return Sigm(Z_T)

def loss_fn(ypred_T, ytrue_T):
    """Compute the mean squared error."""
    err_T = (ypred_T - ytrue_T)**2  # [B, C]
    loss_T = ReduceSum(ReduceSum(err_T, axis=1), axis=0)  # [] 

# training pipeline
opt = SGD(lr=1.0)

train_batch_size = min(32, Xtrain.shape[0])
num_train_batches = Xtrain.shape[0] // train_batch_size
test_batch_size = min(32, Xtest.shape[0])
num_test_batches = Xtest.shape[0] // test_batch_size

# training loop
for epoch in range(rounds):

    opt.lr = 0.9 ** epoch # cool learning rate

    # minibatch training
    train_loss = 0.0
    for i in range(num_train_batches):

        # get minibatch
        Xbatch = Xtrain[i*train_batch_size:(i+1)*train_batch_size]
        ybatch = Ytrain[i*train_batch_size:(i+1)*train_batch_size]
        X_T = Var(val=Xbatch, trainable=False)  # [B, N]
        ytrue_T = Var(val=ybatch, trainable=False)  # [B, C]

        # compute loss
        ypred_T = NN(X_T)  # [B, C]
        loss_T = loss_fn(ypred_T, ytrue_T)  # []

        opt.minimize(loss_T) # update weights
        train_loss += loss_T.val # accumulate loss

    # print training stats
    print(f'Epoch {epoch+1}/{rounds} train loss: {train_loss/num_train_batches}')

Limitations and Future Work

Version 2.0 is under development (see the dev branch) and will feature:

• static execution graphs
• a keras-style neural network API with fit, evaluate, and predict and premade layers
• richer collections of optimizers, metrics, and losses
• more examples

Also maybe for the future:

• custom backends (i.e.: tensorflow or pytorch instead of numpy)

License

All code in this repository is licensed under the MIT license. No restrictions, but no warranties. See the LICENSE file for details.

Contributing

This is a small project. I don't plan to grow it unless necesary. You are welcome to fork and contribute or email me jacob [dot] valdez [at] limboid [dot] ai if you would like to take over.