picograd 1.0.1

A lightweight machine learning framework

picograd

A lightweight machine learning framework

Description • Features • Examples • References • License

Description

A PyTorch-like lightweight deep learning framework written from scratch in Python.

The library has a built-in auto-differentiation engine that dynamically builds a computational graph. The framework is built with basic features to train neural nets: optimizers, training API, data utilities, metrics and loss functions. Additional tools are developed to visualize forward computational graph.

Features

• PyTorch-like auto-differentiation engine (dynamically constructed computational graph)
• Keras-like simple training API
• Neural networks API
• Activations: ReLU, Sigmoid, tanh
• Optimizers: SGD, Adam
• Loss: Mean squared error
• Accuracy: Binary accuracy
• Data utilities
• Computational graph visualizer

Examples

The demo notebook showcases what picograd is all about.

Example Usage

from picograd.engine import Var
from picograd.graph_viz import ForwardGraphViz

graph_builder = ForwardGraphViz()

x = Var(1.0, label='x')
y = (x * 2 + 1).relu(); y.label='y'
y.backward()

graph_builder.create_graph(y)

Output:

Training MLP

import numpy as np
import matplotlib
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.datasets import make_moons

from picograd.nn import MLP
from picograd.engine import Var
from picograd.data import BatchIterator
from picograd.trainer import Trainer
from picograd.optim import SGD, Adam
from picograd.metrics import binary_accuracy, mean_squared_error

# Generate moon-shaped, non-linearly separable data
x_train, y_train = make_moons(n_samples=200, noise=0.10, random_state=0)

model = MLP(in_features=2, layers=[16, 16, 1], activations=['relu', 'relu', 'linear']) # 2 hidden layers
print(model)
print(f"Number of parameters: {len(model.parameters())}")

optimizer = SGD(model.parameters(), lr=0.05)
data_iterator = BatchIterator(x_train, list(map(Var, y_train)))
trainer = Trainer(model, optimizer, loss=mean_squared_error, acc_metric=binary_accuracy)

history = trainer.fit(data_iterator, num_epochs=70, verbose=True)

Decision boundary:

References

• Andrej Karpathy's micrograd library and intro explanation on training neural nets, which is the foundation of picograd's autograd engine.
• Baptiste Pesquet's pyfit library, from which training API was borrowed.

License

MIT