eztorch

点击此处跳转至中文版本

Minimal, NumPy-first deep learning playground with PyTorch-inspired APIs. Includes a tiny Transformer stack, simple trainers, and pluggable tokenization/vectorization for NLP demos.

Features

• Core building blocks: Sequential, Linear, Conv2d, BatchNorm1d, LayerNorm, Residual, activations (ReLU, LeakyReLU, Tanh, Sigmoid, Softmax).
• Training utilities: Trainer with SGD/Adam, CE/MSE losses; build models with Sequential（MLP/CNN 等）。
• Attention/Transformer: multi-head self-attention layer, encoder/decoder layers, and a DefaultTransformer (encoder-decoder with sinusoidal PE and generator head).
• Tests/demos: classification/regression MLPs, toy translation demo (test/test_nlp.py).

Installation

Use your preferred virtual environment and install only the core by default:

# pip
pip install -e .
# or uv
uv pip install -e .

Optional extras to run demos/tests without pulling heavy deps into the core:

# Visualization (matplotlib)
pip install -e .[viz]
# scikit-learn (datasets, demos)
pip install -e .[sklearn]
# tokenizers (modern NLP tokenization)
pip install -e .[nlp]

# everything used by examples/tests
pip install -e .[all]

Quickstart

MLP classification (logits + cross entropy)

import numpy as np
from eztorch.layers.linear import Linear
from eztorch.layers.sequential import Sequential
from eztorch.functions.activations import ReLU
from eztorch.models.sequential_model import SequentialModel
from eztorch.optim.adam import Adam
from eztorch.utils.trainer import Trainer

X = np.random.randn(128, 2)
y = np.random.randint(0, 3, size=128)
mlp = SequentialModel(Sequential([Linear(2, 16), ReLU(), Linear(16, 3)]))
trainer = Trainer(model=mlp.model, forward=mlp.forward, optimizer=Adam(lr=0.01))
losses = trainer.fit(X, y, batch_size=32, max_steps=200, log_every=50)
print("final loss:", losses[-1])

Transformer translation demo (see test/test_nlp.py)

Uses the simple whitespace tokenizer from src/external/ and the built-in DefaultTransformer:

python test/test_nlp.py

This trains on a few tiny source/target pairs and prints decoded predictions, plus an inference on an unseen source sentence.

Tests / demos

• test/test_mlp_classification.py – moons dataset classification with norms + ReLU.
• test/test_mlp_regression.py – synthetic regression with MSE.
• test/test_nlp.py – tiny translation toy using the default Transformer and external tokenizer/vectorizer.