🌊 PaddleOcean

A high-level PaddlePaddle framework inspired by PyTorch Lightning

Trainer · Model · Callbacks · Loggers · DDP · Gear · VisualDL

---

Why PaddleOcean?

PaddleOcean maps every core component of PyTorch Lightning to PaddlePaddle's native API, with zero PyTorch dependencies. If you know Lightning, you already know PaddleOcean.

Lightning-style hooks:

import paddle
import ocean

class MyModel(ocean.Model):
    def __init__(self):
        super().__init__()
        self.linear = paddle.nn.Linear(28, 10)

    def training_step(self, batch, batch_idx):
        x, y = batch
        loss = paddle.nn.functional.cross_entropy(self(x), y)
        self.log("train_loss", loss)
        return loss

    def configure_optimizers(self):
        return paddle.optimizer.Adam(learning_rate=1e-3, parameters=self.parameters())

model = MyModel()
trainer = ocean.Trainer(max_epochs=10)
trainer.fit(model, train_loader, val_loader)

Keras-style quick prototyping:

net = paddle.nn.Sequential(...)
model = ocean.Model(__model__=net)
model.prepare(optimizer=opt, loss=loss_fn, metrics=[acc])
model.fit(train_loader, epochs=10)

---

Features

Category	Coverage
Model (Keras + Lightning dual-mode)	✅
Trainer (fit / validate / test / predict)	✅
DataModule (data lifecycle)	✅
Gear (Fabric equivalent, manual training)	✅
distributed (70+ Paddle distributed APIs)	✅
Callbacks (18 kinds: Checkpoint, EarlyStopping, Timer, SWA, SpikeDetection, ...)	✅
Loggers (9 kinds: CSV, VisualDL, TensorBoard, Wandb, MLFlow, Comet, Ocelogger)	✅
Strategies (6: SingleDevice, DDP, DeepSpeed, FSDP, ModelParallel)	✅
Accelerators (7 devices: CPU, CUDA, ROCm, XPU, IPU, CustomDevice)	✅
Precision (4: 32, AMP O1/O2, Half, Double)	✅
CI (3 OS × 4 Python versions × lint)	✅

~98% feature parity with pytorch-lightning/src/lightning/pytorch/ (140+ source files).

---

Quick Start

pip install paddlepaddle
pip install -e .
pytest tests/ -v --timeout=120

import ocean
print(ocean.__all__)  # 60+ exported symbols

---

Project Structure

paddleOcean/
├── ocean/
│   ├── __init__.py, model.py, datamodule.py
│   ├── trainer/              # Full engine + connectors
│   ├── gear.py               # Manual training API
│   ├── distributed.py        # 70+ Paddle distributed API wrapper
│   ├── callbacks/            # 18 callbacks
│   ├── loggers/              # 9 loggers
│   ├── strategies/           # 6 strategies
│   ├── accelerators/         # 7 device backends
│   ├── plugins/              # precision / IO / environments
│   ├── profilers/            # profiling tools
│   ├── core/                 # hooks, mixins, saving, optimizer
│   └── utils/                # types, seed, rank_zero, exceptions, ...
├── tests/                    # 44 tests (pytest)
├── .github/workflows/        # CI: ubuntu + windows + macOS
├── QWEN.md                   # Full architecture documentation
└── pyproject.toml

---

License

Apache 2.0