{"id":5648375,"date":"2023-07-13T08:00:44","date_gmt":"2023-07-13T12:00:44","guid":{"rendered":"https:\/\/lightning.ai\/pages\/?p=5648375"},"modified":"2023-07-24T12:31:51","modified_gmt":"2023-07-24T16:31:51","slug":"tabular-classification-with-lightning","status":"publish","type":"post","link":"https:\/\/lightning.ai\/pages\/blog\/tabular-classification-with-lightning\/","title":{"rendered":"Tabular Classification with Lightning"},"content":{"rendered":"

Takeaways<\/h3>Learn how to gain back research time by leveraging PyTorch Lightning for over 100 inbuilt methods, hooks, and flags that save you engineering hours on heavy lifts like distributed training in multi-GPU and multi-node environments. <\/div>\n

What is Lightning?<\/h2>\n

\"\"<\/a><\/p>\n

The framework known as Lightning<\/a> is PyTorch Lightning<\/a>. Or perhaps it is better to simply say that \u2013 Lightning contains<\/em> PyTorch Lightning. This changed in mid-2022 when PyTorch Lightning was unified with Lightning Apps<\/a> under a single framework and rebranded as Lightning. As of early 2023, the Lightning repository also includes Lightning Fabric<\/a> \u2013 a lightweight way to scale PyTorch models while returning control over the training loop back to the engineer.<\/p>\n

Neither PyTorch Lightning or Lightning Fabric are meant to replace PyTorch. Meaning we still need to implement algorithms with PyTorch. Lightning\u2019s PyTorch Lightning and Fabric help us to manage the training process of those PyTorch implementations. Code examples shared later in this post will clarify what this relationship between PyTorch and Lightning looks like.<\/p>\n

Building with Lightning<\/h3>\n

Now that Lightning AI\u2019s frameworks are unified into a single repository and framework known as Lightning, this means that when researchers install Lightning, each of PyTorch Lightning, Lightning Apps, and Lightning Fabric are installed, along with our metrics library \u2013\u00a0TorchMetrics<\/a>.<\/p>\n

Researchers, ML Engineers, and Data Scientists also have the option to install each framework individually with one of: pip install pytorch-lightning<\/code>, pip install lightning-apps<\/code>, or pip install lightning-fabric<\/code><\/p><\/blockquote><\/div><\/blockquote>\n

The key takeaway here is that Lightning is still the trustworthy PyTorch Lightning framework that allows you to easily scale your models without having to write your own distributed training code or write your own training and evaluation loops \u2013 winning you back precious time that can be allocated back to research.<\/p>\n

Why Lightning<\/h2>\n

Lightning AI saves you time by devoting engineering hours to maintaining Lightning and tackling the tasks you shouldn\u2019t have to handle as a domain researcher, like distributed training on CUDA devices. PyTorch Lightning and Lightning Fabric each enable researchers to focus on the research aspect of their codebase instead of implementing custom methods and properties<\/a>, callbacks<\/a>, or plugins<\/a> on their own.<\/p>\n

How does this compare to NumPy and Pandas? Using PyTorch instead of some combination of NumPy, Pandas, and Python standard means that you won\u2019t have to write your own autograd engine<\/a>. Using PyTorch Lightning or Lightning Fabric with PyTorch means you won\u2019t have to write your own distributed training code with Python and possibly CUDA C++<\/a>.<\/p>\n

\"\"<\/p>\n

Using Lightning with PyTorch<\/h2>\n

What does \u201cUsing Lightning with PyTorch\u201d mean exactly? It means that researchers and engineers can focus on writing their model as Child Modules<\/a> in PyTorch before writing a custom LightningModule<\/a> or adding Lightning Fabric<\/a> into a vanilla PyTorch training loop. A Child Module is the algorithm that has been implemented in PyTorch and will be trained with PyTorch Lightning or Lightning Fabric. A high-level example of this concept is shown below in the code block.<\/p>\n

\n\n
import lightning as L\r\nimport torch\r\nfrom torch import nn\r\n\r\n\r\nclass MyCustomTorchModule(nn.Module):\r\n    def __init__(self, in_features: int, num_classes: int):\r\n        super().__init__()\r\n        linear1 = nn.Linear(\r\n            in_features=in_features,\r\n            out_features=in_features,\r\n        )\r\n        relu = nn.ReLU()\r\n        linear2 = nn.Linear(\r\n            in_features=in_features,\r\n            out_features=num_classes,\r\n        )\r\n        self.sequential = nn.Sequential([linear1, relu, linear2])\r\n\r\n    def forward(self, x):\r\n        return self.sequential(x)\r\n\r\n\r\nclass MyCustomLightningModule(L.LightningModule):\r\n    def __init__(self, in_features: int, num_classes: int):\r\n        super().__init__()\r\n        self.model = MyCustomTorchModule(in_features=in_features, num_classes=num_classes)\r\n\r\n    def forward(self, x: torch.Tensor):\r\n        return self.model(x)\r\n<\/pre>\n<\/code>