TylerYep/torchinfo

What it solves

PyTorch's default print(model) provides limited information about a network's architecture. torchinfo provides a detailed, formatted summary of a PyTorch model, similar to Keras/TensorFlow's model.summary(), which is essential for debugging and verifying the structure of neural networks.

How it works

The tool takes a PyTorch nn.Module and either a sample input size or actual input data. It performs a forward pass to track the shapes of tensors as they flow through the layers, calculating the number of parameters, output shapes, and the number of multiply-accumulate operations (Mult-Adds).

Who it’s for

Deep learning practitioners and researchers using PyTorch who need to visualize their model architecture, verify tensor shapes, and estimate the memory footprint of their models.

Highlights

• Comprehensive Metrics: Displays layer names, input/output shapes, parameter counts, and Mult-Adds.
• Flexible Input: Supports providing either the input shape (as a tuple/list) or actual input tensors, including multiple inputs with different data types.
• Advanced Architecture Support: Handles RNNs, LSTMs, recursive layers, nn.Sequential, and nn.ModuleList.
• Customizable Output: Allows users to configure column visibility, row settings, depth of nested layers to display, and verbosity levels.
• Memory Estimation: Calculates the estimated total size of the model, including input size, parameter size, and forward/backward pass size.