Netron is a free, open source viewer for neural network, deep learning, and machine learning models. It runs on Windows, macOS, Linux, and the web, with a Python package that can host the viewer. The program presents model structure as an interactive graph and exposes details about its operators, connections, and numerical data.
Model files
Users choose a local model through Open Model or drag its files into the browser viewer. Netron identifies the format and loads the model’s graph or weight structure. The browser’s file selection can include accompanying files that the model loader needs.
Supported formats span several frameworks and types of model data:
- ONNX, TensorFlow, TensorFlow Lite, Keras, and Core ML.
- PyTorch, TorchScript, torch.export, ExecuTorch, and OpenVINO.
- Caffe, Darknet, Safetensors, and NumPy.
Netron also provides experimental readers for formats including MLIR, JAX, GGUF, and scikit-learn.
Following operators and connections
The graph represents operators as nodes and links their inputs and outputs through connections. Clicking a node opens its properties, including the operator type and any available name, attributes, input values, and output values. Attribute entries expose the model’s settings for that operation.
A connection has its own properties panel. It identifies the source node and destination nodes alongside the value’s available type information. Selecting linked objects directs attention to their positions in the graph.
When an operator has a definition, Netron opens documentation from its type entry. This panel explains available attributes, inputs, outputs, and type constraints. Models that contain nested graphs or functions also expose navigation into their internal structure.
Weights and tensor values
Expanding a node’s input can reveal a tensor’s type, shape, quantization details, and numerical values. The Show Tensor control opens a separate panel for an initializer or weight. That panel also displays available descriptions, layouts, and other tensor properties.
Netron can calculate statistics for eligible floating-point tensors, including minimum, maximum, mean, standard deviation, and the proportion of zero values. These calculations depend on the tensor’s data type and available contents.
A save control exports eligible tensor data as a NumPy array file. Netron decodes the values and writes the array in NPY format. Unsupported encodings or layouts, missing shape information, and unavailable data can prevent value inspection or export.
Graph search
Find searches the active graph and separates results into nodes, connections, and weights. Users can toggle those categories to control which objects appear. Selecting a result scrolls the graph to the corresponding object.
Ordinary searches match text without regard to letter case. Quotation marks request an exact match instead. Tensor searches can also match data types and shapes, while node searches include operator types and identifiers.
Model Properties supplies broader context, such as the model format, producer, description, domain, and imports when the file provides them. Graph Properties exposes the active graph’s inputs and outputs separately from individual node details.
Adjusting and exporting the diagram
The View controls change graph direction between horizontal and vertical layouts. Separate switches show or hide attributes, weights, and names. Users can zoom in, zoom out, return to actual size, or choose whether the mouse wheel scrolls or zooms.
Netron exports the active graph as PNG or SVG. It carries the graph drawing and its styles into the output, using the diagram’s bounds for the image. PNG produces a raster image, while SVG retains a vector drawing.
Python viewer server
The Python interface accepts a model filename through the command line or a call to netron.start. It starts an HTTP server that supplies the viewer and model content, then opens the browser by default.
Python callers can supply a host and port instead of the default localhost address. They can also disable automatic browser opening. The server returns its address, and the Python interface provides controls to check its status or stop serving the model.
