← CLI reference
Command

vraven explain

Explain one saved PyTorch model invocation and write one report

Minimal form

vraven explain <model>

Arguments and options

OptionMeaningDefault
modelPyTorch model fileNone
--model-formatModel storage format. VRAVEN instrumentation supports state_dict and trusted pickle models; TorchScript can be loaded but cannot expose PyTorch hooks.state_dict
--model-factory, --factoryImport path such as my_project.models:ClassifierNone
--factory-kwargsJSON object passed to the model factory{}
--state-dict-key—None
--strip-prefix—None
--device—cpu
--non-strict—False
--trusted-modelAllow Python-pickle model loadingFalse
--inputModel input as .npy, .npz or a trusted tensor-only .pt/.pth fileNone
--targetTarget class indexNone
--mode—gradient
--class-namesComma-separated class labelsNone
--module-policy—leaf
--include-module-typesComma-separated module typesNone
--exclude-module-typesComma-separated module typesNone
--include-name-patternsComma-separated layer-name patternsNone
--exclude-name-patternsComma-separated layer-name patternsNone
--max-layers—256
--max-units—128
--representation-size—64
--max-attention-tokens—128
--histogram-bins—32
--no-auto-move-inputs—False
--no-entropy—False
--no-representations—False
--no-unit-profiles—False
--no-attention—False
--no-input-attribution—False
--capture-full-tensors—False
--no-timing—False
--causal-top-k—3
--causal-strategy—zero
--causal-scale—0.0
--seed—17
--output-pathDot path to logits in structured outputNone
--non-strict-captureContinue when an individual capture operation is unsupportedFalse
--report-formatWrite exactly one report in the selected format (default: html)html
--outputSingle report destination; its extension must match --report-formatNone
--output-dirCompatibility option: place the single report in this directoryNone
Exact local helpRun vraven explain --help against your installed version before automating a release workflow.