-m yobx stats … compute statistics on an onnx model

The command analyses an ONNX model and reports the number of nodes per operator type together with an estimation of the computational cost (FLOPs). Results can be printed to the terminal or saved to a CSV or Excel file.

Description

See yobx.helpers.stats_helper.model_statistics().

    usage: stats [-h] [-o OUTPUT] [-v VERBOSE] input
    
    Computes statistics on an ONNX model:
    number of nodes per op_type and estimation of computational cost.
    
    positional arguments:
      input                 onnx model to analyze
    
    options:
      -h, --help            show this help message and exit
      -o OUTPUT, --output OUTPUT
                            output file (.xlsx or .csv) or empty to print on standard output
      -v VERBOSE, --verbose VERBOSE
                            verbosity

Examples

Print statistics to the terminal:

python -m yobx stats model.onnx

Save statistics to a CSV file:

python -m yobx stats model.onnx -o stats.csv

Save statistics to an Excel workbook:

python -m yobx stats model.onnx -o stats.xlsx

Output on a Dummy Model

The following block builds a small MatMul + Relu model on the fly and runs the command to show what the output actually looks like.

    Number of nodes: 2
    
    Report 1 — Node counts per op_type:
      MatMul: 1
      Relu: 1
    
    Report 2 — Estimated FLOPs per op_type:
      MatMul: 256
      Relu: 16
    
    Total estimated FLOPs: 272