_torch_to_plotly.md

Plot pytorch module parameters

A simple implementation.

example.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import plotly.express as px\n",
    "import transformers\n",
    "\n",
    "from module_to_plotly import parse_module_into_plotly_data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = transformers.BertModel.from_pretrained(\"ctheodoris/Geneformer\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "fig = px.treemap(\n",
    "    parse_module_into_plotly_data(model),\n",
    "    names=\"name\",\n",
    "    parents=\"parent\",\n",
    "    values=\"params\",\n",
    ")\n",
    "fig.update_traces(root_color=\"lightgrey\")\n",
    "fig.write_html(\"test.html\")"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "adiformer",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}

module_to_plotly.py

from collections import defaultdict
from typing import Literal
from torch.nn import Module


def parse_module_into_plotly_data(
    module: Module, max_depth: int = -1
) -> dict[Literal["name", "parent", "params"], list[str | int]]:

    def object_basename(obj: object) -> str:
        return obj.__class__.__name__.split(".")[-1]

    def module_parameters(module: Module) -> int:
        return sum(p.numel() for p in module.parameters())

    data = dict(
        name=[object_basename(module)], parent=[""], params=[module_parameters(module)]
    )

    names = defaultdict(int)

    def _parse_module(module: Module, parent: str, max_depth: int):
        if max_depth == 0:
            return
        for name, child in module.named_children():
            params = module_parameters(child)
            name = object_basename(child)
            names[name] += 1
            data["name"].append(f"{name}-{names[name]}")
            data["parent"].append(parent)
            data["params"].append(params)
            _parse_module(child, f"{name}-{names[name]}", max_depth - 1)

    _parse_module(module, object_basename(module), max_depth)

    return data