notebooks/onnx_model_export.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "26947f7e",
   "metadata": {},
   "source": [
    "# ONNX model export\n",
    "\n",
    "In this notebook we'll cover how to export a setfit model to ONNX format.  This is useful for deploying models in production and speeding up inference on CPUs."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "39c7df94",
   "metadata": {},
   "source": [
    "## Setup"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ca808828",
   "metadata": {},
   "source": [
    "If you're running this Notebook on Colab or some other cloud platform, you will need to install a few libraries namely `setfit`, `onnxruntime`, `onnx`, and `skl2onnx`. Uncomment the following cell and run it:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d5d6c9f2",
   "metadata": {},
   "outputs": [],
   "source": [
    "# %pip install setfit[onnx]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "47c3060f",
   "metadata": {},
   "source": [
    "## Convert model to ONNX\n",
    "\n",
    "There are two types of heads available for setfit models. Sklearn based estimators and `torch.nn.Module`. We'll convert the example from `text-classification.ipynb` which uses an `sklearn` head and a base model that hasn't been trained which defaults to a `nn.Module` head.  We start by loading in the models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1a4c320d",
   "metadata": {},
   "outputs": [],
   "source": [
    "from setfit import SetFitModel\n",
    "\n",
    "model = SetFitModel.from_pretrained(\"lewtun/my-awesome-setfit-model\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6080ea85",
   "metadata": {},
   "source": [
    "Next we import the onnx_export onnx function from `setfit.onnx` and run the export.  The export function takes in the `model_body` and the `model_head` from the setfit model loaded above."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "89d24aad",
   "metadata": {},
   "outputs": [],
   "source": [
    "from setfit.exporters.onnx import export_onnx\n",
    "\n",
    "# Export the sklearn based model\n",
    "output_path = \"sklearn_model.onnx\"\n",
    "export_onnx(model.model_body,\n",
    "            model.model_head,\n",
    "            opset=12,\n",
    "            output_path=output_path)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "640ff34a",
   "metadata": {},
   "source": [
    "Excellent! Let's check to see that this worked. If we list our current directory we should see a `model.onnx` file if everything succeeded. If you're interested in looking at the internals of this file check out [netron](https://github.com/lutzroeder/netron). "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a5c62283",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "assert output_path in os.listdir() "
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e2d4707c",
   "metadata": {},
   "source": [
    "With our model successfully converted to `onnx` let's run some initial predictions and make sure we are getting the same results as our earlier model.  We'll start by predicting using the original pytorch pipeline."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a9a00c30",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Run inference using the original model\n",
    "input_text = [\"i loved the spiderman movie!\", \"pineapple on pizza is the worst 🤮\"]\n",
    "pytorch_preds = model(input_text)\n",
    "pytorch_preds "
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4fb8f677",
   "metadata": {},
   "source": [
    "Now let's make some predictions using the onnx model that we just created!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a151b3ba",
   "metadata": {},
   "outputs": [],
   "source": [
    "import onnxruntime\n",
    "from transformers import AutoTokenizer\n",
    "tokenizer = AutoTokenizer.from_pretrained(\"lewtun/my-awesome-setfit-model\")\n",
    "inputs = tokenizer(\n",
    "    input_text,\n",
    "    padding=True,\n",
    "    truncation=True,\n",
    "    return_attention_mask=True,\n",
    "    return_token_type_ids=True,\n",
    "    return_tensors=\"np\",\n",
    ")\n",
    "\n",
    "session = onnxruntime.InferenceSession(output_path)\n",
    "\n",
    "onnx_preds = session.run(None, dict(inputs))[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b1c9c59c",
   "metadata": {},
   "outputs": [],
   "source": [
    "onnx_preds"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c1abc081",
   "metadata": {},
   "source": [
    "The predicted outputs should look the same let's verify using numpy's `array_equal` function."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c552118b",
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "assert np.array_equal(onnx_preds, pytorch_preds)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "41d34da9",
   "metadata": {},
   "source": [
    "Nice! The output from our original pytorch model and our new onnx optimized model is the same!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "80766b89",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "dcc67aa3",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.8.13"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}