fastinference

fastinference/.ipynb_checkpoints/01_Intro-checkpoint.ipynb

@@ -0,0 +1,6 @@
+{
+ "cells": [],
+ "metadata": {},
+ "nbformat": 4,
+ "nbformat_minor": 4
+}

fastinference/.ipynb_checkpoints/02 ONNX-checkpoint.ipynb

@@ -0,0 +1,255 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {
+    "scrolled": true
+   },
+   "outputs": [],
+   "source": [
+    "!pip install git+https://github.com/muellerzr/fastinference.git --quiet"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# ONNX\n",
+    "\n",
+    "Commonly code is run on ONNX, and trying to utilize `fastai` as best you can can be a challenge. The beneifits of ONNX is it uses C++ so it can be a faster runtime, and recently ONNX came out with `CUDA` support as well! How hard is it to integrate? Let's make a `tabular` problem again, this time importing `ONNX`:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Name: fastinference\r\n",
+      "Version: 0.0.13\r\n",
+      "Summary: A collection of inference modules\r\n",
+      "Home-page: https://github.com/muellerzr/fastinference/tree/master/\r\n",
+      "Author: Zachary Mueller\r\n",
+      "Author-email: [email protected]\r\n",
+      "License: Apache Software License 2.0\r\n",
+      "Location: /home/ml1/anaconda3/envs/fastai2/lib/python3.7/site-packages\r\n",
+      "Requires: onnxruntime-gpu, fastai2, shap\r\n",
+      "Required-by: \r\n"
+     ]
+    }
+   ],
+   "source": [
+    "!pip show fastinference"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from fastai2.tabular.all import *\n",
+    "from fastinference.onnx import *"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "path = untar_data(URLs.ADULT_SAMPLE)\n",
+    "df = pd.read_csv(path/'adult.csv')\n",
+    "splits = RandomSplitter()(range_of(df))\n",
+    "cat_names = ['workclass', 'education', 'marital-status', 'occupation', 'relationship', 'race']\n",
+    "cont_names = ['age', 'fnlwgt', 'education-num']\n",
+    "procs = [Categorify, FillMissing, Normalize]\n",
+    "y_names = 'salary'"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "to = TabularPandas(df, procs=procs, cat_names=cat_names, cont_names=cont_names,\n",
+    "                   y_names=y_names, splits=splits)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "dls = to.dataloaders()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "learn = tabular_learner(dls, layers=[200,100])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "I've made a special `fastONNX` wrapper, which will take your `Learner` and export both your model and the `DataLoaders` so ONNX can use them via `learn.to_onnx()`"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "learn.to_onnx('tabular')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Now let's load it in:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "onnx_learn = fastONNX('tabular')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "What all can we do here? We can still do everything exactly the same minus one change: `predict`\n",
+    "\n",
+    "`predict` requires the raw inputs, so instead *always* build a `test_dl` and pass to `get_preds` (just made it simpler to code for me). Let's run a few examples also testing the times:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "single_dl = onnx_learn.test_dl(df.iloc[:1])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "3.46 ms ± 5.44 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)\n"
+     ]
+    }
+   ],
+   "source": [
+    "%%timeit\n",
+    "preds = onnx_learn.get_preds(dl=single_dl)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Wow! 3.46 ms even beats our previous `predict` with 25 ms! what is it's output?"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 12,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "name, preds = onnx_learn.get_preds(dl=single_dl)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "(['<50k'], array([[[0.5036924, 0.4963076]]], dtype=float32))"
+      ]
+     },
+     "execution_count": 13,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "name, preds"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Currently it doesn't support returning the raw inputs, I need to work on that some more"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "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": "3.7.7"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": false,
+   "sideBar": true,
+   "skip_h1_title": true,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": true
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}