PyTorch + scikit-learn benchmark

pytorch_array_api_switch.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "91d24835-d13b-4ac1-b4f2-47925ddb4a2a",
   "metadata": {
    "tags": []
   },
   "source": [
    "# Array API example for LinearDiscriminantAnalysis\n",
    "\n",
    "### This benchmark is ran on a Nvidia GTX 3090 and a AMD 5950x."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "a849ca35-90d9-4783-810d-18250bc021ee",
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.datasets import make_classification\n",
    "from sklearn.discriminant_analysis import LinearDiscriminantAnalysis\n",
    "import numpy as np\n",
    "\n",
    "X_np, y_np = make_classification(random_state=0,\n",
    "                                 n_samples=500_000, n_features=300)\n",
    "X_np, y_np = X_np.astype(np.float32), y_np.astype(np.float32)\n",
    "lda_np = LinearDiscriminantAnalysis()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8eefb285-86e8-4472-ac98-bd292e7e0546",
   "metadata": {},
   "source": [
    "## Fit runtime for NumPy array"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "a1831332-4c8e-4dcf-abdc-f202e747dad1",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 1min 1s, sys: 19.9 s, total: 1min 21s\n",
      "Wall time: 10.3 s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "_ = lda_np.fit(X_np, y_np)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "fbbcd4c2-0622-452c-b1e8-d6e210b70821",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "61.1 ms ± 2.33 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "_ = lda_np.predict(X_np)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d1049837-e4b0-4494-84f7-b5e13f08d979",
   "metadata": {},
   "source": [
    "## Runtime for PyTorch Tensor on CPU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "533eb684-e901-4f02-a69a-cde47f9194db",
   "metadata": {},
   "outputs": [],
   "source": [
    "import sklearn\n",
    "\n",
    "# Enable ArrayAPI dispatching\n",
    "sklearn.set_config(array_api_dispatch=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "4abbfdc0-1462-46a7-b010-67d3541dba14",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "\n",
    "X_torch_cpu = torch.asarray(X_np)\n",
    "y_torch_cpu = torch.asarray(y_np)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "e26bb8a1-1351-44c2-bb44-134b503a6bdd",
   "metadata": {},
   "outputs": [],
   "source": [
    "lda_torch_cpu = LinearDiscriminantAnalysis()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "2874e1f6-b3d8-43df-87eb-c86ea00bc00f",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "2.03 s ± 6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "lda_torch_cpu.fit(X_torch_cpu, y_torch_cpu)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "a84c01b8-31d1-41ee-a569-aafa17b978dc",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "44.9 ms ± 471 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "_ = lda_torch_cpu.predict(X_torch_cpu)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "63084e19-21cd-4069-8f71-1ecabd29081b",
   "metadata": {},
   "source": [
    "### On CPU, PyTorch is more 5x faster during training and 1.5x faster during prediciton."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9f8e7614-3b48-4955-8c02-4883c91fd327",
   "metadata": {
    "tags": []
   },
   "source": [
    "## Runtime for PyTorch Tensor on CUDA"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "56eae76f-1837-48f4-a8f7-ba29c96b7d34",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "\n",
    "X_torch_cuda = torch.asarray(X_np, device=\"cuda\")\n",
    "y_torch_cuda = torch.asarray(y_np, device=\"cuda\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "ea84e268-4ca3-46fe-b5b8-cb5b90799bbe",
   "metadata": {},
   "outputs": [],
   "source": [
    "lda_torch_cuda = LinearDiscriminantAnalysis()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "f828fcfc-ffc2-4ded-960f-d8151b624def",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "145 ms ± 354 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "_ = lda_torch_cuda.fit(X_torch_cuda, y_torch_cuda)\n",
    "torch.cuda.synchronize(device=\"cuda\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "ba0ba5c8-4c16-4627-8e01-f31141e43d2a",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1.54 ms ± 175 ns per loop (mean ± std. dev. of 7 runs, 1,000 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "_ = lda_torch_cuda.predict(X_torch_cuda)\n",
    "torch.cuda.synchronize(device=\"cuda\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "612bd70d-3a90-41d0-b423-2bbb87b17628",
   "metadata": {},
   "source": [
    "### On GPU, PyTorch is 13x faster during training and 30x faster during prediciton"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3102ddf8-608e-4351-8db5-1cd94881631f",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c0980ae9-9c39-4c81-99eb-b7424464cc46",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ba1b9b46-c558-4710-af1b-adbaf1aacba0",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "markdown",
   "id": "24b73a36-a142-48c0-967c-f2e40f14df83",
   "metadata": {},
   "source": [
    "## Check coefs are the same"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "92674019-00ae-4cd7-96ab-1ada7f8dd0c0",
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "\n",
    "np_coef_ = np.asarray(lda_np.coef_)\n",
    "\n",
    "# Is there a bettery way to convert cupy.array_api to a np.ndarray?\n",
    "cu_coef_ = lda_torch_cuda.coef_.cpu().numpy()\n",
    "\n",
    "np.testing.assert_allclose(np_coef_, cu_coef_, atol=1e-3)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "sk1-pytorch (python3)",
   "language": "python",
   "name": "conda-env-sk1-pytorch-py"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}