examples/Ammonia_attack.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Adversarial attacks on ammonia using the pretrained models\n",
    "\n",
    "In this notebook, we perform an adversarial attack on zeolites using the [SchNet NN potential](https://github.com/learningmatter-mit/NeuralForceField). We will be using the third generation of ammonia models, as shown in [our paper](https://arxiv.org/abs/2101.11588).\n",
    "\n",
    "The utilities at `nff` will be used to perform the energy/force predictions. `nglview` will be used to visualize the generated trajectories. A few utility functions from this repo, `robust`, will be used as well. For the sake of generality, all steps for performing the adversarial attack are shown in this notebook."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "4cc7fe9e11ed47d4b1f8f6ae8806068d",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": []
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "import os\n",
    "import sys\n",
    "sys.path.append('..')\n",
    "\n",
    "import robust as rb\n",
    "import torch as ch\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from nff.io import NeuralFF, AtomsBatch, EnsembleNFF\n",
    "from nff.data import Dataset\n",
    "from nff.train import load_model\n",
    "\n",
    "from ase.io import Trajectory, read\n",
    "\n",
    "import nglview as nv"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Loading the dataset, models and initial geometry"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "DEVICE = 2\n",
    "\n",
    "dset = Dataset.from_file('../data/ammonia.pth.tar')\n",
    "\n",
    "PATH = '../models/ammonia'\n",
    "models = []\n",
    "for model_name in sorted(os.listdir(PATH)):\n",
    "    m = NeuralFF.from_file(os.path.join(PATH, model_name), device=DEVICE).model\n",
    "    models.append(m)\n",
    "\n",
    "ensemble = EnsembleNFF(models, device=DEVICE)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "CUTOFF = 5\n",
    "\n",
    "def get_atoms(props):\n",
    "    atoms = AtomsBatch(\n",
    "        positions=props['nxyz'][:, 1:],\n",
    "        numbers=props['nxyz'][:, 0],\n",
    "        cutoff=CUTOFF,\n",
    "        props={'energy': 0, 'energy_grad': []},\n",
    "        calculator=ensemble,\n",
    "        nbr_torch=False,\n",
    "        device=DEVICE,\n",
    "    )\n",
    "    _ = atoms.update_nbr_list()\n",
    "    \n",
    "    return atoms\n",
    "\n",
    "initial = get_atoms(dset[np.argmin(dset.props['energy'])])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Defining the adversarial attack\n",
    "\n",
    "The `Attacker` class allows one to perform an adversarial attack using an ensemble of SchNet models."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "../robust/data.py:76: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
      "  return ch.tensor(tensor).reshape(len(tensor), -1)\n",
      "../robust/data.py:77: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
      "  return ch.tensor(tensor)\n"
     ]
    }
   ],
   "source": [
    "energy_dset = rb.PotentialDataset(\n",
    "    ch.zeros_like(dset.props['energy']),\n",
    "    dset.props['energy'],\n",
    "    ch.zeros_like(dset.props['energy']),\n",
    ")\n",
    "\n",
    "loss_fn = rb.loss.AdvLoss(\n",
    "    train=energy_dset,\n",
    "    temperature=20,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "attacker = rb.schnet.Attacker(\n",
    "    initial,\n",
    "    ensemble,\n",
    "    loss_fn,\n",
    "    device=DEVICE,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 60/60 [00:08<00:00,  7.05it/s]\n"
     ]
    }
   ],
   "source": [
    "results = attacker.attack()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Post-processing the results: visualizing and plotting\n",
    "\n",
    "After the adversarial attack is performed, we can now visualize it and post-process the results. We start by recalculating the variance in forces and energies."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "df = pd.DataFrame(results)\n",
    "df['forces_var'] = [f.var(-1).mean() for f in df['forces']]\n",
    "df['energy_var'] = [e.var() for e in df['energy']]\n",
    "df['energy_avg'] = [e.mean() for e in df['energy']]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Then, we reconstruct the trajectory of the adversarial attack using the values of `delta` along the attack."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "newatoms = []\n",
    "for transl in df.delta:\n",
    "    at = initial.copy()\n",
    "    at.translate(transl)\n",
    "    newatoms.append(at)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "b9932ac05483433cadaf0301d957a2e5",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "NGLWidget(max_frame=59)"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "view = nv.show_asetraj(newatoms)\n",
    "view.add_unitcell()\n",
    "\n",
    "view"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Finally, we can plot the trajectory of the adversarial attack based on the sampled properties shown before (see Fig. S10 of the paper). Notice how the geometry that maximizes the adversarial loss is not necessarily the one with highest energy."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/dskoda/.conda/envs/nff/lib/python3.7/site-packages/ipykernel_launcher.py:13: UserWarning: FixedFormatter should only be used together with FixedLocator\n",
      "  del sys.path[0]\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 288x360 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig, ax = plt.subplots(figsize=(4, 5))\n",
    "ax.spines['right'].set_visible(True)\n",
    "\n",
    "ax.plot(-df.loss, color='k')\n",
    "\n",
    "ax.set_ylabel('Adversarial Loss')\n",
    "\n",
    "COLOR_TAX = '#d62728'\n",
    "tax = ax.twinx()\n",
    "tax.plot(df.energy_avg, color=COLOR_TAX)\n",
    "tax.set_yticklabels(\n",
    "    tax.get_yticks(),\n",
    "    color=COLOR_TAX\n",
    ")\n",
    "tax.set_ylabel('Mean energy (kcal/mol)', color=COLOR_TAX)\n",
    "\n",
    "ax.set_xlabel('Attack step')\n",
    "\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "nff",
   "language": "python",
   "name": "nff"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}