large_model.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "args = {}\n",
    "args['model_path'] = './results/'\n",
    "args['no_env'] = 1\n",
    "args['no_motion_paths'] = 10000\n",
    "args['total_input_size'] = 8\n",
    "args['AE_input_size'] = 0\n",
    "args['mlp_input_size'] = 8\n",
    "args['output_size'] = 4\n",
    "args['learning_rate'] = 0.01\n",
    "args['device'] = 0\n",
    "args['num_epochs'] = 5\n",
    "args['batch_size'] = 100\n",
    "args['data_path'] = './data/cartpole/'\n",
    "args['data_file'] = 'train.pkl'\n",
    "args['start_epoch'] = 0\n",
    "args['env_type'] = 'cartpole'\n",
    "args['world_size'] = [30., 40., 3.141592653589793, 2.]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "loading...\n",
      "training...\n",
      "epoch1\n",
      "epoch2\n",
      "epoch3\n",
      "epoch4\n",
      "epoch5\n"
     ]
    }
   ],
   "source": [
    "\"\"\"\n",
    "This implements the Kinodynamic Planning using MPNet, by using MPNet\n",
    "to generate random samples, that will guide the SST algorithm.\n",
    "\"\"\"\n",
    "import torch\n",
    "import model.AE.identity as cae_identity\n",
    "from model.mlp import MLP\n",
    "from model.mpnet import KMPNet\n",
    "from tools import data_loader\n",
    "from tools.utility import *\n",
    "from plan_utility import cart_pole\n",
    "import argparse\n",
    "import numpy as np\n",
    "import random\n",
    "import os\n",
    "\n",
    "\n",
    "if torch.cuda.is_available():\n",
    "    torch.cuda.set_device(args['device'])\n",
    "# environment setting\n",
    "if args['env_type'] == 'cartpole':\n",
    "    normalize = cart_pole.normalize\n",
    "    unnormalize = cart_pole.unnormalize\n",
    "\n",
    "cae = cae_identity\n",
    "mlp = MLP\n",
    "mpnet = KMPNet(args['total_input_size'], args['AE_input_size'], args['mlp_input_size'], args['output_size'],\n",
    "               cae, mlp)\n",
    "# load net\n",
    "# load previously trained model if start epoch > 0\n",
    "model_path='kmpnet_epoch_%d.pkl' %(args['start_epoch'])\n",
    "torch_seed, np_seed, py_seed = 0, 0, 0\n",
    "if args['start_epoch'] > 0:\n",
    "    load_net_state(mpnet, os.path.join(args['model_path'], model_path))\n",
    "    torch_seed, np_seed, py_seed = load_seed(os.path.join(args['model_path'], model_path))\n",
    "    # set seed after loading\n",
    "    torch.manual_seed(torch_seed)\n",
    "    np.random.seed(np_seed)\n",
    "    random.seed(py_seed)\n",
    "\n",
    "if torch.cuda.is_available():\n",
    "    mpnet.cuda()\n",
    "    mpnet.mlp.cuda()\n",
    "    mpnet.encoder.cuda()\n",
    "    # here we use Adagrad because previous MPNet performs well under it\n",
    "    mpnet.set_opt(torch.optim.Adagrad, lr=args['learning_rate'])\n",
    "if args['start_epoch'] > 0:\n",
    "    load_opt_state(mpnet, os.path.join(args['model_path'], model_path))\n",
    "\n",
    "# load train and test data\n",
    "print('loading...')\n",
    "obs, dataset, targets, env_indices = data_loader.load_train_dataset(N=args['no_env'], \\\n",
    "                                            NP=args['no_motion_paths'], folder=args['data_path'], \\\n",
    "                                            fname=args['data_file'])\n",
    "data=list(zip(dataset,targets,env_indices))\n",
    "random.shuffle(data)\n",
    "dataset,targets,env_indices=list(zip(*data))\n",
    "dataset = list(dataset)\n",
    "targets = list(targets)\n",
    "env_indices = list(env_indices)\n",
    "dataset = np.array(dataset)\n",
    "targets = np.array(targets)\n",
    "env_indices = np.array(env_indices)\n",
    "\n",
    "# Train the Models\n",
    "print('training...')\n",
    "losses = []\n",
    "for epoch in range(args['start_epoch']+1,args['num_epochs']+1):\n",
    "    print('epoch' + str(epoch))\n",
    "    for i in range(0,len(dataset),args['batch_size']):\n",
    "        #print('epoch: %d, training... path: %d' % (epoch, i+1))\n",
    "        dataset_i = dataset[i:i+args['batch_size']]\n",
    "        targets_i = targets[i:i+args['batch_size']]\n",
    "        env_indices_i = env_indices[i:i+args['batch_size']]\n",
    "        # record\n",
    "        bi = np.concatenate( (obs[env_indices_i, :args['AE_input_size']], dataset_i), axis=1).astype(np.float32)\n",
    "        bt = targets_i\n",
    "        bi = torch.FloatTensor(bi)\n",
    "        bt = torch.FloatTensor(bt)\n",
    "        bi, bt = normalize(bi, args['world_size']), normalize(bt, args['world_size'])\n",
    "        mpnet.zero_grad()\n",
    "        bi=to_var(bi)\n",
    "        bt=to_var(bt)\n",
    "        #print('before training losses:')\n",
    "        #print(mpnet.loss(mpnet(bi), bt))\n",
    "        losses.append(mpnet.loss(mpnet(bi), bt).cpu().data)\n",
    "        mpnet.step(bi, bt)\n",
    "        #print('after training losses:')\n",
    "        #print(mpnet.loss(mpnet(bi), bt))\n",
    "    # Save the models\n",
    "    if epoch > 0:\n",
    "        model_path='kmpnet_epoch_%d.pkl' %(epoch)\n",
    "        save_state(mpnet, torch_seed, np_seed, py_seed, os.path.join(args['model_path'],model_path))\n",
    "        # test\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "plt.plot(list(range(len(losses))), losses)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "path 0\n",
      "path 1\n",
      "path 2\n",
      "path 3\n",
      "path 4\n",
      "path 5\n",
      "path 6\n",
      "path 7\n",
      "path 8\n",
      "path 9\n",
      "path 10\n",
      "path 11\n",
      "path 12\n",
      "path 13\n",
      "path 14\n",
      "path 15\n",
      "path 16\n",
      "path 17\n",
      "path 18\n",
      "path 19\n",
      "path 20\n",
      "path 21\n",
      "path 22\n",
      "path 23\n",
      "path 24\n",
      "path 25\n",
      "path 26\n",
      "path 27\n",
      "path 28\n",
      "path 29\n",
      "path 30\n",
      "path 31\n",
      "path 32\n",
      "path 33\n",
      "path 34\n",
      "path 35\n",
      "path 36\n",
      "path 37\n",
      "path 38\n",
      "path 39\n",
      "path 40\n",
      "path 41\n",
      "path 42\n",
      "path 43\n",
      "path 44\n",
      "path 45\n",
      "path 46\n",
      "path 47\n",
      "path 48\n",
      "path 49\n",
      "path 50\n",
      "path 51\n",
      "path 52\n",
      "path 53\n",
      "path 54\n",
      "path 55\n",
      "path 56\n",
      "path 57\n",
      "path 58\n",
      "path 59\n",
      "path 60\n",
      "path 61\n",
      "path 62\n",
      "path 63\n",
      "path 64\n",
      "path 65\n",
      "path 66\n",
      "path 67\n",
      "path 68\n",
      "path 69\n",
      "path 70\n",
      "path 71\n",
      "path 72\n",
      "path 73\n",
      "path 74\n",
      "path 75\n",
      "path 76\n",
      "path 77\n",
      "path 78\n",
      "path 79\n",
      "path 80\n",
      "path 81\n",
      "path 82\n",
      "path 83\n",
      "path 84\n",
      "path 85\n",
      "path 86\n",
      "path 87\n",
      "path 88\n",
      "path 89\n",
      "path 90\n",
      "path 91\n",
      "path 92\n",
      "path 93\n",
      "path 94\n",
      "path 95\n",
      "path 96\n",
      "path 97\n",
      "path 98\n",
      "path 99\n",
      "path 100\n",
      "path 101\n",
      "path 102\n",
      "path 103\n",
      "path 104\n",
      "path 105\n",
      "path 106\n",
      "path 107\n",
      "path 108\n",
      "path 109\n",
      "path 110\n",
      "path 111\n",
      "path 112\n",
      "path 113\n",
      "path 114\n",
      "path 115\n",
      "path 116\n",
      "path 117\n",
      "path 118\n",
      "path 119\n",
      "path 120\n",
      "path 121\n",
      "path 122\n",
      "path 123\n",
      "path 124\n",
      "path 125\n",
      "path 126\n",
      "path 127\n",
      "path 128\n",
      "path 129\n",
      "path 130\n",
      "path 131\n",
      "path 132\n",
      "path 133\n",
      "path 134\n",
      "path 135\n",
      "path 136\n",
      "path 137\n",
      "path 138\n",
      "path 139\n",
      "path 140\n",
      "path 141\n",
      "path 142\n",
      "path 143\n",
      "path 144\n",
      "path 145\n",
      "path 146\n",
      "path 147\n",
      "path 148\n",
      "path 149\n",
      "path 150\n",
      "path 151\n",
      "path 152\n",
      "path 153\n",
      "path 154\n",
      "path 155\n",
      "path 156\n",
      "path 157\n",
      "path 158\n",
      "path 159\n",
      "path 160\n",
      "path 161\n",
      "path 162\n",
      "path 163\n",
      "path 164\n",
      "path 165\n",
      "path 166\n",
      "path 167\n",
      "path 168\n",
      "path 169\n",
      "path 170\n",
      "path 171\n",
      "path 172\n",
      "path 173\n",
      "path 174\n",
      "path 175\n",
      "path 176\n",
      "path 177\n",
      "path 178\n",
      "path 179\n",
      "path 180\n",
      "path 181\n",
      "path 182\n",
      "path 183\n",
      "path 184\n",
      "path 185\n",
      "path 186\n",
      "path 187\n",
      "path 188\n",
      "path 189\n",
      "path 190\n",
      "path 191\n",
      "path 192\n",
      "path 193\n",
      "path 194\n",
      "path 195\n",
      "path 196\n",
      "path 197\n",
      "path 198\n",
      "path 199\n",
      "path 200\n",
      "path 201\n",
      "path 202\n",
      "path 203\n",
      "path 204\n",
      "path 205\n",
      "path 206\n",
      "path 207\n",
      "path 208\n",
      "path 209\n",
      "path 210\n",
      "path 211\n",
      "path 212\n",
      "path 213\n",
      "path 214\n",
      "path 215\n",
      "path 216\n",
      "path 217\n",
      "path 218\n",
      "path 219\n",
      "path 220\n",
      "path 221\n",
      "path 222\n",
      "path 223\n",
      "path 224\n",
      "path 225\n",
      "path 226\n",
      "path 227\n",
      "path 228\n",
      "path 229\n",
      "path 230\n",
      "path 231\n",
      "path 232\n",
      "path 233\n",
      "path 234\n",
      "path 235\n",
      "path 236\n",
      "path 237\n",
      "path 238\n",
      "path 239\n",
      "path 240\n",
      "path 241\n",
      "path 242\n",
      "path 243\n",
      "path 244\n",
      "path 245\n",
      "path 246\n",
      "path 247\n",
      "path 248\n",
      "path 249\n",
      "path 250\n",
      "path 251\n",
      "path 252\n",
      "path 253\n",
      "path 254\n",
      "path 255\n",
      "path 256\n",
      "path 257\n",
      "path 258\n",
      "path 259\n",
      "path 260\n",
      "path 261\n",
      "path 262\n",
      "path 263\n",
      "path 264\n",
      "path 265\n",
      "path 266\n",
      "path 267\n",
      "path 268\n",
      "path 269\n",
      "path 270\n",
      "path 271\n",
      "path 272\n",
      "path 273\n",
      "path 274\n",
      "path 275\n",
      "path 276\n",
      "path 277\n",
      "path 278\n",
      "path 279\n",
      "path 280\n",
      "path 281\n",
      "path 282\n",
      "path 283\n",
      "path 284\n",
      "path 285\n",
      "path 286\n",
      "path 287\n",
      "path 288\n",
      "path 289\n",
      "path 290\n",
      "path 291\n",
      "path 292\n",
      "path 293\n",
      "path 294\n",
      "path 295\n",
      "path 296\n",
      "path 297\n",
      "path 298\n",
      "path 299\n",
      "path 300\n",
      "path 301\n",
      "path 302\n",
      "path 303\n",
      "path 304\n",
      "path 305\n",
      "path 306\n",
      "path 307\n",
      "path 308\n",
      "path 309\n",
      "path 310\n",
      "path 311\n",
      "path 312\n",
      "path 313\n",
      "path 314\n",
      "path 315\n",
      "path 316\n",
      "path 317\n",
      "path 318\n",
      "path 319\n",
      "path 320\n",
      "path 321\n",
      "path 322\n",
      "path 323\n",
      "path 324\n",
      "path 325\n",
      "path 326\n",
      "path 327\n",
      "path 328\n",
      "path 329\n",
      "path 330\n",
      "path 331\n",
      "path 332\n",
      "path 333\n",
      "path 334\n",
      "path 335\n",
      "path 336\n",
      "path 337\n",
      "path 338\n",
      "path 339\n",
      "path 340\n",
      "path 341\n",
      "path 342\n",
      "path 343\n",
      "path 344\n",
      "path 345\n",
      "path 346\n",
      "path 347\n",
      "path 348\n",
      "path 349\n",
      "path 350\n",
      "path 351\n",
      "path 352\n",
      "path 353\n",
      "path 354\n",
      "path 355\n",
      "path 356\n",
      "path 357\n",
      "path 358\n",
      "path 359\n",
      "path 360\n",
      "path 361\n",
      "path 362\n",
      "path 363\n",
      "path 364\n",
      "path 365\n",
      "path 366\n",
      "path 367\n",
      "path 368\n",
      "path 369\n",
      "path 370\n",
      "path 371\n",
      "path 372\n",
      "path 373\n",
      "path 374\n",
      "path 375\n",
      "path 376\n",
      "path 377\n",
      "path 378\n",
      "path 379\n",
      "path 380\n",
      "path 381\n",
      "path 382\n",
      "path 383\n",
      "path 384\n",
      "path 385\n",
      "path 386\n",
      "path 387\n",
      "path 388\n",
      "path 389\n",
      "path 390\n",
      "path 391\n",
      "path 392\n",
      "path 393\n",
      "path 394\n",
      "path 395\n",
      "path 396\n",
      "path 397\n",
      "path 398\n",
      "path 399\n",
      "path 400\n",
      "path 401\n",
      "path 402\n",
      "path 403\n",
      "path 404\n",
      "path 405\n",
      "path 406\n",
      "path 407\n",
      "path 408\n",
      "path 409\n",
      "path 410\n",
      "path 411\n",
      "path 412\n",
      "path 413\n",
      "path 414\n",
      "path 415\n",
      "path 416\n",
      "path 417\n",
      "path 418\n",
      "path 419\n",
      "path 420\n",
      "path 421\n",
      "path 422\n",
      "path 423\n",
      "path 424\n",
      "path 425\n",
      "path 426\n",
      "path 427\n",
      "path 428\n",
      "path 429\n",
      "path 430\n",
      "path 431\n",
      "path 432\n",
      "path 433\n",
      "path 434\n",
      "path 435\n",
      "path 436\n",
      "path 437\n",
      "path 438\n",
      "path 439\n",
      "path 440\n",
      "path 441\n",
      "path 442\n",
      "path 443\n",
      "path 444\n",
      "path 445\n",
      "path 446\n",
      "path 447\n",
      "path 448\n",
      "path 449\n",
      "path 450\n",
      "path 451\n",
      "path 452\n",
      "path 453\n",
      "path 454\n",
      "path 455\n",
      "path 456\n",
      "path 457\n",
      "path 458\n",
      "path 459\n",
      "path 460\n",
      "path 461\n",
      "path 462\n",
      "path 463\n",
      "path 464\n",
      "path 465\n",
      "path 466\n",
      "path 467\n",
      "path 468\n",
      "path 469\n",
      "path 470\n",
      "path 471\n",
      "path 472\n",
      "path 473\n",
      "path 474\n",
      "path 475\n",
      "path 476\n",
      "path 477\n",
      "path 478\n",
      "path 479\n",
      "path 480\n",
      "path 481\n",
      "path 482\n",
      "path 483\n",
      "path 484\n",
      "path 485\n",
      "path 486\n",
      "path 487\n",
      "path 488\n",
      "path 489\n",
      "path 490\n",
      "path 491\n",
      "path 492\n",
      "path 493\n",
      "path 494\n",
      "path 495\n",
      "path 496\n",
      "path 497\n",
      "path 498\n",
      "path 499\n",
      "path 500\n",
      "path 501\n",
      "path 502\n",
      "path 503\n",
      "path 504\n",
      "path 505\n",
      "path 506\n",
      "path 507\n",
      "path 508\n",
      "path 509\n",
      "path 510\n",
      "path 511\n",
      "path 512\n",
      "path 513\n",
      "path 514\n",
      "path 515\n",
      "path 516\n",
      "path 517\n",
      "path 518\n",
      "path 519\n",
      "path 520\n",
      "path 521\n",
      "path 522\n",
      "path 523\n",
      "path 524\n",
      "path 525\n",
      "path 526\n",
      "path 527\n",
      "path 528\n",
      "path 529\n",
      "path 530\n",
      "path 531\n",
      "path 532\n",
      "path 533\n",
      "path 534\n",
      "path 535\n",
      "path 536\n",
      "path 537\n",
      "path 538\n",
      "path 539\n",
      "path 540\n",
      "path 541\n",
      "path 542\n",
      "path 543\n",
      "path 544\n",
      "path 545\n",
      "path 546\n",
      "path 547\n",
      "path 548\n",
      "path 549\n",
      "path 550\n",
      "path 551\n",
      "path 552\n",
      "path 553\n",
      "path 554\n",
      "path 555\n",
      "path 556\n",
      "path 557\n",
      "path 558\n",
      "path 559\n",
      "path 560\n",
      "path 561\n",
      "path 562\n",
      "path 563\n",
      "path 564\n",
      "path 565\n",
      "path 566\n",
      "path 567\n",
      "path 568\n",
      "path 569\n",
      "path 570\n",
      "path 571\n",
      "path 572\n",
      "path 573\n",
      "path 574\n",
      "path 575\n",
      "path 576\n",
      "path 577\n",
      "path 578\n",
      "path 579\n",
      "path 580\n",
      "path 581\n",
      "path 582\n",
      "path 583\n",
      "path 584\n",
      "path 585\n",
      "path 586\n",
      "path 587\n",
      "path 588\n",
      "path 589\n",
      "path 590\n",
      "path 591\n",
      "path 592\n",
      "path 593\n",
      "path 594\n",
      "path 595\n",
      "path 596\n",
      "path 597\n",
      "path 598\n",
      "path 599\n",
      "path 600\n",
      "path 601\n",
      "path 602\n",
      "path 603\n",
      "path 604\n",
      "path 605\n",
      "path 606\n",
      "path 607\n",
      "path 608\n",
      "path 609\n",
      "path 610\n",
      "path 611\n",
      "path 612\n",
      "path 613\n",
      "path 614\n",
      "path 615\n",
      "path 616\n",
      "path 617\n",
      "path 618\n",
      "path 619\n",
      "path 620\n",
      "path 621\n",
      "path 622\n",
      "path 623\n",
      "path 624\n",
      "path 625\n",
      "path 626\n",
      "path 627\n",
      "path 628\n",
      "path 629\n",
      "path 630\n",
      "path 631\n",
      "path 632\n",
      "path 633\n",
      "path 634\n",
      "path 635\n",
      "path 636\n",
      "path 637\n",
      "path 638\n",
      "path 639\n",
      "path 640\n",
      "path 641\n",
      "path 642\n",
      "path 643\n",
      "path 644\n",
      "path 645\n",
      "path 646\n",
      "path 647\n",
      "path 648\n",
      "path 649\n",
      "path 650\n",
      "path 651\n",
      "path 652\n",
      "path 653\n",
      "path 654\n",
      "path 655\n",
      "path 656\n",
      "path 657\n",
      "path 658\n",
      "path 659\n",
      "path 660\n",
      "path 661\n",
      "path 662\n",
      "path 663\n",
      "path 664\n",
      "path 665\n",
      "path 666\n",
      "path 667\n",
      "path 668\n",
      "path 669\n",
      "path 670\n",
      "path 671\n",
      "path 672\n",
      "path 673\n",
      "path 674\n",
      "path 675\n",
      "path 676\n",
      "path 677\n",
      "path 678\n",
      "path 679\n",
      "path 680\n",
      "path 681\n",
      "path 682\n",
      "path 683\n",
      "path 684\n",
      "path 685\n",
      "path 686\n",
      "path 687\n",
      "path 688\n",
      "path 689\n",
      "path 690\n",
      "path 691\n",
      "path 692\n",
      "path 693\n",
      "path 694\n",
      "path 695\n",
      "path 696\n",
      "path 697\n",
      "path 698\n",
      "path 699\n",
      "path 700\n",
      "path 701\n",
      "path 702\n",
      "path 703\n",
      "path 704\n",
      "path 705\n",
      "path 706\n",
      "path 707\n",
      "path 708\n",
      "path 709\n",
      "path 710\n",
      "path 711\n",
      "path 712\n",
      "path 713\n",
      "path 714\n",
      "path 715\n",
      "path 716\n",
      "path 717\n",
      "path 718\n",
      "path 719\n",
      "path 720\n",
      "path 721\n",
      "path 722\n",
      "path 723\n",
      "path 724\n",
      "path 725\n",
      "path 726\n",
      "path 727\n",
      "path 728\n",
      "path 729\n",
      "path 730\n",
      "path 731\n",
      "path 732\n",
      "path 733\n",
      "path 734\n",
      "path 735\n",
      "path 736\n",
      "path 737\n",
      "path 738\n",
      "path 739\n",
      "path 740\n",
      "path 741\n",
      "path 742\n",
      "path 743\n",
      "path 744\n",
      "path 745\n",
      "path 746\n",
      "path 747\n",
      "path 748\n",
      "path 749\n",
      "path 750\n",
      "path 751\n",
      "path 752\n",
      "path 753\n",
      "path 754\n",
      "path 755\n",
      "path 756\n",
      "path 757\n",
      "path 758\n",
      "path 759\n",
      "path 760\n",
      "path 761\n",
      "path 762\n",
      "path 763\n",
      "path 764\n",
      "path 765\n",
      "path 766\n",
      "path 767\n",
      "path 768\n",
      "path 769\n",
      "path 770\n",
      "path 771\n",
      "path 772\n",
      "path 773\n",
      "path 774\n",
      "path 775\n",
      "path 776\n",
      "path 777\n",
      "path 778\n",
      "path 779\n",
      "path 780\n",
      "path 781\n",
      "path 782\n",
      "path 783\n",
      "path 784\n",
      "path 785\n",
      "path 786\n",
      "path 787\n",
      "path 788\n",
      "path 789\n",
      "path 790\n",
      "path 791\n",
      "path 792\n",
      "path 793\n",
      "path 794\n",
      "path 795\n",
      "path 796\n",
      "path 797\n",
      "path 798\n",
      "path 799\n",
      "path 800\n",
      "path 801\n",
      "path 802\n",
      "path 803\n",
      "path 804\n",
      "path 805\n",
      "path 806\n",
      "path 807\n",
      "path 808\n",
      "path 809\n",
      "path 810\n",
      "path 811\n",
      "path 812\n",
      "path 813\n",
      "path 814\n",
      "path 815\n",
      "path 816\n",
      "path 817\n",
      "path 818\n",
      "path 819\n",
      "path 820\n",
      "path 821\n",
      "path 822\n",
      "path 823\n",
      "path 824\n",
      "path 825\n",
      "path 826\n",
      "path 827\n",
      "path 828\n",
      "path 829\n",
      "path 830\n",
      "path 831\n",
      "path 832\n",
      "path 833\n",
      "path 834\n",
      "path 835\n",
      "path 836\n",
      "path 837\n",
      "path 838\n",
      "path 839\n",
      "path 840\n",
      "path 841\n",
      "path 842\n",
      "path 843\n",
      "path 844\n",
      "path 845\n",
      "path 846\n",
      "path 847\n",
      "path 848\n",
      "path 849\n",
      "path 850\n",
      "path 851\n",
      "path 852\n",
      "path 853\n",
      "path 854\n",
      "path 855\n",
      "path 856\n",
      "path 857\n",
      "path 858\n",
      "path 859\n",
      "path 860\n",
      "path 861\n",
      "path 862\n",
      "path 863\n",
      "path 864\n",
      "path 865\n",
      "path 866\n",
      "path 867\n",
      "path 868\n",
      "path 869\n",
      "path 870\n",
      "path 871\n",
      "path 872\n",
      "path 873\n",
      "path 874\n",
      "path 875\n",
      "path 876\n",
      "path 877\n",
      "path 878\n",
      "path 879\n",
      "path 880\n",
      "path 881\n",
      "path 882\n",
      "path 883\n",
      "path 884\n",
      "path 885\n",
      "path 886\n",
      "path 887\n",
      "path 888\n",
      "path 889\n",
      "path 890\n",
      "path 891\n",
      "path 892\n",
      "path 893\n",
      "path 894\n",
      "path 895\n",
      "path 896\n",
      "path 897\n",
      "path 898\n",
      "path 899\n",
      "path 900\n",
      "path 901\n",
      "path 902\n",
      "path 903\n",
      "path 904\n",
      "path 905\n",
      "path 906\n",
      "path 907\n",
      "path 908\n",
      "path 909\n",
      "path 910\n",
      "path 911\n",
      "path 912\n",
      "path 913\n",
      "path 914\n",
      "path 915\n",
      "path 916\n",
      "path 917\n",
      "path 918\n",
      "path 919\n",
      "path 920\n",
      "path 921\n",
      "path 922\n",
      "path 923\n",
      "path 924\n",
      "path 925\n",
      "path 926\n",
      "path 927\n",
      "path 928\n",
      "path 929\n",
      "path 930\n",
      "path 931\n",
      "path 932\n",
      "path 933\n",
      "path 934\n",
      "path 935\n",
      "path 936\n",
      "path 937\n",
      "path 938\n",
      "path 939\n",
      "path 940\n",
      "path 941\n",
      "path 942\n",
      "path 943\n",
      "path 944\n",
      "path 945\n",
      "path 946\n",
      "path 947\n",
      "path 948\n",
      "path 949\n",
      "path 950\n",
      "path 951\n",
      "path 952\n",
      "path 953\n",
      "path 954\n",
      "path 955\n",
      "path 956\n",
      "path 957\n",
      "path 958\n",
      "path 959\n",
      "path 960\n",
      "path 961\n",
      "path 962\n",
      "path 963\n",
      "path 964\n",
      "path 965\n",
      "path 966\n",
      "path 967\n",
      "path 968\n",
      "path 969\n",
      "path 970\n",
      "path 971\n",
      "path 972\n",
      "path 973\n",
      "path 974\n",
      "path 975\n",
      "path 976\n",
      "path 977\n",
      "path 978\n",
      "path 979\n",
      "path 980\n",
      "path 981\n",
      "path 982\n",
      "path 983\n",
      "path 984\n",
      "path 985\n",
      "path 986\n",
      "path 987\n",
      "path 988\n",
      "path 989\n",
      "path 990\n",
      "path 991\n",
      "path 992\n",
      "path 993\n",
      "path 994\n",
      "path 995\n",
      "path 996\n",
      "path 997\n",
      "path 998\n",
      "path 999\n"
     ]
    }
   ],
   "source": [
    "# load training paths\n",
    "import pickle\n",
    "from sparse_rrt import _sst_module\n",
    "from sparse_rrt.systems import standard_cpp_systems\n",
    "import time\n",
    "def informed_sample(env, planner, start, goal):\n",
    "    # repeat until finding a path\n",
    "    max_iter = 500\n",
    "    min_time_steps = 20\n",
    "    max_time_steps = 200\n",
    "    integration_step = 0.002\n",
    "    x_new = np.array(start)\n",
    "    sample_time = []\n",
    "    plan_time_single = []\n",
    "    for i in range(max_iter):\n",
    "        # use MPNet to find the informed sample\n",
    "        time1 = time.time()\n",
    "        if i % 100 == 0:\n",
    "            # use goal from time to time\n",
    "            x_rand = goal\n",
    "        else:\n",
    "            x_in = np.concatenate([x_new, goal]).reshape(1, -1)\n",
    "            x_in = torch.from_numpy(x_in).type(torch.FloatTensor)\n",
    "            x_in = normalize(x_in, args['world_size']).cuda()\n",
    "            x_rand = mpnet(x_in).cpu()[0]\n",
    "            x_rand = unnormalize(x_rand, args['world_size'])\n",
    "            x_rand = x_rand.cpu().data.numpy()\n",
    "        sample_time.append(time.time() - time1)\n",
    "        x_new = planner.step_with_sample(env, x_rand, min_time_steps, max_time_steps, integration_step)\n",
    "        plan_time_single.append(time.time() - time1)\n",
    "        # if x_new near goal it will terminate automatically\n",
    "        #TODO: better iterative sampling method\n",
    "        solution = planner.get_solution()\n",
    "        if solution is not None:\n",
    "            break\n",
    "    solution = planner.get_solution()\n",
    "    return solution, i, sample_time, plan_time_single\n",
    "\n",
    "file = open('data/cartpole/train.pkl', 'rb')\n",
    "paths = pickle.load(file)\n",
    "#N = 100\n",
    "max_iter = 500\n",
    "# for each path, compare the number of samples needed for finding a \"good\" path\n",
    "env = standard_cpp_systems.CartPole()\n",
    "# for each BVP\n",
    "min_time_steps = 20\n",
    "max_time_steps = 200\n",
    "integration_step = 0.002\n",
    "low = []\n",
    "high = []\n",
    "state_bounds = env.get_state_bounds()\n",
    "for i in range(len(state_bounds)):\n",
    "    low.append(state_bounds[i][0])\n",
    "    high.append(state_bounds[i][1])\n",
    "\n",
    "    \n",
    "uniform_suc = []\n",
    "inform_suc = []\n",
    "uniform_iter = []\n",
    "inform_iter = []\n",
    "uniform_cost = []\n",
    "inform_cost = []\n",
    "uniform_sample_time = []\n",
    "uniform_plan_time_single = []\n",
    "uniform_plan_time = []\n",
    "inform_sample_time = []\n",
    "inform_plan_time_single = []\n",
    "inform_plan_time = []\n",
    "num_paths = 1000\n",
    "for i in range(num_paths):\n",
    "    print(\"path %d\" % (i))\n",
    "    # randomly sample collision-free start and goal\n",
    "    #print('random sampling trial:')\n",
    "    start = np.random.uniform(low=low, high=high)\n",
    "    end = np.random.uniform(low=low, high=high)\n",
    "    planner = _sst_module.SSTWrapper(\n",
    "        state_bounds=env.get_state_bounds(),\n",
    "        control_bounds=env.get_control_bounds(),\n",
    "        distance=env.distance_computer(),\n",
    "        start_state=start,\n",
    "        goal_state=end,\n",
    "        goal_radius=0.5,\n",
    "        random_seed=0,\n",
    "        sst_delta_near=0.4,\n",
    "        sst_delta_drain=0.2\n",
    "    )\n",
    "    # generate a path by using SST to plan for some maximal iterations\n",
    "    time0 = time.time()\n",
    "    for iter in range(max_iter):\n",
    "        time1 = time.time()\n",
    "        if iter % 100 == 0:\n",
    "            # from time to time use the goal\n",
    "            sample = end\n",
    "        else:\n",
    "            sample = np.random.uniform(low=low, high=high)\n",
    "        uniform_sample_time.append(time.time() - time1)\n",
    "        planner.step_with_sample(env, sample, min_time_steps, max_time_steps, integration_step)\n",
    "        uniform_plan_time_single.append(time.time() - time1)\n",
    "        # check if we find the solution, if so, stop planning\n",
    "        solution = planner.get_solution()\n",
    "        if solution is not None:\n",
    "            break    \n",
    "    solution = planner.get_solution()\n",
    "    uniform_plan_time.append(time.time() - time0)\n",
    "\n",
    "    if solution is None:\n",
    "        #print('path %d: Uniform sampling failed.' % (i))\n",
    "        uniform_suc.append(0)\n",
    "        uniform_iter.append(np.inf)\n",
    "        uniform_cost.append(np.inf)\n",
    "        \n",
    "    else:\n",
    "        #print('path %d: Uniform succeeded with %d iterations.' % (i, iter))\n",
    "        path, controls, costs = solution\n",
    "        #print(path)\n",
    "        #print('uniform sampling costs:')\n",
    "        #print(np.sum(costs))\n",
    "        uniform_suc.append(1)\n",
    "        uniform_iter.append(iter)\n",
    "        uniform_cost.append(np.sum(costs))\n",
    "\n",
    "        \n",
    "    # informed sampling\n",
    "    # create a new planner\n",
    "    planner = _sst_module.SSTWrapper(\n",
    "        state_bounds=env.get_state_bounds(),\n",
    "        control_bounds=env.get_control_bounds(),\n",
    "        distance=env.distance_computer(),\n",
    "        start_state=start,\n",
    "        goal_state=end,\n",
    "        goal_radius=0.5,\n",
    "        random_seed=0,\n",
    "        sst_delta_near=0.4,\n",
    "        sst_delta_drain=0.2\n",
    "    )    \n",
    "    time0 = time.time()\n",
    "    solution, iter, sample_time, plan_time_single = informed_sample(env, planner, start, end)\n",
    "    inform_plan_time.append(time.time() - time0)\n",
    "    inform_sample_time += sample_time\n",
    "    inform_plan_time_single += plan_time_single\n",
    "    \n",
    "    if solution is not None:\n",
    "        #print('path %d: Informed sampling succeeded with %d iterations.' % (i, iter))\n",
    "        path, controls, costs = solution\n",
    "        #print(path)\n",
    "        #print('Informed sampling costs:')\n",
    "        #print(np.sum(costs))      \n",
    "        inform_suc.append(1)\n",
    "        inform_iter.append(iter)\n",
    "        inform_cost.append(np.sum(costs))\n",
    "    else:\n",
    "        #print('path %d: Informed sampling failed.' % (i))\n",
    "        inform_suc.append(0)\n",
    "        inform_iter.append(np.inf)\n",
    "        inform_cost.append(np.inf)\n",
    "        "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "uniform sampling success rate:\n",
      "954\n",
      "informed sampling success rate:\n",
      "998\n"
     ]
    }
   ],
   "source": [
    "# printout success rate\n",
    "print('uniform sampling success rate:')\n",
    "print(np.sum(uniform_suc))\n",
    "print('informed sampling success rate:')\n",
    "print(np.sum(inform_suc))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "smaller iterations:\n",
      "382\n",
      "equal iterations:\n",
      "581\n",
      "larger iterations:\n",
      "37\n"
     ]
    }
   ],
   "source": [
    "# plot iteration comparsion\n",
    "iter_smaller = []\n",
    "iter_larger = []\n",
    "iter_equal = []\n",
    "for i in range(len(inform_iter)):\n",
    "    if inform_iter[i] < uniform_iter[i]:\n",
    "        iter_smaller.append(1)\n",
    "    elif inform_iter[i] == uniform_iter[i]:\n",
    "        iter_equal.append(1)\n",
    "    else:\n",
    "        iter_larger.append(1)\n",
    "print('smaller iterations:')\n",
    "print(np.sum(iter_smaller))\n",
    "print('equal iterations:')\n",
    "print(np.sum(iter_equal))\n",
    "print('larger iterations:')\n",
    "print(np.sum(iter_larger))\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "smaller iterations:\n",
      "125\n",
      "equal iterations:\n",
      "193\n",
      "larger iterations:\n",
      "682\n",
      "2.032\n",
      "2.3480000000000003\n"
     ]
    }
   ],
   "source": [
    "# plot cost comparsion\n",
    "cost_smaller = []\n",
    "cost_equal = []\n",
    "cost_larger = []\n",
    "for i in range(len(inform_iter)):\n",
    "    if inform_cost[i] < uniform_cost[i]:\n",
    "        cost_smaller.append(1)\n",
    "    elif inform_cost[i] == uniform_cost[i]:\n",
    "        cost_equal.append(1)\n",
    "    else:\n",
    "        cost_larger.append(1)\n",
    "print('smaller iterations:')\n",
    "print(np.sum(cost_smaller))\n",
    "print('equal iterations:')\n",
    "print(np.sum(cost_equal))\n",
    "print('larger iterations:')\n",
    "print(np.sum(cost_larger))\n",
    "\n",
    "print(np.mean(uniform_cost[uniform_cost!=np.inf]))\n",
    "print(np.mean(inform_cost[uniform_cost!=np.inf]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "uniform average time for sampling:\n",
      "4.5457681831403256e-05\n",
      "informed average time for sampling:\n",
      "0.0011924057883996146\n",
      "uniform average time for planning one iteration:\n",
      "0.0007855385739078168\n",
      "informed average time for planning one iteration:\n",
      "0.0018842272827478371\n"
     ]
    }
   ],
   "source": [
    "# compare the time\n",
    "print('uniform average time for sampling:')\n",
    "print(np.mean(uniform_sample_time))\n",
    "print('informed average time for sampling:')\n",
    "print(np.mean(inform_sample_time))\n",
    "\n",
    "print('uniform average time for planning one iteration:')\n",
    "print(np.mean(uniform_plan_time_single))\n",
    "print('informed average time for planning one iteration:')\n",
    "print(np.mean(inform_plan_time_single))\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "smaller iterations:\n",
      "208\n",
      "equal iterations:\n",
      "0.0\n",
      "larger iterations:\n",
      "792\n",
      "0.1190761935710907\n",
      "0.18591996693611146\n"
     ]
    }
   ],
   "source": [
    "# plot cost comparsion\n",
    "time_smaller = []\n",
    "time_equal = []\n",
    "time_larger = []\n",
    "for i in range(len(inform_iter)):\n",
    "    if inform_plan_time[i] < uniform_plan_time[i]:\n",
    "        time_smaller.append(1)\n",
    "    elif inform_plan_time[i] == uniform_plan_time[i]:\n",
    "        time_equal.append(1)\n",
    "    else:\n",
    "        time_larger.append(1)\n",
    "print('smaller iterations:')\n",
    "print(np.sum(time_smaller))\n",
    "print('equal iterations:')\n",
    "print(np.sum(time_equal))\n",
    "print('larger iterations:')\n",
    "print(np.sum(time_larger))\n",
    "\n",
    "print(np.mean(uniform_plan_time))\n",
    "print(np.mean(inform_plan_time))"
   ]
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "kernelspec": {
   "display_name": "Python [conda env:ylmiao_kino]",
   "language": "python",
   "name": "conda-env-ylmiao_kino-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": "3.6.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}