├── .gitmodules ├── .idea ├── E2C.iml ├── encodings.xml ├── misc.xml ├── modules.xml ├── vcs.xml └── workspace.xml ├── .ipynb_checkpoints └── test_plot-checkpoint.ipynb ├── README.md ├── __pycache__ ├── datasets.cpython-36.pyc ├── e2c_model.cpython-36.pyc ├── networks.cpython-36.pyc ├── normal.cpython-36.pyc └── train_e2c.cpython-36.pyc ├── data ├── __pycache__ │ ├── sample_cartpole_data.cpython-36.pyc │ ├── sample_pendulum_data.cpython-36.pyc │ ├── sample_planar_data.cpython-36.pyc │ ├── sample_planar_data_2.cpython-36.pyc │ └── sample_planar_partial.cpython-36.pyc ├── draw_planar_env.py ├── env.npy ├── env.png ├── sample_cartpole_data.py ├── sample_pendulum_data.py └── sample_planar.py ├── datasets.py ├── draw_latent_map.ipynb ├── e2c.yml ├── e2c_model.py ├── evaluate_saved_model.ipynb ├── map.png ├── networks.py ├── normal.py ├── test_plot.ipynb └── train_e2c.py /.gitmodules: -------------------------------------------------------------------------------- 1 | [submodule "gym"] 2 | path = gym 3 | url = https://github.com/ethanluoyc/gym/tree/pendulum_internal 4 | -------------------------------------------------------------------------------- /.idea/E2C.iml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 14 | -------------------------------------------------------------------------------- /.idea/encodings.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | -------------------------------------------------------------------------------- /.idea/misc.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 6 | 7 | -------------------------------------------------------------------------------- /.idea/modules.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | -------------------------------------------------------------------------------- /.idea/vcs.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | -------------------------------------------------------------------------------- /.idea/workspace.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 | 40 | 41 | 42 | 43 | 44 | 45 | 46 | 47 | 57 | 58 | 198 | 199 | 200 | 201 | 202 | 203 | 204 | 205 | 206 | 207 | 208 | 209 | 210 | 211 | 221 | 222 | 223 | 224 | 225 | 245 | 246 | 247 | 267 | 268 | 269 | 289 | 290 | 291 | 311 | 312 | 313 | 333 | 334 | 335 | 336 | 337 | 338 | 339 | 340 | 341 | 342 | 343 | 344 | 345 | 346 | 347 | 348 | 349 | 1568356268398 350 | 362 | 363 | 364 | 365 | 367 | 368 | 369 | 370 | 371 | 372 | 373 | 374 | 375 | 376 | 377 | 378 | 379 | 380 | 381 | 382 | 383 | 384 | 385 | 386 | 387 | 388 | 389 | 390 | 391 | 392 | 393 | 394 | 395 | 396 | 397 | 398 | 399 | 400 | 401 | 402 | 403 | 404 | 405 | 406 | 407 | 408 | 409 | 410 | 411 | 412 | 413 | 414 | 415 | 416 | 417 | 419 | 420 | 421 | 422 | 423 | 424 | 425 | 426 | 427 | 428 | 429 | 430 | 431 | 432 | 433 | 434 | 435 | 436 | 437 | 438 | 439 | 440 | 441 | 442 | 443 | 444 | 445 | 446 | 447 | 448 | 449 | 450 | 451 | 452 | 453 | 454 | 455 | 456 | 457 | 458 | 459 | 460 | 461 | 462 | 463 | 464 | 465 | 466 | 467 | 468 | 469 | 470 | 471 | 472 | 473 | 474 | 475 | 476 | 477 | 478 | 479 | 480 | 481 | 482 | 483 | 484 | 485 | 486 | 487 | 488 | 489 | 490 | 491 | 492 | 493 | 494 | 495 | 496 | 497 | 498 | 499 | 500 | 501 | 502 | 503 | 504 | 505 | 506 | 507 | 508 | 509 | 510 | 511 | 512 | 513 | 514 | 515 | 516 | 517 | 518 | 519 | 520 | 521 | 522 | 523 | 524 | 525 | 526 | 527 | 528 | 529 | 530 | 531 | 532 | 533 | 534 | 535 | 536 | 537 | 538 | 539 | 540 | 541 | 542 | 543 | 544 | 545 | 546 | 547 | 548 | 549 | 550 | 551 | 552 | 553 | 554 | 555 | 556 | 557 | 558 | 559 | 560 | 561 | 562 | 563 | 564 | 565 | 566 | 567 | 568 | 569 | 570 | 571 | 572 | 573 | 574 | 575 | 576 | 577 | 578 | 579 | 580 | 581 | 582 | 583 | 584 | 585 | 586 | 587 | 588 | 589 | 590 | 591 | 592 | 593 | 594 | 595 | 596 | 597 | 598 | 599 | 600 | 601 | 602 | 603 | 604 | 605 | 606 | 607 | 608 | 609 | 610 | 611 | 612 | 613 | 614 | 615 | 616 | 617 | 618 | 619 | 620 | 621 | 622 | 623 | 624 | 625 | 626 | 627 | 628 | 629 | 630 | 631 | 632 | 633 | 634 | 635 | 636 | -------------------------------------------------------------------------------- /.ipynb_checkpoints/test_plot-checkpoint.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": 1, 6 | "metadata": {}, 7 | "outputs": [], 8 | "source": [ 9 | "from tensorboardX import SummaryWriter\n", 10 | "import matplotlib.pyplot as plt\n", 11 | "import numpy as np" 12 | ] 13 | }, 14 | { 15 | "cell_type": "code", 16 | "execution_count": 28, 17 | "metadata": {}, 18 | "outputs": [ 19 | { 20 | "data": { 21 | "image/png": "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\n", 22 | "text/plain": [ 23 | "

" 24 | ] 25 | }, 26 | "metadata": {}, 27 | "output_type": "display_data" 28 | } 29 | ], 30 | "source": [ 31 | "writer = SummaryWriter('test')\n", 32 | "fig, ax = plt.subplots(nrows=1, ncols=2)\n", 33 | "plt.setp(ax, xticks=[], yticks=[])\n", 34 | "# ax[0].set_ylabel('test', rotation=0,size='large')\n", 35 | "ax[0].annotate(row, xy=(0, 0.5), xytext=(-ax.yaxis.labelpad - pad, 0),\n", 36 | " xycoords=ax.yaxis.label, textcoords='offset points',\n", 37 | " size='large', ha='right', va='center')\n", 38 | "ax[0].imshow(np.random.randn(10,10), cmap='Greys')\n", 39 | "fig.tight_layout()" 40 | ] 41 | }, 42 | { 43 | "cell_type": "code", 44 | "execution_count": 26, 45 | "metadata": {}, 46 | "outputs": [], 47 | "source": [ 48 | "writer.add_figure('test', fig, 0)" 49 | ] 50 | }, 51 | { 52 | "cell_type": "code", 53 | "execution_count": null, 54 | "metadata": {}, 55 | "outputs": [], 56 | "source": [] 57 | } 58 | ], 59 | "metadata": { 60 | "kernelspec": { 61 | "display_name": "Python 3", 62 | "language": "python", 63 | "name": "python3" 64 | }, 65 | "language_info": { 66 | "codemirror_mode": { 67 | "name": "ipython", 68 | "version": 3 69 | }, 70 | "file_extension": ".py", 71 | "mimetype": "text/x-python", 72 | "name": "python", 73 | "nbconvert_exporter": "python", 74 | "pygments_lexer": "ipython3", 75 | "version": "3.6.8" 76 | } 77 | }, 78 | "nbformat": 4, 79 | "nbformat_minor": 2 80 | } 81 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | ## Embed to Control 2 | 3 | This is a pytorch implementation of the paper "[Embed to Control: A Locally Linear Latent Dynamics Model for Control from Raw Images](https://arxiv.org/abs/1506.07365)", NIPS, 2015. 4 | 5 | **Note: This is not and official implementation.** 6 | 7 | ### Installing 8 | 9 | First, clone the repository: 10 | 11 | ``` 12 | git clone https://github.com/tungnd1705/E2C-pytorch.git 13 | ``` 14 | 15 | Install the dependencies as listed in `env.yml` and activate the environment 16 | 17 | ``` 18 | conda env create -f env.yml 19 | 20 | conda activate e2c 21 | ``` 22 | 23 | Then install the patch version of gym in order to sample the pendulum data 24 | 25 | ``` 26 | cd gym 27 | 28 | python setup.py install 29 | ``` 30 | 31 | ### Simulate training data 32 | 33 | Currently the code supports simulating 3 environments: `planar`, `pendulum` and `cartpole`. 34 | 35 | In order to generate data, simply run `python sample_{env_name}_data.py --sample_size={sample_size}`. 36 | 37 | **Note: the sample size is equal to the total number of training and test data** 38 | 39 | For the planar task, we base on [this](https://github.com/ethanluoyc/e2c-pytorch) implementation and modify for our needs. 40 | 41 | ### Training 42 | 43 | Run the ``train_e2c.py`` with your own settings. Example: 44 | 45 | ``` 46 | python train_e2c.py \ 47 | --env=planar \ 48 | --propor=3/4 \ 49 | --batch_size=128 \ 50 | --lr=0.0001 \ 51 | --lam=0.25 \ 52 | --num_iter=5000 \ 53 | --iter_save=1000 54 | ``` 55 | 56 | You can visualize the training process by running ``tensorboard --logdir=logs``. 57 | 58 | ### Citation 59 | 60 | If you find E2C useful in your research, please consider citing: 61 | 62 | ``` 63 | @inproceedings{watter2015embed, 64 | title={Embed to control: A locally linear latent dynamics model for control from raw images}, 65 | author={Watter, Manuel and Springenberg, Jost and Boedecker, Joschka and Riedmiller, Martin}, 66 | booktitle={Advances in neural information processing systems}, 67 | pages={2746--2754}, 68 | year={2015} 69 | } 70 | ``` 71 | -------------------------------------------------------------------------------- /__pycache__/datasets.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/__pycache__/datasets.cpython-36.pyc -------------------------------------------------------------------------------- /__pycache__/e2c_model.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/__pycache__/e2c_model.cpython-36.pyc -------------------------------------------------------------------------------- /__pycache__/networks.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/__pycache__/networks.cpython-36.pyc -------------------------------------------------------------------------------- /__pycache__/normal.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/__pycache__/normal.cpython-36.pyc -------------------------------------------------------------------------------- /__pycache__/train_e2c.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/__pycache__/train_e2c.cpython-36.pyc -------------------------------------------------------------------------------- /data/__pycache__/sample_cartpole_data.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/data/__pycache__/sample_cartpole_data.cpython-36.pyc -------------------------------------------------------------------------------- /data/__pycache__/sample_pendulum_data.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/data/__pycache__/sample_pendulum_data.cpython-36.pyc -------------------------------------------------------------------------------- /data/__pycache__/sample_planar_data.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/data/__pycache__/sample_planar_data.cpython-36.pyc -------------------------------------------------------------------------------- /data/__pycache__/sample_planar_data_2.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/data/__pycache__/sample_planar_data_2.cpython-36.pyc -------------------------------------------------------------------------------- /data/__pycache__/sample_planar_partial.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/data/__pycache__/sample_planar_partial.cpython-36.pyc -------------------------------------------------------------------------------- /data/draw_planar_env.py: -------------------------------------------------------------------------------- 1 | from PIL import Image, ImageDraw 2 | import numpy as np 3 | width, height = 40, 40 4 | r = 2.5 5 | obstacles_center = np.array([[20.5, 5.5], [20.5, 12.5], [20.5, 27.5], [20.5, 35.5], [10.5, 20.5], [30.5, 20.5]]) 6 | 7 | def generate_env(): 8 | print ('Making the environment...') 9 | img_arr = np.zeros(shape=(width,height)) 10 | 11 | img_env = Image.fromarray(img_arr) 12 | draw = ImageDraw.Draw(img_env) 13 | for y, x in obstacles_center: 14 | draw.ellipse((int(x)-int(r), int(y)-int(r), int(x)+int(r), int(y)+int(r)), fill=255) 15 | img_env = img_env.convert('L') 16 | img_env.save('env.png') 17 | 18 | img_arr = np.array(img_env) / 255. 19 | np.save('./env.npy', img_arr) 20 | return img_arr 21 | 22 | env = generate_env() -------------------------------------------------------------------------------- /data/env.npy: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/data/env.npy -------------------------------------------------------------------------------- /data/env.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/data/env.png -------------------------------------------------------------------------------- /data/sample_cartpole_data.py: -------------------------------------------------------------------------------- 1 | import os 2 | from os import path 3 | from tqdm import trange 4 | import matplotlib.pyplot as plt 5 | import numpy as np 6 | import gym 7 | import json 8 | from datetime import datetime 9 | import argparse 10 | 11 | def sample_cartpole(sample_size, output_dir='data/cartpole', step_size=5, apply_control=True, num_shards=10): 12 | env = gym.make('CartPole-v0').env 13 | assert sample_size % num_shards == 0 14 | 15 | samples = [] 16 | 17 | if not path.exists(output_dir): 18 | os.makedirs(output_dir) 19 | 20 | state = env.reset() 21 | for i in trange(sample_size): 22 | """ 23 | for each sample: 24 | - draw a random state 25 | - render x_t 26 | - draw a random action u_t and apply 27 | - render x_t+1 after applying u_t 28 | """ 29 | # state = env.reset() 30 | 31 | initial_state = state 32 | before = env.render(mode = 'rgb_array') 33 | 34 | # apply the same control over a few timesteps 35 | if apply_control: 36 | u = env.action_space.sample() 37 | else: 38 | u = np.zeros((1,)) 39 | 40 | for _ in range(step_size): 41 | state, reward, done, info = env.step(u) 42 | 43 | after_state = state 44 | after = env.render(mode = 'rgb_array') 45 | 46 | shard_no = i // (sample_size // num_shards) 47 | 48 | shard_path = path.join('{:03d}-of-{:03d}'.format(shard_no, num_shards)) 49 | 50 | if not path.exists(path.join(output_dir, shard_path)): 51 | os.makedirs(path.join(output_dir, shard_path)) 52 | 53 | before_file = path.join(shard_path, 'before-{:05d}.jpg'.format(i)) 54 | plt.imsave(path.join(output_dir, before_file), before) 55 | 56 | after_file = path.join(shard_path, 'after-{:05d}.jpg'.format(i)) 57 | plt.imsave(path.join(output_dir, after_file), after) 58 | 59 | samples.append({ 60 | 'before_state': initial_state.tolist(), 61 | 'after_state': after_state.tolist(), 62 | 'before': before_file, 63 | 'after': after_file, 64 | 'control': [u], 65 | }) 66 | 67 | with open(path.join(output_dir, 'data.json'), 'wt') as outfile: 68 | json.dump( 69 | { 70 | 'metadata': { 71 | 'num_samples': sample_size, 72 | 'step_size': step_size, 73 | 'apply_control': apply_control, 74 | 'time_created': str(datetime.now()), 75 | 'version': 1 76 | }, 77 | 'samples': samples 78 | }, outfile, indent=2) 79 | 80 | env.viewer.close() 81 | 82 | def main(args): 83 | sample_size = args.sample_size 84 | 85 | sample_cartpole(sample_size=sample_size) 86 | 87 | if __name__ == "__main__": 88 | parser = argparse.ArgumentParser(description='sample data') 89 | 90 | parser.add_argument('--sample_size', required=True, type=int, help='the number of samples') 91 | 92 | args = parser.parse_args() 93 | 94 | main(args) -------------------------------------------------------------------------------- /data/sample_pendulum_data.py: -------------------------------------------------------------------------------- 1 | import os 2 | from os import path 3 | from tqdm import trange 4 | import matplotlib.pyplot as plt 5 | import numpy as np 6 | import gym 7 | import json 8 | from datetime import datetime 9 | import argparse 10 | 11 | env = gym.make('Pendulum-v0').env 12 | width, height = 48 * 2, 48 13 | 14 | def render(state): 15 | # need two observations to restore the Markov property 16 | before1 = state 17 | before2 = env.step_from_state(state, np.array([0])) 18 | return map(env.render_state, (before1[0], before2[0])) 19 | 20 | def sample_pendulum(sample_size, output_dir='data/pendulum', step_size=1, apply_control=True, num_shards=10): 21 | assert sample_size % num_shards == 0 22 | 23 | samples = [] 24 | 25 | if not path.exists(output_dir): 26 | os.makedirs(output_dir) 27 | 28 | for i in trange(sample_size): 29 | """ 30 | for each sample: 31 | - draw a random state (theta, theta dot) 32 | - render x_t (including 2 images) 33 | - draw a random action u_t and apply 34 | - render x_t+1 after applying u_t 35 | """ 36 | # th (theta) and thdot (theta dot) represent a state in Pendulum env 37 | th = np.random.uniform(0, np.pi * 2) 38 | thdot = np.random.uniform(-8, 8) 39 | 40 | state = np.array([th, thdot]) 41 | 42 | initial_state = np.copy(state) 43 | before1, before2 = render(state) 44 | 45 | # apply the same control over a few timesteps 46 | if apply_control: 47 | u = np.random.uniform(-2, 2, size=(1,)) 48 | else: 49 | u = np.zeros((1,)) 50 | 51 | for _ in range(step_size): 52 | state = env.step_from_state(state, u) 53 | 54 | after_state = np.copy(state) 55 | after1, after2 = render(state) 56 | 57 | before = np.hstack((before1, before2)) 58 | after = np.hstack((after1, after2)) 59 | 60 | shard_no = i // (sample_size // num_shards) 61 | 62 | shard_path = path.join('{:03d}-of-{:03d}'.format(shard_no, num_shards)) 63 | 64 | if not path.exists(path.join(output_dir, shard_path)): 65 | os.makedirs(path.join(output_dir, shard_path)) 66 | 67 | before_file = path.join(shard_path, 'before-{:05d}.jpg'.format(i)) 68 | plt.imsave(path.join(output_dir, before_file), before) 69 | 70 | after_file = path.join(shard_path, 'after-{:05d}.jpg'.format(i)) 71 | plt.imsave(path.join(output_dir, after_file), after) 72 | 73 | samples.append({ 74 | 'before_state': initial_state.tolist(), 75 | 'after_state': after_state.tolist(), 76 | 'before': before_file, 77 | 'after': after_file, 78 | 'control': u.tolist(), 79 | }) 80 | 81 | with open(path.join(output_dir, 'data.json'), 'wt') as outfile: 82 | json.dump( 83 | { 84 | 'metadata': { 85 | 'num_samples': sample_size, 86 | 'step_size': step_size, 87 | 'apply_control': apply_control, 88 | 'time_created': str(datetime.now()), 89 | 'version': 1 90 | }, 91 | 'samples': samples 92 | }, outfile, indent=2) 93 | 94 | env.viewer.close() 95 | 96 | def main(args): 97 | sample_size = args.sample_size 98 | 99 | sample_pendulum(sample_size=sample_size) 100 | 101 | if __name__ == "__main__": 102 | parser = argparse.ArgumentParser(description='sample data') 103 | 104 | parser.add_argument('--sample_size', required=True, type=int, help='the number of samples') 105 | 106 | args = parser.parse_args() 107 | 108 | main(args) -------------------------------------------------------------------------------- /data/sample_planar.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import os 3 | from os import path 4 | from tqdm import trange 5 | import json 6 | from datetime import datetime 7 | import argparse 8 | from PIL import Image, ImageDraw 9 | 10 | np.random.seed(1) 11 | 12 | width, height = 40, 40 13 | obstacles_center = np.array([[20.5, 5.5], [20.5, 12.5], [20.5, 27.5], [20.5, 35.5], [10.5, 20.5], [30.5, 20.5]]) 14 | 15 | r_overlap = 0.5 # agent cannot be in any rectangular area with obstacles as centers and half-width = 0.5 16 | r = 2.5 # radius of the obstacles when rendered 17 | rw = 3 # robot half-width 18 | rw_rendered = 2 # robot half-width when rendered 19 | max_step_len = 3 20 | env_path = os.path.dirname(os.path.abspath(__file__)) 21 | env = np.load(env_path + '/env.npy') 22 | 23 | def get_pixel_location(s): 24 | # return the location of agent when rendered 25 | center_x, center_y = int(round(s[0])), int(round(s[1])) 26 | top = center_x - rw_rendered 27 | bottom = center_x + rw_rendered 28 | left = center_y - rw_rendered 29 | right = center_y + rw_rendered 30 | return top, bottom, left, right 31 | 32 | def render(s): 33 | top, bottom, left, right = get_pixel_location(s) 34 | x = np.copy(env) 35 | x[top:bottom, left:right] = 1. # robot is white on black background 36 | return x 37 | 38 | def is_valid(s, u, s_next, epsilon = 0.1): 39 | # if the difference between the action and the actual distance between x and x_next are in range(0,epsilon) 40 | top, bottom, left, right = get_pixel_location(s) 41 | top_next, bottom_next, left_next, right_next = get_pixel_location(s_next) 42 | x_diff = np.array([top_next - top, left_next - left], dtype=np.float) 43 | return (not np.sqrt(np.sum((x_diff - u)**2)) > epsilon) 44 | 45 | def is_colliding(s): 46 | """ 47 | :param s: the continuous coordinate (x, y) of the agent center 48 | :return: if agent body overlaps with obstacles 49 | """ 50 | if np.any([s - rw < 0, s + rw > height]): 51 | return True 52 | x, y = s[0], s[1] 53 | for obs in obstacles_center: 54 | if np.abs(obs[0] - x) <= r_overlap and np.abs(obs[1] - y) <= r_overlap: 55 | return True 56 | return False 57 | 58 | def random_step(s): 59 | # draw a random step until it doesn't collidie with the obstacles 60 | while True: 61 | u = np.random.uniform(low = -max_step_len, high = max_step_len, size = 2) 62 | s_next = s + u 63 | if (not is_colliding(s_next) and is_valid(s, u, s_next)): 64 | return u, s_next 65 | 66 | def sample(sample_size): 67 | """ 68 | return [(s, u, s_next)] 69 | """ 70 | state_samples = [] 71 | for i in trange(sample_size, desc = 'Sampling data'): 72 | while True: 73 | s_x = np.random.uniform(low = rw, high = height - rw) 74 | s_y = np.random.uniform(low = rw, high = width - rw) 75 | s = np.array([s_x, s_y]) 76 | if not is_colliding(s): 77 | break 78 | u, s_next = random_step(s) 79 | state_samples.append((s, u, s_next)) 80 | obs_samples = [(render(s), u, render(s_next)) for s, u, s_next in state_samples] 81 | return state_samples, obs_samples 82 | 83 | def write_to_file(sample_size, output_dir = './data/planar'): 84 | """ 85 | write [(x, u, x_next)] to output dir 86 | """ 87 | if not path.exists(output_dir): 88 | os.makedirs(output_dir) 89 | 90 | state_samples, obs_samples = sample(sample_size) 91 | 92 | samples = [] 93 | 94 | for i, (before, u, after) in enumerate(obs_samples): 95 | before_file = 'before-{:05d}.png'.format(i) 96 | Image.fromarray(before * 255.).convert('L').save(path.join(output_dir, before_file)) 97 | 98 | after_file = 'after-{:05d}.png'.format(i) 99 | Image.fromarray(after * 255.).convert('L').save(path.join(output_dir, after_file)) 100 | 101 | initial_state = state_samples[i][0] 102 | after_state = state_samples[i][2] 103 | 104 | samples.append({ 105 | 'before_state': initial_state.tolist(), 106 | 'after_state': after_state.tolist(), 107 | 'before': before_file, 108 | 'after': after_file, 109 | 'control': u.tolist(), 110 | }) 111 | 112 | with open(path.join(output_dir, 'data.json'), 'wt') as outfile: 113 | json.dump( 114 | { 115 | 'metadata': { 116 | 'num_samples': sample_size, 117 | 'max_distance': max_step_len, 118 | 'time_created': str(datetime.now()), 119 | 'version': 1 120 | }, 121 | 'samples': samples 122 | }, outfile, indent=2) 123 | 124 | def main(args): 125 | sample_size = args.sample_size 126 | 127 | write_to_file(sample_size=sample_size) 128 | 129 | if __name__ == "__main__": 130 | parser = argparse.ArgumentParser(description='sample data') 131 | 132 | parser.add_argument('--sample_size', required=True, type=int, help='the number of samples') 133 | 134 | args = parser.parse_args() 135 | 136 | main(args) -------------------------------------------------------------------------------- /datasets.py: -------------------------------------------------------------------------------- 1 | import os 2 | from os import path 3 | from PIL import Image 4 | import numpy as np 5 | import json 6 | from torchvision.transforms import ToTensor 7 | from torch.utils.data import Dataset 8 | from tqdm import tqdm 9 | import pickle 10 | import torch 11 | 12 | torch.set_default_dtype(torch.float64) 13 | 14 | class PlanarDataset(Dataset): 15 | width = 40 16 | height = 40 17 | action_dim = 2 18 | 19 | def __init__(self, dir): 20 | self.dir = dir 21 | with open(path.join(dir, 'data.json')) as f: 22 | self._data = json.load(f) 23 | self._process() 24 | 25 | def __len__(self): 26 | return len(self._data['samples']) 27 | 28 | def __getitem__(self, index): 29 | return self._processed[index] 30 | 31 | @staticmethod 32 | def _process_image(img): 33 | return ToTensor()((img.convert('L'). 34 | resize((PlanarDataset.width, 35 | PlanarDataset.height)))) 36 | 37 | def _process(self): 38 | preprocessed_file = os.path.join(self.dir, 'processed.pkl') 39 | if not os.path.exists(preprocessed_file): 40 | processed = [] 41 | for sample in tqdm(self._data['samples'], desc='processing data'): 42 | before = Image.open(os.path.join(self.dir, sample['before'])) 43 | after = Image.open(os.path.join(self.dir, sample['after'])) 44 | 45 | processed.append((self._process_image(before), 46 | np.array(sample['control']), 47 | self._process_image(after))) 48 | 49 | with open(preprocessed_file, 'wb') as f: 50 | pickle.dump(processed, f) 51 | self._processed = processed 52 | else: 53 | with open(preprocessed_file, 'rb') as f: 54 | self._processed = pickle.load(f) 55 | 56 | class GymPendulumDatasetV2(Dataset): 57 | width = 48 * 2 58 | height = 48 59 | action_dim = 1 60 | 61 | def __init__(self, dir): 62 | self.dir = dir 63 | with open(path.join(dir, 'data.json')) as f: 64 | self._data = json.load(f) 65 | self._process() 66 | 67 | def __len__(self): 68 | return len(self._data['samples']) 69 | 70 | def __getitem__(self, index): 71 | return self._processed[index] 72 | 73 | @staticmethod 74 | def _process_image(img): 75 | return ToTensor()((img.convert('L'). 76 | resize((GymPendulumDatasetV2.width, 77 | GymPendulumDatasetV2.height)))) 78 | 79 | def _process(self): 80 | preprocessed_file = os.path.join(self.dir, 'processed.pkl') 81 | if not os.path.exists(preprocessed_file): 82 | processed = [] 83 | for sample in tqdm(self._data['samples'], desc='processing data'): 84 | before = Image.open(os.path.join(self.dir, sample['before'])) 85 | after = Image.open(os.path.join(self.dir, sample['after'])) 86 | 87 | processed.append((self._process_image(before), 88 | np.array(sample['control']), 89 | self._process_image(after))) 90 | 91 | with open(preprocessed_file, 'wb') as f: 92 | pickle.dump(processed, f) 93 | self._processed = processed 94 | else: 95 | with open(preprocessed_file, 'rb') as f: 96 | self._processed = pickle.load(f) -------------------------------------------------------------------------------- /draw_latent_map.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": 1, 6 | "metadata": {}, 7 | "outputs": [], 8 | "source": [ 9 | "import matplotlib.pyplot as plt\n", 10 | "from PIL import Image, ImageDraw\n", 11 | "import colour\n", 12 | "from random import randint as rint\n", 13 | "import numpy as np\n", 14 | "from colour import Color\n", 15 | "from data.sample_planar_partial import *" 16 | ] 17 | }, 18 | { 19 | "cell_type": "code", 20 | "execution_count": 2, 21 | "metadata": {}, 22 | "outputs": [ 23 | { 24 | "data": { 25 | "text/plain": [ 26 | "[[10, 20],\n", 27 | " [10, 21],\n", 28 | " [11, 20],\n", 29 | " [11, 21],\n", 30 | " [20, 5],\n", 31 | " [20, 6],\n", 32 | " [20, 12],\n", 33 | " [20, 13],\n", 34 | " [20, 27],\n", 35 | " [20, 28],\n", 36 | " [20, 35],\n", 37 | " [20, 36],\n", 38 | " [21, 5],\n", 39 | " [21, 6],\n", 40 | " [21, 12],\n", 41 | " [21, 13],\n", 42 | " [21, 27],\n", 43 | " [21, 28],\n", 44 | " [21, 35],\n", 45 | " [21, 36],\n", 46 | " [30, 20],\n", 47 | " [30, 21],\n", 48 | " [31, 20],\n", 49 | " [31, 21]]" 50 | ] 51 | }, 52 | "execution_count": 2, 53 | "metadata": {}, 54 | "output_type": "execute_result" 55 | } 56 | ], 57 | "source": [ 58 | "start, end = 3, 37\n", 59 | "unvalid_pos = []\n", 60 | "for x in range(start, end + 1):\n", 61 | " for y in range(start, end + 1):\n", 62 | " s = [x,y]\n", 63 | " if is_colliding(np.array(s)):\n", 64 | " unvalid_pos.append(s)\n", 65 | "unvalid_pos" 66 | ] 67 | }, 68 | { 69 | "cell_type": "code", 70 | "execution_count": 3, 71 | "metadata": {}, 72 | "outputs": [], 73 | "source": [ 74 | "red = Color('blue')\n", 75 | "# red_rgb = np.array(Color('red').rgb)\n", 76 | "# blue_rgb = np.array(Color('blue').rgb)\n", 77 | "colors = list(red.range_to(Color(\"red\"),36))\n", 78 | "colors_rgb = [color.rgb for color in colors]\n", 79 | "# print (colors_rgb)" 80 | ] 81 | }, 82 | { 83 | "cell_type": "code", 84 | "execution_count": 4, 85 | "metadata": {}, 86 | "outputs": [], 87 | "source": [ 88 | "def random_gradient():\n", 89 | " img = Image.new(\"RGB\", (width,height), \"#FFFFFF\")\n", 90 | " draw = ImageDraw.Draw(img)\n", 91 | " \n", 92 | " for i, color in zip(range(start, end+1), colors_rgb):\n", 93 | " r1, g1, b1 = color[0] * 255., color[1] * 255., color[2] * 255.\n", 94 | " draw.line((i,start,i,end), fill=(int(r1), int(g1), int(b1)))\n", 95 | " \n", 96 | " # obstacles as white\n", 97 | "# for y, x in obstacles_center:\n", 98 | "# draw.rectangle(((x - r, y - r), (x + r, y + r)), fill=\"white\")\n", 99 | "\n", 100 | " img_arr = np.array(img)\n", 101 | " for x, y in unvalid_pos:\n", 102 | " img_arr[x, y] = 255.\n", 103 | " return img_arr / 255., img" 104 | ] 105 | }, 106 | { 107 | "cell_type": "code", 108 | "execution_count": 5, 109 | "metadata": {}, 110 | "outputs": [], 111 | "source": [ 112 | "img_arr, img = random_gradient()" 113 | ] 114 | }, 115 | { 116 | "cell_type": "code", 117 | "execution_count": 6, 118 | "metadata": { 119 | "scrolled": false 120 | }, 121 | "outputs": [ 122 | { 123 | "data": { 124 | "image/png": "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\n", 125 | "text/plain": [ 126 | "

" 127 | ] 128 | }, 129 | "metadata": { 130 | "needs_background": "light" 131 | }, 132 | "output_type": "display_data" 133 | } 134 | ], 135 | "source": [ 136 | "plt.imshow(img_arr)\n", 137 | "# plt.axis('off')\n", 138 | "plt.show()" 139 | ] 140 | }, 141 | { 142 | "cell_type": "code", 143 | "execution_count": 7, 144 | "metadata": {}, 145 | "outputs": [ 146 | { 147 | "data": { 148 | "image/png": "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\n", 149 | "text/plain": [ 150 | "

" 151 | ] 152 | }, 153 | "metadata": { 154 | "needs_background": "light" 155 | }, 156 | "output_type": "display_data" 157 | } 158 | ], 159 | "source": [ 160 | "img_scaled = Image.new(\"RGB\", (width * 10,height*10), \"#FFFFFF\")\n", 161 | "draw = ImageDraw.Draw(img_scaled)\n", 162 | "for y in range(start, end + 1):\n", 163 | " for x in range(start, end + 1):\n", 164 | " if [y, x] in unvalid_pos:\n", 165 | " continue\n", 166 | " else:\n", 167 | " x_scaled, y_scaled = x * 10, y * 10\n", 168 | " draw.ellipse((x_scaled-2, y_scaled-2, x_scaled+2, y_scaled+2), fill = img.getpixel((x,y)))\n", 169 | "img_scaled.save('map.png', 'PNG')\n", 170 | "img_arr_scaled = np.array(img_scaled) / 255.\n", 171 | "plt.imshow(img_arr_scaled)\n", 172 | "plt.show()" 173 | ] 174 | }, 175 | { 176 | "cell_type": "code", 177 | "execution_count": 31, 178 | "metadata": {}, 179 | "outputs": [], 180 | "source": [ 181 | "import torch\n", 182 | "from e2c_model import E2C" 183 | ] 184 | }, 185 | { 186 | "cell_type": "code", 187 | "execution_count": 32, 188 | "metadata": {}, 189 | "outputs": [ 190 | { 191 | "data": { 192 | "text/plain": [ 193 | "E2C(\n", 194 | " (encoder): PlanarEncoder(\n", 195 | " (net): Sequential(\n", 196 | " (0): Linear(in_features=1600, out_features=150, bias=True)\n", 197 | " (1): BatchNorm1d(150, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n", 198 | " (2): ReLU()\n", 199 | " (3): Linear(in_features=150, out_features=150, bias=True)\n", 200 | " (4): BatchNorm1d(150, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n", 201 | " (5): ReLU()\n", 202 | " (6): Linear(in_features=150, out_features=150, bias=True)\n", 203 | " (7): BatchNorm1d(150, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n", 204 | " (8): ReLU()\n", 205 | " (9): Linear(in_features=150, out_features=4, bias=True)\n", 206 | " )\n", 207 | " )\n", 208 | " (decoder): PlanarDecoder(\n", 209 | " (net): Sequential(\n", 210 | " (0): Linear(in_features=2, out_features=200, bias=True)\n", 211 | " (1): BatchNorm1d(200, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n", 212 | " (2): ReLU()\n", 213 | " (3): Linear(in_features=200, out_features=200, bias=True)\n", 214 | " (4): BatchNorm1d(200, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n", 215 | " (5): ReLU()\n", 216 | " (6): Linear(in_features=200, out_features=1600, bias=True)\n", 217 | " (7): Sigmoid()\n", 218 | " )\n", 219 | " )\n", 220 | " (trans): PlanarTransition(\n", 221 | " (net): Sequential(\n", 222 | " (0): Linear(in_features=2, out_features=100, bias=True)\n", 223 | " (1): BatchNorm1d(100, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n", 224 | " (2): ReLU()\n", 225 | " (3): Linear(in_features=100, out_features=100, bias=True)\n", 226 | " (4): BatchNorm1d(100, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n", 227 | " (5): ReLU()\n", 228 | " )\n", 229 | " (fc_A): Sequential(\n", 230 | " (0): Linear(in_features=100, out_features=4, bias=True)\n", 231 | " (1): Sigmoid()\n", 232 | " )\n", 233 | " (fc_B): Linear(in_features=100, out_features=4, bias=True)\n", 234 | " (fc_o): Linear(in_features=100, out_features=2, bias=True)\n", 235 | " )\n", 236 | ")" 237 | ] 238 | }, 239 | "execution_count": 32, 240 | "metadata": {}, 241 | "output_type": "execute_result" 242 | } 243 | ], 244 | "source": [ 245 | "model = E2C(1600,2,2,'planar_partial').cuda(0)\n", 246 | "model.load_state_dict(torch.load('result/planar_partial/log_check_valid/model_5000'))\n", 247 | "model.eval()" 248 | ] 249 | }, 250 | { 251 | "cell_type": "code", 252 | "execution_count": 33, 253 | "metadata": {}, 254 | "outputs": [], 255 | "source": [ 256 | "all_z = []\n", 257 | "for x in range(start, end + 1):\n", 258 | " for y in range(start, end + 1):\n", 259 | " s = np.array([x,y])\n", 260 | " if [x,y] in unvalid_pos:\n", 261 | " all_z.append(np.zeros(2))\n", 262 | " else:\n", 263 | " obs = render(s)\n", 264 | " with torch.no_grad():\n", 265 | " obs = torch.Tensor(render(s)).unsqueeze(0).view(-1,1600).double().cuda(0)\n", 266 | " mu, sigma = model.encode(obs)\n", 267 | " z = mu.squeeze().cpu().numpy()\n", 268 | " all_z.append(np.copy(z))\n", 269 | "all_z = np.array(all_z)" 270 | ] 271 | }, 272 | { 273 | "cell_type": "code", 274 | "execution_count": 34, 275 | "metadata": {}, 276 | "outputs": [ 277 | { 278 | "name": "stdout", 279 | "output_type": "stream", 280 | "text": [ 281 | "(3, 3): [23 3]\n", 282 | "(3, 4): [23 3]\n", 283 | "(3, 5): [22 3]\n", 284 | "(3, 6): [22 1]\n", 285 | "(3, 7): [21 2]\n", 286 | "(3, 8): [21 1]\n", 287 | "(3, 9): [20 3]\n", 288 | "(3, 10): [20 8]\n", 289 | "(3, 11): [20 10]\n", 290 | "(3, 12): [20 12]\n", 291 | "(3, 13): [20 13]\n", 292 | "(3, 14): [20 13]\n", 293 | "(3, 15): [20 14]\n", 294 | "(3, 16): [20 14]\n", 295 | "(3, 17): [20 15]\n", 296 | "(3, 18): [19 15]\n", 297 | "(3, 19): [19 15]\n", 298 | "(3, 20): [19 16]\n", 299 | "(3, 21): [19 24]\n", 300 | "(3, 22): [19 24]\n", 301 | "(3, 23): [20 24]\n", 302 | "(3, 24): [20 24]\n", 303 | "(3, 25): [20 24]\n", 304 | "(3, 26): [20 24]\n", 305 | "(3, 27): [21 24]\n", 306 | "(3, 28): [21 23]\n", 307 | "(3, 29): [21 23]\n", 308 | "(3, 30): [22 23]\n", 309 | "(3, 31): [22 23]\n", 310 | "(3, 32): [22 15]\n", 311 | "(3, 33): [22 14]\n", 312 | "(3, 34): [23 13]\n", 313 | "(3, 35): [23 12]\n", 314 | "(3, 36): [23 11]\n", 315 | "(3, 37): [23 10]\n", 316 | "(4, 3): [23 6]\n", 317 | "(4, 4): [22 5]\n", 318 | "(4, 5): [22 5]\n", 319 | "(4, 6): [22 3]\n", 320 | "(4, 7): [21 4]\n", 321 | "(4, 8): [21 2]\n", 322 | "(4, 9): [20 2]\n", 323 | "(4, 10): [20 5]\n", 324 | "(4, 11): [20 8]\n", 325 | "(4, 12): [20 11]\n", 326 | "(4, 13): [20 12]\n", 327 | "(4, 14): [19 13]\n", 328 | "(4, 15): [19 13]\n", 329 | "(4, 16): [19 14]\n", 330 | "(4, 17): [19 14]\n", 331 | "(4, 18): [19 15]\n", 332 | "(4, 19): [19 15]\n", 333 | "(4, 20): [19 16]\n", 334 | "(4, 21): [19 25]\n", 335 | "(4, 22): [19 25]\n", 336 | "(4, 23): [20 25]\n", 337 | "(4, 24): [20 24]\n", 338 | "(4, 25): [20 24]\n", 339 | "(4, 26): [21 24]\n", 340 | "(4, 27): [21 24]\n", 341 | "(4, 28): [21 24]\n", 342 | "(4, 29): [22 24]\n", 343 | "(4, 30): [22 24]\n", 344 | "(4, 31): [22 23]\n", 345 | "(4, 32): [23 15]\n", 346 | "(4, 33): [23 14]\n", 347 | "(4, 34): [23 13]\n", 348 | "(4, 35): [23 12]\n", 349 | "(4, 36): [23 11]\n", 350 | "(4, 37): [24 9]\n", 351 | "(5, 3): [22 10]\n", 352 | "(5, 4): [22 10]\n", 353 | "(5, 5): [22 8]\n", 354 | "(5, 6): [21 6]\n", 355 | "(5, 7): [21 7]\n", 356 | "(5, 8): [20 4]\n", 357 | "(5, 9): [20 4]\n", 358 | "(5, 10): [20 6]\n", 359 | "(5, 11): [19 7]\n", 360 | "(5, 12): [19 9]\n", 361 | "(5, 13): [19 12]\n", 362 | "(5, 14): [19 12]\n", 363 | "(5, 15): [19 13]\n", 364 | "(5, 16): [19 14]\n", 365 | "(5, 17): [19 14]\n", 366 | "(5, 18): [19 14]\n", 367 | "(5, 19): [19 15]\n", 368 | "(5, 20): [19 15]\n", 369 | "(5, 21): [19 25]\n", 370 | "(5, 22): [19 25]\n", 371 | "(5, 23): [20 25]\n", 372 | "(5, 24): [20 25]\n", 373 | "(5, 25): [20 25]\n", 374 | "(5, 26): [21 25]\n", 375 | "(5, 27): [21 24]\n", 376 | "(5, 28): [21 24]\n", 377 | "(5, 29): [22 24]\n", 378 | "(5, 30): [22 24]\n", 379 | "(5, 31): [22 24]\n", 380 | "(5, 32): [23 15]\n", 381 | "(5, 33): [23 14]\n", 382 | "(5, 34): [23 13]\n", 383 | "(5, 35): [23 12]\n", 384 | "(5, 36): [24 12]\n", 385 | "(5, 37): [24 10]\n", 386 | "(6, 3): [22 11]\n", 387 | "(6, 4): [22 10]\n", 388 | "(6, 5): [22 10]\n", 389 | "(6, 6): [21 8]\n", 390 | "(6, 7): [21 8]\n", 391 | "(6, 8): [20 8]\n", 392 | "(6, 9): [20 5]\n", 393 | "(6, 10): [19 5]\n", 394 | "(6, 11): [19 8]\n", 395 | "(6, 12): [19 10]\n", 396 | "(6, 13): [19 11]\n", 397 | "(6, 14): [19 12]\n", 398 | "(6, 15): [19 12]\n", 399 | "(6, 16): [19 13]\n", 400 | "(6, 17): [18 14]\n", 401 | "(6, 18): [18 14]\n", 402 | "(6, 19): [18 15]\n", 403 | "(6, 20): [18 15]\n", 404 | "(6, 21): [19 26]\n", 405 | "(6, 22): [19 26]\n", 406 | "(6, 23): [20 26]\n", 407 | "(6, 24): [20 26]\n", 408 | "(6, 25): [20 25]\n", 409 | "(6, 26): [21 25]\n", 410 | "(6, 27): [21 25]\n", 411 | "(6, 28): [21 25]\n", 412 | "(6, 29): [22 24]\n", 413 | "(6, 30): [22 24]\n", 414 | "(6, 31): [22 24]\n", 415 | "(6, 32): [23 15]\n", 416 | "(6, 33): [23 14]\n", 417 | "(6, 34): [24 13]\n", 418 | "(6, 35): [24 12]\n", 419 | "(6, 36): [24 11]\n", 420 | "(6, 37): [24 10]\n", 421 | "(7, 3): [22 12]\n", 422 | "(7, 4): [22 12]\n", 423 | "(7, 5): [21 11]\n", 424 | "(7, 6): [21 10]\n", 425 | "(7, 7): [21 9]\n", 426 | "(7, 8): [20 9]\n", 427 | "(7, 9): [20 7]\n", 428 | "(7, 10): [19 4]\n", 429 | "(7, 11): [18 6]\n", 430 | "(7, 12): [19 10]\n", 431 | "(7, 13): [18 11]\n", 432 | "(7, 14): [18 11]\n", 433 | "(7, 15): [18 12]\n", 434 | "(7, 16): [18 13]\n", 435 | "(7, 17): [18 13]\n", 436 | "(7, 18): [18 15]\n", 437 | "(7, 19): [18 15]\n", 438 | "(7, 20): [18 15]\n", 439 | "(7, 21): [19 26]\n", 440 | "(7, 22): [19 26]\n", 441 | "(7, 23): [20 26]\n", 442 | "(7, 24): [20 26]\n", 443 | "(7, 25): [21 26]\n", 444 | "(7, 26): [21 26]\n", 445 | "(7, 27): [21 25]\n", 446 | "(7, 28): [22 25]\n", 447 | "(7, 29): [22 25]\n", 448 | "(7, 30): [22 25]\n", 449 | "(7, 31): [23 24]\n", 450 | "(7, 32): [24 15]\n", 451 | "(7, 33): [24 15]\n", 452 | "(7, 34): [24 13]\n", 453 | "(7, 35): [24 12]\n", 454 | "(7, 36): [24 11]\n", 455 | "(7, 37): [25 11]\n", 456 | "(8, 3): [22 13]\n", 457 | "(8, 4): [22 13]\n", 458 | "(8, 5): [21 12]\n", 459 | "(8, 6): [21 11]\n", 460 | "(8, 7): [21 11]\n", 461 | "(8, 8): [21 10]\n", 462 | "(8, 9): [20 8]\n", 463 | "(8, 10): [ 9 11]\n", 464 | "(8, 11): [18 5]\n", 465 | "(8, 12): [10 13]\n", 466 | "(8, 13): [18 9]\n", 467 | "(8, 14): [18 9]\n", 468 | "(8, 15): [18 12]\n", 469 | "(8, 16): [18 12]\n", 470 | "(8, 17): [18 13]\n", 471 | "(8, 18): [18 14]\n", 472 | "(8, 19): [18 14]\n", 473 | "(8, 20): [18 27]\n", 474 | "(8, 21): [19 27]\n", 475 | "(8, 22): [19 27]\n", 476 | "(8, 23): [20 27]\n", 477 | "(8, 24): [20 27]\n", 478 | "(8, 25): [20 26]\n", 479 | "(8, 26): [21 26]\n", 480 | "(8, 27): [21 26]\n", 481 | "(8, 28): [22 26]\n", 482 | "(8, 29): [22 25]\n", 483 | "(8, 30): [22 25]\n", 484 | "(8, 31): [23 25]\n", 485 | "(8, 32): [24 15]\n", 486 | "(8, 33): [24 14]\n", 487 | "(8, 34): [24 13]\n", 488 | "(8, 35): [25 13]\n", 489 | "(8, 36): [25 11]\n", 490 | "(8, 37): [25 10]\n", 491 | "(9, 3): [22 14]\n", 492 | "(9, 4): [21 13]\n", 493 | "(9, 5): [21 13]\n", 494 | "(9, 6): [15 12]\n", 495 | "(9, 7): [12 11]\n", 496 | "(9, 8): [11 12]\n", 497 | "(9, 9): [11 12]\n", 498 | "(9, 10): [ 9 12]\n", 499 | "(9, 11): [10 14]\n", 500 | "(9, 12): [ 9 14]\n", 501 | "(9, 13): [ 9 15]\n", 502 | "(9, 14): [17 10]\n", 503 | "(9, 15): [17 11]\n", 504 | "(9, 16): [17 12]\n", 505 | "(9, 17): [17 13]\n", 506 | "(9, 18): [17 14]\n", 507 | "(9, 19): [17 14]\n", 508 | "(9, 20): [17 14]\n", 509 | "(9, 21): [17 15]\n", 510 | "(9, 22): [19 28]\n", 511 | "(9, 23): [20 27]\n", 512 | "(9, 24): [20 27]\n", 513 | "(9, 25): [21 27]\n", 514 | "(9, 26): [21 27]\n", 515 | "(9, 27): [21 27]\n", 516 | "(9, 28): [22 26]\n", 517 | "(9, 29): [22 26]\n", 518 | "(9, 30): [23 26]\n", 519 | "(9, 31): [23 25]\n", 520 | "(9, 32): [24 15]\n", 521 | "(9, 33): [25 14]\n", 522 | "(9, 34): [25 13]\n", 523 | "(9, 35): [25 12]\n", 524 | "(9, 36): [25 11]\n", 525 | "(9, 37): [26 10]\n", 526 | "(10, 3): [21 14]\n", 527 | "(10, 4): [21 14]\n", 528 | "(10, 5): [13 12]\n", 529 | "(10, 6): [13 12]\n", 530 | "(10, 7): [12 12]\n", 531 | "(10, 8): [12 13]\n", 532 | "(10, 9): [11 13]\n", 533 | "(10, 10): [11 14]\n", 534 | "(10, 11): [10 15]\n", 535 | "(10, 12): [10 15]\n", 536 | "(10, 13): [ 9 16]\n", 537 | "(10, 14): [ 9 16]\n", 538 | "(10, 15): [17 11]\n", 539 | "(10, 16): [17 12]\n", 540 | "(10, 17): [17 12]\n", 541 | "(10, 18): [17 13]\n", 542 | "(10, 19): [17 13]\n", 543 | "(10, 20): [20 20]\n", 544 | "(10, 21): [20 20]\n", 545 | "(10, 22): [19 28]\n", 546 | "(10, 23): [20 28]\n", 547 | "(10, 24): [20 28]\n", 548 | "(10, 25): [21 28]\n", 549 | "(10, 26): [21 27]\n", 550 | "(10, 27): [22 27]\n", 551 | "(10, 28): [22 27]\n", 552 | "(10, 29): [22 26]\n", 553 | "(10, 30): [23 26]\n", 554 | "(10, 31): [23 26]\n", 555 | "(10, 32): [24 25]\n", 556 | "(10, 33): [25 14]\n", 557 | "(10, 34): [25 13]\n", 558 | "(10, 35): [26 12]\n", 559 | "(10, 36): [26 11]\n", 560 | "(10, 37): [26 10]\n", 561 | "(11, 3): [15 11]\n", 562 | "(11, 4): [14 12]\n", 563 | "(11, 5): [14 12]\n", 564 | "(11, 6): [14 13]\n", 565 | "(11, 7): [13 13]\n", 566 | "(11, 8): [12 14]\n", 567 | "(11, 9): [12 14]\n", 568 | "(11, 10): [11 15]\n", 569 | "(11, 11): [11 15]\n", 570 | "(11, 12): [11 16]\n", 571 | "(11, 13): [10 16]\n", 572 | "(11, 14): [ 9 17]\n", 573 | "(11, 15): [17 10]\n", 574 | "(11, 16): [ 8 19]\n", 575 | "(11, 17): [ 7 20]\n", 576 | "(11, 18): [ 4 22]\n", 577 | "(11, 19): [ 5 22]\n", 578 | "(11, 20): [20 20]\n", 579 | "(11, 21): [20 20]\n", 580 | "(11, 22): [19 29]\n", 581 | "(11, 23): [20 29]\n", 582 | "(11, 24): [21 28]\n", 583 | "(11, 25): [21 28]\n", 584 | "(11, 26): [21 28]\n", 585 | "(11, 27): [22 28]\n", 586 | "(11, 28): [22 27]\n", 587 | "(11, 29): [23 27]\n", 588 | "(11, 30): [23 26]\n", 589 | "(11, 31): [23 26]\n", 590 | "(11, 32): [24 26]\n", 591 | "(11, 33): [25 14]\n", 592 | "(11, 34): [26 13]\n", 593 | "(11, 35): [26 12]\n", 594 | "(11, 36): [26 11]\n", 595 | "(11, 37): [27 10]\n", 596 | "(12, 3): [15 12]\n", 597 | "(12, 4): [15 13]\n", 598 | "(12, 5): [14 13]\n", 599 | "(12, 6): [14 14]\n", 600 | "(12, 7): [13 14]\n", 601 | "(12, 8): [13 14]\n", 602 | "(12, 9): [12 15]\n", 603 | "(12, 10): [12 15]\n", 604 | "(12, 11): [12 16]\n", 605 | "(12, 12): [11 16]\n", 606 | "(12, 13): [11 17]\n", 607 | "(12, 14): [10 17]\n", 608 | "(12, 15): [10 18]\n", 609 | "(12, 16): [ 9 19]\n", 610 | "(12, 17): [ 8 20]\n", 611 | "(12, 18): [ 5 22]\n", 612 | "(12, 19): [ 5 23]\n", 613 | "(12, 20): [ 4 25]\n", 614 | "(12, 21): [ 4 26]\n", 615 | "(12, 22): [ 4 27]\n", 616 | "(12, 23): [20 29]\n", 617 | "(12, 24): [20 29]\n", 618 | "(12, 25): [21 29]\n", 619 | "(12, 26): [21 29]\n", 620 | "(12, 27): [22 28]\n", 621 | "(12, 28): [23 28]\n", 622 | "(12, 29): [23 27]\n", 623 | "(12, 30): [23 27]\n", 624 | "(12, 31): [24 26]\n", 625 | "(12, 32): [24 26]\n", 626 | "(12, 33): [25 20]\n", 627 | "(12, 34): [26 13]\n", 628 | "(12, 35): [27 12]\n", 629 | "(12, 36): [27 11]\n", 630 | "(12, 37): [27 10]\n", 631 | "(13, 3): [15 13]\n", 632 | "(13, 4): [15 13]\n", 633 | "(13, 5): [15 14]\n", 634 | "(13, 6): [14 14]\n", 635 | "(13, 7): [14 15]\n", 636 | "(13, 8): [13 15]\n", 637 | "(13, 9): [13 15]\n", 638 | "(13, 10): [13 16]\n", 639 | "(13, 11): [12 16]\n", 640 | "(13, 12): [12 17]\n", 641 | "(13, 13): [12 17]\n", 642 | "(13, 14): [11 18]\n", 643 | "(13, 15): [11 18]\n", 644 | "(13, 16): [10 19]\n", 645 | "(13, 17): [10 20]\n", 646 | "(13, 18): [10 21]\n", 647 | "(13, 19): [ 7 23]\n", 648 | "(13, 20): [ 6 24]\n", 649 | "(13, 21): [ 6 25]\n", 650 | "(13, 22): [ 6 27]\n", 651 | "(13, 23): [20 30]\n", 652 | "(13, 24): [20 30]\n", 653 | "(13, 25): [21 30]\n", 654 | "(13, 26): [22 29]\n", 655 | "(13, 27): [22 29]\n", 656 | "(13, 28): [23 28]\n", 657 | "(13, 29): [23 28]\n", 658 | "(13, 30): [24 27]\n", 659 | "(13, 31): [24 27]\n", 660 | "(13, 32): [24 26]\n", 661 | "(13, 33): [25 26]\n", 662 | "(13, 34): [25 26]\n", 663 | "(13, 35): [25 23]\n", 664 | "(13, 36): [27 11]\n", 665 | "(13, 37): [28 10]\n", 666 | "(14, 3): [16 14]\n", 667 | "(14, 4): [15 14]\n", 668 | "(14, 5): [15 15]\n", 669 | "(14, 6): [15 15]\n", 670 | "(14, 7): [14 15]\n", 671 | "(14, 8): [14 16]\n", 672 | "(14, 9): [14 16]\n", 673 | "(14, 10): [13 16]\n", 674 | "(14, 11): [13 17]\n", 675 | "(14, 12): [12 17]\n", 676 | "(14, 13): [12 18]\n", 677 | "(14, 14): [12 18]\n", 678 | "(14, 15): [11 19]\n", 679 | "(14, 16): [11 20]\n", 680 | "(14, 17): [11 20]\n", 681 | "(14, 18): [ 9 21]\n", 682 | "(14, 19): [ 9 22]\n", 683 | "(14, 20): [ 6 25]\n", 684 | "(14, 21): [ 5 26]\n", 685 | "(14, 22): [ 5 27]\n", 686 | "(14, 23): [ 7 27]\n", 687 | "(14, 24): [ 6 28]\n", 688 | "(14, 25): [21 31]\n", 689 | "(14, 26): [22 30]\n", 690 | "(14, 27): [23 30]\n", 691 | "(14, 28): [23 29]\n", 692 | "(14, 29): [24 29]\n", 693 | "(14, 30): [24 28]\n", 694 | "(14, 31): [24 28]\n", 695 | "(14, 32): [25 27]\n", 696 | "(14, 33): [25 26]\n", 697 | "(14, 34): [25 26]\n", 698 | "(14, 35): [26 26]\n", 699 | "(14, 36): [26 25]\n", 700 | "(14, 37): [28 10]\n", 701 | "(15, 3): [16 15]\n", 702 | "(15, 4): [15 15]\n", 703 | "(15, 5): [15 15]\n", 704 | "(15, 6): [15 15]\n", 705 | "(15, 7): [15 16]\n", 706 | "(15, 8): [14 16]\n", 707 | "(15, 9): [14 16]\n", 708 | "(15, 10): [14 17]\n", 709 | "(15, 11): [13 17]\n", 710 | "(15, 12): [13 18]\n", 711 | "(15, 13): [13 18]\n", 712 | "(15, 14): [12 19]\n", 713 | "(15, 15): [12 19]\n", 714 | "(15, 16): [12 20]\n", 715 | "(15, 17): [11 21]\n", 716 | "(15, 18): [11 21]\n", 717 | "(15, 19): [10 22]\n", 718 | "(15, 20): [ 9 24]\n", 719 | "(15, 21): [10 24]\n", 720 | "(15, 22): [ 8 26]\n", 721 | "(15, 23): [10 26]\n", 722 | "(15, 24): [ 8 28]\n", 723 | "(15, 25): [22 31]\n", 724 | "(15, 26): [22 31]\n", 725 | "(15, 27): [23 30]\n", 726 | "(15, 28): [24 30]\n", 727 | "(15, 29): [24 29]\n", 728 | "(15, 30): [25 29]\n", 729 | "(15, 31): [25 28]\n", 730 | "(15, 32): [25 27]\n", 731 | "(15, 33): [26 27]\n", 732 | "(15, 34): [26 26]\n", 733 | "(15, 35): [26 26]\n", 734 | "(15, 36): [27 25]\n", 735 | "(15, 37): [28 18]\n", 736 | "(16, 3): [16 15]\n", 737 | "(16, 4): [16 15]\n", 738 | "(16, 5): [16 15]\n", 739 | "(16, 6): [15 16]\n", 740 | "(16, 7): [15 16]\n", 741 | "(16, 8): [15 16]\n", 742 | "(16, 9): [14 17]\n", 743 | "(16, 10): [14 17]\n", 744 | "(16, 11): [14 18]\n", 745 | "(16, 12): [13 18]\n", 746 | "(16, 13): [13 19]\n", 747 | "(16, 14): [13 19]\n", 748 | "(16, 15): [13 19]\n", 749 | "(16, 16): [12 20]\n", 750 | "(16, 17): [12 21]\n", 751 | "(16, 18): [12 21]\n", 752 | "(16, 19): [12 22]\n", 753 | "(16, 20): [12 23]\n", 754 | "(16, 21): [11 24]\n", 755 | "(16, 22): [12 24]\n", 756 | "(16, 23): [11 25]\n", 757 | "(16, 24): [ 9 28]\n", 758 | "(16, 25): [22 32]\n", 759 | "(16, 26): [23 32]\n", 760 | "(16, 27): [24 31]\n", 761 | "(16, 28): [24 30]\n", 762 | "(16, 29): [25 30]\n", 763 | "(16, 30): [25 29]\n", 764 | "(16, 31): [26 29]\n", 765 | "(16, 32): [26 28]\n", 766 | "(16, 33): [26 27]\n", 767 | "(16, 34): [27 27]\n", 768 | "(16, 35): [27 24]\n", 769 | "(16, 36): [28 26]\n", 770 | "(16, 37): [28 26]\n", 771 | "(17, 3): [16 15]\n", 772 | "(17, 4): [16 15]\n", 773 | "(17, 5): [16 16]\n", 774 | "(17, 6): [15 16]\n", 775 | "(17, 7): [15 16]\n", 776 | "(17, 8): [15 17]\n", 777 | "(17, 9): [15 17]\n", 778 | "(17, 10): [15 17]\n", 779 | "(17, 11): [14 18]\n", 780 | "(17, 12): [14 18]\n", 781 | "(17, 13): [14 19]\n", 782 | "(17, 14): [13 19]\n", 783 | "(17, 15): [13 20]\n", 784 | "(17, 16): [13 20]\n", 785 | "(17, 17): [13 21]\n", 786 | "(17, 18): [13 21]\n", 787 | "(17, 19): [13 22]\n", 788 | "(17, 20): [13 23]\n", 789 | "(17, 21): [12 23]\n", 790 | "(17, 22): [12 24]\n", 791 | "(17, 23): [11 26]\n", 792 | "(17, 24): [ 9 28]\n", 793 | "(17, 25): [12 27]\n", 794 | "(17, 26): [23 33]\n", 795 | "(17, 27): [24 32]\n", 796 | "(17, 28): [25 31]\n", 797 | "(17, 29): [26 31]\n", 798 | "(17, 30): [26 30]\n", 799 | "(17, 31): [27 29]\n", 800 | "(17, 32): [27 28]\n", 801 | "(17, 33): [27 28]\n", 802 | "(17, 34): [28 28]\n", 803 | "(17, 35): [27 27]\n", 804 | "(17, 36): [28 26]\n", 805 | "(17, 37): [29 26]\n", 806 | "(18, 3): [16 15]\n", 807 | "(18, 4): [16 16]\n", 808 | "(18, 5): [16 16]\n", 809 | "(18, 6): [16 16]\n", 810 | "(18, 7): [16 17]\n", 811 | "(18, 8): [15 17]\n", 812 | "(18, 9): [15 17]\n", 813 | "(18, 10): [15 18]\n", 814 | "(18, 11): [15 18]\n", 815 | "(18, 12): [14 19]\n", 816 | "(18, 13): [14 19]\n", 817 | "(18, 14): [14 19]\n", 818 | "(18, 15): [14 20]\n", 819 | "(18, 16): [14 20]\n", 820 | "(18, 17): [13 21]\n", 821 | "(18, 18): [13 21]\n", 822 | "(18, 19): [13 22]\n", 823 | "(18, 20): [13 23]\n", 824 | "(18, 21): [13 23]\n", 825 | "(18, 22): [13 24]\n", 826 | "(18, 23): [13 25]\n", 827 | "(18, 24): [13 26]\n", 828 | "(18, 25): [13 27]\n", 829 | "(18, 26): [24 34]\n", 830 | "(18, 27): [25 33]\n", 831 | "(18, 28): [26 32]\n", 832 | "(18, 29): [26 32]\n", 833 | "(18, 30): [27 30]\n", 834 | "(18, 31): [28 30]\n", 835 | "(18, 32): [29 30]\n", 836 | "(18, 33): [30 30]\n", 837 | "(18, 34): [32 30]\n", 838 | "(18, 35): [29 27]\n", 839 | "(18, 36): [31 28]\n", 840 | "(18, 37): [32 28]\n", 841 | "(19, 3): [17 16]\n", 842 | "(19, 4): [17 16]\n", 843 | "(19, 5): [16 16]\n", 844 | "(19, 6): [16 17]\n", 845 | "(19, 7): [16 17]\n", 846 | "(19, 8): [16 17]\n", 847 | "(19, 9): [16 18]\n", 848 | "(19, 10): [15 18]\n", 849 | "(19, 11): [15 18]\n", 850 | "(19, 12): [15 19]\n", 851 | "(19, 13): [15 19]\n", 852 | "(19, 14): [15 20]\n", 853 | "(19, 15): [14 20]\n", 854 | "(19, 16): [14 21]\n", 855 | "(19, 17): [14 21]\n", 856 | "(19, 18): [14 22]\n", 857 | "(19, 19): [14 22]\n", 858 | "(19, 20): [14 23]\n", 859 | "(19, 21): [14 23]\n", 860 | "(19, 22): [14 24]\n", 861 | "(19, 23): [14 25]\n", 862 | "(19, 24): [14 26]\n", 863 | "(19, 25): [14 26]\n", 864 | "(19, 26): [18 31]\n", 865 | "(19, 27): [27 34]\n", 866 | "(19, 28): [29 34]\n", 867 | "(19, 29): [30 34]\n", 868 | "(19, 30): [31 33]\n", 869 | "(19, 31): [33 25]\n", 870 | "(19, 32): [31 31]\n", 871 | "(19, 33): [32 31]\n", 872 | "(19, 34): [33 30]\n", 873 | "(19, 35): [33 30]\n", 874 | "(19, 36): [32 28]\n", 875 | "(19, 37): [32 28]\n", 876 | "(20, 3): [17 16]\n", 877 | "(20, 4): [17 16]\n", 878 | "(20, 5): [20 20]\n", 879 | "(20, 6): [20 20]\n", 880 | "(20, 7): [16 17]\n", 881 | "(20, 8): [16 18]\n", 882 | "(20, 9): [16 18]\n", 883 | "(20, 10): [16 18]\n", 884 | "(20, 11): [16 19]\n", 885 | "(20, 12): [20 20]\n", 886 | "(20, 13): [20 20]\n", 887 | "(20, 14): [15 20]\n", 888 | "(20, 15): [15 20]\n", 889 | "(20, 16): [15 21]\n", 890 | "(20, 17): [15 21]\n", 891 | "(20, 18): [14 22]\n", 892 | "(20, 19): [14 22]\n", 893 | "(20, 20): [14 23]\n", 894 | "(20, 21): [14 23]\n", 895 | "(20, 22): [14 24]\n", 896 | "(20, 23): [14 25]\n", 897 | "(20, 24): [14 26]\n", 898 | "(20, 25): [14 26]\n", 899 | "(20, 26): [14 27]\n", 900 | "(20, 27): [20 20]\n", 901 | "(20, 28): [20 20]\n", 902 | "(20, 29): [32 25]\n", 903 | "(20, 30): [32 25]\n", 904 | "(20, 31): [33 25]\n", 905 | "(20, 32): [36 28]\n", 906 | "(20, 33): [34 25]\n", 907 | "(20, 34): [33 30]\n", 908 | "(20, 35): [20 20]\n", 909 | "(20, 36): [20 20]\n", 910 | "(20, 37): [33 28]\n", 911 | "(21, 3): [17 16]\n", 912 | "(21, 4): [17 16]\n", 913 | "(21, 5): [20 20]\n", 914 | "(21, 6): [20 20]\n", 915 | "(21, 7): [17 17]\n", 916 | "(21, 8): [16 18]\n", 917 | "(21, 9): [16 18]\n", 918 | "(21, 10): [16 18]\n", 919 | "(21, 11): [16 19]\n", 920 | "(21, 12): [20 20]\n", 921 | "(21, 13): [20 20]\n", 922 | "(21, 14): [15 20]\n", 923 | "(21, 15): [15 20]\n", 924 | "(21, 16): [15 21]\n", 925 | "(21, 17): [15 21]\n", 926 | "(21, 18): [15 22]\n", 927 | "(21, 19): [15 22]\n", 928 | "(21, 20): [15 23]\n", 929 | "(21, 21): [15 24]\n", 930 | "(21, 22): [15 24]\n", 931 | "(21, 23): [15 25]\n", 932 | "(21, 24): [15 26]\n", 933 | "(21, 25): [15 26]\n", 934 | "(21, 26): [16 26]\n", 935 | "(21, 27): [20 20]\n", 936 | "(21, 28): [20 20]\n", 937 | "(21, 29): [28 23]\n", 938 | "(21, 30): [29 23]\n", 939 | "(21, 31): [30 23]\n", 940 | "(21, 32): [32 24]\n", 941 | "(21, 33): [32 23]\n", 942 | "(21, 34): [33 23]\n", 943 | "(21, 35): [20 20]\n", 944 | "(21, 36): [20 20]\n", 945 | "(21, 37): [34 22]\n", 946 | "(22, 3): [17 16]\n", 947 | "(22, 4): [17 16]\n", 948 | "(22, 5): [17 17]\n", 949 | "(22, 6): [17 17]\n", 950 | "(22, 7): [17 18]\n", 951 | "(22, 8): [17 18]\n", 952 | "(22, 9): [17 18]\n", 953 | "(22, 10): [16 19]\n", 954 | "(22, 11): [16 19]\n", 955 | "(22, 12): [16 19]\n", 956 | "(22, 13): [16 20]\n", 957 | "(22, 14): [16 20]\n", 958 | "(22, 15): [16 20]\n", 959 | "(22, 16): [16 21]\n", 960 | "(22, 17): [15 21]\n", 961 | "(22, 18): [15 22]\n", 962 | "(22, 19): [15 22]\n", 963 | "(22, 20): [15 23]\n", 964 | "(22, 21): [15 24]\n", 965 | "(22, 22): [15 24]\n", 966 | "(22, 23): [15 25]\n", 967 | "(22, 24): [16 25]\n", 968 | "(22, 25): [15 26]\n", 969 | "(22, 26): [16 27]\n", 970 | "(22, 27): [27 23]\n", 971 | "(22, 28): [27 23]\n", 972 | "(22, 29): [28 23]\n", 973 | "(22, 30): [28 22]\n", 974 | "(22, 31): [29 22]\n", 975 | "(22, 32): [31 22]\n", 976 | "(22, 33): [31 22]\n", 977 | "(22, 34): [33 22]\n", 978 | "(22, 35): [34 22]\n", 979 | "(22, 36): [35 22]\n", 980 | "(22, 37): [36 22]\n", 981 | "(23, 3): [18 16]\n", 982 | "(23, 4): [18 17]\n", 983 | "(23, 5): [17 17]\n", 984 | "(23, 6): [17 17]\n", 985 | "(23, 7): [17 18]\n", 986 | "(23, 8): [17 18]\n", 987 | "(23, 9): [17 18]\n", 988 | "(23, 10): [17 19]\n", 989 | "(23, 11): [17 19]\n", 990 | "(23, 12): [17 19]\n", 991 | "(23, 13): [16 20]\n", 992 | "(23, 14): [16 20]\n", 993 | "(23, 15): [16 21]\n", 994 | "(23, 16): [16 21]\n", 995 | "(23, 17): [16 21]\n", 996 | "(23, 18): [16 22]\n", 997 | "(23, 19): [16 22]\n", 998 | "(23, 20): [16 23]\n", 999 | "(23, 21): [16 24]\n", 1000 | "(23, 22): [16 24]\n", 1001 | "(23, 23): [16 25]\n", 1002 | "(23, 24): [18 24]\n", 1003 | "(23, 25): [16 26]\n", 1004 | "(23, 26): [26 23]\n", 1005 | "(23, 27): [27 23]\n", 1006 | "(23, 28): [27 22]\n", 1007 | "(23, 29): [27 22]\n", 1008 | "(23, 30): [28 22]\n", 1009 | "(23, 31): [28 22]\n", 1010 | "(23, 32): [29 21]\n", 1011 | "(23, 33): [30 21]\n", 1012 | "(23, 34): [30 21]\n", 1013 | "(23, 35): [31 21]\n", 1014 | "(23, 36): [31 20]\n", 1015 | "(23, 37): [33 20]\n", 1016 | "(24, 3): [18 16]\n", 1017 | "(24, 4): [18 17]\n", 1018 | "(24, 5): [18 17]\n", 1019 | "(24, 6): [18 17]\n", 1020 | "(24, 7): [18 18]\n", 1021 | "(24, 8): [17 18]\n", 1022 | "(24, 9): [17 19]\n", 1023 | "(24, 10): [17 19]\n", 1024 | "(24, 11): [17 19]\n", 1025 | "(24, 12): [17 20]\n", 1026 | "(24, 13): [17 20]\n", 1027 | "(24, 14): [17 20]\n", 1028 | "(24, 15): [16 21]\n", 1029 | "(24, 16): [16 21]\n", 1030 | "(24, 17): [16 22]\n", 1031 | "(24, 18): [16 22]\n", 1032 | "(24, 19): [16 23]\n", 1033 | "(24, 20): [16 23]\n", 1034 | "(24, 21): [16 24]\n", 1035 | "(24, 22): [16 24]\n", 1036 | "(24, 23): [16 25]\n", 1037 | "(24, 24): [25 23]\n", 1038 | "(24, 25): [25 23]\n", 1039 | "(24, 26): [26 23]\n", 1040 | "(24, 27): [26 22]\n", 1041 | "(24, 28): [27 22]\n", 1042 | "(24, 29): [27 22]\n", 1043 | "(24, 30): [27 22]\n", 1044 | "(24, 31): [28 21]\n", 1045 | "(24, 32): [29 21]\n", 1046 | "(24, 33): [29 21]\n", 1047 | "(24, 34): [30 20]\n", 1048 | "(24, 35): [30 20]\n", 1049 | "(24, 36): [31 19]\n", 1050 | "(24, 37): [34 20]\n", 1051 | "(25, 3): [18 17]\n", 1052 | "(25, 4): [18 17]\n", 1053 | "(25, 5): [18 17]\n", 1054 | "(25, 6): [18 18]\n", 1055 | "(25, 7): [18 18]\n", 1056 | "(25, 8): [18 18]\n", 1057 | "(25, 9): [18 19]\n", 1058 | "(25, 10): [17 19]\n", 1059 | "(25, 11): [17 19]\n", 1060 | "(25, 12): [17 20]\n", 1061 | "(25, 13): [17 20]\n", 1062 | "(25, 14): [17 20]\n", 1063 | "(25, 15): [17 21]\n", 1064 | "(25, 16): [17 21]\n", 1065 | "(25, 17): [17 22]\n", 1066 | "(25, 18): [17 22]\n", 1067 | "(25, 19): [17 23]\n", 1068 | "(25, 20): [17 23]\n", 1069 | "(25, 21): [17 24]\n", 1070 | "(25, 22): [24 23]\n", 1071 | "(25, 23): [24 23]\n", 1072 | "(25, 24): [25 23]\n", 1073 | "(25, 25): [25 22]\n", 1074 | "(25, 26): [25 22]\n", 1075 | "(25, 27): [26 22]\n", 1076 | "(25, 28): [26 22]\n", 1077 | "(25, 29): [27 21]\n", 1078 | "(25, 30): [27 21]\n", 1079 | "(25, 31): [28 21]\n", 1080 | "(25, 32): [28 20]\n", 1081 | "(25, 33): [29 20]\n", 1082 | "(25, 34): [29 20]\n", 1083 | "(25, 35): [30 19]\n", 1084 | "(25, 36): [30 19]\n", 1085 | "(25, 37): [31 19]\n", 1086 | "(26, 3): [19 17]\n", 1087 | "(26, 4): [19 17]\n", 1088 | "(26, 5): [18 17]\n", 1089 | "(26, 6): [18 18]\n", 1090 | "(26, 7): [18 18]\n", 1091 | "(26, 8): [18 18]\n", 1092 | "(26, 9): [18 19]\n", 1093 | "(26, 10): [18 19]\n", 1094 | "(26, 11): [18 19]\n", 1095 | "(26, 12): [17 20]\n", 1096 | "(26, 13): [17 20]\n", 1097 | "(26, 14): [17 21]\n", 1098 | "(26, 15): [17 21]\n", 1099 | "(26, 16): [17 21]\n", 1100 | "(26, 17): [17 22]\n", 1101 | "(26, 18): [17 22]\n", 1102 | "(26, 19): [17 23]\n", 1103 | "(26, 20): [17 23]\n", 1104 | "(26, 21): [17 24]\n", 1105 | "(26, 22): [24 23]\n", 1106 | "(26, 23): [24 23]\n", 1107 | "(26, 24): [24 22]\n", 1108 | "(26, 25): [25 22]\n", 1109 | "(26, 26): [25 22]\n", 1110 | "(26, 27): [25 21]\n", 1111 | "(26, 28): [26 21]\n", 1112 | "(26, 29): [26 21]\n", 1113 | "(26, 30): [27 21]\n", 1114 | "(26, 31): [27 20]\n", 1115 | "(26, 32): [28 20]\n", 1116 | "(26, 33): [28 20]\n", 1117 | "(26, 34): [29 19]\n", 1118 | "(26, 35): [29 19]\n", 1119 | "(26, 36): [30 19]\n", 1120 | "(26, 37): [31 18]\n", 1121 | "(27, 3): [19 17]\n", 1122 | "(27, 4): [19 17]\n", 1123 | "(27, 5): [19 17]\n", 1124 | "(27, 6): [19 18]\n", 1125 | "(27, 7): [18 18]\n", 1126 | "(27, 8): [18 19]\n", 1127 | "(27, 9): [18 19]\n", 1128 | "(27, 10): [18 19]\n", 1129 | "(27, 11): [18 20]\n", 1130 | "(27, 12): [18 20]\n", 1131 | "(27, 13): [18 20]\n", 1132 | "(27, 14): [18 21]\n", 1133 | "(27, 15): [18 21]\n", 1134 | "(27, 16): [17 21]\n", 1135 | "(27, 17): [17 22]\n", 1136 | "(27, 18): [17 22]\n", 1137 | "(27, 19): [17 23]\n", 1138 | "(27, 20): [17 23]\n", 1139 | "(27, 21): [17 24]\n", 1140 | "(27, 22): [17 24]\n", 1141 | "(27, 23): [24 22]\n", 1142 | "(27, 24): [24 22]\n", 1143 | "(27, 25): [24 22]\n", 1144 | "(27, 26): [25 21]\n", 1145 | "(27, 27): [25 21]\n", 1146 | "(27, 28): [25 21]\n", 1147 | "(27, 29): [26 21]\n", 1148 | "(27, 30): [26 20]\n", 1149 | "(27, 31): [27 20]\n", 1150 | "(27, 32): [27 20]\n", 1151 | "(27, 33): [28 19]\n", 1152 | "(27, 34): [28 19]\n", 1153 | "(27, 35): [29 18]\n", 1154 | "(27, 36): [30 18]\n", 1155 | "(27, 37): [31 18]\n", 1156 | "(28, 3): [19 17]\n", 1157 | "(28, 4): [19 17]\n", 1158 | "(28, 5): [19 18]\n", 1159 | "(28, 6): [19 18]\n", 1160 | "(28, 7): [19 18]\n", 1161 | "(28, 8): [19 19]\n", 1162 | "(28, 9): [18 19]\n", 1163 | "(28, 10): [18 19]\n", 1164 | "(28, 11): [18 20]\n", 1165 | "(28, 12): [18 20]\n", 1166 | "(28, 13): [18 20]\n", 1167 | "(28, 14): [18 21]\n", 1168 | "(28, 15): [18 21]\n", 1169 | "(28, 16): [18 21]\n", 1170 | "(28, 17): [18 22]\n", 1171 | "(28, 18): [18 22]\n", 1172 | "(28, 19): [18 23]\n", 1173 | "(28, 20): [18 23]\n", 1174 | "(28, 21): [18 24]\n", 1175 | "(28, 22): [23 22]\n", 1176 | "(28, 23): [23 22]\n", 1177 | "(28, 24): [24 22]\n", 1178 | "(28, 25): [24 21]\n", 1179 | "(28, 26): [24 21]\n", 1180 | "(28, 27): [25 21]\n", 1181 | "(28, 28): [25 21]\n", 1182 | "(28, 29): [25 20]\n", 1183 | "(28, 30): [26 20]\n", 1184 | "(28, 31): [26 20]\n", 1185 | "(28, 32): [27 19]\n", 1186 | "(28, 33): [27 19]\n", 1187 | "(28, 34): [28 18]\n", 1188 | "(28, 35): [28 18]\n", 1189 | "(28, 36): [29 18]\n", 1190 | "(28, 37): [30 17]\n", 1191 | "(29, 3): [20 17]\n", 1192 | "(29, 4): [19 17]\n", 1193 | "(29, 5): [19 18]\n", 1194 | "(29, 6): [19 18]\n", 1195 | "(29, 7): [19 18]\n", 1196 | "(29, 8): [19 19]\n", 1197 | "(29, 9): [19 19]\n", 1198 | "(29, 10): [19 19]\n", 1199 | "(29, 11): [19 20]\n", 1200 | "(29, 12): [18 20]\n", 1201 | "(29, 13): [18 20]\n", 1202 | "(29, 14): [18 21]\n", 1203 | "(29, 15): [18 21]\n", 1204 | "(29, 16): [18 22]\n", 1205 | "(29, 17): [18 22]\n", 1206 | "(29, 18): [18 22]\n", 1207 | "(29, 19): [18 23]\n", 1208 | "(29, 20): [18 23]\n", 1209 | "(29, 21): [18 24]\n", 1210 | "(29, 22): [23 22]\n", 1211 | "(29, 23): [23 22]\n", 1212 | "(29, 24): [23 21]\n", 1213 | "(29, 25): [24 21]\n", 1214 | "(29, 26): [24 21]\n", 1215 | "(29, 27): [24 21]\n", 1216 | "(29, 28): [25 20]\n", 1217 | "(29, 29): [25 20]\n", 1218 | "(29, 30): [26 20]\n", 1219 | "(29, 31): [26 19]\n", 1220 | "(29, 32): [26 19]\n", 1221 | "(29, 33): [27 19]\n", 1222 | "(29, 34): [28 18]\n", 1223 | "(29, 35): [28 18]\n", 1224 | "(29, 36): [29 17]\n", 1225 | "(29, 37): [30 17]\n", 1226 | "(30, 3): [20 17]\n", 1227 | "(30, 4): [20 17]\n", 1228 | "(30, 5): [20 18]\n", 1229 | "(30, 6): [20 18]\n", 1230 | "(30, 7): [19 18]\n", 1231 | "(30, 8): [19 19]\n", 1232 | "(30, 9): [19 19]\n", 1233 | "(30, 10): [19 20]\n", 1234 | "(30, 11): [19 20]\n", 1235 | "(30, 12): [19 20]\n", 1236 | "(30, 13): [19 21]\n", 1237 | "(30, 14): [19 21]\n", 1238 | "(30, 15): [18 21]\n", 1239 | "(30, 16): [18 22]\n", 1240 | "(30, 17): [18 22]\n", 1241 | "(30, 18): [18 22]\n", 1242 | "(30, 19): [18 23]\n", 1243 | "(30, 20): [20 20]\n", 1244 | "(30, 21): [20 20]\n", 1245 | "(30, 22): [23 22]\n", 1246 | "(30, 23): [23 22]\n", 1247 | "(30, 24): [23 21]\n", 1248 | "(30, 25): [23 21]\n", 1249 | "(30, 26): [24 21]\n", 1250 | "(30, 27): [24 20]\n", 1251 | "(30, 28): [24 20]\n", 1252 | "(30, 29): [25 20]\n", 1253 | "(30, 30): [25 19]\n", 1254 | "(30, 31): [26 19]\n", 1255 | "(30, 32): [26 19]\n", 1256 | "(30, 33): [27 18]\n", 1257 | "(30, 34): [27 18]\n", 1258 | "(30, 35): [28 17]\n", 1259 | "(30, 36): [28 17]\n", 1260 | "(30, 37): [29 16]\n", 1261 | "(31, 3): [20 17]\n", 1262 | "(31, 4): [20 17]\n", 1263 | "(31, 5): [20 18]\n", 1264 | "(31, 6): [20 18]\n", 1265 | "(31, 7): [20 18]\n", 1266 | "(31, 8): [20 19]\n", 1267 | "(31, 9): [19 19]\n", 1268 | "(31, 10): [19 20]\n", 1269 | "(31, 11): [19 20]\n", 1270 | "(31, 12): [19 20]\n", 1271 | "(31, 13): [19 21]\n", 1272 | "(31, 14): [19 21]\n", 1273 | "(31, 15): [19 21]\n", 1274 | "(31, 16): [19 22]\n", 1275 | "(31, 17): [19 22]\n", 1276 | "(31, 18): [19 22]\n", 1277 | "(31, 19): [18 23]\n", 1278 | "(31, 20): [20 20]\n", 1279 | "(31, 21): [20 20]\n", 1280 | "(31, 22): [23 22]\n", 1281 | "(31, 23): [23 21]\n", 1282 | "(31, 24): [23 21]\n", 1283 | "(31, 25): [23 21]\n", 1284 | "(31, 26): [23 20]\n", 1285 | "(31, 27): [24 20]\n", 1286 | "(31, 28): [24 20]\n", 1287 | "(31, 29): [24 19]\n", 1288 | "(31, 30): [25 19]\n", 1289 | "(31, 31): [25 19]\n", 1290 | "(31, 32): [26 18]\n", 1291 | "(31, 33): [26 18]\n", 1292 | "(31, 34): [27 18]\n", 1293 | "(31, 35): [27 17]\n", 1294 | "(31, 36): [28 17]\n", 1295 | "(31, 37): [29 16]\n", 1296 | "(32, 3): [21 17]\n", 1297 | "(32, 4): [21 17]\n", 1298 | "(32, 5): [20 18]\n", 1299 | "(32, 6): [20 18]\n", 1300 | "(32, 7): [20 19]\n", 1301 | "(32, 8): [20 19]\n", 1302 | "(32, 9): [20 19]\n", 1303 | "(32, 10): [20 20]\n", 1304 | "(32, 11): [19 20]\n", 1305 | "(32, 12): [19 20]\n", 1306 | "(32, 13): [19 21]\n", 1307 | "(32, 14): [19 21]\n", 1308 | "(32, 15): [19 21]\n", 1309 | "(32, 16): [19 22]\n", 1310 | "(32, 17): [19 22]\n", 1311 | "(32, 18): [19 23]\n", 1312 | "(32, 19): [19 23]\n", 1313 | "(32, 20): [19 23]\n", 1314 | "(32, 21): [22 22]\n", 1315 | "(32, 22): [22 21]\n", 1316 | "(32, 23): [22 21]\n", 1317 | "(32, 24): [23 21]\n", 1318 | "(32, 25): [23 21]\n", 1319 | "(32, 26): [23 20]\n", 1320 | "(32, 27): [23 20]\n", 1321 | "(32, 28): [24 19]\n", 1322 | "(32, 29): [24 19]\n", 1323 | "(32, 30): [25 19]\n", 1324 | "(32, 31): [25 18]\n", 1325 | "(32, 32): [25 18]\n", 1326 | "(32, 33): [26 18]\n", 1327 | "(32, 34): [26 17]\n", 1328 | "(32, 35): [27 17]\n", 1329 | "(32, 36): [28 16]\n", 1330 | "(32, 37): [28 16]\n", 1331 | "(33, 3): [21 17]\n", 1332 | "(33, 4): [21 17]\n", 1333 | "(33, 5): [21 18]\n", 1334 | "(33, 6): [20 18]\n", 1335 | "(33, 7): [20 19]\n", 1336 | "(33, 8): [20 19]\n", 1337 | "(33, 9): [20 19]\n", 1338 | "(33, 10): [20 20]\n", 1339 | "(33, 11): [20 20]\n", 1340 | "(33, 12): [20 20]\n", 1341 | "(33, 13): [19 21]\n", 1342 | "(33, 14): [19 21]\n", 1343 | "(33, 15): [19 21]\n", 1344 | "(33, 16): [19 22]\n", 1345 | "(33, 17): [19 22]\n", 1346 | "(33, 18): [19 23]\n", 1347 | "(33, 19): [21 22]\n", 1348 | "(33, 20): [22 22]\n", 1349 | "(33, 21): [22 22]\n", 1350 | "(33, 22): [22 21]\n", 1351 | "(33, 23): [22 21]\n", 1352 | "(33, 24): [22 21]\n", 1353 | "(33, 25): [23 20]\n", 1354 | "(33, 26): [23 20]\n", 1355 | "(33, 27): [23 20]\n", 1356 | "(33, 28): [24 19]\n", 1357 | "(33, 29): [24 19]\n", 1358 | "(33, 30): [24 19]\n", 1359 | "(33, 31): [25 18]\n", 1360 | "(33, 32): [25 18]\n", 1361 | "(33, 33): [26 17]\n", 1362 | "(33, 34): [26 17]\n", 1363 | "(33, 35): [27 17]\n", 1364 | "(33, 36): [27 16]\n", 1365 | "(33, 37): [28 16]\n", 1366 | "(34, 3): [21 17]\n", 1367 | "(34, 4): [21 18]\n", 1368 | "(34, 5): [21 18]\n", 1369 | "(34, 6): [21 18]\n", 1370 | "(34, 7): [21 19]\n", 1371 | "(34, 8): [20 19]\n", 1372 | "(34, 9): [20 20]\n", 1373 | "(34, 10): [20 20]\n", 1374 | "(34, 11): [20 20]\n", 1375 | "(34, 12): [20 21]\n", 1376 | "(34, 13): [20 21]\n", 1377 | "(34, 14): [20 21]\n", 1378 | "(34, 15): [19 22]\n", 1379 | "(34, 16): [19 22]\n", 1380 | "(34, 17): [19 22]\n", 1381 | "(34, 18): [21 22]\n", 1382 | "(34, 19): [21 22]\n", 1383 | "(34, 20): [21 22]\n", 1384 | "(34, 21): [22 21]\n", 1385 | "(34, 22): [22 21]\n", 1386 | "(34, 23): [22 21]\n", 1387 | "(34, 24): [22 20]\n", 1388 | "(34, 25): [23 20]\n", 1389 | "(34, 26): [23 20]\n", 1390 | "(34, 27): [23 19]\n", 1391 | "(34, 28): [23 19]\n", 1392 | "(34, 29): [24 19]\n", 1393 | "(34, 30): [24 18]\n", 1394 | "(34, 31): [24 18]\n", 1395 | "(34, 32): [25 18]\n", 1396 | "(34, 33): [25 17]\n", 1397 | "(34, 34): [26 17]\n", 1398 | "(34, 35): [26 16]\n", 1399 | "(34, 36): [27 16]\n", 1400 | "(34, 37): [27 15]\n", 1401 | "(35, 3): [22 17]\n", 1402 | "(35, 4): [21 18]\n", 1403 | "(35, 5): [21 18]\n", 1404 | "(35, 6): [21 18]\n", 1405 | "(35, 7): [21 19]\n", 1406 | "(35, 8): [21 19]\n", 1407 | "(35, 9): [21 20]\n", 1408 | "(35, 10): [20 20]\n", 1409 | "(35, 11): [20 20]\n", 1410 | "(35, 12): [20 21]\n", 1411 | "(35, 13): [20 21]\n", 1412 | "(35, 14): [20 21]\n", 1413 | "(35, 15): [20 22]\n", 1414 | "(35, 16): [20 22]\n", 1415 | "(35, 17): [20 22]\n", 1416 | "(35, 18): [20 23]\n", 1417 | "(35, 19): [19 23]\n", 1418 | "(35, 20): [21 22]\n", 1419 | "(35, 21): [21 21]\n", 1420 | "(35, 22): [22 21]\n", 1421 | "(35, 23): [22 21]\n", 1422 | "(35, 24): [22 20]\n", 1423 | "(35, 25): [22 20]\n", 1424 | "(35, 26): [22 19]\n", 1425 | "(35, 27): [23 19]\n", 1426 | "(35, 28): [23 19]\n", 1427 | "(35, 29): [23 18]\n", 1428 | "(35, 30): [24 18]\n", 1429 | "(35, 31): [24 18]\n", 1430 | "(35, 32): [24 17]\n", 1431 | "(35, 33): [25 17]\n", 1432 | "(35, 34): [25 17]\n", 1433 | "(35, 35): [26 16]\n", 1434 | "(35, 36): [26 16]\n", 1435 | "(35, 37): [27 15]\n", 1436 | "(36, 3): [22 17]\n", 1437 | "(36, 4): [22 18]\n", 1438 | "(36, 5): [22 18]\n", 1439 | "(36, 6): [21 19]\n", 1440 | "(36, 7): [21 19]\n", 1441 | "(36, 8): [21 19]\n", 1442 | "(36, 9): [21 20]\n", 1443 | "(36, 10): [21 20]\n", 1444 | "(36, 11): [20 20]\n", 1445 | "(36, 12): [20 21]\n", 1446 | "(36, 13): [20 21]\n", 1447 | "(36, 14): [20 21]\n", 1448 | "(36, 15): [20 22]\n", 1449 | "(36, 16): [20 22]\n", 1450 | "(36, 17): [20 22]\n", 1451 | "(36, 18): [20 23]\n", 1452 | "(36, 19): [21 22]\n", 1453 | "(36, 20): [21 22]\n", 1454 | "(36, 21): [21 21]\n", 1455 | "(36, 22): [21 21]\n", 1456 | "(36, 23): [22 20]\n", 1457 | "(36, 24): [22 20]\n", 1458 | "(36, 25): [22 20]\n", 1459 | "(36, 26): [22 19]\n", 1460 | "(36, 27): [22 19]\n", 1461 | "(36, 28): [23 19]\n", 1462 | "(36, 29): [23 18]\n", 1463 | "(36, 30): [23 18]\n", 1464 | "(36, 31): [24 17]\n", 1465 | "(36, 32): [24 17]\n", 1466 | "(36, 33): [24 17]\n", 1467 | "(36, 34): [25 16]\n", 1468 | "(36, 35): [26 16]\n", 1469 | "(36, 36): [26 15]\n", 1470 | "(36, 37): [27 15]\n", 1471 | "(37, 3): [22 17]\n", 1472 | "(37, 4): [22 18]\n", 1473 | "(37, 5): [22 18]\n", 1474 | "(37, 6): [22 19]\n", 1475 | "(37, 7): [21 19]\n", 1476 | "(37, 8): [21 19]\n", 1477 | "(37, 9): [21 20]\n", 1478 | "(37, 10): [21 20]\n", 1479 | "(37, 11): [21 20]\n", 1480 | "(37, 12): [21 21]\n", 1481 | "(37, 13): [21 21]\n", 1482 | "(37, 14): [20 21]\n", 1483 | "(37, 15): [20 22]\n", 1484 | "(37, 16): [20 22]\n", 1485 | "(37, 17): [20 22]\n", 1486 | "(37, 18): [20 22]\n", 1487 | "(37, 19): [21 22]\n", 1488 | "(37, 20): [21 21]\n", 1489 | "(37, 21): [21 21]\n", 1490 | "(37, 22): [21 21]\n", 1491 | "(37, 23): [21 20]\n", 1492 | "(37, 24): [22 20]\n", 1493 | "(37, 25): [22 20]\n", 1494 | "(37, 26): [22 19]\n", 1495 | "(37, 27): [22 19]\n", 1496 | "(37, 28): [23 18]\n", 1497 | "(37, 29): [23 18]\n", 1498 | "(37, 30): [23 18]\n", 1499 | "(37, 31): [24 17]\n", 1500 | "(37, 32): [24 17]\n", 1501 | "(37, 33): [24 17]\n", 1502 | "(37, 34): [25 16]\n", 1503 | "(37, 35): [25 16]\n", 1504 | "(37, 36): [26 15]\n", 1505 | "(37, 37): [26 15]\n" 1506 | ] 1507 | } 1508 | ], 1509 | "source": [ 1510 | "z_min = np.min(all_z, axis = 0)\n", 1511 | "z_max = np.max(all_z, axis = 0)\n", 1512 | "all_z = np.round(5 * all_z + 20).astype(np.int)\n", 1513 | "# all_z = (all_z - z_min)\n", 1514 | "latent_map = {}\n", 1515 | "i = 0\n", 1516 | "for x in range(start, end + 1):\n", 1517 | " for y in range(start, end + 1):\n", 1518 | " latent_map[(x,y)] = all_z[i]\n", 1519 | " i += 1\n", 1520 | "for k in latent_map:\n", 1521 | " print (str(k) + ': ' + str(latent_map[k]))" 1522 | ] 1523 | }, 1524 | { 1525 | "cell_type": "code", 1526 | "execution_count": 35, 1527 | "metadata": {}, 1528 | "outputs": [ 1529 | { 1530 | "data": { 1531 | "text/plain": [ 1532 | "(1.0746268656716418, 1.0831858407079646)" 1533 | ] 1534 | }, 1535 | "execution_count": 35, 1536 | "metadata": {}, 1537 | "output_type": "execute_result" 1538 | } 1539 | ], 1540 | "source": [ 1541 | "count = 0\n", 1542 | "sum_diff_x = 0.0\n", 1543 | "sum_diff_y = 0.0\n", 1544 | "for i in range(0, all_z.shape[0] - 1):\n", 1545 | " sum_diff_x += np.abs(all_z[i][0] - all_z[i+1][0])\n", 1546 | " sum_diff_y += np.abs(all_z[i][1] - all_z[i+1][1])\n", 1547 | " count += 1\n", 1548 | "avg_diff_x = sum_diff_x / count\n", 1549 | "avg_diff_y = sum_diff_y / count\n", 1550 | "1 / avg_diff_x, 1 / avg_diff_y" 1551 | ] 1552 | }, 1553 | { 1554 | "cell_type": "code", 1555 | "execution_count": 43, 1556 | "metadata": {}, 1557 | "outputs": [ 1558 | { 1559 | "data": { 1560 | "text/plain": [ 1561 | "" 1562 | ] 1563 | }, 1564 | "execution_count": 43, 1565 | "metadata": {}, 1566 | "output_type": "execute_result" 1567 | }, 1568 | { 1569 | "data": { 1570 | "image/png": "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\n", 1571 | "text/plain": [ 1572 | "

" 1573 | ] 1574 | }, 1575 | "metadata": { 1576 | "needs_background": "light" 1577 | }, 1578 | "output_type": "display_data" 1579 | } 1580 | ], 1581 | "source": [ 1582 | "scale = 5\n", 1583 | "img_latent = Image.new(\"RGB\", (width * 10, height * 10), \"#FFFFFF\")\n", 1584 | "draw = ImageDraw.Draw(img_latent)\n", 1585 | "for k in latent_map:\n", 1586 | " x, y = k\n", 1587 | " if [x, y] in unvalid_pos:\n", 1588 | " continue\n", 1589 | " else:\n", 1590 | " x_scaled, y_scaled = latent_map[k][1] * 10, latent_map[k][0] * 10\n", 1591 | " draw.ellipse((x_scaled-2, y_scaled-2, x_scaled+2, y_scaled+2), fill = img.getpixel((y, x)))\n", 1592 | "img_latent.save('latent_map.png', 'PNG')\n", 1593 | "img_latent_scaled = np.array(img_latent) / 255.\n", 1594 | "# plt.imshow(img_arr_scaled)\n", 1595 | "# plt.show()\n", 1596 | "# plt.imshow(img_latent_scaled)\n", 1597 | "plt.show()\n", 1598 | "f, axarr = plt.subplots(1,2, figsize=(15,15))\n", 1599 | "# plt.setp(axarr, xticks=[], yticks=[])\n", 1600 | "axarr[0].imshow(img_arr_scaled)\n", 1601 | "axarr[1].imshow(img_latent_scaled)" 1602 | ] 1603 | }, 1604 | { 1605 | "cell_type": "code", 1606 | "execution_count": null, 1607 | "metadata": {}, 1608 | "outputs": [], 1609 | "source": [] 1610 | } 1611 | ], 1612 | "metadata": { 1613 | "kernelspec": { 1614 | "display_name": "Python 3", 1615 | "language": "python", 1616 | "name": "python3" 1617 | }, 1618 | "language_info": { 1619 | "codemirror_mode": { 1620 | "name": "ipython", 1621 | "version": 3 1622 | }, 1623 | "file_extension": ".py", 1624 | "mimetype": "text/x-python", 1625 | "name": "python", 1626 | "nbconvert_exporter": "python", 1627 | "pygments_lexer": "ipython3", 1628 | "version": "3.6.8" 1629 | } 1630 | }, 1631 | "nbformat": 4, 1632 | "nbformat_minor": 2 1633 | } 1634 | -------------------------------------------------------------------------------- /e2c.yml: -------------------------------------------------------------------------------- 1 | name: e2c 2 | channels: 3 | - pytorch 4 | - conda-forge 5 | - anaconda 6 | - defaults 7 | dependencies: 8 | - _libgcc_mutex=0.1=main 9 | - _tflow_select=2.1.0=gpu 10 | - absl-py=0.7.1=py36_0 11 | - asn1crypto=0.24.0=py36_0 12 | - astor=0.8.0=py36_0 13 | - blas=1.0=mkl 14 | - c-ares=1.15.0=h7b6447c_1001 15 | - ca-certificates=2019.5.15=1 16 | - certifi=2019.9.11=py36_0 17 | - cffi=1.12.3=py36h2e261b9_0 18 | - chardet=3.0.4=py36_1003 19 | - cryptography=2.7=py36h1ba5d50_0 20 | - cudatoolkit=10.0.130=0 21 | - cudnn=7.6.0=cuda10.0_0 22 | - cupti=10.0.130=0 23 | - cycler=0.10.0=py_1 24 | - dbus=1.13.6=h746ee38_0 25 | - expat=2.2.6=he6710b0_0 26 | - fontconfig=2.13.1=he4413a7_1000 27 | - freetype=2.9.1=h8a8886c_1 28 | - gast=0.2.2=py36_0 29 | - gettext=0.19.8.1=hc5be6a0_1002 30 | - glib=2.56.2=had28632_1001 31 | - google-pasta=0.1.7=py_0 32 | - grpcio=1.16.1=py36hf8bcb03_1 33 | - gst-plugins-base=1.14.0=hbbd80ab_1 34 | - gstreamer=1.14.0=hb453b48_1 35 | - h5py=2.9.0=py36h7918eee_0 36 | - hdf5=1.10.4=hb1b8bf9_0 37 | - icu=58.2=hf484d3e_1000 38 | - idna=2.8=py36_0 39 | - intel-openmp=2019.4=243 40 | - jpeg=9b=h024ee3a_2 41 | - keras-applications=1.0.8=py_0 42 | - keras-preprocessing=1.1.0=py_1 43 | - kiwisolver=1.1.0=py36hc9558a2_0 44 | - libedit=3.1.20181209=hc058e9b_0 45 | - libffi=3.2.1=hd88cf55_4 46 | - libgcc-ng=9.1.0=hdf63c60_0 47 | - libgfortran-ng=7.3.0=hdf63c60_0 48 | - libiconv=1.15=h516909a_1005 49 | - libpng=1.6.37=hbc83047_0 50 | - libprotobuf=3.8.0=hd408876_0 51 | - libstdcxx-ng=9.1.0=hdf63c60_0 52 | - libtiff=4.0.10=h2733197_2 53 | - libuuid=2.32.1=h14c3975_1000 54 | - libxcb=1.13=h14c3975_1002 55 | - libxml2=2.9.9=h13577e0_2 56 | - markdown=3.1.1=py36_0 57 | - matplotlib=3.1.1=py36h5429711_0 58 | - mkl=2019.4=243 59 | - mkl-service=2.3.0=py36he904b0f_0 60 | - mkl_fft=1.0.14=py36ha843d7b_0 61 | - mkl_random=1.0.2=py36hd81dba3_0 62 | - ncurses=6.1=he6710b0_1 63 | - ninja=1.9.0=py36hfd86e86_0 64 | - numpy=1.16.5=py36h7e9f1db_0 65 | - numpy-base=1.16.5=py36hde5b4d6_0 66 | - olefile=0.46=py36_0 67 | - openssl=1.1.1=h7b6447c_0 68 | - pcre=8.43=he6710b0_0 69 | - pillow=6.1.0=py36h34e0f95_0 70 | - pip=19.2.2=py36_0 71 | - protobuf=3.8.0=py36he6710b0_0 72 | - pthread-stubs=0.4=h14c3975_1001 73 | - pycparser=2.19=py36_0 74 | - pyopenssl=19.0.0=py36_0 75 | - pyparsing=2.4.2=py_0 76 | - pyqt=5.9.2=py36hcca6a23_4 77 | - pysocks=1.7.1=py36_0 78 | - python=3.6.8=h0371630_0 79 | - python-dateutil=2.8.0=py_0 80 | - pytorch=1.2.0=py3.6_cuda10.0.130_cudnn7.6.2_0 81 | - pytz=2019.2=py_0 82 | - qt=5.9.7=h5867ecd_1 83 | - readline=7.0=h7b6447c_5 84 | - requests=2.22.0=py36_0 85 | - scipy=1.3.1=py36h7c811a0_0 86 | - setuptools=41.2.0=py36_0 87 | - sip=4.19.8=py36hf484d3e_1000 88 | - six=1.12.0=py36_0 89 | - sqlite=3.29.0=h7b6447c_0 90 | - tensorboard=1.14.0=py36hf484d3e_0 91 | - tensorboardx=1.8=py_0 92 | - tensorflow=1.14.0=gpu_py36h57aa796_0 93 | - tensorflow-base=1.14.0=gpu_py36h8d69cac_0 94 | - tensorflow-estimator=1.14.0=py_0 95 | - tensorflow-gpu=1.14.0=h0d30ee6_0 96 | - termcolor=1.1.0=py36_1 97 | - tk=8.6.8=hbc83047_0 98 | - torchvision=0.4.0=py36_cu100 99 | - tornado=6.0.3=py36h516909a_0 100 | - tqdm=4.36.1=py_0 101 | - urllib3=1.24.2=py36_0 102 | - werkzeug=0.15.5=py_0 103 | - wheel=0.33.6=py36_0 104 | - wrapt=1.11.2=py36h7b6447c_0 105 | - xorg-libxau=1.0.9=h14c3975_0 106 | - xorg-libxdmcp=1.1.3=h516909a_0 107 | - xz=5.2.4=h14c3975_4 108 | - zlib=1.2.11=h7b6447c_3 109 | - zstd=1.3.7=h0b5b093_0 110 | - pip: 111 | - future==0.17.1 112 | - gym==0.9.1 113 | - pyglet==1.4.4 114 | prefix: /home/tungnd13/miniconda3/envs/e2c 115 | 116 | -------------------------------------------------------------------------------- /e2c_model.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | 3 | from normal import * 4 | from networks import * 5 | 6 | torch.set_default_dtype(torch.float64) 7 | 8 | class E2C(nn.Module): 9 | def __init__(self, obs_dim, z_dim, u_dim, env = 'planar'): 10 | super(E2C, self).__init__() 11 | enc, dec, trans = load_config(env) 12 | 13 | self.obs_dim = obs_dim 14 | self.z_dim = z_dim 15 | self.u_dim = u_dim 16 | 17 | self.encoder = enc(obs_dim=obs_dim, z_dim=z_dim) 18 | # self.encoder.apply(init_weights) 19 | self.decoder = dec(z_dim=z_dim, obs_dim=obs_dim) 20 | # self.decoder.apply(init_weights) 21 | self.trans = trans(z_dim=z_dim, u_dim=u_dim) 22 | # self.trans.apply(init_weights) 23 | 24 | def encode(self, x): 25 | """ 26 | :param x: 27 | :return: mean and log variance of q(z | x) 28 | """ 29 | return self.encoder(x) 30 | 31 | def decode(self, z): 32 | """ 33 | :param z: 34 | :return: bernoulli distribution p(x | z) 35 | """ 36 | return self.decoder(z) 37 | 38 | def transition(self, z_bar, q_z, u): 39 | """ 40 | :param z_bar: 41 | :param q_z: 42 | :param u: 43 | :return: samples z_hat_next and Q(z_hat_next) 44 | """ 45 | return self.trans(z_bar, q_z, u) 46 | 47 | def reparam(self, mean, logvar): 48 | sigma = (logvar / 2).exp() 49 | epsilon = torch.randn_like(sigma) 50 | return mean + torch.mul(epsilon, sigma) 51 | 52 | def forward(self, x, u, x_next): 53 | mu, logvar = self.encode(x) 54 | z = self.reparam(mu, logvar) 55 | q_z = NormalDistribution(mu, logvar) 56 | 57 | x_recon = self.decode(z) 58 | 59 | z_next, q_z_next_pred = self.transition(z, q_z, u) 60 | 61 | x_next_pred = self.decode(z_next) 62 | 63 | mu_next, logvar_next = self.encode(x_next) 64 | q_z_next = NormalDistribution(mean=mu_next, logvar=logvar_next) 65 | 66 | return x_recon, x_next_pred, q_z, q_z_next_pred, q_z_next 67 | 68 | def predict(self, x, u): 69 | mu, logvar = self.encoder(x) 70 | z = self.reparam(mu, logvar) 71 | q_z = NormalDistribution(mu, logvar) 72 | 73 | z_next, q_z_next_pred = self.transition(z, q_z, u) 74 | 75 | x_next_pred = self.decode(z_next) 76 | return x_next_pred -------------------------------------------------------------------------------- /evaluate_saved_model.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": 34, 6 | "metadata": {}, 7 | "outputs": [ 8 | { 9 | "data": { 10 | "text/plain": [ 11 | "IncompatibleKeys(missing_keys=[], unexpected_keys=[])" 12 | ] 13 | }, 14 | "execution_count": 34, 15 | "metadata": {}, 16 | "output_type": "execute_result" 17 | } 18 | ], 19 | "source": [ 20 | "import torch\n", 21 | "from torch.utils.data import random_split\n", 22 | "from torch.utils.data import DataLoader\n", 23 | "import matplotlib.pyplot as plt\n", 24 | "import random\n", 25 | "from datasets import *\n", 26 | "from e2c_model import E2C\n", 27 | "from train_e2c import evaluate\n", 28 | "model = E2C(1600,2,2,'planar_partial').cuda()\n", 29 | "model.load_state_dict(torch.load('result/planar_partial/log_check_valid/model_3000'))" 30 | ] 31 | }, 32 | { 33 | "cell_type": "code", 34 | "execution_count": 35, 35 | "metadata": {}, 36 | "outputs": [], 37 | "source": [ 38 | "batch_size = 128\n", 39 | "propor = 3/4\n", 40 | "dataset = PlanarDataset('./data/data/planar_partial')\n", 41 | "train_set, test_set = dataset[:int(len(dataset) * propor)], dataset[int(len(dataset) * propor):]\n", 42 | "test_loader = DataLoader(test_set, batch_size=batch_size, shuffle=False, drop_last=False, num_workers=8)" 43 | ] 44 | }, 45 | { 46 | "cell_type": "code", 47 | "execution_count": 36, 48 | "metadata": {}, 49 | "outputs": [ 50 | { 51 | "name": "stdout", 52 | "output_type": "stream", 53 | "text": [ 54 | "State loss: 12.65210761933395\n", 55 | "Next state loss: 12.661921219078062\n" 56 | ] 57 | } 58 | ], 59 | "source": [ 60 | "state_loss, next_state_loss = evaluate(model, test_loader)\n", 61 | "print ('State loss: ' + str(state_loss))\n", 62 | "print ('Next state loss: ' + str(next_state_loss))" 63 | ] 64 | }, 65 | { 66 | "cell_type": "code", 67 | "execution_count": 67, 68 | "metadata": {}, 69 | "outputs": [ 70 | { 71 | "data": { 72 | "text/plain": [ 73 | "" 74 | ] 75 | }, 76 | "execution_count": 67, 77 | "metadata": {}, 78 | "output_type": "execute_result" 79 | }, 80 | { 81 | "data": { 82 | "image/png": "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\n", 83 | "text/plain": [ 84 | "

" 85 | ] 86 | }, 87 | "metadata": {}, 88 | "output_type": "display_data" 89 | } 90 | ], 91 | "source": [ 92 | "rand_idx = random.randint(0, len(test_set))\n", 93 | "x, u, x_next = test_set[rand_idx]\n", 94 | "with torch.no_grad():\n", 95 | " x_next_pred = model.predict(x.unsqueeze(0).view(-1,1600).double().cuda(),\n", 96 | " torch.Tensor(u).unsqueeze(0).double().cuda())\n", 97 | "plt.show()\n", 98 | "f, axarr = plt.subplots(1,2)\n", 99 | "plt.setp(axarr, xticks=[], yticks=[])\n", 100 | "axarr[0].imshow(x_next.squeeze(), cmap='gray')\n", 101 | "axarr[1].imshow(x_next_pred.squeeze().view(40,40).cpu(), cmap='gray')" 102 | ] 103 | }, 104 | { 105 | "cell_type": "code", 106 | "execution_count": null, 107 | "metadata": {}, 108 | "outputs": [], 109 | "source": [] 110 | } 111 | ], 112 | "metadata": { 113 | "kernelspec": { 114 | "display_name": "Python 3", 115 | "language": "python", 116 | "name": "python3" 117 | }, 118 | "language_info": { 119 | "codemirror_mode": { 120 | "name": "ipython", 121 | "version": 3 122 | }, 123 | "file_extension": ".py", 124 | "mimetype": "text/x-python", 125 | "name": "python", 126 | "nbconvert_exporter": "python", 127 | "pygments_lexer": "ipython3", 128 | "version": "3.6.8" 129 | } 130 | }, 131 | "nbformat": 4, 132 | "nbformat_minor": 2 133 | } 134 | -------------------------------------------------------------------------------- /map.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/tung-nd/E2C-pytorch/8af375b58e1ddef3cd041d518b7061faece2f102/map.png -------------------------------------------------------------------------------- /networks.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | from normal import NormalDistribution 4 | 5 | torch.set_default_dtype(torch.float64) 6 | 7 | def weights_init(m): 8 | if type(m) in [nn.Conv2d, nn.Linear, nn.ConvTranspose2d]: 9 | torch.nn.init.orthogonal_(m.weight) 10 | 11 | class Encoder(nn.Module): 12 | def __init__(self, net, obs_dim, z_dim): 13 | super(Encoder, self).__init__() 14 | self.net = net 15 | self.net.apply(weights_init) 16 | self.img_dim = obs_dim 17 | self.z_dim = z_dim 18 | 19 | def forward(self, x): 20 | """ 21 | :param x: observation 22 | :return: the parameters of distribution q(z|x) 23 | """ 24 | return self.net(x).chunk(2, dim = 1) # first half is mean, second half is logvar 25 | 26 | class Decoder(nn.Module): 27 | def __init__(self, net, z_dim, obs_dim): 28 | super(Decoder, self).__init__() 29 | self.net = net 30 | self.net.apply(weights_init) 31 | self.z_dim = z_dim 32 | self.obs_dim = obs_dim 33 | 34 | def forward(self, z): 35 | """ 36 | :param z: sample from q(z|x) 37 | :return: reconstructed x 38 | """ 39 | return self.net(z) 40 | 41 | class Transition(nn.Module): 42 | def __init__(self, net, z_dim, u_dim): 43 | super(Transition, self).__init__() 44 | self.net = net # network to output the last layer before predicting A_t, B_t and o_t 45 | self.net.apply(weights_init) 46 | self.h_dim = self.net[-3].out_features 47 | self.z_dim = z_dim 48 | self.u_dim = u_dim 49 | 50 | self.fc_A = nn.Sequential( 51 | nn.Linear(self.h_dim, self.z_dim * 2), # v_t and r_t 52 | nn.Sigmoid() 53 | ) 54 | self.fc_A.apply(weights_init) 55 | 56 | self.fc_B = nn.Linear(self.h_dim, self.z_dim * self.u_dim) 57 | torch.nn.init.orthogonal_(self.fc_B.weight) 58 | self.fc_o = nn.Linear(self.h_dim, self.z_dim) 59 | torch.nn.init.orthogonal_(self.fc_o.weight) 60 | 61 | def forward(self, z_bar_t, q_z_t, u_t): 62 | """ 63 | :param z_bar_t: the reference point 64 | :param Q_z_t: the distribution q(z|x) 65 | :param u_t: the action taken 66 | :return: the predicted q(z^_t+1 | z_t, z_bar_t, u_t) 67 | """ 68 | h_t = self.net(z_bar_t) 69 | B_t = self.fc_B(h_t) 70 | o_t = self.fc_o(h_t) 71 | 72 | v_t, r_t = self.fc_A(h_t).chunk(2, dim=1) 73 | v_t = torch.unsqueeze(v_t, dim=-1) 74 | r_t = torch.unsqueeze(r_t, dim=-2) 75 | 76 | A_t = torch.eye(self.z_dim).repeat(z_bar_t.size(0), 1, 1).cuda() + torch.bmm(v_t, r_t) 77 | 78 | B_t = B_t.view(-1, self.z_dim, self.u_dim) 79 | 80 | mu_t = q_z_t.mean 81 | 82 | mean = A_t.bmm(mu_t.unsqueeze(-1)).squeeze(-1) + B_t.bmm(u_t.unsqueeze(-1)).squeeze(-1) + o_t 83 | 84 | return mean, NormalDistribution(mean, logvar=q_z_t.logvar, v=v_t.squeeze(), r=r_t.squeeze(), A=A_t) 85 | 86 | class PlanarEncoder(Encoder): 87 | def __init__(self, obs_dim = 1600, z_dim = 2): 88 | net = nn.Sequential( 89 | nn.Linear(obs_dim, 150), 90 | nn.BatchNorm1d(150), 91 | nn.ReLU(), 92 | 93 | nn.Linear(150, 150), 94 | nn.BatchNorm1d(150), 95 | nn.ReLU(), 96 | 97 | nn.Linear(150, 150), 98 | nn.BatchNorm1d(150), 99 | nn.ReLU(), 100 | 101 | nn.Linear(150, z_dim * 2) 102 | ) 103 | super(PlanarEncoder, self).__init__(net, obs_dim, z_dim) 104 | 105 | class PlanarDecoder(Decoder): 106 | def __init__(self, z_dim = 2, obs_dim = 1600): 107 | net = nn.Sequential( 108 | nn.Linear(z_dim, 200), 109 | nn.BatchNorm1d(200), 110 | nn.ReLU(), 111 | 112 | nn.Linear(200, 200), 113 | nn.BatchNorm1d(200), 114 | nn.ReLU(), 115 | 116 | nn.Linear(200, 1600), 117 | nn.Sigmoid() 118 | ) 119 | super(PlanarDecoder, self).__init__(net, z_dim, obs_dim) 120 | 121 | class PlanarTransition(Transition): 122 | def __init__(self, z_dim = 2, u_dim = 2): 123 | net = nn.Sequential( 124 | nn.Linear(z_dim, 100), 125 | nn.BatchNorm1d(100), 126 | nn.ReLU(), 127 | 128 | nn.Linear(100, 100), 129 | nn.BatchNorm1d(100), 130 | nn.ReLU() 131 | ) 132 | super(PlanarTransition, self).__init__(net, z_dim, u_dim) 133 | 134 | class PendulumEncoder(Encoder): 135 | def __init__(self, obs_dim = 4608, z_dim = 3): 136 | net = nn.Sequential( 137 | nn.Linear(obs_dim, 800), 138 | nn.BatchNorm1d(800), 139 | nn.ReLU(), 140 | 141 | nn.Linear(800, 800), 142 | nn.BatchNorm1d(800), 143 | nn.ReLU(), 144 | 145 | nn.Linear(800, z_dim * 2) 146 | ) 147 | super(PendulumEncoder, self).__init__(net, obs_dim, z_dim) 148 | 149 | class PendulumDecoder(Decoder): 150 | def __init__(self, z_dim = 3, obs_dim = 4608): 151 | net = nn.Sequential( 152 | nn.Linear(z_dim, 800), 153 | nn.BatchNorm1d(800), 154 | nn.ReLU(), 155 | 156 | nn.Linear(800, 800), 157 | nn.BatchNorm1d(800), 158 | nn.ReLU(), 159 | 160 | nn.Linear(800, obs_dim) 161 | ) 162 | super(PendulumDecoder, self).__init__(net, z_dim, obs_dim) 163 | 164 | class PendulumTransition(Transition): 165 | def __init__(self, z_dim = 3, u_dim = 1): 166 | net = nn.Sequential( 167 | nn.Linear(z_dim, 100), 168 | nn.BatchNorm1d(100), 169 | nn.ReLU(), 170 | 171 | nn.Linear(100, 100), 172 | nn.BatchNorm1d(100), 173 | nn.ReLU() 174 | ) 175 | super(PendulumTransition, self).__init__(net, z_dim, u_dim) 176 | 177 | CONFIG = { 178 | 'planar': (PlanarEncoder, PlanarDecoder, PlanarTransition), 179 | 'pendulum': (PendulumEncoder, PendulumDecoder, PendulumTransition) 180 | } 181 | 182 | def load_config(name): 183 | return CONFIG[name] 184 | 185 | __all__ = ['load_config'] 186 | 187 | # enc = PendulumEncoder() 188 | # dec = PendulumDecoder() 189 | # trans = PendulumTransition() 190 | # 191 | # x = torch.randn(size=(10, 4608)) 192 | # # print (x.size()) 193 | # mean, logvar = enc(x) 194 | # # print (logvar.size()) 195 | # x_recon = dec(mean) 196 | # # print (x_recon.size()) 197 | # 198 | # q_z_t = NormalDistribution(mean, logvar) 199 | # print (q_z_t.mean.size()) 200 | # print (q_z_t.cov.size()) 201 | # u_t = torch.randn(size=(10, 1)) 202 | # z_t_1 = trans(mean, q_z_t, u_t) 203 | # print (z_t_1[1].mean.size()) 204 | # print (z_t_1[1].cov.size()) 205 | # 206 | # kl = NormalDistribution.KL_divergence(z_t_1[1], q_z_t) 207 | # print (kl) -------------------------------------------------------------------------------- /normal.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | torch.set_default_dtype(torch.float64) 4 | 5 | class NormalDistribution: 6 | def __init__(self, mean, logvar, v=None, r=None, A=None): 7 | """ 8 | :param mean: mu in the paper 9 | :param logvar: \Sigma in the paper 10 | :param v: 11 | :param r: 12 | if A is not None then covariance matrix = A \Sigma A^T, where A = I + v^T r 13 | else the covariance matrix is simply diag(logvar.exp()) 14 | """ 15 | self.mean = mean 16 | self.logvar = logvar 17 | self.v = v 18 | self.r = r 19 | 20 | sigma = torch.diag_embed(torch.exp(logvar)) 21 | if A is None: 22 | self.cov = sigma 23 | else: 24 | self.cov = A.bmm(sigma.bmm(A.transpose(1, 2))) 25 | 26 | 27 | @staticmethod 28 | def KL_divergence(q_z_next_pred, q_z_next): 29 | """ 30 | :param q_z_next_pred: q(z_{t+1} | z_bar_t, q_z_t, u_t) using the transition 31 | :param q_z_next: q(z_t+1 | x_t+1) using the encoder 32 | :return: KL divergence between two distributions 33 | """ 34 | mu_0 = q_z_next_pred.mean 35 | mu_1 = q_z_next.mean 36 | sigma_0 = torch.exp(q_z_next_pred.logvar) 37 | sigma_1 = torch.exp(q_z_next.logvar) 38 | v = q_z_next_pred.v 39 | r = q_z_next_pred.r 40 | k = float(q_z_next_pred.mean.size(1)) 41 | 42 | sum = lambda x: torch.sum(x, dim=1) 43 | 44 | KL = 0.5 * torch.mean(sum((sigma_0 + 2*sigma_0*v*r) / sigma_1) 45 | + sum(r.pow(2) * sigma_0) * sum(v.pow(2) / sigma_1) 46 | + sum(torch.pow(mu_1-mu_0, 2) / sigma_1) - k 47 | + 2 * (sum(q_z_next.logvar - q_z_next_pred.logvar) - torch.log(1 + sum(v*r))) 48 | ) 49 | return KL -------------------------------------------------------------------------------- /test_plot.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": 34, 6 | "metadata": {}, 7 | "outputs": [], 8 | "source": [ 9 | "from tensorboardX import SummaryWriter\n", 10 | "import matplotlib.pyplot as plt\n", 11 | "import numpy as np" 12 | ] 13 | }, 14 | { 15 | "cell_type": "code", 16 | "execution_count": 35, 17 | "metadata": {}, 18 | "outputs": [ 19 | { 20 | "data": { 21 | "image/png": "iVBORw0KGgoAAAANSUhEUgAAAagAAAEYCAYAAAAJeGK1AAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADh0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uMy4xLjEsIGh0dHA6Ly9tYXRwbG90bGliLm9yZy8QZhcZAAAJlklEQVR4nO3dX6jX9R3H8fentbmVWi0PYxB6sChOzIrwokmUF9qaK7qSoG1QdLOBV+4PMXY5dBd1402W7CKaN8Egk0Ujajf748URt0I2Q3Qqupk6kjVpBn12kQMRZ+r5xHkdezxAkK/Hl58fh8OT7++o39Z7LwBIc9VsHwAAzkegAIgkUABEEigAIgkUAJGunu0DAJdu0aJFfXJycraPAUPs3LnzeO994tzrAgVz0OTkZE1PT8/2MWCI1tqB8133Fh8AkQQKgEgCBUAkgQIgkkABEEmgAIgkUABEEigAIgkUAJEECoBIAgVAJIECIJJAARBJoACIJFAARBIoACIJFACRBAqASAIFQCSBAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgkkABEEmgAIgkUABEEigAIgkUAJEECoBIAgVAJIECIJJAARBJoACIJFAARBIoACJdPdsHuBQ33nhjX7x48Yx3PvjggwGnqTp16tSQnQ8//HDIzsKFC4fszJ8/f8jOvn37huxMTEwM2Tl+/PiQnUWLFg3Zeeedd4733se8OLgCzalALV68uN58880Z7+zZs2fAaareeuutITuHDx8esrN69eohO/fee++QnbVr1w7ZWbdu3ZCd5557bsjOk08+OWRn1apVB4YMwRXKW3wARBIoACIJFACRBAqASAIFQCSBAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgkkABEEmgAIg0px5YePLkyXrttddmvPPYY48NOE3Vhg0bhuzceuutQ3befvvtITuvv/76kJ3169cP2RnxkMqqqmXLlg3Zue+++4bsABfmDgqASAIFQCSBAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgkkABEEmgAIgkUABEEigAIgkUAJEECoBIc+qJujfccEOtXbt2yM4IJ0+eHLLz8ssvD9lZunTpkJ158+YN2dm7d++QnRUrVgzZOXTo0JCdHTt2DNkBLswdFACRBAqASAIFQCSBAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgkkABEEmgAIgkUABEEigAIgkUAJHm1BN1jx49Ws8888yMd1auXDnzwwy0fv36ITtTU1NDdp5//vkhO9dcc82QnXfffXfIzqgnDm/fvn3IDnBh7qAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgkkABEEmgAIgkUABEEigAIgkUAJEECoBIAgVAJIECIJJAARBJoACINKeeqHv69Ok6cODAjHd27do14DRVK1asGLKzZMmSITsvvPDCkJ077rhjyM7ChQuH7GzZsmXIzqjXNWoHuDB3UABEEigAIgkUAJEECoBIAgVAJIECIJJAARBJoACIJFAARBIoACIJFACRBAqASBcVqNba31prq2byB7XWHm+t/W4mGwB8driDAiDSJwaqtfZiVS2uqu2ttfdbaz9urd3TWvtDa+291tqfW2srz/r4x1tr+1pr/2qt7W+tfbu1NlVVm6vq62c23vvUXhEAV4RPDFTv/btVdbCqHu69z6+qrVX166r6WVV9uap+WFW/aq1NtNaurapNVfXN3vuCqlpRVX/qvf+lqr5XVX/svc/vvV//6bwcAK4Ul/MW33eq6tXe+6u99496769X1XRVrTnz6x9V1ddaa1/qvf+997571GEB+Oy4nCfqLqmqta21h8+69vmq+m3v/d+ttUfr47uqX7TWfl9VP+i9/3XAWeu6666rNWvWfPIHfoITJ04MOE3V1NTUkJ0HHnhgyM7mzZuH7GzatGnIzsaNG4fsPPvss0N29uzZM2TnqaeeGrIDXNjF3kH1s35+qKpe7L1ff9aPa3vvP6+q6r3/pve+uqq+WlV/raot59kAgAu62EAdraqlZ37+y6p6uLX2jdba51prX2ytrWyt3dRa+0pr7ZEz34v6T1W9Xx+/5fe/jZtaa18Y+goAuCJdbKA2VtVPz/ztu0er6pGq+klVHauP76h+dGbrqqpaX1VHquqfVXV/VX3/zMabVbW7qv7RWjs+6gUAcGW6qO9B9d63VdW2cy7f/38+/LzXe++nq+pbF380AD7L/ENdACIJFACRBAqASAIFQCSBAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgUut97jxHcNmyZX3btnP/U/VLt3Xr1gGnqbr55puH7Ozdu3fIzoIFC4bsPPTQQ0N2XnrppSE7R44cGbIz6vO1atWqITt33nnnzt778sv5vcuXL+/T09NDzgGzrbV23q8Fd1AARBIoACIJFACRBAqASAIFQCSBAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgkkABEEmgAIgkUABEunq2DzAbnnjiiSE7r7zyypCd/fv3D9l58MEHh+wcPnx4yM4tt9wyZOeNN94YsjPqybwnTpwYsgNcmDsoACIJFACRBAqASAIFQCSBAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgkkABEEmgAIgkUABEEigAIs2pJ+qeOnWqdu3aNeOdgwcPDjjNuCfYjnhNVVU7duwYsrNu3bohO8eOHRuyc9tttw3Zuf3224fsTE1NDdnZsGHDkB24UrmDAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgkkABEEmgAIgkUABEEigAIgkUAJEECoBIAgVAJIECINKceqLuvHnzanJycsY7p0+fnvlhqmr37t1Ddu66664hO3ffffeQnaeffnrIzj333DNkZ2JiYsjO6tWrh+yM+rwDF+YOCoBIAgVAJIECIJJAARBJoACIJFAARBIoACIJFACRBAqASAIFQCSBAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEit9z7bZ7horbVjVXVgts8BgyzpvV/W44KXL1/ep6enR58HZkVrbWfvffm51+fUI98v94sZgLnHW3wARBIoACIJFACRBAqASAIFQCSBAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgkkABEEmgAIgkUABEEigAIgkUAJEECoBIAgVAJIECIJJAARBJoACIJFAARBIoACIJFACRBAqASAIFQCSBAiCSQAEQSaAAiCRQAEQSKAAiCRQAkQQKgEgCBUAkgQIgkkABEKn13mf7DMAlaq0dq6oDs30OGGRJ733i3IsCBUAkb/EBEEmgAIgkUABEEigAIgkUAJEECoBIAgVAJIECIJJAARDpv7rGe11cTsUVAAAAAElFTkSuQmCC\n", 22 | "text/plain": [ 23 | "