Skip to content

Instantly share code, notes, and snippets.

@pilipolio

pilipolio/DoomBasic.ipynb

Last active March 4, 2017 11:57
Show Gist options
  • Save pilipolio/3732a420de21778392328d66d10e5778 to your computer and use it in GitHub Desktop.
Save pilipolio/3732a420de21778392328d66d10e5778 to your computer and use it in GitHub Desktop.
Download ZIP
Display the source blob
Display the rendered blob
Raw
DoomBasic.ipynb
{
"cells": [
{
"cell_type": "code",
"execution_count": 2,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"import numpy as np\n",
"import gym\n",
"import ppaquette_gym_doom\n",
"from ppaquette_gym_doom.wrappers import SetResolution, ToDiscrete\n",
"from gym.wrappers import SkipWrapper\n",
"from gym import wrappers\n",
"\n",
"# (see https://github.com/ppaquette/gym-doom/blob/master/ppaquette_gym_doom/doom_basic.py)\n",
"def create_env(seed=None):\n",
" env_spec = gym.spec('ppaquette/DoomBasic-v0')\n",
" env_spec.id = 'DoomBasic-v0'\n",
" env = env_spec.make()\n",
"\n",
" if seed is not None:\n",
" env.seed(seed)\n",
"\n",
" return SetResolution('200x150')(\n",
" SkipWrapper(repeat_count=4)(\n",
" ToDiscrete('minimal')(env)))\n",
"\n",
"env = create_env()\n",
"WIDTH, HEIGHT = env.screen_width, env.screen_height\n",
"\n",
"NOOP, SHOOT, RIGHT, LEFT = 0, 1, 2, 3"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Collecting experiences"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"from collections import namedtuple\n",
"import operator\n",
"\n",
"SARE = namedtuple('SAR', ['state', 'action', 'reward', 'end'])\n",
"\n",
"\n",
"def generate_sares(env, agent, episode_count=100):\n",
" reward = 0\n",
" done = False\n",
"\n",
" for i in range(episode_count):\n",
" observation = env.reset()\n",
" while True:\n",
" action = agent.act(observation, reward, done)\n",
" new_observation, reward, done, _ = env.step(action)\n",
" yield SARE(observation, action, reward, done)\n",
" \n",
" if done:\n",
" break\n",
" else:\n",
" observation = new_observation\n",
"\n",
"def episode_sares(env, agent, episode_count=100):\n",
" sares = list(generate_sares(env, agent, episode_count))\n",
" print('average reward per episode = {}'.format(\n",
" sum(r for _, _, r, _ in sares) / float(sum(e for _, _, _, e in sares))))\n",
" return sares\n",
"\n",
" \n",
"def to_experiences(sares, only_n_misses=100):\n",
" experiences = [\n",
" (previous_s, a, r, next_s, end)\n",
" for (previous_s, a, r, end), (next_s, _, _, _) in zip(sares[:-1], sares[1:])\n",
" ]\n",
"\n",
" # simplistic experience prioritization\n",
" shuffled_exps = experiences if only_n_misses is None\\\n",
" else random.choices(experiences, k=only_n_misses) + [e for e in experiences if e[2] > 0]\n",
" random.shuffle(shuffled_exps)\n",
"\n",
" prev_frames, actions, rewards, next_frames, is_ends = zip(*shuffled_exps)\n",
" prev_frames = np.asarray(prev_frames)\n",
" next_frames = np.asarray(next_frames)\n",
" actions = np.asarray(actions)\n",
" rewards = np.asarray(rewards)\n",
" is_ends = np.asarray(is_ends)\n",
" \n",
" print('Training on {}/{} positive/total out of {} 1-step experiences with actions distribution {}'.format(\n",
" np.sum(rewards>=0),\n",
" len(rewards),\n",
" len(experiences),\n",
" np.bincount(actions)))\n",
" \n",
" return (prev_frames, next_frames, actions, rewards, is_ends)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Deep Q-learning"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {
"collapsed": false
},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"Using TensorFlow backend.\n"
]
}
],
"source": [
"import tensorflow as tf\n",
"from keras import backend as K\n",
"\n",
"from keras.layers import Dense, Convolution2D, Flatten, Activation\n",
"from keras.models import Sequential\n",
"from keras.optimizers import Adam\n",
"\n",
"sess = tf.InteractiveSession()\n",
"K.set_session(sess)\n",
"\n",
"def create_q_model(conv1_weights=None, conv2_weights=None, dense1_weights=None, dense2_weights=None):\n",
" model = Sequential()\n",
"\n",
" model.add(Convolution2D(\n",
" 2, nb_row=6, nb_col=6, border_mode='valid', weights=conv1_weights,\n",
" input_shape=[HEIGHT, WIDTH, 3], dim_ordering='tf'))\n",
" model.add(Activation('relu'))\n",
" model.add(Convolution2D(4, nb_row=2, nb_col=2, weights=conv2_weights))\n",
" model.add(Activation('relu'))\n",
" model.add(Flatten())\n",
" model.add(Dense(64, init='normal', weights=dense1_weights))\n",
" model.add(Activation('relu'))\n",
" model.add(Dense(4, init='normal', weights=dense2_weights))\n",
" model.compile(loss='mse', optimizer=Adam())\n",
" \n",
" return model\n",
"\n",
"acting_model = create_q_model()\n",
"target_model = create_q_model()\n",
"\n",
"def copy_model(model):\n",
" conv1_weights = [w.eval() for w in model.layers[0].weights]\n",
" conv2_weights = [w.eval() for w in model.layers[2].weights]\n",
" dense1_weights = [w.eval() for w in model.layers[5].weights]\n",
" dense2_weights = [w.eval() for w in model.layers[7].weights]\n",
" return create_q_model(conv1_weights, conv2_weights, dense1_weights, dense2_weights)"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"def sares_to_input_targets(model, sares, gamma = .99, reward_clip=5, only_n_misses=100):\n",
" \n",
" prev_frames, next_frames, actions, rewards, is_ends = to_experiences(sares, only_n_misses)\n",
" \n",
" n_samples = len(actions)\n",
" clipped_rewards = np.clip(rewards, -np.inf, reward_clip)\n",
" \n",
" # Transcription of the Q-learning target formula\n",
" targets = clipped_rewards + gamma * (1 - is_ends) * model.predict(next_frames).max(axis=1)\n",
"\n",
" target_action_rewards = model.predict(prev_frames)\n",
" target_action_rewards[np.arange(n_samples), actions] = targets\n",
"\n",
" return prev_frames, target_action_rewards"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Training"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false
},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"[2017-03-04 11:44:01,389] DEPRECATION WARNING: env.spec.timestep_limit has been deprecated. Replace your call to `env.spec.timestep_limit` with `env.spec.tags.get('wrapper_config.TimeLimit.max_episode_steps')`. This change was made 12/28/2016 and is included in version 0.7.0\n",
"[2017-03-04 11:44:01,390] Clearing 27 monitor files from previous run (because force=True was provided)\n",
"[2017-03-04 11:44:01,764] Starting new video recorder writing to /Users/gui/Dev/rl-study/tmp/q_learning/openaigym.video.0.48905.video000000.mp4\n",
"[2017-03-04 11:44:02,552] Starting new video recorder writing to /Users/gui/Dev/rl-study/tmp/q_learning/openaigym.video.0.48905.video000001.mp4\n",
"[2017-03-04 11:44:05,313] Starting new video recorder writing to /Users/gui/Dev/rl-study/tmp/q_learning/openaigym.video.0.48905.video000008.mp4\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"average reward per episode = -295.0\n",
"Training on 4/104 positive/total out of 499 1-step experiences with actions distribution [ 2 91 5 6]\n",
"average reward per episode = -173.8\n",
"Training on 6/106 positive/total out of 349 1-step experiences with actions distribution [ 4 96 5 1]\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"[2017-03-04 11:44:17,532] Starting new video recorder writing to /Users/gui/Dev/rl-study/tmp/q_learning/openaigym.video.0.48905.video000027.mp4\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"average reward per episode = -301.7\n",
"Training on 3/103 positive/total out of 507 1-step experiences with actions distribution [ 4 95 1 3]\n",
"average reward per episode = -186.5\n",
"Training on 5/105 positive/total out of 369 1-step experiences with actions distribution [ 2 99 2 2]\n",
"average reward per episode = -344.1\n",
"Training on 3/103 positive/total out of 572 1-step experiences with actions distribution [ 3 95 2 3]\n",
"average reward per episode = -283.2\n",
"Training on 4/104 positive/total out of 480 1-step experiences with actions distribution [ 4 96 2 2]\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"[2017-03-04 11:44:42,014] Starting new video recorder writing to /Users/gui/Dev/rl-study/tmp/q_learning/openaigym.video.0.48905.video000064.mp4\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"average reward per episode = -212.8\n",
"Training on 5/104 positive/total out of 393 1-step experiences with actions distribution [ 2 97 3 2]\n",
"average reward per episode = -238.4\n",
"Training on 6/105 positive/total out of 432 1-step experiences with actions distribution [ 4 96 1 4]\n",
"average reward per episode = -282.7\n",
"Training on 4/104 positive/total out of 478 1-step experiences with actions distribution [ 1 99 2 2]\n",
"average reward per episode = -243.1\n",
"Training on 5/105 positive/total out of 437 1-step experiences with actions distribution [ 0 100 2 3]\n",
"average reward per episode = -326.9\n",
"Training on 4/103 positive/total out of 546 1-step experiences with actions distribution [ 1 98 2 2]\n",
"average reward per episode = -173.7\n",
"Training on 7/105 positive/total out of 349 1-step experiences with actions distribution [ 4 98 2 1]\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"[2017-03-04 11:45:20,178] Starting new video recorder writing to /Users/gui/Dev/rl-study/tmp/q_learning/openaigym.video.0.48905.video000125.mp4\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"average reward per episode = -328.8\n",
"Training on 4/104 positive/total out of 547 1-step experiences with actions distribution [ 2 96 3 3]\n",
"average reward per episode = -319.3\n",
"Training on 4/103 positive/total out of 533 1-step experiences with actions distribution [ 1 97 2 3]\n",
"average reward per episode = -284.6\n",
"Training on 4/104 positive/total out of 483 1-step experiences with actions distribution [ 1 99 0 4]\n",
"average reward per episode = -250.8\n",
"Training on 5/104 positive/total out of 432 1-step experiences with actions distribution [ 2 99 2 1]\n",
"average reward per episode = -362.6\n",
"Training on 2/102 positive/total out of 567 1-step experiences with actions distribution [ 2 96 4]\n",
"average reward per episode = -117.1\n",
"Training on 7/107 positive/total out of 282 1-step experiences with actions distribution [ 0 98 2 7]\n",
"average reward per episode = -343.6\n",
"Training on 2/102 positive/total out of 555 1-step experiences with actions distribution [ 3 96 2 1]\n",
"average reward per episode = -226.5\n",
"Training on 5/105 positive/total out of 412 1-step experiences with actions distribution [ 2 98 2 3]\n",
"average reward per episode = -333.1\n",
"Training on 3/103 positive/total out of 540 1-step experiences with actions distribution [ 2 92 5 4]\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"[2017-03-04 11:46:19,153] Starting new video recorder writing to /Users/gui/Dev/rl-study/tmp/q_learning/openaigym.video.0.48905.video000216.mp4\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"average reward per episode = -250.3\n",
"Training on 6/105 positive/total out of 447 1-step experiences with actions distribution [ 2 93 8 2]\n",
"average reward per episode = -281.4\n",
"Training on 3/103 positive/total out of 480 1-step experiences with actions distribution [ 6 93 3 1]\n",
"average reward per episode = -297.9\n",
"Training on 4/104 positive/total out of 502 1-step experiences with actions distribution [ 4 87 10 3]\n",
"average reward per episode = -415.4\n",
"Training on 1/101 positive/total out of 630 1-step experiences with actions distribution [ 1 92 7 1]\n",
"average reward per episode = -394.4\n",
"Training on 1/101 positive/total out of 615 1-step experiences with actions distribution [ 5 93 0 3]\n",
"average reward per episode = -200.4\n",
"Training on 7/105 positive/total out of 373 1-step experiences with actions distribution [ 5 91 3 6]\n",
"average reward per episode = -87.6\n",
"Training on 7/106 positive/total out of 237 1-step experiences with actions distribution [ 4 101 0 1]\n",
"average reward per episode = -255.3\n",
"Training on 5/104 positive/total out of 440 1-step experiences with actions distribution [ 3 95 3 3]\n",
"average reward per episode = -301.7\n",
"Training on 3/102 positive/total out of 492 1-step experiences with actions distribution [ 0 98 2 2]\n",
"average reward per episode = -170.0\n",
"Training on 6/106 positive/total out of 342 1-step experiences with actions distribution [ 0 99 3 4]\n",
"average reward per episode = -294.0\n",
"Training on 4/104 positive/total out of 513 1-step experiences with actions distribution [ 3 93 4 4]\n",
"average reward per episode = -343.5\n",
"Training on 4/103 positive/total out of 557 1-step experiences with actions distribution [ 6 94 1 2]\n",
"average reward per episode = -171.3\n",
"Training on 8/105 positive/total out of 347 1-step experiences with actions distribution [ 5 91 6 3]\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"[2017-03-04 11:47:45,261] Starting new video recorder writing to /Users/gui/Dev/rl-study/tmp/q_learning/openaigym.video.0.48905.video000343.mp4\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"average reward per episode = -155.1\n",
"Training on 11/107 positive/total out of 340 1-step experiences with actions distribution [ 2 102 2 1]\n",
"average reward per episode = -388.8\n",
"Training on 2/101 positive/total out of 604 1-step experiences with actions distribution [ 2 95 2 2]\n",
"average reward per episode = -403.0\n",
"Training on 1/101 positive/total out of 628 1-step experiences with actions distribution [ 4 91 2 4]\n",
"average reward per episode = -331.6\n",
"Training on 2/102 positive/total out of 537 1-step experiences with actions distribution [ 2 97 1 2]\n",
"average reward per episode = -425.8\n",
"Training on 4/102 positive/total out of 661 1-step experiences with actions distribution [ 4 94 3 1]\n",
"average reward per episode = -205.2\n",
"Training on 8/106 positive/total out of 396 1-step experiences with actions distribution [ 5 98 1 2]\n",
"average reward per episode = -157.5\n",
"Training on 5/105 positive/total out of 324 1-step experiences with actions distribution [ 3 96 2 4]\n",
"average reward per episode = -328.3\n",
"Training on 4/103 positive/total out of 547 1-step experiences with actions distribution [ 2 95 2 4]\n",
"average reward per episode = -378.3\n",
"Training on 3/103 positive/total out of 608 1-step experiences with actions distribution [ 6 94 1 2]\n",
"average reward per episode = -328.8\n",
"Training on 4/104 positive/total out of 550 1-step experiences with actions distribution [ 3 97 3 1]\n",
"average reward per episode = -111.7\n",
"Training on 7/106 positive/total out of 273 1-step experiences with actions distribution [ 3 96 5 2]\n",
"average reward per episode = -227.4\n",
"Training on 5/105 positive/total out of 414 1-step experiences with actions distribution [ 0 96 4 5]\n",
"average reward per episode = -102.0\n",
"Training on 12/106 positive/total out of 258 1-step experiences with actions distribution [ 2 100 4]\n",
"average reward per episode = -255.5\n",
"Training on 4/104 positive/total out of 440 1-step experiences with actions distribution [ 4 94 3 3]\n",
"average reward per episode = -304.3\n",
"Training on 4/104 positive/total out of 514 1-step experiences with actions distribution [ 5 95 0 4]\n",
"average reward per episode = -330.7\n",
"Training on 4/104 positive/total out of 550 1-step experiences with actions distribution [ 4 92 6 2]\n",
"average reward per episode = -180.4\n",
"Training on 6/105 positive/total out of 360 1-step experiences with actions distribution [ 1 98 2 4]\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"[2017-03-04 11:49:41,426] Starting new video recorder writing to /Users/gui/Dev/rl-study/tmp/q_learning/openaigym.video.0.48905.video000512.mp4\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"average reward per episode = -396.8\n",
"Training on 4/103 positive/total out of 634 1-step experiences with actions distribution [ 1 100 2]\n",
"average reward per episode = -230.3\n",
"Training on 6/104 positive/total out of 419 1-step experiences with actions distribution [ 1 97 4 2]\n",
"average reward per episode = -331.2\n",
"Training on 3/103 positive/total out of 536 1-step experiences with actions distribution [ 3 94 4 2]\n",
"average reward per episode = -232.8\n",
"Training on 5/104 positive/total out of 421 1-step experiences with actions distribution [ 2 99 3]\n",
"average reward per episode = -93.8\n",
"Training on 6/106 positive/total out of 247 1-step experiences with actions distribution [ 2 98 3 3]\n",
"average reward per episode = -82.3\n",
"Training on 13/107 positive/total out of 244 1-step experiences with actions distribution [ 2 104 0 1]\n",
"average reward per episode = -81.9\n",
"Training on 8/107 positive/total out of 246 1-step experiences with actions distribution [ 2 100 3 2]\n",
"average reward per episode = -286.0\n",
"Training on 3/103 positive/total out of 485 1-step experiences with actions distribution [ 1 98 2 2]\n",
"average reward per episode = -206.0\n",
"Training on 7/106 positive/total out of 398 1-step experiences with actions distribution [ 0 102 1 3]\n",
"average reward per episode = -149.5\n",
"Training on 6/106 positive/total out of 330 1-step experiences with actions distribution [ 5 93 5 3]\n",
"average reward per episode = -276.5\n",
"Training on 5/104 positive/total out of 469 1-step experiences with actions distribution [ 1 101 0 2]\n",
"average reward per episode = -196.6\n",
"Training on 7/105 positive/total out of 367 1-step experiences with actions distribution [ 3 98 1 3]\n",
"average reward per episode = -192.5\n",
"Training on 9/107 positive/total out of 395 1-step experiences with actions distribution [ 2 100 3 2]\n",
"average reward per episode = -286.9\n",
"Training on 6/104 positive/total out of 486 1-step experiences with actions distribution [ 2 101 1]\n",
"average reward per episode = -303.1\n",
"Training on 6/104 positive/total out of 511 1-step experiences with actions distribution [ 3 94 2 5]\n",
"average reward per episode = -338.8\n",
"Training on 5/103 positive/total out of 551 1-step experiences with actions distribution [ 2 95 4 2]\n",
"average reward per episode = -229.3\n",
"Training on 5/105 positive/total out of 433 1-step experiences with actions distribution [ 4 95 1 5]\n",
"average reward per episode = -149.4\n",
"Training on 12/106 positive/total out of 331 1-step experiences with actions distribution [ 3 96 1 6]\n",
"average reward per episode = -216.5\n",
"Training on 5/104 positive/total out of 398 1-step experiences with actions distribution [ 2 97 2 3]\n",
"average reward per episode = -311.6\n",
"Training on 4/104 positive/total out of 541 1-step experiences with actions distribution [ 4 94 4 2]\n",
"average reward per episode = -309.3\n",
"Training on 3/102 positive/total out of 504 1-step experiences with actions distribution [ 2 97 2 1]\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"[2017-03-04 11:52:07,127] Starting new video recorder writing to /Users/gui/Dev/rl-study/tmp/q_learning/openaigym.video.0.48905.video000729.mp4\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"average reward per episode = -343.0\n",
"Training on 3/103 positive/total out of 555 1-step experiences with actions distribution [ 3 91 2 7]\n",
"average reward per episode = -357.0\n",
"Training on 0/100 positive/total out of 699 1-step experiences with actions distribution [93 2 3 2]\n",
"average reward per episode = -316.3\n",
"Training on 2/102 positive/total out of 659 1-step experiences with actions distribution [92 4 4 2]\n",
"average reward per episode = -358.5\n",
"Training on 0/100 positive/total out of 699 1-step experiences with actions distribution [95 0 1 4]\n"
]
}
],
"source": [
"import random\n",
"\n",
"class EpsilonGreedyQAgent(object):\n",
" def __init__(self, model, epsilon=.1):\n",
" self.model = model\n",
" self.epsilon = epsilon\n",
"\n",
" def act(self, observation, reward, done):\n",
" if random.uniform(0, 1) <= self.epsilon:\n",
" return random.choice([NOOP, SHOOT, LEFT, RIGHT])\n",
" else:\n",
" return self.model.predict(observation[np.newaxis])[0].argmax()\n",
"\n",
"\n",
"N_BATCHES = 100\n",
"N_BATCHED_EPISODES = 10\n",
"UPDATE_TARGET_EVERY_N_BACTHES = 2\n",
"MINI_BATCH_SIZE = 32\n",
"REWARD_CLIP = 5\n",
"ONLY_N_MISSES = 200\n",
"\n",
"env = create_env()\n",
"env = wrappers.Monitor(env, directory='tmp/q_learning', force=True, mode='training')\n",
"\n",
"for _ in range(N_BATCHES):\n",
" for _ in range(UPDATE_TARGET_EVERY_N_BACTHES):\n",
" sares = episode_sares(env, EpsilonGreedyQAgent(acting_model, epsilon=.1), N_BATCHED_EPISODES)\n",
" prev_frames, target_action_rewards = sares_to_input_targets(target_model, sares, reward_clip=REWARD_CLIP, only_n_misses=ONLY_N_MISSES)\n",
" acting_model.fit(x=prev_frames, y=target_action_rewards, batch_size=MINI_BATCH_SIZE, nb_epoch=1, verbose=0)\n",
" \n",
" target_model = copy_model(acting_model)\n",
"\n",
"\n",
"# final greedy episodes\n",
"sares = episode_sares(env, EpsilonGreedyQAgent(acting_model, epsilon=0), episode_count=1000)\n",
"\n",
"plt.plot(np.cumsum(list(map(operator.attrgetter('reward'), sares))));\n",
"plt.xlabel('steps'); plt.ylabel('Cumulated rewards');\n",
"\n",
"env.close()\n",
"gym.upload('tmp/q_learning', api_key='sk_bNZUvCfkTfabQCoKoKbjFA')"
]
},
{
"cell_type": "code",
"execution_count": 70,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"import json\n",
"model_name = 'DoomBasic-v0_res=200x150_skip=4_discrete=minimal_fc64'\n",
"\n",
"acting_model.save(model_name + '.h5')\n",
"\n",
"with open(model_name + '.json', 'w+') as f:\n",
" json.dump(acting_model.to_json(), f)"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.6.0"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment