sim_est_error_single.py

"""
Please contact the author(s) of this library if you have any questions.
Authors: Kai-Chieh Hsu ( kaichieh@princeton.edu )

1. We want to evaluate how well we learned from the data.
2. We compare the DDQN-predicted value vs. the rollout value by DDQN-induced
policies.
3. This script samples the initial states of testing rollouts. We then specify
    the arguments of the rollout (see help section of arguments for more
    details). The rollout results are stored in
    `{args.modelFolder}/data/{args.outFile}.npy`.

EXAMPLES
    toEnd, low turning rate:
        python3 sim_est_error_single.py -te -l -of carOneLow
          -mf models/store_best/car/RA/small/tanh
    TF, high turning rate:
        python3 sim_est_error_single.py -of carOneHighTF
          -mf models/store_best/car/RA/big/tanh
    Array:
        python3 sim_est_error_single.py -sf
          -mf scratch/car/highA/highA-0-2021-02-10-21_08
"""

import os
import time
from warnings import simplefilter
import argparse
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
import torch
from multiprocessing import Pool
import gym

from utils.carOneAnalysis import loadAgent, loadEnv
from gym_reachability import gym_reachability  # Custom Gym env.

simplefilter(action='ignore', category=FutureWarning)
matplotlib.use('Agg')


def multiExp(env, agent, samples, firstIdx, numSample, maxLength, toEnd):
  freeCoordNum = 2
  shapeTmp = np.ones(freeCoordNum, dtype=int) * numSample
  rolloutResult = np.empty(shape=shapeTmp, dtype=int)
  trajLength = np.empty(shape=shapeTmp, dtype=int)
  ddqnValue = np.empty(shape=shapeTmp, dtype=float)
  rolloutValue = np.empty(shape=shapeTmp, dtype=float)
  it = np.nditer(rolloutResult, flags=['multi_index'])

  while not it.finished:
    idx = it.multi_index
    stateIdx = idx + (firstIdx,)
    print(stateIdx, end='\r')
    state = samples[stateIdx, np.arange(3)]
    traj, result, minV, _ = env.simulate_one_trajectory(
        agent.Q_network, T=maxLength, state=state, toEnd=toEnd
    )
    trajLength[idx] = traj.shape[0]
    rolloutResult[idx] = result  # result \in { 1, -1}
    rolloutValue[idx] = minV

    agent.Q_network.eval()
    stateTensor = torch.from_numpy(state).float().to(agent.device)
    state_action_values = agent.Q_network(stateTensor)
    Q_vec = state_action_values.detach().cpu().reshape(-1)
    ddqnValue[idx] = Q_vec.min().item()

    it.iternext()

  carOneDict = {}
  carOneDict['rolloutResult'] = rolloutResult
  carOneDict['trajLength'] = trajLength
  carOneDict['ddqnValue'] = ddqnValue
  carOneDict['rolloutValue'] = rolloutValue

  print()
  return carOneDict


def run(args):
  startTime = time.time()

  # == ENVIRONMENT ==
  env = loadEnv(args)
  state_dim = env.state.shape[0]
  action_num = env.action_space.n
  action_list = np.arange(action_num)
  device = env.device

  # == AGENT ==
  agent = loadAgent(args, device, state_dim, action_num, action_list)

  # == ROLLOUT RESULTS ==
  print("\n== Estimation Error Information ==")
  np.set_printoptions(precision=2, suppress=True)
  numSample = args.numSample
  bounds = np.array([[-1.1, 1.1], [-1.1, 1.1],
                     [0, 2 * np.pi * (1 - 1/numSample)]])
  samples = np.linspace(start=bounds[:, 0], stop=bounds[:, 1], num=numSample)

  maxLength = args.maxLength
  toEnd = args.toEnd
  carPESubDictList = []
  numThread = args.numWorker
  numTurn = int(numSample / (numThread+1e-6)) + 1
  for ith in range(numTurn):
    print('{} / {}'.format(ith + 1, numTurn))
    with Pool(processes=numThread) as pool:
      startIdx = ith * numThread
      endIdx = min(numSample, (ith+1) * numThread)
      firstIdxList = list(range(startIdx, endIdx))
      print(firstIdxList)
      numExp = len(firstIdxList)
      envList = [env] * numExp
      agentList = [agent] * numExp
      samplesList = [samples] * numExp
      numSampleList = [numSample] * numExp
      maxLengthList = [maxLength] * numExp
      toEndList = [toEnd] * numExp

      carPESubDict_i = pool.starmap(
          multiExp,
          zip(
              envList, agentList, samplesList, firstIdxList, numSampleList,
              maxLengthList, toEndList
          )
      )
    carPESubDictList = carPESubDictList + carPESubDict_i

  # == COMBINE RESULTS ==
  shapeTmp = np.ones(3, dtype=int) * numSample
  rolloutResult = np.empty(shape=shapeTmp, dtype=int)
  trajLength = np.empty(shape=shapeTmp, dtype=int)
  ddqnValue = np.empty(shape=shapeTmp, dtype=float)
  rolloutValue = np.empty(shape=shapeTmp, dtype=float)

  for i, carPESubDict_i in enumerate(carPESubDictList):
    rolloutResult[:, :, i] = carPESubDict_i['rolloutResult']
    trajLength[:, :, i] = carPESubDict_i['trajLength']
    ddqnValue[:, :, i] = carPESubDict_i['ddqnValue']
    rolloutValue[:, :, i] = carPESubDict_i['rolloutValue']

  endTime = time.time()
  execTime = endTime - startTime
  print('--> Execution time: {:.1f}'.format(execTime))

  carOneDict = {}
  carOneDict['numSample'] = numSample
  carOneDict['maxLength'] = maxLength
  carOneDict['toEnd'] = toEnd
  carOneDict['rolloutResult'] = rolloutResult
  carOneDict['trajLength'] = trajLength
  carOneDict['ddqnValue'] = ddqnValue
  carOneDict['rolloutValue'] = rolloutValue
  carOneDict['samples'] = samples
  carOneDict['execTime'] = execTime

  outFolder = os.path.join(args.modelFolder, 'data')
  os.makedirs(outFolder, exist_ok=True)
  outFile = os.path.join(outFolder, args.outFile + '.npy')
  np.save('{:s}'.format(outFile), carOneDict)
  print('--> Save to {:s} ...'.format(outFile))

  # == Plot RA Set of the analytic solution and approximate value function ==
  if args.plotFigure or args.storeFigure:
    fig, axes = plt.subplots(1, 2, figsize=(8, 4), sharex=True, sharey=True)
    axStyle = np.array([-1.1, 1.1, -1.1, 1.1])

    ax = axes[0]
    ax.imshow(
        carOneDict['ddqnValue'][:, :, 0].T, interpolation='none',
        extent=axStyle, origin="lower", cmap='seismic', vmin=-1, vmax=1
    )
    env.plot_reach_avoid_set(ax, c='g', lw=3, orientation=0)
    ax.set_xlabel(r'$\theta={:.0f}^\circ$'.format(0), fontsize=24)

    # == Rollout ==
    ax = axes[1]
    ax.imshow(
        carOneDict['rolloutValue'][:, :, 0].T <= 0, interpolation='none',
        extent=axStyle, origin="lower", cmap='coolwarm', vmin=0, vmax=1
    )
    ax.set_xlabel('Rollout', fontsize=24)

    # == Formatting ==
    for ax in axes:
      env.plot_target_failure_set(ax=ax)
      env.plot_formatting(ax=ax)

    if args.storeFigure:
      figureFolder = os.path.join(args.modelFolder, 'figure')
      os.makedirs(figureFolder, exist_ok=True)
      plt.savefig(os.path.join(figureFolder, 'rollout.png'))
    if args.plotFigure:
      plt.show()
      plt.pause(0.001)
      plt.close()


if __name__ == '__main__':
  parser = argparse.ArgumentParser()
  # Environment Parameters
  parser.add_argument(
      "-low", "--low", help="low turning rate", action="store_true"
  )

  # Simulation Parameters
  parser.add_argument(
      "-te", "--toEnd",
      help="continue the rollout until the car crosses the boundary",
      action="store_true"
  )
  parser.add_argument(
      "-f", "--forceCPU", help="force PyTorch to use CPU", action="store_true"
  )
  parser.add_argument(
      "-ns", "--numSample", help="#samples", default=101, type=int
  )
  parser.add_argument(
      "-nw", "--numWorker", help="#workers", default=5, type=int
  )
  parser.add_argument(
      "-ml", "--maxLength", help="maximum length of rollout episodes",
      default=100, type=int
  )

  # File Parameters
  parser.add_argument(
      "-pf", "--plotFigure", help="plot figures", action="store_true"
  )
  parser.add_argument(
      "-sf", "--storeFigure", help="store figures", action="store_true"
  )
  parser.add_argument(
      "-of", "--outFile", help="output file", default='estError', type=str
  )
  parser.add_argument("-mf", "--modelFolder", help="model folder", type=str)

  args = parser.parse_args()
  print("\n== Arguments ==")
  print(args)

  # == Execution ==
  run(args)