# https://iopscience.iop.org/article/10.1088/0143-0807/37/6/065008/pdf # Importantly they depict the threshold # for admissible acceleration onset or jerk as j = 15g/s or ~150m/s^3. jerk_threshold=150.0, # 15g/s incent_win=True, constrain_controls=False, incent_yield_to_oncoming_traffic=True, physics_steps_per_observation=12, discrete_actions=COMFORTABLE_ACTIONS2, ) net_config = dict(hidden_units=(256, 256, 256), activation=torch.nn.Tanh) eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) # eg.add('seed', 0) eg.add( 'resume', '/home/c2/src/tmp/spinningup/data/intersection_discrete_micro_turn_lower_lane_pen2_diag_lane20/intersection_discrete_micro_turn_lower_lane_pen2_diag_lane20_s0_2020_05-16_12-44.32.941234' ) # eg.add('reinitialize_optimizer_on_resume', True) # eg.add('num_inputs_to_add', 0) eg.add('pi_lr', 3e-4) # default pi_lr=3e-4 eg.add('vf_lr', 1e-3) # default vf_lr=1e-3, # eg.add('boost_explore', 5) eg.add('epochs', 20000) eg.add('steps_per_epoch', 8000) eg.add('ac_kwargs:hidden_sizes', net_config['hidden_units'], 'hid') eg.add('ac_kwargs:activation', net_config['activation'], '') eg.add('notes', notes, '')
import torch from gym_match_input_continuous.experiments import utils experiment_name = os.path.basename(__file__)[:-3] notes = """ Trying to shift advantages by one standard deviation vs 90: https://photos.app.goo.gl/MZ9p9sdjFBf3yZba8 """ env_config = dict(env_name='match-input-continuous-v0', ) net_config = dict(hidden_units=(32, 32), activation=torch.nn.Tanh) eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) # eg.add('gamma', 0.999) # Lower gamma so seconds of effective horizon remains at 10s with current physics steps = 12 * 1/60s * 1 / (1-gamma) eg.add('epochs', 1000) eg.add('steps_per_epoch', 500) eg.add('shift_advs_pct', 68) eg.add('ac_kwargs:hidden_sizes', net_config['hidden_units'], 'hid') eg.add('ac_kwargs:activation', net_config['activation'], '') eg.add('notes', notes, '') eg.add('run_filename', os.path.realpath(__file__), '') eg.add('env_config', env_config, '') def train(): eg.run(ppo_pytorch)
def run_vpg_lava(): import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=4) parser.add_argument('--num_runs', type=int, default=3) args = parser.parse_args() eg = ExperimentGrid(name='vpg-pt-bench') eg.add('env_name', 'MiniGrid-LavaCrossingS9N2-v1', '', True) eg.add('seed', [6 * i for i in range(args.num_runs)]) eg.add('epochs', 500) eg.add('steps_per_epoch', 5000) eg.add('max_ep_len', 200) # eg.add('ac_kwargs:hidden_sizes', [(32,), (64,64)], 'hid') eg.add('ac_kwargs:hidden_sizes', [(32, )], 'hid') # eg.add('ac_kwargs:activation', [torch.nn.Tanh, torch.nn.ReLU], '') eg.add('ac_kwargs:activation', [torch.nn.Tanh], '') eg.run(vpg_pytorch, num_cpu=args.cpu, datestamp=True)
gforce_penalty_coeff=0.06, collision_penalty_coeff=4, lane_penalty_coeff=0.02, speed_reward_coeff=0.50, end_on_harmful_gs=False, end_on_lane_violation=False, incent_win=True, constrain_controls=False, incent_yield_to_oncoming_traffic=True, physics_steps_per_observation=12, ) net_config = dict(hidden_units=(256, 256), activation=torch.nn.Tanh) eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) # eg.add('seed', 0) eg.add( 'resume', '/workspace/dd0-data-resume1/intersection_2_agents_fine_tune_add_left_yield_from_scratch/intersection_2_agents_fine_tune_add_left_yield_from_scratch_s0_2020_03-25_21-16.12/best_EpRet/2020_03-27_11-01.22' ) # eg.add('reinitialize_optimizer_on_resume', True) # eg.add('num_inputs_to_add', 0) eg.add('pi_lr', 3e-6) eg.add('vf_lr', 1e-5) # eg.add('boost_explore', 5) eg.add('epochs', 8000) eg.add('steps_per_epoch', 32000) eg.add('ac_kwargs:hidden_sizes', net_config['hidden_units'], 'hid') eg.add('ac_kwargs:activation', net_config['activation'], '') eg.add('notes', notes, '')
expect_normalized_action_deltas=False, jerk_penalty_coeff=0.10 / (60*10), # dropped from 0.10, see notes gforce_penalty_coeff=0.031, collision_penalty_coeff=1, gforce_threshold=None, incent_win=True, constrain_controls=False, ) net_config = dict( hidden_units=(256, 256), activation=torch.nn.Tanh ) eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) # eg.add('seed', 0) eg.add('resume', '/home/c2/src/tmp/spinningup/data/intersection_2_agents_fine_tune_collision_resume_add_comfort/intersection_2_agents_fine_tune_collision_resume_add_comfort_s0_2020_03-13_15-24.07/best_trip_pct/2020_03-13_16-18.59') eg.add('reinitialize_optimizer_on_resume', True) eg.add('pi_lr', 3e-6) # doesn't seem to have an effect, but playing it safe and lowering learning rate since we're not restoring adam rates eg.add('vf_lr', 1e-5) # doesn't seem to have an effect, but playing it safe and lowering learning rate since we're not restoring adam rates eg.add('epochs', 8000) eg.add('steps_per_epoch', 8000) eg.add('ac_kwargs:hidden_sizes', net_config['hidden_units'], 'hid') eg.add('ac_kwargs:activation', net_config['activation'], '') eg.add('notes', notes, '') eg.add('run_filename', os.path.realpath(__file__), '') eg.add('env_config', env_config, '') def train(): eg.run(ppo_pytorch)
def parse_and_execute_grid_search(cmd, args): """Interprets algorithm name and cmd line args into an ExperimentGrid.""" # Parse which algorithm to execute algo = eval('spinup.' + cmd) # Before all else, check to see if any of the flags is 'help'. valid_help = ['--help', '-h', 'help'] if any([arg in valid_help for arg in args]): print('\n\nShowing docstring for spinup.' + cmd + ':\n') print(algo.__doc__) sys.exit() def process(arg): # Process an arg by eval-ing it, so users can specify more # than just strings at the command line (eg allows for # users to give functions as args). try: return eval(arg) except: return arg # Make first pass through args to build base arg_dict. Anything # with a '--' in front of it is an argument flag and everything after, # until the next flag, is a possible value. arg_dict = dict() for i, arg in enumerate(args): assert i > 0 or '--' in arg, \ friendly_err("You didn't specify a first flag.") if '--' in arg: arg_key = arg.lstrip('-') arg_dict[arg_key] = [] else: arg_dict[arg_key].append(process(arg)) # Make second pass through, to catch flags that have no vals. # Assume such flags indicate that a boolean parameter should have # value True. for k, v in arg_dict.items(): if len(v) == 0: v.append(True) # Third pass: check for user-supplied shorthands, where a key has # the form --keyname[kn]. The thing in brackets, 'kn', is the # shorthand. NOTE: modifying a dict while looping through its # contents is dangerous, and breaks in 3.6+. We loop over a fixed list # of keys to avoid this issue. given_shorthands = dict() fixed_keys = list(arg_dict.keys()) for k in fixed_keys: p1, p2 = k.find('['), k.find(']') if p1 >= 0 and p2 >= 0: # Both '[' and ']' found, so shorthand has been given k_new = k[:p1] shorthand = k[p1 + 1:p2] given_shorthands[k_new] = shorthand arg_dict[k_new] = arg_dict[k] del arg_dict[k] # Penultimate pass: sugar. Allow some special shortcuts in arg naming, # eg treat "env" the same as "env_name". This is super specific # to Spinning Up implementations, and may be hard to maintain. # These special shortcuts are described by SUBSTITUTIONS. for special_name, true_name in SUBSTITUTIONS.items(): if special_name in arg_dict: # swap it in arg dict arg_dict[true_name] = arg_dict[special_name] del arg_dict[special_name] if special_name in given_shorthands: # point the shortcut to the right name given_shorthands[true_name] = given_shorthands[special_name] del given_shorthands[special_name] # Final pass: check for the special args that go to the 'run' command # for an experiment grid, separate them from the arg dict, and make sure # that they have unique values. The special args are given by RUN_KEYS. run_kwargs = dict() for k in RUN_KEYS: if k in arg_dict: val = arg_dict[k] assert len(val)==1, \ friendly_err("You can only provide one value for %s."%k) run_kwargs[k] = val[0] del arg_dict[k] # Determine experiment name. If not given by user, will be determined # by the algorithm name. if 'exp_name' in arg_dict: assert len(arg_dict['exp_name'])==1, \ friendly_err("You can only provide one value for exp_name.") exp_name = arg_dict['exp_name'][0] del arg_dict['exp_name'] else: exp_name = 'cmd_' + cmd # Make sure that if num_cpu > 1, the algorithm being used is compatible # with MPI. if 'num_cpu' in run_kwargs and not (run_kwargs['num_cpu'] == 1): assert cmd in MPI_COMPATIBLE_ALGOS, \ friendly_err("This algorithm can't be run with num_cpu > 1.") # Special handling for environment: make sure that env_name is a real, # registered gym environment. valid_envs = [e.id for e in list(gym.envs.registry.all())] assert 'env_name' in arg_dict, \ friendly_err("You did not give a value for --env_name! Add one and try again.") for env_name in arg_dict['env_name']: err_msg = dedent(""" %s is not registered with Gym. Recommendations: * Check for a typo (did you include the version tag?) * View the complete list of valid Gym environments at https://gym.openai.com/envs/ """ % env_name) assert env_name in valid_envs, err_msg # Construct and execute the experiment grid. eg = ExperimentGrid(name=exp_name) for k, v in arg_dict.items(): eg.add(k, v, shorthand=given_shorthands.get(k)) eg.run(algo, **run_kwargs)
logger.log_tabular('Time', time.time() - start_time) logger.dump_tabular() if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--env', type=str, default='HalfCheetah-v2') parser.add_argument('--h', type=int, default=300) parser.add_argument('--l', type=int, default=1) parser.add_argument('--num_runs', '-n', type=int, default=3) parser.add_argument('--steps_per_epoch', '-s', type=int, default=5000) parser.add_argument('--total_steps', '-t', type=int, default=int(5e4)) args = parser.parse_args() def td3_with_actor_critic(**kwargs): td3(ac_kwargs=dict(hidden_sizes=[args.h] * args.l), start_steps=5000, max_ep_len=150, batch_size=64, polyak=0.95, **kwargs) eg = ExperimentGrid(name='ex2-3_td3') eg.add('replay_size', int(args.total_steps)) eg.add('env_name', args.env, '', True) eg.add('seed', [10 * i for i in range(args.num_runs)]) eg.add('epochs', int(args.total_steps / args.steps_per_epoch)) eg.add('steps_per_epoch', args.steps_per_epoch) eg.add('remove_action_clip', [False, True]) eg.run(td3_with_actor_critic, datestamp=True)
from spinup import sac_pytorch import torch import gym TASK_HORIZON = 200 NUM_TASKS = 10 if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=8) parser.add_argument('--num_runs', type=int, default=1) args = parser.parse_args() eg = ExperimentGrid(name='sac-MT10') eg.add('env_name', 'MT10Helper-v0', '', True) eg.add('num_tasks', [10]) eg.add('seed', [10 * i for i in range(args.num_runs)]) eg.add('epochs', 1000) eg.add('steps_per_epoch', TASK_HORIZON * NUM_TASKS) eg.add('lr', [3e-4]) eg.add('start_steps', 1000) eg.add('ac_kwargs:hidden_sizes', [(400, 400)], 'hid') eg.add('ac_kwargs:activation', [torch.nn.ReLU], '') eg.run(sac_pytorch, num_cpu=args.cpu) #from metaworld.benchmarks import MT10 # #env_fn = lambda : MTEnv(MT10.get_train_tasks()) # #ac_kwargs = dict(hidden_sizes=[400,400], activation=torch.nn.ReLU) #
from spinup.utils.run_utils import ExperimentGrid from spinup import dvpg_pytorch import torch if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=2) parser.add_argument('--num_runs', type=int, default=10) args = parser.parse_args() eg = ExperimentGrid(name='dvpg_LunarLander-v2') eg.add('env_name', 'LunarLander-v2', '', True) eg.add('seed', [10 * i for i in range(args.num_runs)]) eg.add('epochs', 500) eg.add('epsilon', [0.01, 0.1, 1, 10]) eg.add('pi_lr', [3e-4, 3e-3, 3e-2]) eg.add('steps_per_epoch', 4000) eg.add('ac_kwargs:hidden_sizes', [(32, 32)], 'hid') eg.add('ac_kwargs:activation', [torch.nn.Tanh], '') eg.run(dvpg_pytorch, num_cpu=args.cpu)
def runExperiment(someEnv,someOptimizer, someActivation): env = ExperimentGrid(name='vpg-trece') env.add('env_name', someEnv, '', True) # eg.add('clip_ratio', [0.1,0.2]) env.add('seed', [10*i for i in range(args.num_runs)]) env.add('epochs', 10) env.add('steps_per_epoch', [4000]) #someOptimizer should be list env.add('optimizer',[someOptimizer]) env.add('ac_kwargs:hidden_sizes', [(32,), (64,64)], 'hid') env.add('ac_kwargs:activation', [someActivation], '') env.run(vpg, num_cpu=args.cpu)
jerk_penalty_coeff=0.20 / (60*100), gforce_penalty_coeff=0.06, collision_penalty_coeff=4, end_on_harmful_gs=False, incent_win=True, constrain_controls=False, incent_yield_to_oncoming_traffic=True, ) net_config = dict( hidden_units=(256, 256), activation=torch.nn.Tanh ) eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) # eg.add('seed', 0) eg.add('resume', '/home/c2/src/tmp/spinningup/data/intersection_2_agents_fine_tune_add_left_yield/intersection_2_agents_fine_tune_add_left_yield_s0_2020_03-23_13-16.15') eg.add('reinitialize_optimizer_on_resume', False) # Old optimizer had only 28 inputs despite NN having 29! eg.add('num_inputs_to_add', 0) eg.add('pi_lr', 3e-6) eg.add('vf_lr', 1e-5) eg.add('boost_explore', 5) eg.add('epochs', 8000) eg.add('steps_per_epoch', 32000) eg.add('ac_kwargs:hidden_sizes', net_config['hidden_units'], 'hid') eg.add('ac_kwargs:activation', net_config['activation'], '') eg.add('notes', notes, '') eg.add('run_filename', os.path.realpath(__file__), '') eg.add('env_config', env_config, '')
help= 'if set to true, then output out the material type first, then condition the material thickness on the material type' ) parser.add_argument('--use_rnn', action='store_true') parser.add_argument('--spectrum_repr', action='store_true') args = parser.parse_args() env_kwargs = { "discrete_thick": args.discrete_thick, 'spectrum_repr': args.spectrum_repr, 'bottom_up': False, 'merit_func': cal_reward } eg = ExperimentGrid(name=args.exp_name) eg.add('env_fn', get_env_fn(args.env, **env_kwargs)) eg.add('seed', [42 * (i + 1) for i in range(args.num_runs)]) eg.add('epochs', 3000) eg.add('steps_per_epoch', 1000) eg.add('ac_kwargs:hidden_sizes', [(64, )], 'hid') eg.add('ac_kwargs:cell_size', 64, '') eg.add('ac_kwargs:not_repeat', [True, False]) eg.add('ac_kwargs:ortho_init', ['on']) eg.add('ac_kwargs:hierarchical', [True, False]) eg.add('ac_kwargs:channels', 16) eg.add('ac_kwargs:act_emb', [True]) eg.add('ac_kwargs:act_emb_dim', 5) eg.add('use_rnn', args.use_rnn) eg.add('gamma', 1) eg.add('beta', [0.01]) eg.add('lam', [0.95])
from spinup.utils.run_utils import ExperimentGrid from spinup import vpg_pytorch import torch if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=8) parser.add_argument('--num_runs', type=int, default=5) args = parser.parse_args() eg = ExperimentGrid(name='vpg_LunarLander-v2') eg.add('env_name', 'LunarLander-v2', '', True) eg.add('seed', [10*i for i in range(args.num_runs)]) eg.add('epochs', 500) eg.add('pi_lr', [3e-4, 3e-3, 3e-2]) eg.add('steps_per_epoch', 4000) eg.add('ac_kwargs:hidden_sizes', [(32,32)], 'hid') eg.add('ac_kwargs:activation', [torch.nn.Tanh], '') eg.run(vpg_pytorch, num_cpu=args.cpu)
from spinup.utils.run_utils import ExperimentGrid from spinup import sac_pytorch import torch import gym if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=4) parser.add_argument('--num_runs', type=int, default=1) args = parser.parse_args() eg = ExperimentGrid(name='sac-goal') eg.add('env_name', 'SawyerPush-v0', '', True) eg.add('seed', [10 * i for i in range(args.num_runs)]) eg.add('epochs', 300) eg.add('steps_per_epoch', 10 * 150) eg.add('lr', [3e-4, 1e-3]) eg.add('start_steps', 1000) eg.add('ac_kwargs:hidden_sizes', [(400, 400)], 'hid') eg.add('ac_kwargs:activation', [torch.nn.ReLU], '') eg.run(sac_pytorch, num_cpu=args.cpu) #from metaworld.benchmarks import MT10 # #env_fn = lambda : MTEnv(MT10.get_train_tasks()) # #ac_kwargs = dict(hidden_sizes=[400,400], activation=torch.nn.ReLU) # #logger_kwargs = dict(output_dir='~/spinup/data/', exp_name='SAC_MT10') # #sac_pytorch(env_fn=env_fn, ac_kwargs=ac_kwargs, steps_per_epoch=128 * 10, epochs=1000, start_steps=1000, lr=3e-4, logger_kwargs=logger_kwargs)
# https://iopscience.iop.org/article/10.1088/0143-0807/37/6/065008/pdf # Importantly they depict the threshold # for admissible acceleration onset or jerk as j = 15g/s or ~150m/s^3. jerk_threshold=150.0, # 15g/s incent_win=True, constrain_controls=False, incent_yield_to_oncoming_traffic=True, physics_steps_per_observation=12, discrete_actions=COMFORTABLE_ACTIONS2, ) net_config = dict(hidden_units=(256, 256), activation=torch.nn.Tanh) eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) # eg.add('seed', 0) eg.add( 'resume', '/home/c2/src/tmp/spinningup/data/intersection_discrete_micro_turn_lower_lane_pen2_diag_lane14_2/intersection_discrete_micro_turn_lower_lane_pen2_diag_lane14_2_s0_2020_05-12_18-26.22.747749' ) # eg.add('reinitialize_optimizer_on_resume', True) # eg.add('num_inputs_to_add', 0) # eg.add('pi_lr', 3e-6) # eg.add('vf_lr', 1e-5) # eg.add('boost_explore', 5) eg.add('epochs', 20000) eg.add('steps_per_epoch', 8000) eg.add('ac_kwargs:hidden_sizes', net_config['hidden_units'], 'hid') eg.add('ac_kwargs:activation', net_config['activation'], '') eg.add('notes', notes, '')
# for admissible acceleration onset or jerk as j = 15g/s or ~150m/s^3. jerk_threshold=150.0, # 15g/s incent_win=True, constrain_controls=False, incent_yield_to_oncoming_traffic=True, physics_steps_per_observation=12, discrete_actions=COMFORTABLE_ACTIONS, ) net_config = dict( hidden_units=(256, 256), activation=torch.nn.Tanh ) eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) # eg.add('seed', 0) eg.add('resume', '/home/c2/src/tmp/spinningup/data/intersection_from_scratch_discrete_steer_and_accel_resume_snapshot/intersection_from_scratch_discrete_steer_and_accel_resume_s0_2020_04-22_13-55.32.354696') # eg.add('reinitialize_optimizer_on_resume', True) # eg.add('num_inputs_to_add', 0) # eg.add('pi_lr', 3e-6) # eg.add('vf_lr', 1e-5) # eg.add('boost_explore', 5) eg.add('epochs', 20000) eg.add('steps_per_epoch', 4000) eg.add('ac_kwargs:hidden_sizes', net_config['hidden_units'], 'hid') eg.add('ac_kwargs:activation', net_config['activation'], '') eg.add('notes', notes, '') eg.add('run_filename', os.path.realpath(__file__), '') eg.add('env_config', env_config, '')
remainder = remainder % division indexes.append(index) total = division actual_setting = {} for j in range(len(indexes)): actual_setting[setting_names[j]] = settings[j][indexes[j]] return indexes, actual_setting indexes, actual_setting = get_setting(args.setting, total, settings, setting_names) eg = ExperimentGrid(name=EXPERIMENT_NAME) # use eg.add to add parameters in the settings or add parameters tha apply to all jobs # we now automated this part, as long as you added settings correctly into the arrays at the start of this program # they should be added to experiment automatically for i in range(len(actual_setting)): setting_name = setting_names[i] if setting_name != 'env_name' and setting_name != 'seed': eg.add(setting_name, actual_setting[setting_name], setting_savename_prefix[i], whether_add_to_savename[i]) eg.add('env_name', actual_setting['env_name'], '', True) eg.add('seed', actual_setting['seed']) eg.run(function_to_run, num_cpu=args.cpu, data_dir=save_data_dir) print( '\n###################################### GRID EXP END ######################################' ) print('total time for grid experiment:', time.time() - start_time)
from spinup.utils.run_utils import ExperimentGrid from spinup import ppo from spinup import vpg import tensorflow as tf if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=4) parser.add_argument('--num_runs', type=int, default=3) args = parser.parse_args() eg = ExperimentGrid(name='vpg-bench') eg.add('env_name', 'Pong-ram-v0', '', True) #eg.add('seed', [10*i for i in range(args.num_runs)]) eg.add('epochs', 10000) #eg.add('steps_per_epoch', 4000) eg.add('ac_kwargs:hidden_sizes', [200], 'hid') eg.add('ac_kwargs:activation', [tf.nn.relu], '') eg.run(vpg, num_cpu=args.cpu)
from spinup.utils.run_utils import ExperimentGrid from spinup import ppo import tensorflow as tf if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=4) parser.add_argument('--num_runs', type=int, default=3) args = parser.parse_args() eg = ExperimentGrid(name='ppo-bench') eg.add('env_name', 'CartPole-v0', '', True) eg.add('seed', [10 * i for i in range(args.num_runs)]) eg.add('epochs', 10) eg.add('steps_per_epoch', 4000) eg.add('ac_kwargs:hidden_sizes', [(32, ), (64, 64)], 'hid') eg.add('ac_kwargs:activation', [tf.tanh, tf.nn.relu], '') eg.run(ppo, num_cpu=args.cpu)
def train(): eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) # eg.add('seed', 0) eg.add( 'resume', '/home/c2/src/tmp/spinningup/data/deepdrive-2d-intersection-no-constrained-controls-example/deepdrive-2d-intersection-no-constrained-controls-example_s0_2020_03-10_13-14.50/best_HorizonReturn/2020_03-11_11-36.27' ) eg.add('reinitialize_optimizer_on_resume', False) eg.add( 'pi_lr', 3e-5 ) # doesn't seem to have an effect, but playing it safe and lowering learning rate since we're not restoring adam rates eg.add( 'vf_lr', 1e-4 ) # doesn't seem to have an effect, but playing it safe and lowering learning rate since we're not restoring adam rates eg.add('epochs', 8000) # eg.add('steps_per_epoch', 4000) eg.add('ac_kwargs:hidden_sizes', (256, 256), 'hid') eg.add('ac_kwargs:activation', torch.nn.Tanh, '') eg.add('notes', notes, '') eg.add('run_filename', os.path.realpath(__file__), '') eg.add('env_config', env_config, '') eg.run(ppo_pytorch)
# Use default hidden sizes in actor_critic function, comment below out eg.add('ac_kwargs:hidden_sizes', [(16,16)], 'hid') eg.add('ac_kwargs:activation', [tf.nn.relu], '') eg.run(algo[i], num_cpu=args.cpu) ''' #Training if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=1) parser.add_argument('--num_runs', type=int, default=1) args = parser.parse_args() eg = ExperimentGrid(name=ex_number) eg.add('env_name', env_name, '', True) eg.add('seed', [10*i for i in range(args.num_runs)]) eg.add('epochs', 5) #eg.add('steps_per_epoch', 4000) #eg.add('max_ep_len', 1500) eg.add('saveCheckpointEpochStep', 1) eg.add('ac_kwargs:activation', [tf.nn.relu], '') eg.add('ac_kwargs:hidden_sizes', [(8,8)], 'hid') eg.run(ddpg, num_cpu=args.cpu)
from spinup.utils.run_utils import ExperimentGrid from spinup import ppo_pytorch, vpg_pytorch import torch if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=4) parser.add_argument('--num_runs', type=int, default=3) parser.add_argument('--algo', type=str, default='vpg', choices=['vpg', 'ppo']) args = parser.parse_args() eg = ExperimentGrid(name=args.algo+'-pyt-bench') eg.add('env_name', 'sorty:sorty-v0', 'n=6', True) eg.add('seed', [10*i for i in range(args.num_runs)]) eg.add('epochs', 10) eg.add('steps_per_epoch', 4000) eg.add('ac_kwargs:hidden_sizes', [(32,), (64,64)], 'hid') eg.add('ac_kwargs:activation', [torch.nn.Tanh, torch.nn.ReLU], '') if args.algo == 'vpg': eg.run(vpg_pytorch, num_cpu=args.cpu) elif args.algo == 'ppo': eg.run(ppo_pytorch, num_cpu=args.cpu) else: raise ValueError("Unsupported Training Algorithm!")
from spinup import vpg import tensorflow as tf # ENVS # BipedalWalkerHardcore-v2 # LunarLanderContinuous-v2 # MontezumaRevenge-ram-v0 # Enduro-ram-v0 # MsPacman-ram-v0 # Ant-v2 # HumanoidStandup-v2 if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=1) parser.add_argument('--num_runs', type=int, default=10) # parser.add_argument('--clip_ratio', type=int, ) args = parser.parse_args() Enduro = ExperimentGrid(name='vpg-nueve-singular') Enduro.add('env_name', 'Enduro-ram-v0', '', True) # eg.add('clip_ratio', [0.1,0.2]) Enduro.add('seed', [10*i for i in range(args.num_runs)]) Enduro.add('epochs', 10) Enduro.add('steps_per_epoch', [4000,100]) Enduro.add('optimizer',['GradientDescentOptimizer', 'MomentumOptimizer', 'ProximalAdagradOptimizer', 'ProximalGradientDescentOptimizer', 'RMSPropOptimizer', 'AdaMaxOptimizer', 'AdamGSOptimizer', 'AdamWOptimizer', 'AddSignOptimizer', 'GGTOptimizer', 'LARSOptimizer', 'LazyAdamGSOptimizer', 'LazyAdamOptimizer', 'MomentumWOptimizer', 'NadamOptimizer', 'PowerSignOptimizer', 'RegAdagradOptimizer', 'ShampooOptimizer']) Enduro.add('ac_kwargs:hidden_sizes', [(32,), (64,64)], 'hid') Enduro.add('ac_kwargs:activation', [tf.nn.tanh], '') Enduro.run(vpg, num_cpu=args.cpu) #, 'ProximalAdagradOptimizer', 'ProximalGradientDescentOptimizer', 'RMSPropOptimizer', 'AdaMaxOptimizer', 'AdamGSOptimizer', 'AdamWOptimizer', 'AddSignOptimizer', 'GGTOptimizer', 'LARSOptimizer', 'LazyAdamGSOptimizer', 'LazyAdamOptimizer', 'MomentumWOptimizer', 'NadamOptimizer', 'PowerSignOptimizer', 'RegAdagradOptimizer', 'ShampooOptimizer' # , , , , tf.nn.selu, tf.nn.softplus, tf.nn.softsign, tf.sigmoid, tf.tanh
from spinup import vpg import tensorflow as tf # ENVS # BipedalWalkerHardcore-v2 # LunarLanderContinuous-v2 # MontezumaRevenge-ram-v0 # Enduro-ram-v0 # MsPacman-ram-v0 # Ant-v2 # HumanoidStandup-v2 if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument('--cpu', type=int, default=4) parser.add_argument('--num_runs', type=int, default=10) # parser.add_argument('--clip_ratio', type=int, ) args = parser.parse_args() HumanoidStandup = ExperimentGrid(name='vpg-ocho') HumanoidStandup.add('env_name', 'HumanoidStandup-v2', '', True) # eg.add('clip_ratio', [0.1,0.2]) HumanoidStandup.add('seed', [10*i for i in range(args.num_runs)]) HumanoidStandup.add('epochs', 10) HumanoidStandup.add('steps_per_epoch', [4000,100]) HumanoidStandup.add('optimizer',['LARSOptimizer', 'LazyAdamGSOptimizer', 'LazyAdamOptimizer', 'MomentumWOptimizer', 'NadamOptimizer', 'PowerSignOptimizer', 'RegAdagradOptimizer', 'ShampooOptimizer']) HumanoidStandup.add('ac_kwargs:hidden_sizes', [(300,), (128,128)], 'hid') HumanoidStandup.add('ac_kwargs:activation', [tf.nn.relu, tf.nn.relu6, tf.nn.crelu, tf.nn.elu, tf.nn.selu, tf.nn.softplus, tf.nn.softsign, tf.sigmoid, tf.tanh], '') HumanoidStandup.run(vpg, num_cpu=args.cpu)
def train(): eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) # eg.add('seed', 0) eg.add('epochs', 8000) eg.add('gamma', 0.95) eg.add('lam', 0.835) # eg.add('steps_per_epoch', 4000) eg.add('ac_kwargs:hidden_sizes', (256, 256), 'hid') eg.add('ac_kwargs:activation', torch.nn.Tanh, '') eg.add('notes', notes, '') eg.add('run_filename', os.path.realpath(__file__), '') eg.add('env_config', env_config, '') eg.run(ppo_pytorch)
def run_experiment(args): def env_fn(): import offload_env # registers custom envs to gym env registry env = gym.make('offload-v0') return env eg = ExperimentGrid(name=args.exp_name) eg.add('env_fn', env_fn) eg.add('seed', [10*i for i in range(args.num_runs)]) eg.add('epochs', 10) eg.add('steps_per_epoch', 10000) eg.add('save_freq', 20) eg.add('ac_kwargs:hidden_sizes', [(32,), (64,64)], 'hid') eg.add('ac_kwargs:activation', [tf.tanh, tf.nn.relu], '') eg.run(vpg, num_cpu=args.cpu)
env_name='deepdrive-2d-intersection-w-gs-allow-decel-v0', is_intersection_map=True, expect_normalized_action_deltas=False, jerk_penalty_coeff=0.20 / (60 * 100), gforce_penalty_coeff=0.06, collision_penalty_coeff=4, gforce_threshold=None, incent_win=True, constrain_controls=False, incent_yield_to_oncoming_traffic=True, ) net_config = dict(hidden_units=(256, 256), activation=torch.nn.Tanh) eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) # eg.add('seed', 0) eg.add( 'resume', '/home/c2/src/tmp/spinningup/data/intersection_2_agents_fine_tune_add_left_yield2/intersection_2_agents_fine_tune_add_left_yield2_s0_2020_03-23_22-40.11' ) eg.add('reinitialize_optimizer_on_resume', True) eg.add('num_inputs_to_add', 0) eg.add('pi_lr', 3e-6) eg.add('vf_lr', 1e-5) # eg.add('boost_explore', 5) eg.add('epochs', 8000) eg.add('steps_per_epoch', 32000) eg.add('ac_kwargs:hidden_sizes', net_config['hidden_units'], 'hid') eg.add('ac_kwargs:activation', net_config['activation'], '') eg.add('notes', notes, '')
def run_experiment(args): # def env_fn(): # import flexibility # register flexibility to gym env registry # return gym.make(args.env_name) eg = ExperimentGrid(name=args.exp_name) eg.add('seed', [10*i for i in range(args.num_runs)] if args.seed is None else args.seed) eg.add('epochs', args.epochs) eg.add('steps_per_epoch', args.steps_per_epoch) eg.add('save_freq', args.save_freq) eg.add('max_ep_len', 200) eg.add('ac_kwargs:activation', eval(args.act), '') eg.add('custom_h', args.custom_h) eg.add('do_checkpoint_eval', args.do_checkpoint_eval) eg.add('eval_episodes', args.eval_episodes) eg.add('train_v_iters', args.train_v_iters) eg.add('eval_temp', args.eval_temp) eg.add('train_starting_temp', args.train_starting_temp) eg.add('gamma', args.gamma) eg.add('env_version', args.env_version) eg.add('env_name', args.env_name) eg.add('env_subtract_full_flex', args.env_subtract_full_flex) eg.add('meta_learning', args.meta_learning) eg.add('finetune', args.finetune) eg.add('finetune_model_path', args.finetune_model_path) eg.add('lam', args.lam) eg.add('early_stop_epochs', args.early_stop_epochs) eg.add('save_all_eval', args.save_all_eval) if args.episodes_per_epoch is not None: eg.add('episodes_per_epoch', args.episodes_per_epoch) if args.env_version >= 3: # args.file_path = "/home/user/git/spinningup/spinup/FlexibilityEnv/input_m8n12_cv0.8.pkl" prefix = os.getcwd().split('RL_flex_design')[0] args.file_path = prefix + "RL_flex_design/spinup/FlexibilityEnv_input/{}".format(args.env_input) m, n, mean_c, mean_d, sd_d, profit_mat, target_arcs, fixed_costs, flex_0 = load_FlexibilityEnv_input(args.file_path) eg.add('env_input', args.file_path) eg.add('env_n_sample', args.env_n_sample) if args.target_arcs is None: eg.add('target_arcs', target_arcs) else: # target_arcs is explicitly specified by the scripts, which overrides the target_arc from the input file eg.add('target_arcs', args.target_arcs) if args.algo == "ppo": eg.add('train_pi_iters', args.train_pi_iters) eg.run(ppo) elif args.algo == "vpg": eg.run(vpg)
jerk_penalty_coeff=3.3e-5, gforce_penalty_coeff=0.006 * 5, collision_penalty_coeff=4, lane_penalty_coeff=0.02, speed_reward_coeff=0.50, gforce_threshold=None, incent_win=True, constrain_controls=False, incent_yield_to_oncoming_traffic=True, physics_steps_per_observation=12, ) net_config = dict(hidden_units=(64, 64), activation=torch.nn.Tanh) eg = ExperimentGrid(name=experiment_name) eg.add('env_name', env_config['env_name'], '', False) pso = env_config['physics_steps_per_observation'] effective_horizon_seconds = 10 eg.add( 'gamma', 1 - pso / (effective_horizon_seconds * FPS) ) # Lower gamma so seconds of effective horizon remains at 10s with current physics steps = 12 * 1/60s * 1 / (1-gamma) eg.add('epochs', 417) eg.add('try_rollouts', 2) eg.add('steps_per_try_rollout', 1) eg.add('take_worst_rollout', True) eg.add('steps_per_epoch', 20000) eg.add('ac_kwargs:hidden_sizes', net_config['hidden_units'], 'hid') eg.add('ac_kwargs:activation', net_config['activation'], '') eg.add('notes', notes, '') eg.add('run_filename', os.path.realpath(__file__), '') eg.add('env_config', env_config, '')
if __name__ == '__main__': import argparse, yaml from dotmap import DotMap parser = argparse.ArgumentParser() parser.add_argument('params_file', type=str, default=None) parser.add_argument('--cloud', default=False, action="store_true") args = parser.parse_args() assert args.params_file is not None params = None with open(args.params_file) as f: params = DotMap(yaml.load(f)) eg = ExperimentGrid(name=params.name) eg.add('env_name', params.env_name, '', True) eg.add('test_env_names', params.test_env_names, "test", False) eg.add('seed', params.seed) eg.add('epochs', params.epochs) eg.add('steps_per_epoch', params.steps_per_epoch) eg.add('update_after', params.update_after) eg.add('alpha', params.alpha, 'alp', True) eg.add('ac_kwargs:hidden_sizes', params.hidden_sizes, 'hid') activation = None if params.activation == "relu": activation = torch.nn.ReLU elif params.activation == "tanh": activation = torch.nn.Tanh if args.cloud: