def launch_experiments(variant_generator):
variants = variant_generator.variants()
num_experiments = len(variants)
print('Launching {} experiments.'.format(num_experiments))
for i, variant in enumerate(variants):
if variant['seed'] == 'random':
variant['seed'] = np.random.randint(1, 100)
print("Experiment: {}/{}".format(i, num_experiments))
experiment_prefix = variant['prefix'] + '/' + args.exp_name
experiment_name = (variant['prefix'] + '-' + args.exp_name + '-' +
str(i).zfill(2))
run_sac_experiment(
run_experiment,
mode=args.mode,
variant=variant,
exp_prefix=experiment_prefix,
exp_name=experiment_name,
n_parallel=1,
seed=variant['seed'],
terminate_machine=True,
log_dir=args.log_dir,
snapshot_mode=variant['snapshot_mode'],
snapshot_gap=variant['snapshot_gap'],
sync_s3_pkl=variant['sync_pkl'],
)
sys.exit()
def launch_experiments(variant_generator):
variants = variant_generator.variants()
num_experiments = len(variants)
print('Launching {} experiments.'.format(num_experiments))
exp_name = 'acer_test1'
for i, variant in enumerate(variants):
print("Experiment: {}/{}".format(i, num_experiments))
experiment_prefix = 'acer_baselines_final/' + variant['env'] + '/'
experiment_name = (exp_name + '-' + str(i).zfill(2))
run_sac_experiment(
run_acer,
mode='local',
variant=variant,
exp_prefix=experiment_prefix,
exp_name=experiment_name,
n_parallel=1,
seed=variant['seed'],
terminate_machine=True,
snapshot_mode='last',
snapshot_gap=1000,
sync_s3_pkl=False,
)
def launch_experiments(variant_generator, args):
variants = variant_generator.variants()
# TODO: Remove unflatten. Our variant generator should support nested params
variants = [unflatten(variant, separator='.') for variant in variants]
num_experiments = len(variants)
print('Launching {} experiments.'.format(num_experiments))
for i, variant in enumerate(variants):
print("Experiment: {}/{}".format(i, num_experiments))
run_params = variant['run_params']
algo_params = variant['algorithm_params']
variant['algorithm_params']['scale'] = args.scale
variant['algorithm_params']['cost_type'] = args.cost_type
experiment_prefix = variant['prefix'] + '/' + args.exp_name
experiment_name = '{prefix}-{exp_name}-{i:02}'.format(
prefix=variant['prefix'], exp_name=args.exp_name, i=i)
run_sac_experiment(
run_experiment,
mode=args.mode,
variant=variant,
exp_prefix=experiment_prefix,
exp_name=experiment_name,
n_parallel=1,
seed=run_params['seed'],
terminate_machine=True,
log_dir=args.log_dir,
snapshot_mode=run_params['snapshot_mode'],
snapshot_gap=run_params['snapshot_gap'],
sync_s3_pkl=run_params['sync_pkl'],
)
def launch_experiments(variant_generator):
variants = variant_generator.variants()
for i, variant in enumerate(variants):
tag = 'finetune__'
print(variant['snapshot_filename'])
tag += variant['snapshot_filename'].split('/')[-2]
tag += '____'
tag += '__'.join(['%s_%s' % (key, variant[key]) for key in TAG_KEYS])
log_dir = os.path.join(args.log_dir, tag)
variant['video_dir'] = os.path.join(log_dir, 'videos')
print('Launching {} experiments.'.format(len(variants)))
run_sac_experiment(
run_experiment,
mode=args.mode,
variant=variant,
exp_prefix=variant['prefix'] + '/' + args.exp_name,
exp_name=variant['prefix'] + '-' + args.exp_name + '-' +
str(i).zfill(2),
n_parallel=1, # Increasing this barely effects performance,
# but breaks learning of hierarchical policy.
seed=variant['seed'],
terminate_machine=True,
log_dir=log_dir,
snapshot_mode=variant['snapshot_mode'],
snapshot_gap=variant['snapshot_gap'],
sync_s3_pkl=variant['sync_pkl'],
)
def launch_experiments(variant_generator, args):
variants = variant_generator.variants()
# TODO: Remove unflatten. Our variant generator should support nested params
variants = [unflatten(variant, separator='.') for variant in variants]
num_experiments = len(variants)
print('Launching {} experiments.'.format(num_experiments))
for i, variant in enumerate(variants):
print("Experiment: {}/{}".format(i, num_experiments))
run_params = variant['run_params']
algo_params = variant['algorithm_params']
experiment_prefix = variant['prefix'] + '/' + args.exp_name
experiment_name = '{prefix}-{exp_name}-{i:02}'.format(
prefix=variant['prefix'], exp_name=args.exp_name, i=args.seed)
## Hacks to get it to work while we figure out code!
variant['algorithm_params']['base_kwargs'][
'n_train_repeat'] = args.n_train_repeat
variant['algorithm_params']['base_kwargs']['n_epochs'] = args.n_epochs
variant['algorithm_params']['base_kwargs'][
'gpu_fraction'] = args.gpu_fraction
#algo_params['base_kwargs']['n_epochs'] = 2000.0
log_dir = os.path.join(args.log_dir, experiment_name)
##
print('run params: {}'.format(variant['run_params']))
print('algorithm_params: {}'.format(variant['algorithm_params']))
print('env_params: {}'.format(variant['env_params']))
print('value_fn_params: {}'.format(variant['value_fn_params']))
print('replay_buffer_params: {}'.format(
variant['replay_buffer_params']))
print('policy_params: {}'.format(variant['policy_params']))
print('sampler_params: {}'.format(variant['sampler_params']))
print('task: {}'.format(variant['task']))
run_sac_experiment(
run_experiment,
mode=args.mode,
variant=variant,
exp_prefix=experiment_prefix,
exp_name=experiment_name,
n_parallel=args.n_parallel,
seed=args.seed, #run_params['seed'],
terminate_machine=True,
log_dir=log_dir, # RC: TODO change back to args.logdir
snapshot_mode=run_params['snapshot_mode'],
snapshot_gap=run_params['snapshot_gap'],
sync_s3_pkl=run_params['sync_pkl'],
)
def launch_experiments(variant_generator, args):
variants = variant_generator.variants()
# TODO: Remove unflatten. Our variant generator should support nested params
variants = [unflatten(variant, separator='.') for variant in variants]
print('Launching seed={} experiment.'.format(args.seed))
variant = variants[args.seed - 1]
variant['lr'] = args.lr
variant['tau'] = args.tau
variant['l1regpi'] = args.l1regpi
variant['l2regpi'] = args.l2regpi
variant['l1regvf'] = args.l1regvf
variant['l2regvf'] = args.l2regvf
variant['wclippi'] = args.wclippi
variant['wclipvf'] = args.wclipvf
variant['dropoutpi'] = args.dropoutpi
variant['dropoutvf'] = args.dropoutvf
variant['ent_coef'] = args.ent_coef
variant['batchnormpi'] = args.batchnormpi
variant['batchnormvf'] = args.batchnormvf
variant['reward_scale'] = args.reward_scale
variant['num_hidden'] = args.num_hidden
variant['policypath'] = args.policypath
variant['valuepath'] = args.valuepath
print("Variant for this experiment:", variant)
run_params = variant['run_params']
algo_params = variant['algorithm_params']
experiment_prefix = variant['prefix'] + '/' + args.exp_name
experiment_name = '{prefix}-{exp_name}-{i:02}'.format(
prefix=variant['prefix'], exp_name=args.exp_name, i=args.seed)
run_sac_experiment(
run_experiment,
mode=args.mode,
variant=variant,
exp_prefix=experiment_prefix,
exp_name=experiment_name,
n_parallel=1,
seed=run_params['seed'],
terminate_machine=True,
log_dir=args.log_dir,
snapshot_mode=run_params['snapshot_mode'],
snapshot_gap=run_params['snapshot_gap'],
sync_s3_pkl=run_params['sync_pkl'],
)
def launch_experiments(args):
num_experiments = 1
print('Launching {} experiments.'.format(num_experiments))
for i in range(num_experiments):
print("Experiment: {}/{}".format(i + 1, num_experiments))
run_sac_experiment(
run_experiment,
mode='local',
n_parallel=1,
terminate_machine=True,
log_dir='/root/code/log/{0}/{1}'.format(args.domain, timestamp()),
snapshot_mode='gap',
snapshot_gap=100,
sync_s3_pkl=True,
)
def launch_experiments(variant_generator):
variants = variant_generator.variants()
for i, variant in enumerate(variants):
print('Launching {} experiments.'.format(len(variants)))
run_sac_experiment(
run_experiment,
mode=args.mode,
variant=variant,
exp_prefix=variant['prefix'] + '/' + args.exp_name,
exp_name=variant['prefix'] + '-' + args.exp_name + '-' + str(i).zfill(2),
n_parallel=1,
seed=variant['seed'],
terminate_machine=True,
log_dir=args.log_dir,
# use_cloudpickle=True,
snapshot_mode=variant['snapshot_mode'],
snapshot_gap=variant['snapshot_gap'],
sync_s3_pkl=variant['sync_pkl'],
)
def launch_experiments(variant_generator):
variants = variant_generator.variants()
for i, variant in enumerate(variants):
log_dir = get_logdir(args, variant)
print('Launching {} experiments.'.format(len(variants)))
run_sac_experiment(
run_experiment,
mode=args.mode,
variant=variant,
exp_prefix=variant['prefix'] + '/' + args.exp_name,
exp_name=variant['prefix'] + '-' + args.exp_name + '-' + str(i).zfill(2),
n_parallel=1, # Increasing this barely effects performance,
# but breaks learning of hierarchical policy.
seed=variant['seed'],
terminate_machine=True,
log_dir=log_dir,
snapshot_mode=variant['snapshot_mode'],
snapshot_gap=variant['snapshot_gap'],
sync_s3_pkl=variant['sync_pkl'],
)
def launch_experiments():
args = arg()
num_experiments = 1
print('Launching {} experiments.'.format(num_experiments))
for i in range(num_experiments):
print("Experiment: {}/{}".format(i + 1, num_experiments))
experiment_prefix = args.domain + '/' + args.exp_name
experiment_name = '{prefix}-{exp_name}-{i:02}'.format(
prefix=args.domain, exp_name=args.exp_name, i=0)
run_sac_experiment(
run_experiment,
mode='local',
exp_prefix=experiment_prefix,
exp_name=experiment_name,
n_parallel=1,
terminate_machine=True,
log_dir='/root/code/log/{0}/{1}'.format(args.domain, timestamp()),
snapshot_mode='gap',
snapshot_gap=100,
sync_s3_pkl=True,
)
def launch_experiments(args):
num_experiments = 5
print('Launching {} experiments.'.format(num_experiments))
i = 0
if i == 0:
print("Experiment: {}/{}".format(i, num_experiments))
experiment_prefix = 'ant/cross-maze' + '/' + args.exp_name
experiment_name = '{prefix}-{exp_name}-{i:02}'.format(
prefix='ant/cross-maze', exp_name=args.exp_name, i=0)
run_sac_experiment(
run_experiment,
mode=args.mode,
exp_prefix=experiment_prefix,
exp_name=experiment_name,
n_parallel=1,
terminate_machine=True,
log_dir=args.log_dir,
snapshot_mode='gap',
snapshot_gap=1000,
sync_s3_pkl=True,
)
env=env,
policy=policy,
initial_exploration_policy=initial_exploration_policy,
pool=pool,
qf1=qf1,
qf2=qf2,
vf=vf,
lr=3e-4,
scale_reward=20,
discount=0.99,
tau=0.005,
reparameterize=True,
target_update_interval=1,
action_prior='uniform',
save_full_state=False,
)
algorithm._sess.run(tf.global_variables_initializer())
algorithm.train()
if __name__ == "__main__":
run_sac_experiment(
run_experiment,
mode='local',
log_dir='/root/code/log/prim/reach/{0}'.format(timestamp()),
snapshot_mode='gap',
snapshot_gap=100,
)
