gae_lambda = 0.95
entropy_coef = 0.01
value_loss_coef = 0.5
max_grad_norm = 0.5
seed = 1 # didnt change
cuda_deterministic = False
num_processes = 1
num_steps = 2500
custom_gym = "growspace"
ppo_epoch = 4
num_mini_batch = 32
clip_param = 0.1
log_interval = 10 # amount of times we save to wandb
save_interval = 100 # amount of times we save internal
eval_interval = None
num_env_steps = 1e6 # no change
env_name = "GrowSpaceEnv-Control-v0" #"GrowSpaceSpotlight-Mnist4-v0"
log_dir = "/tmp/gym/"
save_dir = "./trained_models/"
use_proper_time_limits = False
recurrent_policy = False
use_linear_lr_decay = True
no_cuda = False
cuda = not no_cuda and torch.cuda.is_available()
experiment_buddy.register(locals())
tensorboard = experiment_buddy.deploy("mila",
sweep_yaml="pposweep.yaml",
proc_num=10,
wandb_kwargs={"entity": "growspace"})
use_fa = False
horizon = 6 if use_fa else 20
penalty = 0.15 if use_fa else 1.4 #1.6
eps = 1e-4
gamma = 0.9
eta = 1.
grid_size = 10
agent = "pg_clip"
save_interval = 10
max_steps = int(2e3)
seed = 984
eval_episodes = 10
data = "data"
REMOTE = 1
RUN_SWEEP = REMOTE
NUM_PROCS = 5
sweep_yaml = "sweep_params.yaml" if RUN_SWEEP else False
HOST = "mila" if REMOTE else "" # in host
DEBUG = '_pydev_bundle.pydev_log' in sys.modules.keys()
render = not DEBUG
experiment_buddy.register(locals())
tb = experiment_buddy.deploy(host=HOST, sweep_yaml=sweep_yaml, proc_num=NUM_PROCS,
wandb_kwargs=dict(mode= "disabled" if DEBUG else "online"))
os.makedirs(data, exist_ok=True)
plot_path = os.path.join(tb.objects_path, "plots")
os.makedirs(plot_path, exist_ok=True)
seed = 984
h_dim = 32
# wandb_mode = "online" if DEBUG else "offline"
use_cuda = False
experiment_buddy.register(locals())
device = torch.device("cuda" if use_cuda else "cpu")
################################################################
# Derivative parameters
################################################################
# esh = """
# #SBATCH --mem=24GB
# """
esh = """
#SBATCH --job-name=spython
#SBATCH --output=job_output.txt
#SBATCH --error=job_error.txt
#SBATCH --time=2-00:00
#SBATCH --mem=12GB
#SBATCH --gres=gpu:1
#SBATCH --cpus-per-task=4
#SBATCH --partition=long
#SBATCH --get-user-env=L
"""
tb = experiment_buddy.deploy(host=HOST,
sweep_yaml=sweep_yaml,
extra_slurm_headers=esh,
proc_num=NUM_PROCS)
num_steps = 10000
algo = "ppo"
gail = False
gail_experts_dir = './gail_experts'
gail_batch_size = 128
gail_epoch = 5
alpha = 0.99
seed = np.random.randint(1, 80, size=1) # didnt change
cuda_deterministic = False
num_processes = 4
custom_gym = "growspace"
log_interval = 10
save_interval = 100
eval_interval = None
num_updates = 1e3
env_name = "GrowSpaceEnv-Control-v0" # "GrowSpaceSpotlight-MnistMix-v0"
log_dir = "/tmp/gym/"
save_dir = "./trained_models/"
use_proper_time_limits = False
recurrent_policy = False
no_cuda = False
cuda = not no_cuda and torch.cuda.is_available()
momentum = 0.9 # if sgd is used
experiment_buddy.register(locals())
tensorboard = experiment_buddy.deploy("",
sweep_yaml="",
proc_num=3,
wandb_kwargs={"entity": "growspace"})
import experiment_buddy
# Hyperparameters
ENV_NAME = 'BipedalWalker-v3'
MAX_ITER = 500000
BATCH_SIZE = 64
PPO_EPOCHS = 7
CLIP_GRADIENT = 0.2
CLIP_EPS = 0.2
TRAJECTORY_SIZE = 2049
GAE_LAMBDA = 0.95
GAMMA = 0.99
## Test Hyperparameters
test_episodes = 50
save_video_test = False
N_ITER_TEST = 100
POLICY_LR = 0.0004
VALUE_LR = 0.001
experiment_buddy.register(locals())
tensorboard = experiment_buddy.deploy("",
sweep_yaml="",
proc_num=1,
wandb_kwargs={"entity": "ionelia"})
import experiment_buddy
initial_lr = .0001
decay_steps = 500000
num_hidden = 1024
decay_factor = .5
batch_size = 128
momentum_mass = 0.99
weight_norm = 0.00
num_epochs = 1000
experiment_buddy.register(locals())
################################################################
# Derivative parameters
################################################################
learning_rate = jax.experimental.optimizers.inverse_time_decay(initial_lr,
decay_steps,
decay_factor,
staircase=True)
eval_every = math.ceil(num_epochs / 1000)
HOST = os.environ['DEPLOY_DESTINATION']
host_map = {'cluster': 'mila', 'local': ''}
tensorboard = experiment_buddy.deploy(host=host_map[HOST], sweep_yaml="")
import jax.experimental.optimizers
import experiment_buddy
initial_lr = .0001
decay_steps = 500000
num_hidden = 1024
decay_factor = .5
batch_size = 128
momentum_mass = 0.99
weight_norm = 0.00
num_epochs = 1000
experiment_buddy.register(locals())
################################################################
# Derivative parameters
################################################################
learning_rate = jax.experimental.optimizers.inverse_time_decay(initial_lr,
decay_steps,
decay_factor,
staircase=True)
eval_every = math.ceil(num_epochs / 1000)
tensorboard = experiment_buddy.deploy(host=os.environ.get('BUDDY_HOST', ""),
sweep_yaml=os.environ.get('SWEEP', ""))