)
local_launcher = doodad.LocalMode()
"""
ENV=walker2d-medium-v0
python scripts/brac/train_offline.py \
--alsologtostderr --sub_dir=0 \
--env_name=$ENV --identifier="train_bc" \
--agent_name=bc \
--total_train_steps=300000 \
--n_train=1000000 \
--gin_bindings="train_eval_offline.model_params=((200, 200),)" \
--gin_bindings="train_eval_offline.batch_size=256" \
--gin_bindings="train_eval_offline.optimizers=(('adam', 5e-4),)"
"""
doodad.run_python(
target='brac/train_offline.py',
mode=local_launcher,
mounts=mounts,
docker_image='justinfu/brac_flow:0.3',
verbose=True,
cli_args='--alsologtostderr --save_freq=1000 --sub_dir=0 --env_name=%s --identifier="train_bc" --agent_name=bc --total_train_steps=300000 --n_train=1000000 --model_arch=0 --opt_params=0' % env_name
#'--gin_bindings="train_eval_offline.model_params=((200, 200),)" ' + \
#'--gin_bindings="train_eval_offline.batch_size=256" ' + \
#'--gin_bindings="train_eval_offline.optimizers=((\'adam\', 5e-4),)"'
)
# mount_point='/data/b_ckpt'
# ))
mounts.append(
doodad.MountLocal(
local_dir=
"/data/doodad_results/merge-random/learn/flow-merge-random-v0/train_bc/bc/0/0/",
mount_point="/data/b_ckpt",
))
cli_args = (
"--alsologtostderr --sub_dir=auto --env_name={env_name} --agent_name={agent_type} --total_train_steps=500000 --model_arch=1 --opt_params=1 "
+ "--b_ckpt={b_ckpt} --value_penalty={value_penalty} ")
cli_args = cli_args.format(
env_name=env_name,
agent_type="brac_primal",
alpha=0.3,
value_penalty=1,
divergence="kl",
b_ckpt="/data/b_ckpt/agent_behavior",
)
doodad.run_python(
target="brac/train_offline.py",
mode=local_launcher,
mounts=mounts,
docker_image="justinfu/brac_flow:0.3",
verbose=True,
cli_args=cli_args,
)
pythonpath=True,
filter_dir=('data', '.git', 'awr_env')))
mounts.append(
doodad.MountLocal(local_dir='~/code/d4rl',
mount_point='/code/d4rl',
pythonpath=True,
filter_dir=('data', '.git', 'scripts')))
mounts.append(
doodad.MountLocal(local_dir='~/.d4rl/rlkit/%s' % dirname,
mount_point='/datasets'))
mounts.append(
doodad.MountLocal(local_dir='/data/doodad/awr',
mount_point='/data',
output=True))
gcp_launcher = doodad.GCPMode(gcp_bucket='justin-doodad',
gcp_log_path='doodad/logs/bear',
gcp_project='qlearning000',
instance_type='n1-standard-1',
zone='us-west1-a',
gcp_image='ubuntu-1804-docker-gpu',
gcp_image_project='qlearning000')
local_launcher = doodad.LocalMode()
doodad.run_python(target='scripts/run_flow.py',
mode=local_launcher,
mounts=mounts,
docker_image='justinfu/awr_flow:0.1',
verbose=True,
cli_args='--output_dir=/data --env=' + env_name)