File size: 3,659 Bytes
079c32c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 |
from easydict import EasyDict
walker2d_trex_onppo_config = dict(
exp_name='walker2d_trex_onppo_seed0',
env=dict(
env_id='Walker2d-v3',
norm_obs=dict(use_norm=False, ),
norm_reward=dict(use_norm=False, ),
collector_env_num=8,
evaluator_env_num=10,
n_evaluator_episode=10,
stop_value=3000,
),
reward_model=dict(
min_snippet_length=30,
max_snippet_length=100,
checkpoint_min=10000,
checkpoint_max=90000,
checkpoint_step=10000,
num_snippets=60000,
learning_rate=1e-5,
update_per_collect=1,
# Users should add their own model path here. Model path should lead to a model.
# Absolute path is recommended.
# In DI-engine, it is ``exp_name/ckpt/ckpt_best.pth.tar``.
# However, here in ``expert_model_path``, it is ``exp_name`` of the expert config.
expert_model_path='model_path_placeholder',
# Path where to store the reward model
reward_model_path='data_path_placeholder + /Walker2d.params',
# Users should add their own data path here. Data path should lead to a file to store data or load the stored data.
# Absolute path is recommended.
# In DI-engine, it is usually located in ``exp_name`` directory
# See ding/entry/application_entry_trex_collect_data.py to collect the data
data_path='data_path_placeholder',
),
policy=dict(
cuda=True,
recompute_adv=True,
model=dict(
obs_shape=17,
action_shape=6,
action_space='continuous',
),
action_space='continuous',
learn=dict(
epoch_per_collect=10,
batch_size=64,
learning_rate=3e-4,
value_weight=0.5,
entropy_weight=0.0,
clip_ratio=0.2,
adv_norm=True,
value_norm=True,
),
collect=dict(
n_sample=2048,
unroll_len=1,
discount_factor=0.99,
gae_lambda=0.97,
),
eval=dict(evaluator=dict(eval_freq=5000, )),
),
)
walker2d_trex_onppo_config = EasyDict(walker2d_trex_onppo_config)
main_config = walker2d_trex_onppo_config
walker2d_trex_onppo_create_config = dict(
env=dict(
type='mujoco',
import_names=['dizoo.mujoco.envs.mujoco_env'],
),
env_manager=dict(type='subprocess'),
policy=dict(type='ppo', ),
)
walker2d_trex_onppo_create_config = EasyDict(walker2d_trex_onppo_create_config)
create_config = walker2d_trex_onppo_create_config
if __name__ == '__main__':
# Users should first run ``walker2d_onppo_config.py`` to save models (or checkpoints).
# Note: Users should check that the checkpoints generated should include iteration_'checkpoint_min'.pth.tar, iteration_'checkpoint_max'.pth.tar with the interval checkpoint_step
# where checkpoint_max, checkpoint_min, checkpoint_step are specified above.
import argparse
import torch
from ding.entry import trex_collecting_data
from ding.entry import serial_pipeline_trex_onpolicy
parser = argparse.ArgumentParser()
parser.add_argument('--cfg', type=str, default='please enter abs path for this file')
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--device', type=str, default='cuda' if torch.cuda.is_available() else 'cpu')
args = parser.parse_args()
# The function ``trex_collecting_data`` below is to collect episodic data for training the reward model in trex.
trex_collecting_data(args)
serial_pipeline_trex_onpolicy([main_config, create_config])
|