-
Notifications
You must be signed in to change notification settings - Fork 5
Expand file tree
/
Copy pathil.py
More file actions
executable file
·234 lines (217 loc) · 8.95 KB
/
Copy pathil.py
File metadata and controls
executable file
·234 lines (217 loc) · 8.95 KB
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
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
from typing import List, Dict, Any, Tuple, Union
from collections import namedtuple
import torch
import torch.nn as nn
from ding.torch_utils import Adam, to_device
from ding.model import model_wrap
from ding.utils import POLICY_REGISTRY
from ding.utils.data import default_collate, default_decollate
from .base_policy import Policy
try:
from dizoo.gfootball.model.bots import FootballKaggle5thPlaceModel
except ImportError:
FootballKaggle5thPlaceModel = None
@POLICY_REGISTRY.register('IL')
class ILPolicy(Policy):
r"""
Overview:
Policy class of Imitation learning algorithm
Interface:
__init__, set_setting, __repr__, state_dict_handle
Property:
learn_mode, collect_mode, eval_mode
"""
config = dict(
type='IL',
cuda=True,
# (bool) whether use on-policy training pipeline(behaviour policy and training policy are the same)
on_policy=False,
priority=False,
# (bool) Whether use Importance Sampling Weight to correct biased update. If True, priority must be True.
priority_IS_weight=False,
learn=dict(
multi_gpu=False,
# (int) collect n_episode data, train model n_iteration time
update_per_collect=20,
# (int) the number of data for a train iteration
batch_size=64,
# (float) gradient-descent step size
learning_rate=0.0002,
),
collect=dict(
# (int) collect n_sample data, train model n_iteration time
# n_sample=128,
# (float) discount factor for future reward, defaults int [0, 1]
discount_factor=0.99,
),
eval=dict(evaluator=dict(eval_freq=800, ), ),
other=dict(
replay_buffer=dict(
replay_buffer_size=100000,
# (int) max use count of data, if count is bigger than this value,
# the data will be removed from buffer
max_reuse=10,
),
command=dict(),
),
)
def _init_learn(self) -> None:
r"""
Overview:
Learn mode init method. Called by ``self.__init__``.
Init optimizers, algorithm config, main and target models.
"""
# actor and critic optimizer
self._optimizer = Adam(self._model.parameters(), lr=self._cfg.learn.learning_rate)
# main and target models
self._learn_model = model_wrap(self._model, wrapper_name='base')
self._learn_model.train()
self._learn_model.reset()
self._forward_learn_cnt = 0 # count iterations
def _forward_learn(self, data: dict) -> Dict[str, Any]:
r"""
Overview:
Forward and backward function of learn mode.
Arguments:
- data (:obj:`dict`): Dict type data, including at least ['obs', 'action', 'reward', 'next_obs']
Returns:
- info_dict (:obj:`Dict[str, Any]`): Including at least actor and critic lr, different losses.
"""
data = default_collate(data, cat_1dim=False)
data['done'] = None
if self._cuda:
data = to_device(data, self._device)
loss_dict = {}
# ====================
# imitation learn forward
# ====================
obs = data.get('obs')
logit = data.get('logit')
assert isinstance(obs['processed_obs'], torch.Tensor), obs['processed_obs']
model_action_logit = self._learn_model.forward(obs['processed_obs'])['logit']
supervised_loss = nn.MSELoss(reduction='none')(model_action_logit, logit).mean()
self._optimizer.zero_grad()
supervised_loss.backward()
self._optimizer.step()
loss_dict['supervised_loss'] = supervised_loss
return {
'cur_lr': self._optimizer.defaults['lr'],
**loss_dict,
}
def _state_dict_learn(self) -> Dict[str, Any]:
return {
'model': self._learn_model.state_dict(),
'optimizer': self._optimizer.state_dict(),
}
def _load_state_dict_learn(self, state_dict: Dict[str, Any]) -> None:
self._learn_model.load_state_dict(state_dict['model'])
self._optimizer.load_state_dict(state_dict['optimizer'])
def _init_collect(self) -> None:
r"""
Overview:
Collect mode init method. Called by ``self.__init__``.
Init traj and unroll length, collect model.
"""
self._collect_model = model_wrap(FootballKaggle5thPlaceModel(), wrapper_name='base')
self._gamma = self._cfg.collect.discount_factor
self._collect_model.eval()
self._collect_model.reset()
def _forward_collect(self, data: dict) -> dict:
r"""
Overview:
Forward function of collect mode.
Arguments:
- data (:obj:`Dict[str, Any]`): Dict type data, stacked env data for predicting policy_output(action), \
values are torch.Tensor or np.ndarray or dict/list combinations, keys are env_id indicated by integer.
Returns:
- output (:obj:`Dict[int, Any]`): Dict type data, including at least inferred action according to input obs.
ReturnsKeys
- necessary: ``action``
- optional: ``logit``
"""
data_id = list(data.keys())
data = default_collate(list(data.values()))
if self._cuda:
data = to_device(data, self._device)
with torch.no_grad():
output = self._collect_model.forward(default_decollate(data['obs']['raw_obs']))
if self._cuda:
output = to_device(output, 'cpu')
output = default_decollate(output)
return {i: d for i, d in zip(data_id, output)}
def _process_transition(self, obs: Any, model_output: dict, timestep: namedtuple) -> Dict[str, Any]:
r"""
Overview:
Generate dict type transition data from inputs.
Arguments:
- obs (:obj:`Any`): Env observation
- model_output (:obj:`dict`): Output of collect model, including at least ['action']
- timestep (:obj:`namedtuple`): Output after env step, including at least ['obs', 'reward', 'done'] \
(here 'obs' indicates obs after env step, i.e. next_obs).
Return:
- transition (:obj:`Dict[str, Any]`): Dict type transition data.
"""
transition = {
'obs': obs,
'action': model_output['action'],
'logit': model_output['logit'],
'reward': timestep.reward,
'done': timestep.done,
}
return transition
def _get_train_sample(self, origin_data: list) -> Union[None, List[Any]]:
datas = []
pre_rew = 0
for i in range(len(origin_data) - 1, -1, -1):
data = {}
data['obs'] = origin_data[i]['obs']
data['action'] = origin_data[i]['action']
cur_rew = origin_data[i]['reward']
pre_rew = cur_rew + (pre_rew * self._gamma)
# sample uniformly
data['priority'] = 1
data['logit'] = origin_data[i]['logit']
datas.append(data)
return datas
def _init_eval(self) -> None:
r"""
Overview:
Evaluate mode init method. Called by ``self.__init__``.
Init eval model. Unlike learn and collect model, eval model does not need noise.
"""
self._eval_model = model_wrap(self._model, wrapper_name='argmax_sample')
self._eval_model.reset()
def _forward_eval(self, data: dict) -> dict:
r"""
Overview:
Forward function of eval mode, similar to ``self._forward_collect``.
Arguments:
- data (:obj:`Dict[str, Any]`): Dict type data, stacked env data for predicting policy_output(action), \
values are torch.Tensor or np.ndarray or dict/list combinations, keys are env_id indicated by integer.
Returns:
- output (:obj:`Dict[int, Any]`): The dict of predicting action for the interaction with env.
ReturnsKeys
- necessary: ``action``
- optional: ``logit``
"""
data_id = list(data.keys())
data = default_collate(list(data.values()))
if self._cuda:
data = to_device(data, self._device)
with torch.no_grad():
output = self._eval_model.forward(data['obs']['processed_obs'])
if self._cuda:
output = to_device(output, 'cpu')
output = default_decollate(output)
return {i: d for i, d in zip(data_id, output)}
# TODO different collect model and learn model
def default_model(self) -> Tuple[str, List[str]]:
return 'football_iql', ['dizoo.gfootball.model.iql.iql_network']
def _monitor_vars_learn(self) -> List[str]:
r"""
Overview:
Return variables' name if variables are to used in monitor.
Returns:
- vars (:obj:`List[str]`): Variables' name list.
"""
return ['cur_lr', 'supervised_loss']