-
Notifications
You must be signed in to change notification settings - Fork 5
Expand file tree
/
Copy pathmodel_wrappers.py
More file actions
executable file
·451 lines (396 loc) · 16.1 KB
/
Copy pathmodel_wrappers.py
File metadata and controls
executable file
·451 lines (396 loc) · 16.1 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
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
from typing import Any, Tuple, Callable, Optional, List
from abc import ABC
import numpy as np
import torch
from ding.torch_utils import get_tensor_data
from ding.rl_utils import create_noise_generator
from torch.distributions import Categorical
class IModelWrapper(ABC):
r"""
Overview:
the base class of Model Wrappers
Interfaces:
register
"""
def __init__(self, model: Any) -> None:
self._model = model
def __getattr__(self, key: str) -> Any:
r"""
Overview:
Get the attrbute in model.
Arguments:
- key (:obj:`str`): The key to query.
Returns:
- ret (:obj:`Any`): The queried attribute.
"""
return getattr(self._model, key)
def info(self, attr_name):
r"""
Overview:
get info of attr_name
"""
if attr_name in dir(self):
if isinstance(self._model, IModelWrapper):
return '{} {}'.format(self.__class__.__name__, self._model.info(attr_name))
else:
if attr_name in dir(self._model):
return '{} {}'.format(self.__class__.__name__, self._model.__class__.__name__)
else:
return '{}'.format(self.__class__.__name__)
else:
if isinstance(self._model, IModelWrapper):
return '{}'.format(self._model.info(attr_name))
else:
return '{}'.format(self._model.__class__.__name__)
class BaseModelWrapper(IModelWrapper):
r"""
Overview:
the base class of Model Wrappers
Interfaces:
register
"""
def reset(self, data_id: List[int] = None) -> None:
r"""
Overview
the reset function that the Model Wrappers with states should implement
used to reset the stored states
"""
pass
class HiddenStateWrapper(IModelWrapper):
def __init__(
self, model: Any, state_num: int, save_prev_state: bool = False, init_fn: Callable = lambda: None
) -> None:
"""
Overview:
Maintain the hidden state for RNN-base model. Each sample in a batch has its own state. \
Init the maintain state and state function; Then wrap the ``model.forward`` method with auto \
saved data ['prev_state'] input, and create the ``model.reset`` method.
Arguments:
- model(:obj:`Any`): Wrapped model class, should contain forward method.
- state_num (:obj:`int`): Number of states to process.
- save_prev_state (:obj:`bool`): Whether to output the prev state in output['prev_state'].
- init_fn (:obj:`Callable`): The function which is used to init every hidden state when init and reset. \
Default return None for hidden states.
.. note::
1. This helper must deal with an actual batch with some parts of samples, e.g: 6 samples of state_num 8.
2. This helper must deal with the single sample state reset.
"""
super().__init__(model)
self._state_num = state_num
self._state = {i: init_fn() for i in range(state_num)}
self._save_prev_state = save_prev_state
self._init_fn = init_fn
def forward(self, data, **kwargs):
state_id = kwargs.pop('data_id', None)
valid_id = kwargs.pop('valid_id', None)
data, state_info = self.before_forward(data, state_id)
output = self._model.forward(data, **kwargs)
h = output.pop('next_state', None)
if h:
self.after_forward(h, state_info, valid_id)
if self._save_prev_state:
prev_state = get_tensor_data(data['prev_state'])
output['prev_state'] = prev_state
return output
def reset(self, *args, **kwargs):
state = kwargs.pop('state', None)
state_id = kwargs.get('data_id', None)
self.reset_state(state, state_id)
if hasattr(self._model, 'reset'):
return self._model.reset(*args, **kwargs)
def reset_state(self, state: Optional[list] = None, state_id: Optional[list] = None) -> None:
if state_id is None:
state_id = [i for i in range(self._state_num)]
if state is None:
state = [self._init_fn() for i in range(len(state_id))]
assert len(state) == len(state_id), '{}/{}'.format(len(state), len(state_id))
for idx, s in zip(state_id, state):
self._state[idx] = s
def before_forward(self, data: dict, state_id: Optional[list]) -> Tuple[dict, dict]:
if state_id is None:
state_id = [i for i in range(self._state_num)]
state_info = {idx: self._state[idx] for idx in state_id}
data['prev_state'] = list(state_info.values())
return data, state_info
def after_forward(self, h: Any, state_info: dict, valid_id: Optional[list] = None) -> None:
assert len(h) == len(state_info), '{}/{}'.format(len(h), len(state_info))
for i, idx in enumerate(state_info.keys()):
if valid_id is None:
self._state[idx] = h[i]
else:
if idx in valid_id:
self._state[idx] = h[i]
def sample_action(logit=None, prob=None):
if prob is None:
prob = torch.softmax(logit, dim=-1)
shape = prob.shape
prob += 1e-8
prob = prob.view(-1, shape[-1])
# prob can also be treated as weight in multinomial sample
action = torch.multinomial(prob, 1).squeeze(-1)
action = action.view(*shape[:-1])
return action
class ArgmaxSampleWrapper(IModelWrapper):
r"""
Overview:
Used to help the model to sample argmax action
"""
def forward(self, *args, **kwargs):
output = self._model.forward(*args, **kwargs)
assert isinstance(output, dict), "model output must be dict, but find {}".format(type(output))
logit = output['logit']
assert isinstance(logit, torch.Tensor) or isinstance(logit, list)
if isinstance(logit, torch.Tensor):
logit = [logit]
if 'action_mask' in output:
mask = output['action_mask']
if isinstance(mask, torch.Tensor):
mask = [mask]
logit = [l.sub_(1e8 * (1 - m)) for l, m in zip(logit, mask)]
action = [l.argmax(dim=-1) for l in logit]
if len(action) == 1:
action, logit = action[0], logit[0]
output['action'] = action
return output
class MultinomialSampleWrapper(IModelWrapper):
r"""
Overview:
Used to helper the model get the corresponding action from the output['logits']
Interfaces:
register
"""
def forward(self, *args, **kwargs):
output = self._model.forward(*args, **kwargs)
assert isinstance(output, dict), "model output must be dict, but find {}".format(type(output))
logit = output['logit']
assert isinstance(logit, torch.Tensor) or isinstance(logit, list)
if isinstance(logit, torch.Tensor):
logit = [logit]
if 'action_mask' in output:
mask = output['action_mask']
if isinstance(mask, torch.Tensor):
mask = [mask]
logit = [l.sub_(1e8 * (1 - m)) for l, m in zip(logit, mask)]
action = [sample_action(logit=l) for l in logit]
if len(action) == 1:
action, logit = action[0], logit[0]
output['action'] = action
return output
class EpsGreedySampleWrapper(IModelWrapper):
r"""
Overview:
Epsilon greedy sampler used in collector_model to help balance exploratin and exploitation.
Interfaces:
register
"""
def forward(self, *args, **kwargs):
eps = kwargs.pop('eps')
output = self._model.forward(*args, **kwargs)
assert isinstance(output, dict), "model output must be dict, but find {}".format(type(output))
logit = output['logit']
assert isinstance(logit, torch.Tensor) or isinstance(logit, list)
if isinstance(logit, torch.Tensor):
logit = [logit]
if 'action_mask' in output:
mask = output['action_mask']
if isinstance(mask, torch.Tensor):
mask = [mask]
logit = [l.sub_(1e8 * (1 - m)) for l, m in zip(logit, mask)]
else:
mask = None
action = []
for i, l in enumerate(logit):
if np.random.random() > eps:
action.append(l.argmax(dim=-1))
else:
if mask:
action.append(sample_action(prob=mask[i].float()))
else:
action.append(torch.randint(0, l.shape[-1], size=l.shape[:-1]))
if len(action) == 1:
action, logit = action[0], logit[0]
output['action'] = action
return output
class EpsGreedySampleWrapperSql(IModelWrapper):
r"""
Overview:
Epsilon greedy sampler used in collector_model to help balance exploratin and exploitation.
Interfaces:
register
"""
def forward(self, *args, **kwargs):
eps = kwargs.pop('eps')
alpha = kwargs.pop('alpha')
output = self._model.forward(*args, **kwargs)
assert isinstance(output, dict), "model output must be dict, but find {}".format(type(output))
logit = output['logit']
assert isinstance(logit, torch.Tensor) or isinstance(logit, list)
if isinstance(logit, torch.Tensor):
logit = [logit]
if 'action_mask' in output:
mask = output['action_mask']
if isinstance(mask, torch.Tensor):
mask = [mask]
logit = [l.sub_(1e8 * (1 - m)) for l, m in zip(logit, mask)]
else:
mask = None
action = []
for i, l in enumerate(logit):
if np.random.random() > eps:
prob = torch.softmax(output['logit'] / alpha, dim=-1)
prob = prob / torch.sum(prob, 1, keepdims=True)
pi_action = torch.zeros(prob.shape)
pi_action = Categorical(prob)
pi_action = pi_action.sample()
action.append(pi_action)
else:
if mask:
action.append(sample_action(prob=mask[i].float()))
else:
action.append(torch.randint(0, l.shape[-1], size=l.shape[:-1]))
if len(action) == 1:
action, logit = action[0], logit[0]
output['action'] = action
return output
class ActionNoiseWrapper(IModelWrapper):
r"""
Overview:
Add noise to collector's action output; Do clips on both generated noise and action after adding noise.
Interfaces:
register, __init__, add_noise, reset
Arguments:
- model (:obj:`Any`): Wrapped model class. Should contain ``forward`` method.
- noise_type (:obj:`str`): The type of noise that should be generated, support ['gauss', 'ou'].
- noise_kwargs (:obj:`dict`): Keyword args that should be used in noise init. Depends on ``noise_type``.
- noise_range (:obj:`Optional[dict]`): Range of noise, used for clipping.
- action_range (:obj:`Optional[dict]`): Range of action + noise, used for clip, default clip to [-1, 1].
"""
def __init__(
self,
model: Any,
noise_type: str = 'gauss',
noise_kwargs: dict = {},
noise_range: Optional[dict] = None,
action_range: Optional[dict] = {
'min': -1,
'max': 1
}
) -> None:
super().__init__(model)
self.noise_generator = create_noise_generator(noise_type, noise_kwargs)
self.noise_range = noise_range
self.action_range = action_range
def forward(self, *args, **kwargs):
output = self._model.forward(*args, **kwargs)
assert isinstance(output, dict), "model output must be dict, but find {}".format(type(output))
if 'action' in output:
action = output['action']
assert isinstance(action, torch.Tensor)
action = self.add_noise(action)
output['action'] = action
return output
def add_noise(self, action: torch.Tensor) -> torch.Tensor:
r"""
Overview:
Generate noise and clip noise if needed. Add noise to action and clip action if needed.
Arguments:
- action (:obj:`torch.Tensor`): Model's action output.
Returns:
- noised_action (:obj:`torch.Tensor`): Action processed after adding noise and clipping.
"""
noise = self.noise_generator(action.shape, action.device)
if self.noise_range is not None:
noise = noise.clamp(self.noise_range['min'], self.noise_range['max'])
action += noise
if self.action_range is not None:
action = action.clamp(self.action_range['min'], self.action_range['max'])
return action
def reset(self) -> None:
r"""
Overview:
Reset noise generator.
"""
pass
class TargetNetworkWrapper(IModelWrapper):
r"""
Overview:
Maintain and update the target network
Interfaces:
update, reset
"""
def __init__(self, model: Any, update_type: str, update_kwargs: dict):
super().__init__(model)
assert update_type in ['momentum', 'assign']
self._update_type = update_type
self._update_kwargs = update_kwargs
self._update_count = 0
def reset(self, *args, **kwargs):
self.reset_state()
if hasattr(self._model, 'reset'):
return self._model.reset(*args, **kwargs)
def update(self, state_dict: dict, direct: bool = False) -> None:
r"""
Overview:
Update the target network state dict
Arguments:
- state_dict (:obj:`dict`): the state_dict from learner model
- direct (:obj:`bool`): whether to update the target network directly, \
if ture then will simply call the load_state_dict method of the model
"""
if direct:
self._model.load_state_dict(state_dict, strict=True)
self._update_count = 0
elif self._update_type == 'assign':
if (self._update_count + 1) % self._update_kwargs['freq'] == 0:
self._model.load_state_dict(state_dict, strict=True)
self._update_count += 1
elif self._update_type == 'momentum':
theta = self._update_kwargs['theta']
for name, p in self._model.named_parameters():
# default theta = 0.001
p.data = (1 - theta) * p.data + theta * state_dict[name]
def reset_state(self) -> None:
r"""
Overview:
Reset the update_count
"""
self._update_count = 0
class TeacherNetworkWrapper(IModelWrapper):
r"""
Overview:
Set the teacher Network. Set the model's model.teacher_cfg to the input teacher_cfg
Interfaces:
register
"""
def __init__(self, model, teacher_cfg):
super().__init__(model)
self._model._teacher_cfg = teacher_cfg
wrapper_name_map = {
'base': BaseModelWrapper,
'hidden_state': HiddenStateWrapper,
'argmax_sample': ArgmaxSampleWrapper,
'eps_greedy_sample': EpsGreedySampleWrapper,
'eps_greedy_sample_sql': EpsGreedySampleWrapperSql,
'multinomial_sample': MultinomialSampleWrapper,
'action_noise': ActionNoiseWrapper,
# model wrapper
'target': TargetNetworkWrapper,
'teacher': TeacherNetworkWrapper,
}
def model_wrap(model, wrapper_name: str = None, **kwargs):
if wrapper_name in wrapper_name_map:
if not isinstance(model, IModelWrapper):
model = wrapper_name_map['base'](model)
model = wrapper_name_map[wrapper_name](model, **kwargs)
return model
def register_wrapper(name: str, wrapper_type: type):
r"""
Overview:
Register new wrapper to wrapper_name_map
Arguments:
- name (:obj:`str`): the name of the wrapper
- wrapper_type (subclass of :obj:`IModelWrapper`): the wrapper class added to the plguin_name_map
"""
assert isinstance(name, str)
assert issubclass(wrapper_type, IModelWrapper)
wrapper_name_map[name] = wrapper_type