-
Notifications
You must be signed in to change notification settings - Fork 371
/
gail_irl_model.py
293 lines (266 loc) · 13.2 KB
/
gail_irl_model.py
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
from typing import List, Dict, Any
import pickle
import random
from collections.abc import Iterable
from easydict import EasyDict
import torch
import torch.nn as nn
import torch.optim as optim
from ding.utils import REWARD_MODEL_REGISTRY
from .base_reward_model import BaseRewardModel
import torch.nn.functional as F
from functools import partial
def concat_state_action_pairs(iterator):
"""
Overview:
Concatenate state and action pairs from input.
Arguments:
- iterator (:obj:`Iterable`): Iterables with at least ``obs`` and ``action`` tensor keys.
Returns:
- res (:obj:`Torch.tensor`): State and action pairs.
"""
assert isinstance(iterator, Iterable)
res = []
for item in iterator:
state = item['obs'].flatten() # to allow 3d obs and actions concatenation
action = item['action']
s_a = torch.cat([state, action.float()], dim=-1)
res.append(s_a)
return res
def concat_state_action_pairs_one_hot(iterator, action_size: int):
"""
Overview:
Concatenate state and action pairs from input. Action values are one-hot encoded
Arguments:
- iterator (:obj:`Iterable`): Iterables with at least ``obs`` and ``action`` tensor keys.
Returns:
- res (:obj:`Torch.tensor`): State and action pairs.
"""
assert isinstance(iterator, Iterable)
res = []
for item in iterator:
state = item['obs'].flatten() # to allow 3d obs and actions concatenation
action = item['action']
action = torch.Tensor([int(i == action) for i in range(action_size)])
s_a = torch.cat([state, action], dim=-1)
res.append(s_a)
return res
class RewardModelNetwork(nn.Module):
def __init__(self, input_size: int, hidden_size: int, output_size: int) -> None:
super(RewardModelNetwork, self).__init__()
self.l1 = nn.Linear(input_size, hidden_size)
self.l2 = nn.Linear(hidden_size, output_size)
self.a1 = nn.Tanh()
self.a2 = nn.Sigmoid()
def forward(self, x: torch.Tensor) -> torch.Tensor:
out = x
out = self.l1(out)
out = self.a1(out)
out = self.l2(out)
out = self.a2(out)
return out
class AtariRewardModelNetwork(nn.Module):
def __init__(self, input_size: int, action_size: int) -> None:
super(AtariRewardModelNetwork, self).__init__()
self.input_size = input_size
self.action_size = action_size
self.conv1 = nn.Conv2d(4, 16, 7, stride=3)
self.conv2 = nn.Conv2d(16, 16, 5, stride=2)
self.conv3 = nn.Conv2d(16, 16, 3, stride=1)
self.conv4 = nn.Conv2d(16, 16, 3, stride=1)
self.fc1 = nn.Linear(784, 64)
self.fc2 = nn.Linear(64 + self.action_size, 1) # here we add 1 to take consideration of the action concat
self.a = nn.Sigmoid()
def forward(self, x: torch.Tensor) -> torch.Tensor:
# input: x = [B, 4 x 84 x 84 + self.action_size], last element is action
actions = x[:, -self.action_size:] # [B, self.action_size]
# get observations
x = x[:, :-self.action_size]
x = x.reshape([-1] + self.input_size) # [B, 4, 84, 84]
x = F.leaky_relu(self.conv1(x))
x = F.leaky_relu(self.conv2(x))
x = F.leaky_relu(self.conv3(x))
x = F.leaky_relu(self.conv4(x))
x = x.reshape(-1, 784)
x = F.leaky_relu(self.fc1(x))
x = torch.cat([x, actions], dim=-1)
x = self.fc2(x)
r = self.a(x)
return r
@REWARD_MODEL_REGISTRY.register('gail')
class GailRewardModel(BaseRewardModel):
"""
Overview:
The Gail reward model class (https://arxiv.org/abs/1606.03476)
Interface:
``estimate``, ``train``, ``load_expert_data``, ``collect_data``, ``clear_date``, \
``__init__``, ``state_dict``, ``load_state_dict``, ``learn``
Config:
== ==================== ======== ============= =================================== =======================
ID Symbol Type Default Value Description Other(Shape)
== ==================== ======== ============= =================================== =======================
1 ``type`` str gail | RL policy register name, refer | this arg is optional,
| to registry ``POLICY_REGISTRY`` | a placeholder
2 | ``expert_data_`` str expert_data. | Path to the expert dataset | Should be a '.pkl'
| ``path`` .pkl | | file
3 | ``learning_rate`` float 0.001 | The step size of gradient descent |
4 | ``update_per_`` int 100 | Number of updates per collect |
| ``collect`` | |
5 | ``batch_size`` int 64 | Training batch size |
6 | ``input_size`` int | Size of the input: |
| | obs_dim + act_dim |
7 | ``target_new_`` int 64 | Collect steps per iteration |
| ``data_count`` | |
8 | ``hidden_size`` int 128 | Linear model hidden size |
9 | ``collect_count`` int 100000 | Expert dataset size | One entry is a (s,a)
| | | tuple
10 | ``clear_buffer_`` int 1 | clear buffer per fixed iters | make sure replay
| ``per_iters`` | buffer's data count
| | isn't too few.
| | (code work in entry)
== ==================== ======== ============= =================================== =======================
"""
config = dict(
# (str) RL policy register name, refer to registry ``POLICY_REGISTRY``.
type='gail',
# (float) The step size of gradient descent.
learning_rate=1e-3,
# (int) How many updates(iterations) to train after collector's one collection.
# Bigger "update_per_collect" means bigger off-policy.
# collect data -> update policy-> collect data -> ...
update_per_collect=100,
# (int) How many samples in a training batch.
batch_size=64,
# (int) Size of the input: obs_dim + act_dim.
input_size=4,
# (int) Collect steps per iteration.
target_new_data_count=64,
# (int) Linear model hidden size.
hidden_size=128,
# (int) Expert dataset size.
collect_count=100000,
# (int) Clear buffer per fixed iters.
clear_buffer_per_iters=1,
)
def __init__(self, config: EasyDict, device: str, tb_logger: 'SummaryWriter') -> None: # noqa
"""
Overview:
Initialize ``self.`` See ``help(type(self))`` for accurate signature.
Arguments:
- cfg (:obj:`EasyDict`): Training config
- device (:obj:`str`): Device usage, i.e. "cpu" or "cuda"
- tb_logger (:obj:`SummaryWriter`): Logger, defaultly set as 'SummaryWriter' for model summary
"""
super(GailRewardModel, self).__init__()
self.cfg = config
assert device in ["cpu", "cuda"] or "cuda" in device
self.device = device
self.tb_logger = tb_logger
obs_shape = config.input_size
if isinstance(obs_shape, int) or len(obs_shape) == 1:
self.reward_model = RewardModelNetwork(config.input_size, config.hidden_size, 1)
self.concat_state_action_pairs = concat_state_action_pairs
elif len(obs_shape) == 3:
action_shape = self.cfg.action_size
self.reward_model = AtariRewardModelNetwork(config.input_size, action_shape)
self.concat_state_action_pairs = partial(concat_state_action_pairs_one_hot, action_size=action_shape)
self.reward_model.to(self.device)
self.expert_data = []
self.train_data = []
self.expert_data_loader = None
self.opt = optim.Adam(self.reward_model.parameters(), config.learning_rate)
self.train_iter = 0
self.load_expert_data()
def load_expert_data(self) -> None:
"""
Overview:
Getting the expert data from ``config.data_path`` attribute in self
Effects:
This is a side effect function which updates the expert data attribute \
(i.e. ``self.expert_data``) with ``fn:concat_state_action_pairs``
"""
with open(self.cfg.data_path + '/expert_data.pkl', 'rb') as f:
self.expert_data_loader: list = pickle.load(f)
self.expert_data = self.concat_state_action_pairs(self.expert_data_loader)
def state_dict(self) -> Dict[str, Any]:
return {
'model': self.reward_model.state_dict(),
}
def load_state_dict(self, state_dict: Dict[str, Any]) -> None:
self.reward_model.load_state_dict(state_dict['model'])
def learn(self, train_data: torch.Tensor, expert_data: torch.Tensor) -> float:
"""
Overview:
Helper function for ``train`` which calculates loss for train data and expert data.
Arguments:
- train_data (:obj:`torch.Tensor`): Data used for training
- expert_data (:obj:`torch.Tensor`): Expert data
Returns:
- Combined loss calculated of reward model from using ``train_data`` and ``expert_data``.
"""
# calculate loss, here are some hyper-param
out_1: torch.Tensor = self.reward_model(train_data)
loss_1: torch.Tensor = torch.log(out_1 + 1e-8).mean()
out_2: torch.Tensor = self.reward_model(expert_data)
loss_2: torch.Tensor = torch.log(1 - out_2 + 1e-8).mean()
# log(x) with 0<x<1 is negative, so to reduce this loss we have to minimize the opposite
loss: torch.Tensor = -(loss_1 + loss_2)
self.opt.zero_grad()
loss.backward()
self.opt.step()
return loss.item()
def train(self) -> None:
"""
Overview:
Training the Gail reward model. The training and expert data are randomly sampled with designated\
batch size abstracted from the ``batch_size`` attribute in ``self.cfg`` and \
correspondingly, the ``expert_data`` as well as ``train_data`` attributes initialized ``self`
Effects:
- This is a side effect function which updates the reward model and increment the train iteration count.
"""
for _ in range(self.cfg.update_per_collect):
sample_expert_data: list = random.sample(self.expert_data, self.cfg.batch_size)
sample_train_data: list = random.sample(self.train_data, self.cfg.batch_size)
sample_expert_data = torch.stack(sample_expert_data).to(self.device)
sample_train_data = torch.stack(sample_train_data).to(self.device)
loss = self.learn(sample_train_data, sample_expert_data)
self.tb_logger.add_scalar('reward_model/gail_loss', loss, self.train_iter)
self.train_iter += 1
def estimate(self, data: list) -> List[Dict]:
"""
Overview:
Estimate reward by rewriting the reward key in each row of the data.
Arguments:
- data (:obj:`list`): the list of data used for estimation, with at least \
``obs`` and ``action`` keys.
Effects:
- This is a side effect function which updates the reward values in place.
"""
# NOTE: deepcopy reward part of data is very important,
# otherwise the reward of data in the replay buffer will be incorrectly modified.
train_data_augmented = self.reward_deepcopy(data)
res = self.concat_state_action_pairs(train_data_augmented)
res = torch.stack(res).to(self.device)
with torch.no_grad():
reward = self.reward_model(res).squeeze(-1).cpu()
reward = torch.chunk(reward, reward.shape[0], dim=0)
for item, rew in zip(train_data_augmented, reward):
item['reward'] = -torch.log(rew + 1e-8)
return train_data_augmented
def collect_data(self, data: list) -> None:
"""
Overview:
Collecting training data formatted by ``fn:concat_state_action_pairs``.
Arguments:
- data (:obj:`Any`): Raw training data (e.g. some form of states, actions, obs, etc)
Effects:
- This is a side effect function which updates the data attribute in ``self``
"""
self.train_data.extend(self.concat_state_action_pairs(data))
def clear_data(self) -> None:
"""
Overview:
Clearing training data. \
This is a side effect function which clears the data attribute in ``self``
"""
self.train_data.clear()