-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset.py
183 lines (144 loc) · 7.07 KB
/
dataset.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
import torch
import numpy as np
from typing import List, Tuple
import os
class ReplayBuffer:
def __init__(self,
state_dim: int,
action_dim: int,
buffer_size: int = 1000000) -> None:
self.buffer_size = buffer_size
self.pointer = 0
self.size = 0
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.device = device
self.states = torch.zeros((buffer_size, state_dim), dtype=torch.float32, device=device)
self.actions = torch.zeros((buffer_size, action_dim), dtype=torch.float32, device=device)
self.rewards = torch.zeros((buffer_size, 1), dtype=torch.float32, device=device)
self.next_states = torch.zeros((buffer_size, state_dim), dtype=torch.float32, device=device)
self.dones = torch.zeros((buffer_size, 1), dtype=torch.float32, device=device)
# i/o order: state, action, reward, next_state, done
def from_json(self, json_file: str):
import json
if not json_file.endswith('.json'):
json_file = json_file + '.json'
json_file = os.path.join("json_datasets", json_file)
output = dict()
with open(json_file) as f:
dataset = json.load(f)
for k, v in dataset.items():
v = np.array(v)
if k != "terminals":
v = v.astype(np.float32)
output[k] = v
self.from_d4rl(output)
@staticmethod
def to_tensor(data: np.ndarray, device=None) -> torch.Tensor:
if device is None:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
return torch.tensor(data, dtype=torch.float32, device=device)
def sample(self, batch_size: int):
indexes = np.random.randint(0, self.size, size=batch_size)
return (
self.to_tensor(self.states[indexes], self.device),
self.to_tensor(self.actions[indexes], self.device),
self.to_tensor(self.rewards[indexes], self.device),
self.to_tensor(self.next_states[indexes], self.device),
self.to_tensor(self.dones[indexes], self.device)
)
def from_d4rl(self, dataset):
if self.size:
print("Warning: loading data into non-empty buffer")
n_transitions = dataset["observations"].shape[0]
if n_transitions < self.buffer_size:
self.states[:n_transitions] = self.to_tensor(dataset["observations"][-n_transitions:], self.device)
self.actions[:n_transitions] = self.to_tensor(dataset["actions"][-n_transitions:], self.device)
self.next_states[:n_transitions] = self.to_tensor(dataset["next_observations"][-n_transitions:], self.device)
self.rewards[:n_transitions] = self.to_tensor(dataset["rewards"][-n_transitions:].reshape(-1, 1), self.device)
self.dones[:n_transitions] = self.to_tensor(dataset["terminals"][-n_transitions:].reshape(-1, 1), self.device)
else:
self.buffer_size = n_transitions
self.states = self.to_tensor(dataset["observations"][-n_transitions:], self.device)
self.actions = self.to_tensor(dataset["actions"][-n_transitions:])
self.next_states = self.to_tensor(dataset["next_observations"][-n_transitions:], self.device)
self.rewards = self.to_tensor(dataset["rewards"][-n_transitions:].reshape(-1, 1), self.device)
self.dones = self.to_tensor(dataset["terminals"][-n_transitions:].reshape(-1, 1), self.device)
self.size = n_transitions
self.pointer = n_transitions % self.buffer_size
def from_d4rl_finetune(self, dataset):
raise NotImplementedError()
def normalize_states(self, eps=1e-3):
mean = self.states.mean(0, keepdim=True)
std = self.states.std(0, keepdim=True) + eps
self.states = (self.states - mean) / std
self.next_states = (self.next_states - mean) / std
return mean, std
def clip(self, eps=1e-5):
self.action = torch.clip(self.action, - 1 + eps, 1 - eps)
def add_transition(self,
state: torch.Tensor,
action: torch.Tensor,
reward: torch.Tensor,
next_state: torch.Tensor,
done: torch.Tensor):
if not isinstance(state, torch.Tensor):
state = self.to_tensor(state)
action = self.to_tensor(action)
reward = self.to_tensor(reward)
next_state = self.to_tensor(next_state)
done = self.to_tensor(done)
self.states[self.pointer] = state
self.actions[self.pointer] = action
self.rewards[self.pointer] = reward
self.next_states[self.pointer] = next_state
self.dones[self.pointer] = done
self.pointer = (self.pointer + 1) % self.buffer_size
self.size = min(self.size + 1, self.buffer_size)
def add_batch(self,
states: List[torch.Tensor],
actions: List[torch.Tensor],
rewards: List[torch.Tensor],
next_states: List[torch.Tensor],
dones: List[torch.Tensor]):
for state, action, reward, next_state, done in zip(states, actions, rewards, next_states, dones):
self.add_transition(state, action, reward, next_state, done)
@staticmethod
def dataset_stats(dataset):
episode_returns = []
returns = 0
episode_length = 0
for reward, done in zip(dataset["rewards"], dataset["terminals"]):
if done:
episode_returns.append(returns)
returns = 0
episode_length = 0
else:
episode_length += 1
returns += reward
if episode_length == 1000:
episode_returns.append(returns)
returns = 0
episode_length = 0
episode_returns = np.array(episode_returns)
return episode_returns.mean(), episode_returns.std()
def train_val_split(replay_buffer: ReplayBuffer, val_size: float) -> Tuple[ReplayBuffer, ReplayBuffer]:
data_size = replay_buffer.size
val_size = int(data_size * val_size)
permutation = torch.randperm(data_size)
train_rb = ReplayBuffer(replay_buffer.state_dim, replay_buffer.action_dim)
val_rb = ReplayBuffer(replay_buffer.state_dim, replay_buffer.action_dim)
train_rb.add_batch(
replay_buffer.states[permutation[val_size:]],
replay_buffer.actions[permutation[val_size:]],
replay_buffer.rewards[permutation[val_size:]],
replay_buffer.next_states[permutation[val_size:]],
replay_buffer.dones[permutation[val_size:]]
)
val_rb.add_batch(
replay_buffer.states[permutation[:val_size]],
replay_buffer.actions[permutation[:val_size]],
replay_buffer.rewards[permutation[:val_size]],
replay_buffer.next_states[permutation[:val_size]],
replay_buffer.dones[permutation[:val_size]]
)
return train_rb, val_rb