-
Notifications
You must be signed in to change notification settings - Fork 0
/
dqn_vecenv.py
218 lines (163 loc) · 6.34 KB
/
dqn_vecenv.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
#%%
import os
import torch
from torch import nn
import itertools
from collections import deque
import numpy as np
import random
from utilities.logger import logger
from utilities.mass_wrapper import MASS_Env, MASS_VecEnv
from tqdm import tqdm
import msgpack
from utilities.msgpack_numpy import patch as msgpck_numpy_patch
msgpck_numpy_patch()
#%%
GAMMA = 0.99
BATCH_SIZE = 32
BUFFER_SIZE = int(1e6)
MIN_REPLAY_SIZE = 50000
EPSILON_START = 1.0
EPSILON_END = 0.1
EPSILON_DECAY = int(1e6)
TARGET_UPDATE_FREQ = 10000
NUM_ENVS = 4
LR = 2.5e-4
SAVE_INTERVAL = 10000
LOG_INTERVAL = 1000
PROJECT_NAME = 'MASS_DQN'
PROJECT_ID = '_0_18'
SAVE_PATH = './models/' + PROJECT_NAME + '/' + PROJECT_ID + '_Model.pack'
#%%
def CNN(obs_space, spissitude, final_layer):
n_input_channels = obs_space.shape[0]
cnn = nn.Sequential(
nn.Conv3d(n_input_channels, spissitude[0], kernel_size=8, stride=4),
nn.ReLU(),
nn.Conv3d(spissitude[0], spissitude[1], kernel_size=4, stride=2),
nn.ReLU(),
nn.Conv3d(spissitude[1], spissitude[2], kernel_size=3, stride=1),
nn.ReLU(),
nn.Flatten())
# compute shape by doing one forward pass
with torch.no_grad():
n_flatten = cnn(torch.as_tensor(obs_space.sample()[None]).float()).shape[1]
out = nn.Sequential(cnn, nn.Linear(n_flatten, final_layer), nn.ReLU())
return out
class Network(nn.Module):
def __init__(self, env, device) -> None:
super().__init__()
self.num_act = env.act_space_n
self.device = device
conv_net = CNN(env.obs_space, spissitude=(32, 64, 64), final_layer=512)
self.net = nn.Sequential(conv_net, nn.Linear(512,self.num_act))
def forward(self, x):
return self.net(x)
def act(self, obss, epsilon):
obss_t = torch.as_tensor(obss, dtype=torch.float32, device=self.device)
q_vals = self(obss_t)
max_q_indices = torch.argmax(q_vals, dim=1)
act = max_q_indices.detach().tolist()
for i in range(len(act)):
rnd_sample = random.random()
if rnd_sample <= epsilon:
act[i] = random.randint(0, self.num_act - 1)
return act
def compute_loss(self, transitions, target_network):
obs = [t[0] for t in transitions]
act = np.asarray([t[1] for t in transitions])
rew = np.asarray([t[2] for t in transitions])
done = np.asarray([t[3] for t in transitions])
new_obs = np.asarray([t[4] for t in transitions])
obs_t = torch.as_tensor(obs, dtype=torch.float32, device=self.device).unsqueeze(-1)
act_t = torch.as_tensor(act, dtype=torch.int64, device=self.device).unsqueeze(-1)
rew_t = torch.as_tensor(rew, dtype=torch.float32, device=self.device).unsqueeze(-1)
done_t = torch.as_tensor(done, dtype=torch.float32).unsqueeze(-1)
new_obs_t = torch.as_tensor(new_obs, dtype=torch.float32, device=self.device).unsqueeze(-1)
# Compute Targets
target_q_val = target_network(new_obs_t)
max_target_q_val = target_q_val.max(dim=1, keepdim=True)[0]
targets = rew_t + GAMMA * (1 - done_t) * max_target_q_val
# Compute Loss
q_vals = self(obs_t)
action_q_vals = torch.gather(input=q_vals, dim=1, index=act_t)
loss = nn.functional.smooth_l1_loss(action_q_vals, targets)
return loss
def save(self, save_path):
params = {k: t.detach().cpu().numpy() for k, t in self.state_dict().items()}
params_data = msgpack.dumps(params)
os.mkdir(os.path.dirname(save_path), exist_ok=True)
with open(save_path, 'wb') as f:
f.write(params_data)
def load(self, load_path):
if not os.path.exists(load_path):
raise FileNotFoundError(load_path)
with open(load_path, 'rb') as f:
params_numpy = msgpack.loads(f.read())
params = {k: torch.as_tensor(v, device=self.device) for k, v in params_numpy.items()}
self.load_state_dict(params)
#%%
device = torch.device('cuda:0' if torch.cuda.is_available() else "cpu")
env = MASS_VecEnv(NUM_ENVS)
replay_buffer = deque(maxlen=BUFFER_SIZE)
epinfos_buffer = deque([], maxlen=100)
episode_count = 0.0
# Initialize Networks
online_network = Network(env, device = device)
target_network = Network(env, device = device)
online_network.to(device)
target_network.to(device)
target_network.load_state_dict(online_network.state_dict())
optimizer = torch.optim.Adam(online_network.parameters(), lr=LR)
# Initialize the Replay Buffer
obss = env.reset()
for _ in range(MIN_REPLAY_SIZE):
acts = [e.act_space_sample() for e in env.es]
new_obss, rews, dones, _ = env.step(acts)
for obs, action, rew, done, new_obs in zip(obss, acts, rews, dones, new_obss):
transition = (obs, action, rew, done, new_obs)
replay_buffer.append(transition)
obss = new_obss
#%%
# Main training loop
obss = env.reset()
log = logger(PROJECT_NAME, PROJECT_ID, ['WB', 'LO'])
tq = tqdm()
for step in itertools.count():
# tq.update(1)
epsilon = np.interp(step * NUM_ENVS, [0, EPSILON_DECAY], [EPSILON_START, EPSILON_END])
acts = online_network.act(obss, epsilon)
new_obss, rews, dones, infos = env.step(acts)
for obs, action, rew, done, new_obs, info in zip(obss, acts, rews, dones, new_obss, infos):
transition = (obs, action, rew, done, new_obs)
replay_buffer.append(transition)
if done:
epinfos_buffer.append(info['episode'])
episode_count += 1
obss = new_obss
# Start Gradient Step
transitions = random.sample(replay_buffer, BATCH_SIZE)
loss = online_network.compute_loss(transitions, target_network)
# Gradient Descent
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Update target network
if step % TARGET_UPDATE_FREQ ==0 :
target_network.load_state_dict(online_network.state_dict())
# Logging
if step % LOG_INTERVAL == 0:
rew_mean = np.mean([e['r'] for e in epinfos_buffer]) or 0
len_mean = np.mean([e['l'] for e in epinfos_buffer]) or 0
# TODO: Update tags
log_dict = {
'AvgRew': rew_mean,
'AvgEpLen': len_mean,
'Episodes': episode_count
}
log.update(log_dict, step)
# Saving
if step % SAVE_INTERVAL == 0 and step != 0:
# print('Saving...')
online_network.save(SAVE_PATH)
# %%