-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
358 lines (273 loc) · 13.8 KB
/
train.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
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
import torch
import torch.nn as nn
import torch.optim as optim
from torch.distributions import Categorical
import numpy as np
from collections import deque
import random
from fortnite_env import FortniteEnv
from pushbullet import Pushbullet
from dotenv import load_dotenv
load_dotenv()
import os
access_token = os.getenv("ACCESS_TOKEN")
title = "GAME STOPPED GAME STOPPED GAME STOPPED"
body = "GAME STOPPED GAME STOPPED GAME STOPPED"
# Hyperparameters
LEARNING_RATE = 1e-6
GAMMA = 0.99
GAE_LAMBDA = 0.95
PPO_EPSILON = 0.2
VALUE_LOSS_COEF = 0.5
ENTROPY_COEF = 0.15
TRANSITION_LOSS_COEF = 0.1
MAX_GRAD_NORM = 0.5
NUM_MINI_BATCHES = 4
PPO_EPOCHS = 10
BATCH_SIZE = 256
STEPS_PER_EPISODE = 2048
NO_OF_EPISODES = 1000
FRAME_STACK = 4
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
class FrameStackEnv:
def __init__(self, env, num_stack):
self.env = env
self.num_stack = num_stack
self.frames = deque([], maxlen=num_stack)
def reset(self):
obs, _ = self.env.reset()
for _ in range(self.num_stack):
self.frames.append(obs)
return self.get_obs(), {}
def step(self, action):
obs, reward, terminated, truncated, info = self.env.step(action)
if reward == 696969:
Pushbullet(access_token).push_note(title, body)
print("Ending environment")
exit()
self.frames.append(obs)
return self.get_obs(), reward, terminated, truncated, info
def get_obs(self):
return np.concatenate(list(self.frames), axis=0)
class PerceptionComponent(nn.Module):
def __init__(self, input_shape, num_heads=8):
super(PerceptionComponent, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(input_shape[0] * FRAME_STACK, 32, kernel_size=3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1),
nn.ReLU(),
nn.Flatten()
)
with torch.no_grad():
sample_input = torch.zeros(1, input_shape[0] * FRAME_STACK, *input_shape[1:])
self.feature_size = self.conv(sample_input).shape[1]
self.fc = nn.Linear(self.feature_size, 512)
self.attention = nn.MultiheadAttention(embed_dim=512, num_heads=num_heads)
def forward(self, x):
x = self.conv(x)
x = self.fc(x) # Shape: (batch_size, 512)
x = x.unsqueeze(0) # Shape: (1, batch_size, 512)
attn_output, _ = self.attention(x, x, x) # Shape: (1, batch_size, 512)
attn_output = attn_output.squeeze(0) # Shape: (batch_size, 512)
return attn_output
class ReactivePolicy(nn.Module):
def __init__(self, input_size, discrete_action_dims):
super(ReactivePolicy, self).__init__()
self.input_size = input_size
self.discrete_action_dims = discrete_action_dims
# Common layers
self.common_layers = nn.Sequential(
nn.Linear(input_size, 256),
nn.ReLU(),
nn.Linear(256, 256),
nn.ReLU()
)
# Discrete action heads
self.discrete_action_heads = nn.ModuleDict()
for action_name, action_dim in discrete_action_dims.items():
self.discrete_action_heads[action_name] = nn.Linear(256, action_dim)
def forward(self, x):
common_features = self.common_layers(x) # Shape: (batch_size, 256)
discrete_action_logits = {}
for action_name in self.discrete_action_dims.keys():
discrete_action_logits[action_name] = self.discrete_action_heads[action_name](common_features) # Shape: (batch_size, action_dim)
return discrete_action_logits
def sample_action(self, x):
discrete_action_logits = self.forward(x)
discrete_actions = {}
log_probs = {}
for action_name, logits in discrete_action_logits.items():
dist = Categorical(logits=logits)
action = dist.sample()
discrete_actions[action_name] = action
log_probs[action_name] = dist.log_prob(action)
return discrete_actions, log_probs
def evaluate_actions(self, x, discrete_actions):
discrete_action_logits = self.forward(x)
log_probs = {}
entropies = {}
for action_name, logits in discrete_action_logits.items():
dist = Categorical(logits=logits)
if len(discrete_actions[action_name].shape) > 1:
discrete_actions[action_name] = discrete_actions[action_name].squeeze(-1)
log_probs[action_name] = dist.log_prob(discrete_actions[action_name])
entropies[action_name] = dist.entropy()
return log_probs, entropies
class ValueFunction(nn.Module):
def __init__(self, input_size):
super(ValueFunction, self).__init__()
self.fc = nn.Linear(input_size, 1)
def forward(self, x):
return self.fc(x) # Shape: (batch_size, 1)
class TransitionModel(nn.Module):
def __init__(self, state_dim, action_dim, num_heads=4):
super(TransitionModel, self).__init__()
self.fc = nn.Sequential(
nn.Linear(state_dim + action_dim, 256),
nn.ReLU(),
)
self.attention = nn.MultiheadAttention(embed_dim=256, num_heads=num_heads)
self.fc_out = nn.Linear(256, state_dim)
def forward(self, state, action):
x = torch.cat([state, action.unsqueeze(-1)], dim=-1) # Shape: (batch_size, state_dim + action_dim)
x = self.fc(x) # Shape: (batch_size, 256)
x = x.unsqueeze(0) # Shape: (1, batch_size, 256)
attn_output, _ = self.attention(x, x, x) # Shape: (1, batch_size, 256)
attn_output = attn_output.squeeze(0) # Shape: (batch_size, 256)
return self.fc_out(attn_output)
class FortniteAgent(nn.Module):
def __init__(self, state_dim, discrete_action_dims):
super(FortniteAgent, self).__init__()
self.perception = PerceptionComponent(state_dim)
self.policy = ReactivePolicy(512, discrete_action_dims)
self.value = ValueFunction(512)
total_action_dim = sum([1 for v in discrete_action_dims.keys()])
self.transition_model = TransitionModel(512, total_action_dim)
def forward(self, state):
features = self.perception(state)
discrete_action_logits = self.policy(features)
value = self.value(features)
return discrete_action_logits, value
def predict_next_state(self, state, discrete_actions):
features = self.perception(state)
# print(discrete_actions)
# Convert discrete actions to float and concatenate
discrete_part = torch.cat([action.float().unsqueeze(-1) for action in discrete_actions.values()], dim=-1)
# Use discrete actions only
action = discrete_part.squeeze(-1)
# print(action.shape)
return self.transition_model(features, action)
class PPOAgent:
def __init__(self, state_dim, discrete_action_dims):
self.agent = FortniteAgent(state_dim, discrete_action_dims).to(device)
self.optimizer = optim.Adam(self.agent.parameters(), lr=LEARNING_RATE)
self.memory = deque(maxlen=BATCH_SIZE)
self.discrete_action_dims = discrete_action_dims
self.discrete_action_names = list(discrete_action_dims.keys())
def select_action(self, state):
with torch.no_grad():
features = self.agent.perception(state)
discrete_actions, _ = self.agent.policy.sample_action(features)
log_probs, _ = self.agent.policy.evaluate_actions(features, discrete_actions)
return discrete_actions, log_probs
def update(self):
batch = list(self.memory)
states, discrete_actions,old_log_probs, rewards, next_states = map(np.array, zip(*batch))
states = torch.FloatTensor(states).to(device)
# Handle discrete actions
discrete_actions_dict = {}
for i, action_name in enumerate(self.discrete_action_names):
discrete_actions_dict[action_name] = torch.LongTensor(discrete_actions[:, i]).to(device)
old_log_probs_dict = {action_name: torch.FloatTensor([lp[action_name].item() for lp in old_log_probs]).to(device)
for action_name in self.discrete_action_names}
rewards = torch.FloatTensor(rewards).to(device)
next_states = torch.FloatTensor(next_states).to(device)
with torch.no_grad():
_, values = self.agent(states)
_, next_values = self.agent(next_states)
advantages = self.compute_gae(rewards, values, next_values).to(device)
returns = advantages + values
advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-8)
for _ in range(PPO_EPOCHS):
for indices in self.get_minibatch_indices():
mini_batch_states = states[indices]
mini_batch_discrete_actions = {k: v[indices].unsqueeze(-1) for k, v in discrete_actions_dict.items()}
mini_batch_old_log_probs = {k: v[indices] for k, v in old_log_probs_dict.items()}
mini_batch_advantages = advantages[indices]
mini_batch_returns = returns[indices]
mini_batch_values = values[indices]
features = self.agent.perception(mini_batch_states)
discrete_action_logits, new_values = self.agent(mini_batch_states)
new_log_probs, entropies = self.agent.policy.evaluate_actions(features, mini_batch_discrete_actions)
ratio = torch.exp(sum(new_log_probs.values()) - sum(mini_batch_old_log_probs.values()))
surr1 = ratio * mini_batch_advantages
surr2 = torch.clamp(ratio, 1 - PPO_EPSILON, 1 + PPO_EPSILON) * mini_batch_advantages
actor_loss = -torch.min(surr1, surr2).mean()
value_loss = nn.MSELoss()(new_values, mini_batch_returns)
entropy = sum(entropies.values()).mean()
# Add transition model loss
predicted_next_states = self.agent.predict_next_state(mini_batch_states, mini_batch_discrete_actions)
transition_loss = nn.MSELoss()(predicted_next_states, self.agent.perception(next_states[indices]))
loss = actor_loss + VALUE_LOSS_COEF * value_loss - ENTROPY_COEF * entropy + TRANSITION_LOSS_COEF * transition_loss
self.optimizer.zero_grad()
loss.backward()
nn.utils.clip_grad_norm_(self.agent.parameters(), MAX_GRAD_NORM)
self.optimizer.step()
self.memory.clear()
def compute_gae(self, rewards, values, next_values):
gae = 0
advantages = []
for step in reversed(range(len(rewards))):
delta = rewards[step] + GAMMA * next_values[step] - values[step]
gae = delta + GAMMA * GAE_LAMBDA * gae
advantages.insert(0, gae)
return torch.tensor(advantages)
def get_minibatch_indices(self):
indices = np.arange(BATCH_SIZE)
np.random.shuffle(indices)
return np.array_split(indices, NUM_MINI_BATCHES)
def train(env, agent):
state, _ = env.reset()
episode_reward = 0
episode_steps = 0
episode_count = 0
total_steps = 0
for _ in range(NO_OF_EPISODES):
for _ in range(STEPS_PER_EPISODE):
state_tensor = torch.FloatTensor(state).unsqueeze(0).to(device)
discrete_actions , log_probs = agent.select_action(state_tensor)
# print("discretetete",discrete_actions)
# Convert actions to numpy arrays for the environment
discrete_actions_np = {k: v.cpu().numpy().squeeze() for k, v in discrete_actions.items()}
# Use discrete actions only for the environment step
combined_action = discrete_actions_np
next_state, reward, terminated, truncated, _ = env.step(combined_action)
# Flatten the actions for storage in memory
flat_discrete_actions = np.array([v for v in discrete_actions_np.values()])
agent.memory.append((state, flat_discrete_actions , log_probs, reward, next_state))
episode_reward += reward
state = next_state
total_steps += 1
episode_steps += 1
if len(agent.memory) == BATCH_SIZE:
print("Updating agent")
agent.update()
agent.memory.clear()
torch.save(agent.agent.state_dict(), f'models/model{total_steps}.pth')
if terminated or truncated or episode_steps >= 1024:
print(f"Episode {episode_count}, Steps: {episode_steps}, Total Reward: {episode_reward}")
state, _ = env.reset()
episode_count += 1
episode_reward = 0
episode_steps = 0
# Initialize environment and agent
env = FrameStackEnv(FortniteEnv(), FRAME_STACK)
state_dim = env.env.observation_space.shape
# Simplified action space
discrete_action_dims = {'fire': 2} # Binary fire action (fire or not fire)
agent = PPOAgent(state_dim, discrete_action_dims)
# agent.agent.load_state_dict(torch.load('models/model15360.pth'))
train(env, agent)