-
Notifications
You must be signed in to change notification settings - Fork 0
/
PPO_main.py
292 lines (226 loc) · 10.5 KB
/
PPO_main.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
import torch
import numpy as np
import pandas as pd
import pickle
import ast
from torch.distributions import Categorical
from torch import nn
from langmodel import LanguageModel
from torch.nn import functional as F
from value_network import ValueNetwork
import numpy as np
import datetime as datetime
import time
from GenEnv import GenerationEnv3
import ast
################## Data Imports ########################
def load_set_from_file(filename):
with open(filename, 'rb') as file:
data = pickle.load(file)
return data
unique_words = load_set_from_file('shakespeare_word_set.pkl')
data = load_set_from_file('encoded_data_tensor.pkl')
with open('input.txt', 'r', encoding='utf-8') as f:
text = f.read()
chars = sorted(list(set(text)))
vocab_size = len(chars)
print(f"Vocabulary: {''.join(chars)}")
print(f"Vocabulary Size: {vocab_size}")
stoi = { ch:i for i,ch in enumerate(chars) }
itos = { i:ch for i,ch in enumerate(chars) }
encode = lambda s: [stoi[c] for c in s] # encoder: take a string, output a list of integers
decode = lambda l: ''.join([itos[i] for i in l])
################## Globl Variables ########################
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
env = GenerationEnv3(unique_words, data, decode, 0.65, device=device)
################### Hyperparameters #######################
name = "RLHF"
log_interval = 1 # print avg reward in the interval
max_episodes = 1_000 # max training episodes
max_timesteps = 500 # max timesteps in one episode
update_timestep = 2000 # update policy every n timesteps
vocab_size = 65
embed_dim = 256
hidden_dim = 512
solved_reward = 6 # stop training if avg_reward > solved_reward
block_size = 128
lr = 0.0006 #I will reduce the learning rate
betas = (0.9, 0.999) # same as the default Adam betas
gamma = 1 # discount factor
K_epochs = 4 # update policy for K epochs
eps_clip = 0.2 # clip parameter for PPO
random_seed = None
################### Helper Functions #######################
def save_training_hist(dict):
date = datetime.datetime.now().strftime("%Y-%m-%d")
with open(f'training_log_{date}.pkl', 'wb') as f:
pickle.dump(dict, f)
def save_model(ppo, name, epoch):
# Get the current date
date = datetime.datetime.now().strftime("%Y-%m-%d")
# Save the model
torch.save(ppo.actor.state_dict(), 'PPO_{}_{}_Epoch_{}.pth'.format(name, date, epoch))
def freeze_weights(model):
for param in model.parameters():
param.requires_grad = False
# Then unfreeze the parameters you are interested in
for name, param in model.named_parameters():
if 'blocks.3' in name or name in ['ln_f.weight', 'ln_f.bias', 'lm_head.weight', 'lm_head.bias']:
param.requires_grad = True
return model
#############################################################
class Memory:
def __init__(self):
self.actions = []
self.states = []
self.logprobs = []
self.rewards = []
self.is_terminals = []
def clear_memory(self):
del self.actions[:]
del self.states[:]
del self.logprobs[:]
del self.rewards[:]
del self.is_terminals[:]
class PPO2:
def __init__(self, vocab_size, embed_dim, hidden_dim, lr, betas, gamma, K_epochs, eps_clip, block_size):
self.lr = lr
self.betas = betas
self.gamma = gamma
self.eps_clip = eps_clip
self.K_epochs = K_epochs
self.block_size = block_size
self.actor = LanguageModel().to(device) # our policy given by transformer
self.actor.load_state_dict(torch.load('Best_PPO_RLHF_2023-08-01_Epoch_17.pth', map_location=torch.device(device))) #loading pre-trained weights
self.actor = freeze_weights(self.actor) # freezing all but the last transformer block and outputs (softmax) layer
self.critic = ValueNetwork(vocab_size, embed_dim, hidden_dim).to(device)
self.act_optimizer = torch.optim.Adam(self.actor.parameters(),
lr=lr, betas=betas)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=lr, betas=betas) # our value network to calculate advantage
self.policy_old = LanguageModel().to(device)
self.policy_old.load_state_dict(self.actor.state_dict())
self.MseLoss = nn.MSELoss()
def act(self, state, memory):
state = state[:, -self.block_size:]
logits, _ = self.actor(state)
logits = logits[:, -1, :] # becomes (B, C)
# apply softmax to get probabilities
probs = F.softmax(logits, dim=-1) # (B, C)
log_probs = F.log_softmax(logits, dim=-1) # log_probs used in PPO update
# sample from the distribution
action = torch.multinomial(probs, num_samples=1) # (B, 1)
memory.states.append(state)
memory.actions.append(action)
memory.logprobs.append(log_probs.squeeze(0)[action]) # B
return action # tensor of dim (1, 1)
def update(self, memory):
# Monte Carlo estimate of state rewards
rewards = []
discounted_reward = 0
for reward, is_terminal in zip(reversed(memory.rewards), reversed(memory.is_terminals)):
if is_terminal:
discounted_reward = 0
discounted_reward = reward + (self.gamma * discounted_reward)
rewards.insert(0, discounted_reward) # insert at front to regain original order
# Normalizing the rewards
rewards = torch.tensor(rewards, dtype=torch.float32).to(device).unsqueeze(1)
rewards = (rewards - rewards.mean()) / (rewards.std() + 1e-5)
# Convert list to tensor
old_states = torch.stack(memory.states).to(device).detach().squeeze(1)
old_actions = torch.stack(memory.actions).to(device).detach().squeeze(1)
old_logprobs = torch.stack(memory.logprobs).to(device).detach().squeeze(1)
# Optimization step
for _ in range(self.K_epochs):
# Get policy logits
logits, _ = self.actor(old_states)
logits = logits[:, -1, :] # only care about the last logits
# Get probabilities from logits
probs = F.softmax(logits, dim=-1)
# Get log probabilities
logprobs = F.log_softmax(logits, dim=-1)
# Get the logprob of the taken action
logprobs = torch.gather(logprobs, 1, old_actions)
# Calculating the entropy
dist_entropy = -(probs * logprobs).sum(-1).mean()
# Evaluating old values
state_values = self.critic(old_states)
# Calculating the policy loss
ratios = torch.exp(logprobs - old_logprobs.detach())
advantages = rewards - state_values.detach()
surr1 = ratios * advantages
surr2 = torch.clamp(ratios, 1 - self.eps_clip, 1 + self.eps_clip) * advantages
policy_loss = -torch.min(surr1, surr2) - 0.01 * dist_entropy
# Calculating the value loss
value_loss = 0.5 * self.MseLoss(state_values, rewards)
# Taking a gradient step for policy network
self.act_optimizer.zero_grad()
policy_loss.mean().backward() # get average policy loss across the batch dimension
self.act_optimizer.step()
# Taking a gradient step for value network
self.critic_optimizer.zero_grad()
value_loss.mean().backward()
self.critic_optimizer.step()
def main():
if random_seed:
torch.manual_seed(random_seed)
env.seed(random_seed)
memory = Memory()
ppo = PPO2(vocab_size, embed_dim, hidden_dim, lr, betas, gamma, K_epochs, eps_clip, block_size)
print(f"The Learning Rate is: {lr}, The Betas for Adam Optimizer are: {betas}")
# logging variables
running_reward = 0
avg_length = 0
time_step = 0
training_dict = {'reward_history': [], 'cumulative_reward_history': [], 'time_history': []}
# training loop
for i_episode in range(1, max_episodes+1):
state = env.reset()
episode_rewards = 0
start_time = time.time()
old_weights = ppo.actor.state_dict()
for t in range(max_timesteps):
time_step +=1
# running policy_old:
action = ppo.act(state, memory)
state, reward, done = env.step(state, action)
state.to(device)
episode_rewards += reward
# saving reward and is_terminal:
memory.rewards.append(reward)
memory.is_terminals.append(done)
# update if its time
if time_step % update_timestep == 0:
# print(memory.states)
ppo.update(memory)
memory.clear_memory()
time_step = 0
running_reward += reward
if done:
break
training_dict['cumulative_reward_history'].append(running_reward)
training_dict['reward_history'].append(episode_rewards)
avg_length += t
# stop training if we reach a high level of reward
if running_reward > (log_interval*solved_reward):
print("########## Solved! ##########")
print('Episode {} \t avg length: {} \t reward: {}'.format(i_episode, avg_length, running_reward))
torch.save(ppo.actor.state_dict(), 'PPO_{}.pth'.format(name)) # i do like this here through.
time_to_finish = time.time() - sum(training_dict['time_history'])
break
if i_episode == max_episodes+1:
time_to_finish = time.time() - sum(training_dict['time_history'])
break
# logging
if i_episode % log_interval == 0: # this I also like
avg_length = avg_length/log_interval
update_interval = time.time() - start_time
training_dict['time_history'].append(update_interval)
running_reward = running_reward/log_interval
save_model(ppo, name, i_episode)
save_training_hist(training_dict)
print('Episode {} \t num tokens generated: {} \t reward: {}'.format(i_episode, avg_length, running_reward))
running_reward = 0
avg_length = 0
if __name__ == '__main__':
main()