-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
315 lines (255 loc) · 11.7 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
import Env
import gym
from stable_baselines3.common.vec_env import SubprocVecEnv
import argparse
import configs
import torch
import random
import numpy as np
from time import time
import cv2
from torch.optim import Adam
import os
from VecMonitor import VecMonitor
#Reproducibility
np.random.seed(0)
random.seed(0)
torch.manual_seed(0)
torch.backends.cudnn.benchmark = True
torch.use_deterministic_algorithms(True)
import importlib
def parse_args():
parser = argparse.ArgumentParser(description="Train Tess")
parser.add_argument("--substrate",type=str,default="clean_up")
parser.add_argument("--model",type=str,default="impala")
parser.add_argument("--train-config",type=str,default="ImpalaConfig")
return parser.parse_args()
if __name__ == "__main__":
args = parse_args()
# Getting config information
config = getattr(configs, args.train_config)
#Setting Up Environment
env = SubprocVecEnv([lambda: gym.make("TessEnv-v1",\
render_mode="rgb_array",
name=args.substrate)\
for _ in range(config.num_envs)])
env = VecMonitor(env)
obs_s = env.observation_space.shape
act_s = env.action_space.shape
num_act = env.get_attr("num_act")[0]
#Setting Up Model Instance
model = importlib.import_module(f"models.{args.model}").Model
agent = model(obs_s[1:], num_act)
agent = agent.to("cuda")
# Setting hyperparameters
num_steps = config.num_steps
num_envs = config.num_envs
batch_size = num_envs * num_steps * act_s[0]
minibatch_size = batch_size // config.minibatch
total_timesteps = config.total_timesteps
num_updates = total_timesteps // batch_size
lr = config.lr
gae = config.gae
clip_coef = config.clip_coef
gamma = config.gamma
ent_coef = config.ent_coef
epoch = config.epoch
clip_v_loss=config.clip_v_loss
#Wandb
import wandb
wandb.login()
run = wandb.init(project="Tess",name=f"{num_envs}-{num_steps}-{config.minibatch}-{lr}-{clip_coef}-{ent_coef}-{epoch}")
#Setting Up Optimizer
optimizer = Adam(agent.parameters(), lr=lr)
#Observation Wrapping
obs = env.reset()
done = torch.tensor([0]*num_envs,device="cuda").view(-1,1).expand(-1,act_s[0])
shape = (obs_s[-1],)+obs_s[-3:-1]
obs = torch.from_numpy(obs)\
.permute((0,1,4,2,3))\
.to("cuda")
#Rollout transactions
roll_o = torch.zeros((num_steps, num_envs, act_s[0],)+shape, device="cuda")
roll_a = torch.zeros((num_steps, num_envs, act_s[0]), device="cuda")
roll_lp = torch.zeros((num_steps, num_envs, act_s[0]), device="cuda")
roll_rew = torch.zeros((num_steps, num_envs, act_s[0]), device="cuda")
roll_dones = torch.zeros((num_steps, num_envs, act_s[0]), device="cuda")
roll_val = torch.zeros((num_steps, num_envs, act_s[0]), device="cuda")
#LSTM helper
h_0 = torch.zeros((num_envs, act_s[0], agent.last_layer),device="cuda",dtype=torch.float32)
c_0 = torch.zeros((num_envs, act_s[0], agent.last_layer),device="cuda",dtype=torch.float32)
#Annealing Stuffs
anneal_lr = lambda update: lr * (total_timesteps - update*batch_size) / total_timesteps
#Metrics
last_scores = [0] * num_envs
last_act_scores = np.zeros((num_envs,act_s[0]))
last_episode_scores = [0] * num_envs
mean_act_score = 0
#Metric Helpers
done_envs = [False] * num_envs
for update in range(num_updates):
annealed_lr = anneal_lr(update)
optimizer.param_groups[0]["lr"] = annealed_lr
print((update +1) * batch_size)
for step in range(num_steps):
#Render 1 env
img = env.get_images()[0]
if config.visual:
if update % 10 == 0:
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
cv2.imshow("image",cv2.resize(img, (700,500)))
cv2.waitKey(50)
else:
cv2.destroyAllWindows()
roll_dones[step] = done
roll_o[step] = obs
#Sample action and critic with logprob
with torch.no_grad():
batch_o = obs.view((-1,)+shape)
h_0 = h_0.view((-1,)+(agent.last_layer,))
c_0 = h_0.view((-1,)+(agent.last_layer,))
act, log_prob, value, (h_0, c_0) = agent.sample_act_and_value(batch_o, history=(h_0,c_0))
act = act.view(num_envs,-1)
log_prob = log_prob.view(num_envs,-1)
value = value.view(num_envs,-1)
roll_a[step] = act
roll_lp[step] = log_prob
roll_val[step] = value
h_0 = h_0.view(num_envs,act_s[0],agent.last_layer)
c_0 = c_0.view(num_envs,act_s[0],agent.last_layer)
obs, rew, done, info = env.step(act)
real_rewards = []
episode_rewards = []
log = False
for ind,i in enumerate(done):
if i:
log = True
real_ = info[ind]["real_rewards"]
epis_ = info[ind]["episode_rewards"]
last_act_scores[ind] = real_
last_scores[ind] = real_.sum()
last_episode_scores[ind] = epis_.sum()
h_0[ind] *= 0
c_0[ind] *= 0
done_envs[ind] = True
save = False
if not (False in done_envs):
done_envs = [False] * num_envs
std_epis = np.std(last_scores)
std_acts = np.std(last_act_scores)
mean_act = np.mean(last_act_scores)
if mean_act > mean_act_score:
save = True
mean_act_score = mean_act
mean_epis = np.mean(last_episode_scores)
mean_real = np.mean(last_scores)
run.log({"mean_real":mean_real,"std_scores":std_epis,\
"mean_epis":mean_epis,"std_acts":std_acts,\
"mean_act":mean_act})
last_scores = [0] * num_envs
last_act_scores = np.zeros((num_envs,act_s[0]))
last_episode_scores = [0] * num_envs
obs = torch.from_numpy(obs)\
.permute((0,1,4,2,3))\
.to("cuda")
if save:
try:
os.makedirs(f"./Tess/saved_models/{args.substrate}/",exist_ok=True)
except:
pass
torch.save(agent.state_dict(),f"./Tess/saved_models/{args.substrate}/{args.substrate}-v2.pt")
done = torch.from_numpy(done).to("cuda").view(-1,1).expand(-1,act_s[0])
#Unknown reason sometimes done comes as bool tensor ( probably related with SB3 )
if done.dtype == torch.bool:
done = torch.where(done, 1, 0)
rew = torch.from_numpy(rew).view(num_envs,-1).to("cuda")
roll_rew[step] = rew
with torch.no_grad():
batch_o = obs.view((-1,)+shape)
h_0 = h_0.view(-1,agent.last_layer)
c_0 = c_0.view(-1,agent.last_layer)
val_plus1 = agent.get_value(batch_o, history=(h_0,c_0)).view(num_envs,-1)
advantages = torch.zeros_like(roll_rew, device="cuda")
lastgaelam = 0
for t in reversed(range(num_steps)):
if t == num_steps - 1:
nextnonterminal = (1.0 - done)
nextvalues = val_plus1
else:
nextnonterminal = 1.0 - roll_dones[t + 1]
nextvalues = roll_val[t + 1]
delta = roll_rew[t] + gamma * nextvalues * nextnonterminal - roll_val[t]
advantages[t] = lastgaelam = delta + gamma * gae * nextnonterminal * lastgaelam
returns = advantages + roll_val
b_obs = roll_o.view((-1,)+shape)
b_act = roll_a.view(-1)
b_logprobs = roll_lp.view(-1)
b_returns = returns.view(-1)
b_adv = advantages.view(-1)
b_val = roll_val.view(-1)
inds = np.arange(batch_size,)
for ith_e in range(epoch):
#np.random.shuffle(inds)
for start in range(0, batch_size, minibatch_size):
end = start + minibatch_size
minibatch_ind = inds[start:end]
# Fill LSTM state batches
mb_obs = b_obs[minibatch_ind].view((-1,num_envs, act_s[0],)+shape)
h__0 = [torch.zeros((num_envs*act_s[0],agent.last_layer),dtype=torch.float32,device="cuda")]
c__0 = [torch.zeros((num_envs*act_s[0],agent.last_layer),dtype=torch.float32,device="cuda")]
with torch.no_grad():
for i in range(mb_obs.shape[0] - 1):
_, (h_n,c_n) = agent.forward(mb_obs[i].view((-1,)+shape),history=(h__0[-1],c__0[-1]))
for index,done_info in enumerate(roll_dones[start//act_s[0]//num_envs+i+1]):
if done_info[0] == 1:
h_n = h_n.view(num_envs,act_s[0],agent.last_layer)
c_n = c_n.view(num_envs,act_s[0],agent.last_layer)
h_n[index] *= 0
c_n[index] *= 0
h_n = h_n.view(-1,agent.last_layer)
c_n = c_n.view(-1,agent.last_layer)
h__0.append(h_n)
c__0.append(c_n)
h__0 = torch.concatenate(h__0,dim=0)
c__0 = torch.concatenate(c__0,dim=0)
mb_obs = mb_obs.view((-1,)+shape)
mb_advantages = b_adv[minibatch_ind]
if config.use_advantage_norm:
mb_advantages = (mb_advantages - mb_advantages.mean()) / (mb_advantages.std()+1e-8)
mb_actions = b_act[minibatch_ind]
new_logprob, entropy, value = agent.check_action_and_value(mb_obs,\
mb_actions, history=(h__0,c__0))
mb_logprob = b_logprobs[minibatch_ind]
mb_returns = b_returns[minibatch_ind]
mb_values = b_val[minibatch_ind]
log = new_logprob - mb_logprob
ratio = (log).exp()
with torch.no_grad():
approx_kl = ((ratio - 1) - log).mean()
if ith_e == epoch-1:
pass
#print(approx_kl)
ratio_clip = torch.clamp(ratio, 1 - clip_coef, 1 + clip_coef)
pg_loss = torch.max(-mb_advantages * ratio_clip, -mb_advantages * ratio).mean()
#Clipping Value Loss
if clip_v_loss:
clip_v = config.clip_v
v_loss = (value - mb_returns).square()
v_clipped = mb_values + torch.clamp(
value - mb_values,
-clip_v,
clip_v,
)
v_loss_clipped = (v_clipped - mb_returns).square()
v_loss_max = torch.max(v_loss, v_loss_clipped)
v_loss = 0.5 * v_loss_max.mean()
else:
v_loss = (value - mb_returns).square().mean() * .5
ent_loss = entropy.mean()
loss = pg_loss + v_loss - ent_loss * ent_coef
run.log({"entropy":ent_loss,"policy_loss":pg_loss,\
"v_loss":v_loss,"kl":approx_kl})
optimizer.zero_grad(set_to_none = True)
loss.backward()
torch.nn.utils.clip_grad_norm_(agent.parameters(), .5)
optimizer.step()