-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
365 lines (307 loc) · 13.9 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
359
360
361
362
363
364
365
import argparse
from itertools import count
import signal
import sys
import os
import time
import numpy as np
import gym
import torch
import torch.autograd as autograd
from torch.autograd import Variable
import scipy.optimize
import matplotlib.pyplot as plt
from value import Value
from policy import Policy
from utils import *
from trpo import trpo_step
parser = argparse.ArgumentParser(description='PyTorch actor-critic example')
# Algorithm Parameters
parser.add_argument('--gamma', type=float, default=0.995, metavar='G', help='discount factor (default: 0.995)')
parser.add_argument('--lambda-', type=float, default=0.97, metavar='G', help='gae (default: 0.97)')
# Value Function Learning Parameters
parser.add_argument('--l2-reg', type=float, default=1e-3, metavar='G', help='(NOT USED)l2 regularization regression (default: 1e-3)')
parser.add_argument('--val-opt-iter', type=int, default=200, metavar='G', help='iteration number for value function learning(default: 200)')
parser.add_argument('--lr', type=float, default=1e-3, metavar='G', help='learning rate for value function (default: 1e-3)')
parser.add_argument('--value-memory', type=int, default=1, metavar='G', help='ratio of past value to be used to batch size (default: 1)')
parser.add_argument('--value-memory-shuffle', action='store_true',help='if not shuffled latest memory stay') # TODO: implement
# Policy Optimization parameters
parser.add_argument('--max-kl', type=float, default=1e-2, metavar='G', help='max kl value (default: 1e-2)')
parser.add_argument('--damping', type=float, default=1e-1, metavar='G', help='damping (default: 1e-1)')
parser.add_argument('--fisher-ratio', type=float, default=1, metavar='G', help='ratio of data to calcualte fisher vector product (default: 1)')
# Environment parameters
parser.add_argument('--env-name', default="Pendulum-v0", metavar='G', help='name of the environment to run')
parser.add_argument('--seed', type=int, default=543, metavar='N', help='random seed (default: 1)')
# Training length
parser.add_argument('--batch-size', type=int, default=5000, metavar='N', help='number of steps per iteration')
parser.add_argument('--episode-length', type=int, default=1000, metavar='N', help='max step size for one episode')
parser.add_argument('--max-iteration-number', type=int, default=200, metavar='N', help='max policy iteration number')
# Rendering
parser.add_argument('--render', action='store_true', help='render the environment')
# Logging
parser.add_argument('--log-interval', type=int, default=1, metavar='N', help='interval between training status logs (default: 10)')
parser.add_argument('--log', action='store_true', help='log the results at the end')
parser.add_argument('--log-dir', type=str, default=".", metavar='N', help='log directory')
parser.add_argument('--log-prefix', type=str, default="log", metavar='N', help='log file prefix')
# Load
parser.add_argument('--load', action='store_true', help='load models')
parser.add_argument('--save', action='store_true', help='load models')
parser.add_argument('--load-dir', type=str, default=".", metavar='N', help='')
args = parser.parse_args()
env = gym.make(args.env_name)
env.seed(args.seed)
num_inputs = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
torch.set_printoptions(profile="full")
if args.load:
policy_net = Policy(num_inputs, num_actions,30)
value_net = Value(num_inputs,30)
set_flat_params_to(value_net, loadParameterCsv(args.load_dir+"/ValueNet"))
set_flat_params_to(policy_net, loadParameterCsv(args.load_dir+"/PolicyNet"))
print("Networks are loaded from "+args.load_dir+"/")
else:
policy_net = Policy(num_inputs, num_actions,30)
value_net = Value(num_inputs,30)
def signal_handler(sig, frame):
""" Signal Handler to save the networks when shutting down via ctrl+C
Parameters:
Returns:
"""
if(args.save):
valueParam = get_flat_params_from(value_net)
policyParam = get_flat_params_from(policy_net)
saveParameterCsv(valueParam,args.load_dir+"/ValueNet")
saveParameterCsv(policyParam,args.load_dir+"/PolicyNet")
print("Networks are saved in "+args.load_dir+"/")
print('Closing!!')
env.close()
sys.exit(0)
def prepare_data(batch,valueBatch,previousBatch):
""" Get the batch data and calculate value,return and generalized advantage
Detail: TODO
Parameters:
batch (dict of arrays of numpy) : TODO
valueBatch (dict of arrays of numpy) : TODO
previousBatch (dict of arrays of numpy) : TODO
Returns:
"""
# TODO : more description above
stateList = [ torch.from_numpy(np.concatenate(x,axis=0)) for x in batch["states"]]
actionsList = [torch.from_numpy(np.concatenate(x,axis=0)) for x in batch["actions"]]
for states in stateList:
value = value_net.forward(states)
batch["values"].append(value)
advantagesList = []
returnsList = []
rewardsList = []
for rewards,values,masks in zip(batch["rewards"],batch["values"],batch["mask"]):
returns = torch.Tensor(len(rewards),1)
advantages = torch.Tensor(len(rewards),1)
deltas = torch.Tensor(len(rewards),1)
prev_return = 0
prev_value = 0
prev_advantage = 0
for i in reversed(range(len(rewards))):
returns[i] = rewards[i] + args.gamma * prev_value * masks[i] # TD
# returns[i] = rewards[i] + args.gamma * prev_return * masks[i] # Monte Carlo
deltas[i] = rewards[i] + args.gamma * prev_value * masks[i]- values.data[i]
advantages[i] = deltas[i] + args.gamma * args.lambda_* prev_advantage* masks[i]
prev_return = returns[i, 0]
prev_value = values.data[i, 0]
prev_advantage = advantages[i, 0]
returnsList.append(returns)
advantagesList.append(advantages)
rewardsList.append(torch.Tensor(rewards))
batch["states"] = torch.cat(stateList,0)
batch["actions"] = torch.cat(actionsList,0)
batch["rewards"] = torch.cat(rewardsList,0)
batch["returns"] = torch.cat(returnsList,0)
advantagesList = torch.cat(advantagesList,0)
batch["advantages"] = (advantagesList- advantagesList.mean()) / advantagesList.std()
valueBatch["states"] = torch.cat(( previousBatch["states"],batch["states"]),0)
valueBatch["targets"] = torch.cat((previousBatch["returns"],batch["returns"]),0)
def update_policy(batch):
""" Get advantage , states and action and calls trpo step
Parameters:
batch (dict of arrays of numpy) : TODO (batch is different than prepare_data by structure)
Returns:
"""
advantages = batch["advantages"]
states = batch["states"]
actions = batch["actions"]
trpo_step(policy_net, states,actions,advantages , args.max_kl, args.damping)
def update_value(valueBatch):
""" Get valueBatch and run adam optimizer to learn value function
Parameters:
valueBatch (dict of arrays of numpy) : TODO
Returns:
"""
# shuffle the data
dataSize = valueBatch["targets"].size()[0]
permutation = torch.randperm(dataSize)
input = valueBatch["states"][permutation]
target = valueBatch["targets"][permutation]
iter = args.val_opt_iter
batchSize = int(dataSize/ iter)
loss_fn = torch.nn.MSELoss(reduction='sum')
optimizer = torch.optim.Adam(value_net.parameters(), lr=args.lr)
for t in range(iter):
prediction = value_net(input[t*batchSize:t*batchSize+batchSize])
loss = loss_fn(prediction, target[t*batchSize:t*batchSize+batchSize])
# XXX : Comment out for debug
# if t%100==0:
# print("\t%f"%loss.data)
optimizer.zero_grad()
loss.backward()
optimizer.step()
def save_to_previousBatch(previousBatch,batch):
""" Save previous batch to use in future value optimization
Details: TODO
Parameters:
Returns:
"""
if args.value_memory<0:
print("Value memory should be equal or greater than zero")
elif args.value_memory>0:
if previousBatch["returns"].size() == 0:
previousBatch= {"states":batch["states"],
"returns":batch["returns"]}
else:
previous_size = previousBatch["returns"].size()[0]
size = batch["returns"].size()[0]
if previous_size/size == args.value_memory:
previousBatch["states"] = torch.cat([previousBatch["states"][size:],batch["states"]],0)
previousBatch["returns"] = torch.cat([previousBatch["returns"][size:],batch["returns"]],0)
else:
previousBatch["states"] = torch.cat([previousBatch["states"],batch["states"]],0)
previousBatch["returns"] = torch.cat([previousBatch["returns"],batch["returns"]],0)
if args.value_memory_shuffle:
permutation = torch.randperm(previousBatch["returns"].size()[0])
previousBatch["states"] = previousBatch["states"][permutation]
previousBatch["returns"] = previousBatch["returns"][permutation]
def calculate_loss(reward_sum_mean,reward_sum_std,test_number = 10):
""" Calculate mean cummulative reward for test_nubmer of trials
Parameters:
reward_sum_mean (list): holds the history of the means.
reward_sum_std (list): holds the history of the std.
Returns:
list: new value appended means
list: new value appended stds
"""
rewardSum = []
for i in range(test_number):
state = env.reset()
rewardSum.append(0)
for t in range(args.episode_length):
state, reward, done, _ = env.step(policy_net.get_action(state)[0] )
state = np.transpose(state)
rewardSum[-1] += reward
if done:
break
reward_sum_mean.append(np.array(rewardSum).mean())
reward_sum_std.append(np.array(rewardSum).std())
return reward_sum_mean, reward_sum_std
def log(rewards):
""" Saves mean and std over episodes in log file
Parameters:
Returns:
"""
# TODO : add duration to log
filename = args.log_dir+"/"+ args.log_prefix \
+ "_env_" + args.env_name \
+ "_maxIter_" + str(args.max_iteration_number) \
+ "_batchSize_" + str(args.batch_size) \
+ "_gamma_" + str(args.gamma) \
+ "_lambda_" + str(args.lambda_) \
+ "_lr_" + str(args.lr) \
+ "_valOptIter_" + str(args.val_opt_iter)
if os.path.exists(filename + "_index_0.csv"):
id = 0
file = filename + "_index_" + str(id)
while os.path.exists(file + ".csv"):
id = id +1
file = filename + "_index_" + str(id)
filename = file
else:
filename = filename + "_index_0"
import csv
filename = filename+ ".csv"
pythonVersion = sys.version_info[0]
if pythonVersion == 3:
with open(filename, 'w', newline='') as csvfile:
spamwriter = csv.writer(csvfile, delimiter=' ',
quotechar='|', quoting=csv.QUOTE_MINIMAL)
spamwriter.writerow(rewards)
elif pythonVersion == 2:
with open(filename, 'w', ) as csvfile:
spamwriter = csv.writer(csvfile, delimiter=' ',
quotechar='|', quoting=csv.QUOTE_MINIMAL)
spamwriter.writerow(rewards)
def main():
"""
Parameters:
Returns:
"""
signal.signal(signal.SIGINT, signal_handler)
time_start = time.time()
reward_sum_mean,reward_sum_std = [], []
previousBatch= {"states":torch.Tensor(0) ,
"returns":torch.Tensor(0)}
reward_sum_mean,reward_sum_std = calculate_loss(reward_sum_mean,reward_sum_std)
print("Initial loss \n\tloss | mean : %6.4f / std : %6.4f"%(reward_sum_mean[-1],reward_sum_std[-1]) )
for i_episode in range(args.max_iteration_number):
time_episode_start = time.time()
# reset batches
batch = {"states":[] ,
"actions":[],
"next_states":[] ,
"rewards":[],
"returns":[],
"values":[],
"advantages":[],
"mask":[]}
valueBatch = {"states" :[],
"targets" : []}
num_steps = 0
while num_steps < args.batch_size:
state = env.reset()
reward_sum = 0
states,actions,rewards,next_states,masks = [],[],[],[],[]
steps = 0
for t in range(args.episode_length):
action = policy_net.get_action(state)[0] # agent
next_state, reward, done, info = env.step(action)
next_state = np.transpose(next_state)
mask = 0 if done else 1
masks.append(mask)
states.append(state)
actions.append(action)
next_states.append(next_state)
rewards.append(reward)
state = next_state
reward_sum += reward
steps+=1
if args.render:
env.render()
if done:
break
batch["states"].append(np.expand_dims(states, axis=1) )
batch["actions"].append(actions)
batch["next_states"].append(np.expand_dims(next_states, axis=1))
batch["rewards"].append(rewards)
batch["mask"].append(masks)
num_steps += steps
prepare_data(batch,valueBatch,previousBatch)
update_policy(batch) # First policy update to avoid overfitting
update_value(valueBatch)
save_to_previousBatch(previousBatch,batch)
print("episode %d | total: %.4f "%( i_episode, time.time()-time_episode_start))
reward_sum_mean,reward_sum_std = calculate_loss(reward_sum_mean,reward_sum_std)
print("\tloss | mean : %6.4f / std : %6.4f"%(reward_sum_mean[-1],reward_sum_std[-1]) )
if args.log:
print("Data is logged in "+args.log_dir+"/")
log(reward_sum_mean)
print("Total training duration: %.4f "%(time.time()-time_start))
env.close()
if __name__ == '__main__':
main()