-
Notifications
You must be signed in to change notification settings - Fork 5
/
training.py
412 lines (340 loc) · 15.1 KB
/
training.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
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
"""
Description: Main training code
Author: Tae-woo Kim
Contact: [email protected]
"""
from CommonObject import *
import glob
import argparse
import os
import time
import datetime
import copy
import torch
# GPU selection in case of multi-gpu
os.environ['CUDA_VISIBLE_DEVICES'] = '0, 1'
facNum = 1
dev_num = (facNum-1) % 2
device = torch.device("cuda:" + str(dev_num)) # torch.cuda.set_device(facNum % 2 - 1)
print('device:: ', device, ' dev_num: ', dev_num)
torch.cuda.set_device(dev_num)
# imports for V-REP
from ExpCollector import ExpCollector
try:
import vrep
except:
print('--------------------------------------------------------------')
print('"vrep.py" could not be imported. This means very probably that')
print('either "vrep.py" or the remoteApi library could not be found.')
print('Make sure both are in the same folder as this file,')
print('or appropriately adjust the file "vrep.py"')
print('--------------------------------------------------------------')
print('')
from model import MLPPolicy
from storage import RolloutStorage
import algo
from robots.NAO_MIMIC import NAO_MIMIC
from robots.BAXTER_MIMIC import BAXTER_MIMIC
from robots.C3PO_MIMIC import C3PO_MIMIC
#### settings for RL initial values ####
assert args.algo in ['a2c', 'ppo', 'acktr']
if args.recurrent_policy: # False is default
assert args.algo in ['a2c', 'ppo'], \
'Recurrent policy is not implemented for ACKTR'
action_class_list = [22, 23, 31, 37, 38, 39] # 23: hand waving, 38: salute
str_action_class = 'A' + 'A'.join('{0:03}'.format(action_class_list[i]) for i in range(len(action_class_list)))
str_action_class = 'UNI' if len(action_class_list) >= 6 else str_action_class
print(str_action_class)
"""
=================================================================
User Parameters for robot selection.
Pick a robot name in the "robotDict" and set it to "robot_name"
=================================================================
"""
robotDict = {'NAO': NAO_MIMIC(), 'Baxter': BAXTER_MIMIC(), 'C3PO': C3PO_MIMIC()}
robot_name = 'NAO'
try:
myrobot = robotDict[robot_name]
except KeyError:
print('The robot is NOT supported..')
exit()
args.env_name = myrobot.robot_name + '_' + str_action_class
args.num_processes = 1
args.num_stack = 1
args.algo = 'ppo'
args.use_gae = True # Generalized Advantage Estimation
args.vis_interval = 100
args.log_interval = 10 # TODO
args.num_steps = 2048 # This should be greater than num_mini_batch
args.a_lr = 1e-4
args.c_lr = 2e-4
args.entropy_coef = 0.005
args.value_loss_coef = 1
args.cuda = True
args.ppo_epoch = 5
args.num_mini_batch = 32
args.gamma = 0.98
args.tau = .95
args.port = 6006 # tensorboard port [default: 6006]
args.vis = True
args.num_frames = 1000000
args.save_interval = 10
args.cont_learning = False
args.itr_num = 487
# Learning Mode variables
cyclic_policy = True
symm_policy = False
args.discount_mode = 'local' # [single, local, global]
args.wnd_size = 3 # [1, 3, 5]
print('args: ', args)
num_updates = int(args.num_frames) // args.num_steps // args.num_processes
torch.manual_seed(args.seed)
if args.cuda:
print('cuda is vailable..1')
torch.cuda.manual_seed(args.seed)
# print("#######")
# print("WARNING: We note that all rewards are clipped and normalized)
# print("#######")
# sym, asym, c3po
obs_shape = myrobot.observation_space.shape
obs_shape = (obs_shape[0] * args.num_stack, *obs_shape[1:])
state_shape = myrobot.state_space.shape
state_shape = (state_shape[0] * args.num_stack, *state_shape[1:])
action_shape = myrobot.action_space.shape
full_state_shape = (state_shape[0] * args.num_stack, obs_shape[1] + state_shape[1])
if symm_policy:
full_state_shape = obs_shape
state_shape = obs_shape
from gym.spaces import Box
myrobot.state_space = Box(low=float('-inf'), high=float('inf'), shape=state_shape, dtype=np.float32)
print('obs shape: ', obs_shape)
print('state shape: ', state_shape)
print('action shape: ', action_shape)
print('full state shape: ', full_state_shape)
portNum = 88880+facNum
env = ExpCollector(myrobot, portNum, 0, motion_sampling=False)
# load or new
if args.cont_learning == True:
print('#### continue learning... ####')
actor_critic = MLPPolicy(obs_shape[1], full_state_shape[1], env.robot.action_space, symm_policy=symm_policy, use_seq=False, cuda_use=args.cuda)
exist_mode_name = '(1M_3L_512_ReLU)'
exist_mode_name += '(cyclic)' if cyclic_policy else '(acyclic)'
exist_mode_name += '(sym)' if symm_policy else '(asym)'
print('Load Policy:::: ', os.path.join(args.load_dir + args.algo
+ '/', args.env_name + exist_mode_name
+ '(' + str(args.itr_num) + ')' + ".pt"))
state_dict, ob_rms, st_rms, ret_rms = \
torch.load(os.path.join(args.load_dir + args.algo
+ '/', args.env_name + exist_mode_name
+ '(' + str(args.itr_num) + ')' + ".pt"))
actor_critic.load_state_dict(state_dict)
env.robot.ob_rms = ob_rms
env.robot.st_rms = st_rms
env.robot.ret_rms = ret_rms
args.env_name = exist_mode_name
else:
print('$$$$ new learning... $$$$')
actor_critic = MLPPolicy(obs_shape[1], full_state_shape[1], env.robot.action_space, symm_policy=symm_policy, use_seq=False, cuda_use=args.cuda)
print('model desc: ', actor_critic)
args.env_name += '(2M_3L_512_ReLU_inf)'
args.env_name += '(cyclic)' if cyclic_policy else '(acyclic)'
args.env_name += '(sym)' if symm_policy else '(asym)'
if args.discount_mode == 'single':
args.env_name += '(single)'
args.gamma = 0.0
args.use_gae = False
elif args.discount_mode == 'local':
args.env_name += '(local_' + str(args.wnd_size) + ')'
args.use_gae = False
elif args.discount_mode == 'global':
args.env_name += '(global)'
args.use_gae = True
else:
raise ValueError
if hasattr(actor_critic, 'actor_lstm'):
args.env_name += '(LSTM)'
print(args.env_name)
####################################
# visualization
from torch.utils.tensorboard import SummaryWriter
from datetime import datetime
writer = SummaryWriter('runs/C-3PO_Task')
now = str(datetime.now().strftime('%Y-%m-%d@%H:%M:%S'))
if args.cuda:
actor_critic.cuda()
print('cuda is available..2')
agent = algo.PPO(actor_critic, args.clip_param, args.ppo_epoch, args.num_mini_batch,
args.value_loss_coef, args.entropy_coef, a_lr=args.a_lr, c_lr=args.a_lr,
eps=args.eps, max_grad_norm=args.max_grad_norm)
rollouts = RolloutStorage(args.num_steps, args.num_processes, obs_shape, full_state_shape,
env.robot.action_space)
current_obs = torch.zeros(args.num_processes, *obs_shape)
current_state = torch.zeros(args.num_processes, *full_state_shape)
######################
# Load skeleton data
from SkeletonDataManager import SkeletonDataManager
all_train_skeleton_data = np.empty((0, 76))
all_train_skeleton_fc = []
for i in action_class_list:
np_train_skel_data, shape, fc_list = load_skeleton_data(skeleton_train_dict[i])
all_train_skeleton_data = np.vstack((all_train_skeleton_data, np_train_skel_data))
all_train_skeleton_fc += fc_list
print('Finished the frame data Load of: ', i)
sdm = SkeletonDataManager(all_train_skeleton_data, all_train_skeleton_fc)
print('shape: ', all_train_skeleton_data.shape, 'skel: ', all_train_skeleton_data)
print('len: ', len(all_train_skeleton_fc), 'frame count: ', all_train_skeleton_fc)
# Load skeleton vae model
from vae_model import VAE, VAE_SK
load_path = os.path.join('./trained_models', 'vae')
load_name = 'Skel_vae_Superset'
state_dict = torch.load(os.path.join(load_path, load_name + '.pt'))
# Skeleton Model(input dim, output dim)
sModel = VAE_SK(75, 75, use_batch_norm=False, activation='ReLU')
sModel.load_state_dict(state_dict)
sModel.eval()
# Robot Model
# load_name = 'Motion_vae_' + str_action_class # make_class_prefix([22, 23, 31, 37, 38, 39])
if myrobot.robot_name == 'BAXTER_MIMIC':
load_name = '[Baxter]Motion_vae_Superset'
elif myrobot.robot_name == 'NAO_MIMIC':
load_name = '[NAO]Motion_vae_Superset'
elif myrobot.robot_name == 'C3PO_MIMIC':
load_name = '[C3PO]Motion_vae_Superset'
state_dict = torch.load(os.path.join(load_path, load_name + '.pt'))
rModel = VAE(14, latent_dim=np.prod(action_shape), use_batch_norm=False, activation='Tanh') # action: 10*20
rModel.load_state_dict(state_dict)
rModel.eval()
# Set rModel to myrobot
if not symm_policy:
myrobot.rModel = rModel
reward_list = []
if env.simStart(gui_on=True, autoStart=False, autoQuit=True, epiNum=args.num_steps) != -1:
print('sim start!!----------')
time.sleep(1)
data, last_scene, first_scene = sdm.get_random_scene_frame()
with torch.no_grad():
mu, logvar = sModel.encode(torch.from_numpy(data[:-1]).unsqueeze(0).float())
zs = sModel.reparameterize(mu, logvar)
if cyclic_policy:
with torch.no_grad():
rSkel = sModel.decode(zs).reshape(75, -1).squeeze()
rSkel = rSkel.numpy()
else:
rSkel = data
# zs = zs.cpu().numpy()
obs, st = env.reset(zs.cpu().numpy(), rSkel)
update_current_obs(obs, current_obs)
full_state = np.concatenate((current_obs.numpy().flatten(), # initial obs
st[0])) # initial state
if symm_policy:
full_state = st
update_current_state(full_state, current_state)
rollouts.observations[0].copy_(current_obs)
rollouts.states[0].copy_(current_state)
if args.cuda:
print('cuda is available..3')
current_obs = current_obs.cuda()
current_state = current_state.cuda()
rollouts.cuda()
start = time.time()
for j in range(num_updates):
print('num_updates: ', j)
episode_rewards = torch.zeros([args.num_processes, 1])
final_rewards = torch.zeros([args.num_processes, 1])
itr_time = time.time()
for step in range(args.num_steps): # PPO steps
with torch.no_grad():
value, action, action_log_prob, states, _, _ = actor_critic.act(
rollouts.observations[step],
rollouts.states[step],
rollouts.masks[step])
# get next observation
data, last_scene, first_scene = sdm.get_random_scene_frame()
# get latent vector
with torch.no_grad():
mu, logvar = sModel.encode(torch.from_numpy(data[:-1]).unsqueeze(0).float())
zs = sModel.reparameterize(mu, logvar)
# cpu_actions = action.data.squeeze(1).cpu().numpy()
# get robot motor value by decoding
with torch.no_grad():
x_hat = rModel.decode(action.cpu()).reshape(-1, 14)
if cyclic_policy:
with torch.no_grad():
rSkel = sModel.decode(zs).reshape(75, -1).squeeze()
rSkel = rSkel.numpy()
else:
rSkel = data
next_obs, next_state, reward, done, info, true_rew = env.step(x_hat, zs.cpu().numpy(), rSkel)
# To separate the Q-value of the current scene from others
# if last_scene:
# done[0] = 1 # done is np array, [0]
# obs, next_state = env.reset(zs, rSkel)
# local q-function
if args.discount_mode == 'local' and j % args.wnd_size == 0:
done[0] = 1 # done is np array, [0]
reward = torch.from_numpy(np.expand_dims(np.stack(reward), 1)).float()
true_rew = torch.from_numpy(np.expand_dims(np.stack(true_rew), 1)).float()
final_rewards += true_rew
episode_rewards += reward # scalar value
masks = torch.FloatTensor([[0.0] if done_ else [1.0] for done_ in done])
# final_rewards *= masks
# final_rewards += (1 - masks) * episode_rewards
episode_rewards *= masks
if args.cuda:
masks = masks.cuda()
if current_obs.dim == 4: # This means image
current_obs *= masks.unsqueeze(2).unsqueeze(2)
else:
current_obs *= masks
current_state *= masks
update_current_obs(next_obs, current_obs)
# make full state
full_state = np.concatenate((current_obs.cpu().numpy().flatten(),
next_state.flatten())) # current state!!
if symm_policy:
full_state = states.cpu().numpy()
update_current_state(full_state, current_state)
rollouts.insert(current_obs, current_state, action.data, action_log_prob.data, value.data, reward, masks)
rollout_time = time.time() - itr_time
itr_time = time.time()
with torch.no_grad():
next_value = actor_critic.get_value(rollouts.observations[-1],
rollouts.states[-1],
rollouts.masks[-1]).detach()
rollouts.compute_returns(next_value, args.use_gae, args.gamma, args.tau, args.discount_mode)
value_loss, action_loss, dist_entropy = agent.update(rollouts)
rollouts.after_update() # observation, state, mask의 0번째 값을 마지막에 얻은 값으로 대체
print('itr rewards: ', final_rewards, ' rollout_time: ', rollout_time, ' update_time: ', time.time() - itr_time)
reward_list.append(final_rewards.item())
try:
writer.add_scalar(args.env_name + '/' + now, reward_list[-1], len(reward_list) + 1)
# reconst = reconst.squeeze(0)
# b, g, r = torch.split(reconst, 1, dim=0)
# reconst = torch.cat((r, g, b))
# writer.add_image(args.env_name + '_IMG' + '/' + now, reconst, global_step=len(reward_list)+1)
pass
except NameError:
pass
# save learned network only for root node
if (j % args.save_interval == 0 and args.save_dir != "") or j >= num_updates-1:
print('save model!!')
save_path = os.path.join(args.save_dir, args.algo, 'phase2', args.env_name)
try:
os.makedirs(save_path)
except OSError:
pass
save_state_dict = actor_critic.state_dict()
# if args.cuda:
# save_model = copy.deepcopy(actor_critic).cpu()
save_model = [save_state_dict,
hasattr(env.robot, 'ob_rms') and env.robot.ob_rms or None,
hasattr(env.robot, 'st_rms') and env.robot.st_rms or None,
hasattr(env.robot, 'ret_rms') and env.robot.ret_rms or None]
# save_model = [save_model, None]
str_itr_num = '(' + str(args.itr_num + j) + ')' if args.cont_learning else '(' + str(j) + ')'
torch.save(save_model, os.path.join(save_path, args.env_name + str_itr_num + ".pt"))
# save list file
with open(os.path.join(save_path, 'Rewards_Train_' + args.env_name + '.txt'), 'w') as f:
for item in reward_list:
f.write("%s\n" % item)
print('total elapsed time {:.2f}'.format(time.time()-start))