-
Notifications
You must be signed in to change notification settings - Fork 54
/
train_MPER.py
288 lines (249 loc) · 11.5 KB
/
train_MPER.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
import torch, numpy, argparse, pdb, os, time, math, random, re
import utils
from dataloader import DataLoader
import torch.nn.functional as F
import torch.optim as optim
import models, planning
import importlib
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
#################################################
# Train an action-conditional forward model
#################################################
parser = argparse.ArgumentParser()
# data params
parser.add_argument('-seed', type=int, default=1)
parser.add_argument('-dataset', type=str, default='i80')
parser.add_argument('-v', type=int, default=4)
parser.add_argument('-model', type=str, default='fwd-cnn')
parser.add_argument('-policy', type=str, default='policy-deterministic')
parser.add_argument('-model_dir', type=str, default='models/')
parser.add_argument('-ncond', type=int, default=20)
parser.add_argument('-npred', type=int, default=16)
parser.add_argument('-batch_size', type=int, default=8)
parser.add_argument('-layers', type=int, default=3)
parser.add_argument('-nfeature', type=int, default=256)
parser.add_argument('-n_hidden', type=int, default=256)
parser.add_argument('-beta', type=float, default=0.0, help='weight coefficient of prior loss')
parser.add_argument('-p_dropout', type=float, default=0.0, help='set z=0 with this probability')
parser.add_argument('-dropout', type=float, default=0.0, help='regular dropout')
parser.add_argument('-nz', type=int, default=2)
parser.add_argument('-n_mixture', type=int, default=10)
parser.add_argument('-context_dim', type=int, default=2)
parser.add_argument('-actions_subsample', type=int, default=4)
parser.add_argument('-lrt', type=float, default=0.0001)
parser.add_argument('-grad_clip', type=float, default=1.0)
parser.add_argument('-epoch_size', type=int, default=500)
parser.add_argument('-curriculum_length', type=int, default=16)
parser.add_argument('-zeroact', type=int, default=0)
parser.add_argument('-warmstart', type=int, default=0)
parser.add_argument('-targetprop', type=int, default=0)
parser.add_argument('-loss_c', type=int, default=0)
parser.add_argument('-lambda_c', type=float, default=0.0)
parser.add_argument('-lambda_h', type=float, default=0.0)
parser.add_argument('-lambda_lane', type=float, default=0.1)
parser.add_argument('-lrt_traj', type=float, default=0.5)
parser.add_argument('-niter_traj', type=int, default=20)
parser.add_argument('-gamma', type=float, default=1.0)
#parser.add_argument('-mfile', type=str, default='model=fwd-cnn-vae-fp-layers=3-bsize=64-ncond=20-npred=20-lrt=0.0001-nfeature=256-dropout=0.1-nz=32-beta=1e-06-zdropout=0.5-gclip=5.0-warmstart=1-seed=1.step200000.model')
parser.add_argument('-mfile', type=str, default='model=fwd-cnn-layers=3-bsize=64-ncond=20-npred=20-lrt=0.0001-nfeature=256-dropout=0.1-gclip=5.0-warmstart=0-seed=1.step200000.model')
parser.add_argument('-load_model_file', type=str, default='')
parser.add_argument('-combine', type=str, default='add')
parser.add_argument('-debug', action='store_true')
parser.add_argument('-test_only', type=int, default=0)
parser.add_argument('-enable_tensorboard', action='store_true',
help='Enables tensorboard logging.')
parser.add_argument('-tensorboard_dir', type=str, default='models/policy_networks',
help='path to the directory where to save tensorboard log. If passed empty path' \
' no logs are saved.')
opt = parser.parse_args()
opt.n_inputs = 4
opt.n_actions = 2
opt.height = 117
opt.width = 24
opt.h_height = 14
opt.h_width = 3
opt.hidden_size = opt.nfeature*opt.h_height*opt.h_width
os.system('mkdir -p ' + opt.model_dir + '/policy_networks/')
random.seed(opt.seed)
numpy.random.seed(opt.seed)
torch.manual_seed(opt.seed)
torch.cuda.manual_seed(opt.seed)
opt.model_file = f'{opt.model_dir}/policy_networks/'
opt.model_file += f'mbil-{opt.policy}-nfeature={opt.nfeature}-npred={opt.npred}-lambdac={opt.lambda_c}-gamma={opt.gamma}-seed={opt.seed}'
if 'vae' in opt.mfile:
opt.model_file += f'-model=vae'
model_type = 'vae'
elif 'ten' in opt.mfile:
opt.model_file += f'-model=ten'
model_type = 'ten'
elif 'model=fwd-cnn-layers' in opt.mfile:
model_type = 'det'
opt.model_file += '-deterministic'
if 'zdropout=0.5' in opt.mfile:
opt.model_file += '-zdropout=0.5'
elif 'zdropout=0.0' in opt.mfile:
opt.model_file += '-zdropout=0.0'
print(f'[will save as: {opt.model_file}]')
if os.path.isfile(opt.model_file + '.model') and False:
print('[found previous checkpoint, loading]')
checkpoint = torch.load(opt.model_file + '.model')
model = checkpoint['model']
optimizer = optim.Adam(model.policy_net.parameters(), opt.lrt)
optimizer.load_state_dict(checkpoint['optimizer'])
n_iter = checkpoint['n_iter']
if opt.test_only == 0:
utils.log(opt.model_file + '.log', '[resuming from checkpoint]')
else:
# load the model
model = torch.load(opt.model_dir + opt.mfile)
if type(model) is dict: model = model['model']
model.create_policy_net(opt)
model.opt.actions_subsample = opt.actions_subsample
optimizer = optim.Adam(model.policy_net.parameters(), opt.lrt)
n_iter = 0
# stats = torch.load('/misc/vlgscratch4/LecunGroup/nvidia-collab/traffic-data-atcold/data_i80_v0/data_stats.pth')
# model.stats=stats
if 'ten' in opt.mfile:
pzfile = opt.model_dir + opt.mfile + '.pz'
p_z = torch.load(pzfile)
model.p_z = p_z
if opt.actions_subsample == -1:
opt.context_dim = 0
model.intype('gpu')
model.cuda()
print('[loading data]')
dataloader = DataLoader(None, opt, opt.dataset)
# training and testing functions. We will compute several losses:
# loss_i: images
# loss_s: states
# loss_c: costs
# loss_p: prior (optional)
def compute_loss(targets, predictions, gamma=1.0, r=True):
target_images, target_states, target_costs = targets
pred_images, pred_states, pred_costs, loss_p = predictions
loss_i = F.mse_loss(pred_images, target_images, reduce=False).mean(4).mean(3).mean(2)
loss_s = F.mse_loss(pred_states, target_states, reduce=False).mean(2)
# loss_c = F.mse_loss(pred_costs, target_costs, reduce=False).mean(2)
if gamma < 1.0:
loss_i *= gamma_mask
loss_s *= gamma_mask
loss_c *= gamma_mask
return loss_i.mean(), loss_s.mean(), torch.zeros(1), loss_p.mean()
def train(nbatches, npred):
gamma_mask = torch.Tensor([opt.gamma**t for t in range(npred)]).view(1, -1).cuda()
model.eval()
model.policy_net.train()
total_loss_i, total_loss_s, total_loss_c, total_loss_policy, total_loss_p, n_updates = 0, 0, 0, 0, 0, 0
for i in range(nbatches):
optimizer.zero_grad()
inputs, actions, targets, _, _ = dataloader.get_batch_fm('train', npred)
pred, _ = planning.train_policy_net_mper(model, inputs, targets, dropout=opt.p_dropout, model_type=model_type)
loss_i, loss_s, loss_c_, loss_p = compute_loss(targets, pred)
# proximity_cost, lane_cost = pred[2][:, :, 0], pred[2][:, :, 1]
# proximity_cost = proximity_cost * gamma_mask
# lane_cost = lane_cost * gamma_mask
# loss_c = proximity_cost.mean() + opt.lambda_lane * lane_cost.mean()
loss_policy = loss_i + loss_s + opt.lambda_h*loss_p
if opt.loss_c == 1:
loss_policy += loss_c_
if not math.isnan(loss_policy.item()):
loss_policy.backward()
torch.nn.utils.clip_grad_norm(model.policy_net.parameters(), opt.grad_clip)
optimizer.step()
total_loss_i += loss_i.item()
total_loss_s += loss_s.item()
total_loss_p += loss_p.item()
total_loss_policy += loss_policy.item()
n_updates += 1
else:
print('warning, NaN')
del inputs, actions, targets, pred
total_loss_i /= n_updates
total_loss_s /= n_updates
total_loss_c /= n_updates
total_loss_policy /= n_updates
total_loss_p /= n_updates
return total_loss_i, total_loss_s, total_loss_c, total_loss_policy, total_loss_p
def test(nbatches, npred):
gamma_mask = torch.Tensor([opt.gamma**t for t in range(npred)]).view(1, -1).cuda()
model.eval()
total_loss_i, total_loss_s, total_loss_c, total_loss_policy, total_loss_p, n_updates = 0, 0, 0, 0, 0, 0
for i in range(nbatches):
inputs, actions, targets, _, _ = dataloader.get_batch_fm('test', npred)
pred, pred_actions = planning.train_policy_net_mper(model, inputs, targets, targetprop = opt.targetprop, dropout=0.0, model_type = model_type)
loss_i, loss_s, loss_c_, loss_p = compute_loss(targets, pred)
loss_policy = loss_i + loss_s
if opt.loss_c == 1:
loss_policy += loss_c_
if not math.isnan(loss_policy.item()):
total_loss_i += loss_i.item()
total_loss_s += loss_s.item()
total_loss_p += loss_p.item()
total_loss_policy += loss_policy.item()
n_updates += 1
del inputs, actions, targets, pred
total_loss_i /= n_updates
total_loss_s /= n_updates
total_loss_c /= n_updates
total_loss_policy /= n_updates
total_loss_p /= n_updates
return total_loss_i, total_loss_s, total_loss_c, total_loss_policy, total_loss_p
# set by hand to fit on 12gb GPU
def get_batch_size(npred):
if npred <= 15:
return 64
elif npred <= 50:
return 32
elif npred <= 100:
return 16
elif npred <= 200:
return 8
elif npred <= 400:
return 4
elif npred <= 800:
return 2
else:
return 1
if opt.test_only == 1:
print('[testing]')
valid_losses = test(10, 200)
else:
writer = utils.create_tensorboard_writer(opt)
print('[training]')
utils.log(opt.model_file + '.log', f'[job name: {opt.model_file}]')
npred = opt.npred if opt.npred != -1 else 16
for i in range(500):
bsize = get_batch_size(npred)
dataloader.opt.batch_size = bsize
train_losses = train(opt.epoch_size, npred)
valid_losses = test(int(opt.epoch_size / 2), npred)
n_iter += opt.epoch_size
model.intype('cpu')
torch.save({'model': model,
'optimizer': optimizer.state_dict(),
'opt': opt,
'npred': npred,
'n_iter': n_iter},
opt.model_file + '.model')
model.intype('gpu')
if writer is not None:
writer.add_scalar('Loss/train_state_img', train_losses[0], i)
writer.add_scalar('Loss/train_state_vct', train_losses[1], i)
writer.add_scalar('Loss/train_costs', train_losses[2], i)
writer.add_scalar('Loss/train_policy', train_losses[3], i)
writer.add_scalar('Loss/train_relative_entropy', train_losses[4], i)
writer.add_scalar('Loss/validation_state_img', valid_losses[0], i)
writer.add_scalar('Loss/validation_state_vct', valid_losses[1], i)
writer.add_scalar('Loss/validation_costs', valid_losses[2], i)
writer.add_scalar('Loss/validation_policy', valid_losses[3], i)
writer.add_scalar('Loss/validation_relative_entropy', valid_losses[4], i)
log_string = f'step {n_iter} | npred {npred} | bsize {bsize} | esize {opt.epoch_size} | '
log_string += utils.format_losses(train_losses[0], train_losses[1], split='train')
log_string += utils.format_losses(valid_losses[0], valid_losses[1], split='valid')
print(log_string)
utils.log(opt.model_file + '.log', log_string)
if i > 0 and(i % opt.curriculum_length == 0) and (opt.npred == -1) and npred < 400:
npred += 8
if writer is not None:
writer.close()