-
Notifications
You must be signed in to change notification settings - Fork 0
/
model_base.py
301 lines (254 loc) · 12.8 KB
/
model_base.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
import time
import numpy as np
import hyperparams as hyp
import torch
from tensorboardX import SummaryWriter
from backend import saverloader, inputs
from backend import inputs as load_inputs
# from backend.double_pool import DoublePool
from torchvision import datasets, transforms
from DoublePool import DoublePool_O
from DoublePool import MOC_DICT,MOC_QUEUE_NORMAL
from DoublePool import ClusterPool
from DoublePool import DetPool
import utils_basic
import socket
import time
import torch.nn.functional as F
import pickle
import utils_eval
import utils_improc
import utils_basic
import ipdb
st = ipdb.set_trace
from collections import defaultdict
import cross_corr
np.set_printoptions(precision=2)
EPS = 1e-6
np.random.seed(0)
MAX_QUEUE = 10 # how many items before the summaryWriter flushes
class Model(object):
def __init__(self, checkpoint_dir, log_dir):
print('------ CREATING NEW MODEL ------')
print(hyp.name)
self.checkpoint_dir = checkpoint_dir
self.log_dir = log_dir
self.all_inputs = inputs.get_inputs()
print("------ Done getting inputs ------")
if hyp.moc:
self.poolvox_moc = MOC_QUEUE_NORMAL(hyp.moc_qsize)
if hyp.moc_2d:
self.poolvox_moc_2d = MOC_QUEUE_NORMAL(hyp.moc_qsize)
if hyp.offline_cluster:
self.cluster_pool = ClusterPool(hyp.offline_cluster_pool_size)
if hyp.offline_cluster_eval:
self.info_dict = defaultdict(lambda:[])
if hyp.max.hardmining or hyp.hard_eval or hyp.hard_vis:
self.mbr = cross_corr.meshgrid_based_rotation(hyp.BOX_SIZE-2*hyp.max.margin, hyp.BOX_SIZE-2*hyp.max.margin, hyp.BOX_SIZE-2*hyp.max.margin)
self.mbr16 = cross_corr.meshgrid_based_rotation(hyp.BOX_SIZE, hyp.BOX_SIZE, hyp.BOX_SIZE)
self.mbr_unpr = cross_corr.meshgrid_based_rotation(32,32,32)
self.hpm = hardPositiveMiner.HardPositiveMiner(self.mbr,self.mbr16,self.mbr_unpr)
elif hyp.do_orientation:
self.mbr16 = cross_corr.meshgrid_based_rotation(hyp.BOX_SIZE, hyp.BOX_SIZE, hyp.BOX_SIZE)
self.mbr_unpr = cross_corr.meshgrid_based_rotation(32,32,32)
self.hpm = None
else:
self.mbr16 = None
self.hpm = None
if hyp.self_improve_iterate:
self.pool_det = DetPool(hyp.det_pool_size)
if hyp.do_eval_recall:
self.eval_dicts = {}
if hyp.dataset_name=="bigbird":
self.recalls = [3, 5, 10]
else:
self.recalls = [10, 20, 30]
if hyp.do_debug or hyp.low_dict_size:
self.pool_size = 11
else:
self.pool_size = hyp.pool_size
self.eval_dicts["eval"] = {}
F = 3
if hyp.eval_recall_o:
self.eval_dicts["eval"]['pool3D_e'] = DoublePool_O(self.pool_size)
self.eval_dicts["eval"]['pool3D_g'] = DoublePool_O(self.pool_size)
else:
self.eval_dicts["eval"]['pool3D_e'] = DoublePool(self.pool_size)
self.eval_dicts["eval"]['pool3D_g'] = DoublePool(self.pool_size)
self.eval_dicts["eval"]['precision3D'] = np.nan*np.array([0.0, 0.0, 0.0], np.float32)
self.eval_dicts["eval"]['neighbors3D'] = np.zeros((F*10, F*11, 3), np.float32)
self.device = torch.device("cuda")
def infer(self):
pass
def momentum_update(self,model_q, model_k, beta = 0.999):
param_k = model_k.state_dict()
param_q = model_q.named_parameters()
for n, q in param_q:
if n in param_k:
param_k[n].data.copy_(beta*param_k[n].data + (1-beta)*q.data)
model_k.load_state_dict(param_k)
def go(self):
self.start_time = time.time()
self.infer()
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=hyp.lr, weight_decay=hyp.weight_decay)
print("------ Done creating models ------")
# st()
self.start_iter = saverloader.load_weights(self.model, self.optimizer)
if hyp.self_improve_once or hyp.filter_boxes:
self.model.detnet_target.load_state_dict(self.model.detnet.state_dict())
if hyp.moc:
self.init_model_k(self.model,self.model_key)
if hyp.sets_to_run["test"]:
self.start_iter = 0
print("------ Done loading weights ------")
set_nums = []
set_names = []
set_inputs = []
set_writers = []
set_log_freqs = []
set_do_backprops = []
set_dicts = []
set_loaders = []
# st()
for set_name in hyp.set_names:
if hyp.sets_to_run[set_name]:
set_nums.append(hyp.set_nums[set_name])
set_names.append(set_name)
set_inputs.append(self.all_inputs[set_name])
set_writers.append(SummaryWriter(self.log_dir + '/' + set_name, max_queue=MAX_QUEUE, flush_secs=60))
set_log_freqs.append(hyp.log_freqs[set_name])
set_do_backprops.append(hyp.sets_to_backprop[set_name])
set_dicts.append({})
set_loaders.append(iter(set_inputs[-1]))
# st()
for step in range(self.start_iter+1, hyp.max_iters+1):
for i, (set_input) in enumerate(set_inputs):
if step % len(set_input) == 0: #restart after one epoch. Note this does nothing for the tfrecord loader
set_loaders[i] = iter(set_input)
for (set_num,
set_name,
set_input,
set_writer,
set_log_freq,
set_do_backprop,
set_dict,
set_loader,
) in zip(
set_nums,
set_names,
set_inputs,
set_writers,
set_log_freqs,
set_do_backprops,
set_dicts,
set_loaders,
):
log_this = np.mod(step, set_log_freq)==0
total_time, read_time, iter_time = 0.0, 0.0, 0.0
if log_this or set_do_backprop or hyp.break_constraint:
read_start_time = time.time()
feed = next(set_loader)
# st()
feed_cuda = {}
# st()
if hyp.do_clevr_sta:
tree_seq_filename = feed.pop('tree_seq_filename')
filename_e = feed.pop('filename_e')
filename_g = feed.pop('filename_g')
# st()
# st()
if hyp.dataset_name == "replica" or hyp.dataset_name == "carla_mix" or hyp.dataset_name == "clevr_vqa" or hyp.dataset_name == "carla_det":
if hyp.debug_match or hyp.do_match_det:
feed['classes'] = np.array(feed['classes']).reshape([hyp.B*2,hyp.N])
classes = feed.pop('classes')
for k in feed:
#feed_cuda[k] = feed[k].to(self.device)
if hyp.typeVal == "content" or hyp.debug_match or hyp.do_match_det:
feed_cuda[k] = feed[k].cuda(non_blocking=True).float().squeeze(0)
else:
feed_cuda[k] = feed[k].cuda(non_blocking=True).float()
# feed_cuda = next(iter(set_input))
read_time = time.time() - read_start_time
# st()
if hyp.do_clevr_sta:
if hyp.typeVal == "content" or hyp.debug_match or hyp.do_match_det:
feed_cuda['tree_seq_filename'] = np.array(tree_seq_filename).squeeze(1)
else:
feed_cuda['tree_seq_filename'] = tree_seq_filename
feed_cuda['filename_e'] = filename_e
feed_cuda['filename_g'] = filename_g
if hyp.dataset_name == "replica" or hyp.dataset_name == "carla_mix" or hyp.dataset_name == "clevr_vqa" or hyp.dataset_name == "carla_det":
if not hyp.debug_match and not hyp.do_match_det:
classes = np.transpose(np.array(classes))
feed_cuda['classes'] = classes
feed_cuda['filename_e'] = filename_e
feed_cuda['filename_g'] = filename_g
feed_cuda['writer'] = set_writer
feed_cuda['global_step'] = step
feed_cuda['set_num'] = set_num
feed_cuda['set_name'] = set_name
iter_start_time = time.time()
if hyp.typeVal == "content" or hyp.debug_match or hyp.do_match_det:
hyp.B = hyp.B*2
if set_do_backprop:
start_time = time.time()
self.model.train()
# st()
loss, results = self.model(feed_cuda)
if hyp.profile_time:
print("forwardpass time",time.time()-start_time)
else:
self.model.eval()
with torch.no_grad():
loss, results = self.model(feed_cuda)
if hyp.halucinate_vals !=1:
if hal_num == 0:
maps = []
filenames = []
maps.append(results['maps'])
filenames.append(results['filenames'])
if (hal_num+1) == (hyp.halucinate_vals):
maps_avg = np.mean(np.stack(maps),axis=0)
for ind, overlap in enumerate(results['ious']):
results['summ'].summ_scalar('ap_avg/%.2f_iou' % overlap, maps_avg[ind])
if hyp.typeVal == "content":
hyp.B = hyp.B//2
loss_vis = loss.cpu().item()
summ_writer = utils_improc.Summ_writer(writer=feed_cuda['writer'],
global_step=feed_cuda['global_step'],
set_name=feed_cuda['set_name'],
fps=8)
summ_writer.summ_scalar('loss',loss_vis)
if set_do_backprop:
if hyp.accumulate_grad:
loss = loss / hyp.accumulation_steps # Normalize our loss (if averaged)
loss.backward() # Backward pass
if (step) % hyp.accumulation_steps == 0: # Wait for several backward steps
self.optimizer.step() # Now we can do an optimizer step
self.optimizer.zero_grad()
else:
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
iter_time = time.time()-iter_start_time
total_time = time.time()-self.start_time
print("%s; [%4d/%4d]; ttime: %.0f (%.2f, %.2f); loss: %.3f (%s)" % (hyp.name,
step,
hyp.max_iters,
total_time,
read_time,
iter_time,
loss_vis,
set_name))
# st()
if np.mod(step, hyp.snap_freq) == 0:
saverloader.save(self.model, self.checkpoint_dir, step, self.optimizer)
for writer in set_writers: #close writers to flush cache into file
writer.close()
def set_requires_grad(self, nets, requires_grad=False):
if not isinstance(nets, list):
nets = [nets]
for net in nets:
if net is not None:
for param in net.parameters():
param.requires_grad = requires_grad