-
Notifications
You must be signed in to change notification settings - Fork 0
/
clients.py
191 lines (143 loc) · 8.52 KB
/
clients.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
import threading
import configparser
from contextlib import nullcontext
import random
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import numpy as np
import time
import copy
from models import *
import sys
sys.path.append('dataloaders')
import tar_datasets as datasets
import tqdm
class StereoClient(threading.Thread):
def __init__(self, cfg, args, idx, server=None):
threading.Thread.__init__(self)
config = configparser.ConfigParser(converters={'list': lambda x: [i.strip() for i in x.split(',')]})
config.read(cfg)
self.idx = idx
self.gpu = int(config['adaptation']['gpu'])
self.server = server
self.args=args
self.adapt_mode = config['adaptation']['adapt_mode']
self.sample_mode = config['adaptation']['sample_mode']
self.model = config['network']['model']
self.sender = config['federated'].getboolean('sender')
self.listener = config['federated'].getboolean('listener')
self.bootstrap_countdown = 0 if self.server is None else self.server.bootstrap_countdown
self.bootstrapped = False if (self.listener and self.server is not None and self.bootstrap_countdown != 0) else True
self.runs = []
for i in range(len(config['environment'].getlist('dataset'))):
dataset = config['environment'].getlist('dataset')[i]
datapath = config['environment'].getlist('datapath')[i]
domain = config['environment'].getlist('domain')[i]
subs = int(config['environment'].getlist('subs')[i])
proxy16 = config['environment'].getboolean('proxy16')
self.runs.append( {'loader': datasets.fetch_single_dataloader(dataset,datapath,domain,subs,proxy16),
'dataset':dataset,
'domain':domain,
})
self.net = models_lut[self.model](args)
self.net = nn.DataParallel(self.net)
self.net.to('cuda:%d'%self.gpu)
self.net.load_state_dict(torch.load(config['network']['checkpoint'], torch.device('cuda:%d'%self.gpu))['state_dict'])
self.net = self.net.module
self.optimizer = optim.Adam(self.net.parameters(), lr=float(config['adaptation']['lr']), betas=(0.9, 0.999))
self.current_run = self.runs[0]
self.loader = self.current_run['loader']
self.accumulator = {}
def run(self):
args=self.args
if self.listener or self.args.verbose:
self.pbar = tqdm.tqdm(total=self.current_run['loader'].__len__, file=sys.stdout)
while not self.bootstrapped:
time.sleep(0.01)
self.net.eval()
with torch.no_grad() if (self.adapt_mode == 'none') else nullcontext():
for batch_idx, data in enumerate(self.loader):
if self.server is not None:
ret = self.server.gpu_locks[self.gpu].acquire()
while not ret:
time.sleep(0.01)
ret = self.server.gpu_locks[self.gpu].acquire()
if self.adapt_mode != 'none':
self.optimizer.zero_grad()
data['image_02.jpg'], data['image_03.jpg'] = data['image_02.jpg'].to('cuda:%d'%self.gpu), data['image_03.jpg'].to('cuda:%d'%self.gpu)
if 'proxy.png' in data:
data['validpr'] = (data['proxy.png']>0).float()
data['proxy.png'], data['validpr'] = data['proxy.png'].to('cuda:%d'%self.gpu), data['validpr'].to('cuda:%d'%self.gpu)
if data['image_02.jpg'].shape[-1] != data['proxy.png'].shape[-1]:
data['proxy.png'] = data['proxy.png'][...,:data['image_02.jpg'].shape[-1]]
data['validpr'] = data['validpr'][...,:data['image_02.jpg'].shape[-1]]
# pad images
ht, wt = data['image_02.jpg'].shape[-2], data['image_02.jpg'].shape[-1]
pad_ht = (((ht // 128) + 1) * 128 - ht) % 128
pad_wd = (((wt // 128) + 1) * 128 - wt) % 128
_pad = [pad_wd//2, pad_wd - pad_wd//2, pad_ht//2, pad_ht - pad_ht//2]
data['image_02.jpg'] = F.pad(data['image_02.jpg'], _pad, mode='replicate')
data['image_03.jpg'] = F.pad(data['image_03.jpg'], _pad, mode='replicate')
pred_disps = self.net(data['image_02.jpg'], data['image_03.jpg'], mad = 'mad' in self.adapt_mode)
# upsample and remove padding for final prediction
pred_disp = F.interpolate( pred_disps[0], scale_factor=4., mode='bilinear')[0]*-20.
ht, wd = pred_disp.shape[-2:]
c = [_pad[2], ht-_pad[3], _pad[0], wd-_pad[1]]
pred_disp = pred_disp[..., c[0]:c[1], c[2]:c[3]]
# upsample and remove padding from all predictions (if needed for adaptation)
if self.adapt_mode != 'none':
pred_disps = [F.interpolate( pred_disps[i], scale_factor=2**(i+2))*-20. for i in range(len(pred_disps))]
pred_disps = [pred_disps[i][..., c[0]:c[1], c[2]:c[3]] for i in range(len(pred_disps))]
data['image_02.jpg'] = data['image_02.jpg'][..., c[0]:c[1], c[2]:c[3]]
data['image_03.jpg'] = data['image_03.jpg'][..., c[0]:c[1], c[2]:c[3]]
if self.adapt_mode != 'none':
block = self.net.sample_block(self.sample_mode, seed=batch_idx) if ('mad' in self.adapt_mode) else self.net.sample_all()
loss = self.net.compute_loss(data['image_02.jpg'], data['image_03.jpg'], pred_disps, data['proxy.png'], data['validpr'], adapt_mode=self.adapt_mode, idx=block)
loss.backward()
self.optimizer.step()
pred_disp = pred_disp.detach()
result = {}
if 'groundtruth.png' in data:
data['validgt'] = (data['groundtruth.png'] > 0).float()
result = kitti_metrics(pred_disp.cpu().numpy(), data['groundtruth.png'].numpy(), data['validgt'].numpy())
result['disp'] = pred_disp
for k in result:
if k != 'disp' and k!= 'errormap':
if k not in self.accumulator:
self.accumulator[k] = []
self.accumulator[k].append(result[k])
if self.listener or self.args.verbose:
self.pbar.set_description("Thread %d, Seq: %s/%s, Frame %s, bad3: %2.2f"%(self.idx, self.current_run['dataset'], self.current_run['domain'], data['__key__'][0], result['bad 3'] if 'bad 3' in result else np.nan))
self.pbar.update(1)
if self.server is not None and len(self.server._listening_clients) == 0:
self.server.gpu_locks[self.gpu].release()
break
if self.sender and self.server is not None:
if (batch_idx > 0 and batch_idx % self.server.interval == 0):
ret = self.server.pushing_model.acquire(blocking=False)
while not ret:
time.sleep(0.01)
ret = self.server.pushing_model.acquire(blocking=False)
block = -1
if self.server.fed_mode == 'fedmad':
block = self.net.get_block_to_send(seed=self.idx+(batch_idx//self.server.interval))
self.server.push_model(self.idx, self.net, block)
self.server.pushing_model.release()
if self.server is not None:
self.server.gpu_locks[self.gpu].release()
if self.listener or self.args.verbose:
self.pbar.close()
if not self.listener and len(self.server._listening_clients) != 0:
self.run()
if self.server is None:
self.print_stats()
elif self.server is not None and self.listener:
self.print_stats()
self.server.remove_listening_client(self)
def print_stats(self):
metrs = ''
for k in self.accumulator:
metrs += '& %.2f '%np.array(self.accumulator[k]).mean()
print("\nThread %d results on Seq %s:\\\\ \n%s \\\\"%(self.idx,self.current_run['domain'],str(metrs)))