-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_one_epoch.py
147 lines (119 loc) · 5.46 KB
/
train_one_epoch.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
import math
import sys
import utils
import torch
import torch.distributed as dist
from util.grouping_utils import generate_grid
def train_one_epoch(model, optimizer, data_loader, device, epoch, print_freq, args):
model.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
header = 'Epoch: [{}]'.format(epoch)
lr_scheduler = None
if epoch == 0:
warmup_factor = 1. / 1000
warmup_iters = min(1000, len(data_loader) - 1)
lr_scheduler = utils.warmup_lr_scheduler(optimizer, warmup_iters, warmup_factor)
i = 0
for images, targets_spp, targets_gt in metric_logger.log_every(data_loader, print_freq, header):
i+=1
# metric_logger.update(loss=0.3)
# metric_logger.update(lr=optimizer.param_groups[0]["lr"])
# if i > 10:
# break;
images = list(image.to(device) for image in images)
targets_spp = [{k: v.to(device) for k, v in t.items()} for t in targets_spp]
targets_gt = [{k: v.to(device) for k, v in t.items()} for t in targets_gt]
if args.spp == "grid":
for tgt in targets_spp:
boxes_spp = tgt["boxes"]
H,W = tgt["masks"].shape[-2:]
boxes, masks = generate_grid(H,W)
tgt["boxes"] = boxes.to(boxes_spp)
tgt["masks"] = masks.to(boxes_spp)
tgt["labels"] = torch.ones(len(boxes), dtype=torch.int64, device=device)
loss_dict = model(images, targets_spp, targets_gt)
# aout = [None for _ in range(8)]
# dist.all_gather_object(aout, image_id)
# image_id = [a for sublist in aout for a in sublist]
# torch.save(image_id , "image_id.pth")
losses = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = utils.reduce_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
loss_value = losses_reduced.item()
if not math.isfinite(loss_value) or torch.isnan(losses_reduced):
print("Loss is {}, stopping training".format(loss_value))
print(loss_dict_reduced)
sys.exit(1)
optimizer.zero_grad()
losses.backward()
# torch.nn.utils.clip_grad_norm_(model.parameters(), 3.0)
optimizer.step()
if lr_scheduler is not None:
lr_scheduler.step()
metric_logger.update(loss=losses_reduced, **loss_dict_reduced)
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
return metric_logger
# def _get_iou_types(model):
# model_without_ddp = model
# if isinstance(model, torch.nn.parallel.DistributedDataParallel):
# model_without_ddp = model.module
# iou_types = ["bbox"]
# if isinstance(model_without_ddp, torchvision.models.detection.MaskRCNN):
# iou_types.append("segm")
# if isinstance(model_without_ddp, torchvision.models.detection.KeypointRCNN):
# iou_types.append("keypoints")
# return iou_types
# @torch.no_grad()
# def evaluate(model, data_loader, device, toBinary=False):
# n_threads = torch.get_num_threads()
# toBinary = True
# # FIXME remove this and make paste_masks_in_image run on the GPU
# torch.set_num_threads(1)
# cpu_device = torch.device("cpu")
# model.eval()
# metric_logger = utils.MetricLogger(delimiter=" ")
# header = 'Test:'
# coco = get_coco_api_from_dataset(data_loader.dataset)
# # iou_types = _get_iou_types(model)
# iou_types = ["bbox" , "segm"]
# coco_evaluator = CocoEvaluator(coco, iou_types, toBinary=toBinary)
# for iou_type in iou_types:
# coco_evaluator.coco_eval[iou_type].params.maxDets = [300,500,1000]
# it = 0
# for images, targets_gt in metric_logger.log_every(data_loader, 100, header):
# import pdb; pdb.set_trace()
# it += 1
# # if it>100:
# # break
# images = list(img.to(device) for img in images)
# # targets_spp = [{k: v.to(device) for k, v in t.items()} for t in targets_spp]
# targets_gt = [{k: v.to(device) for k, v in t.items()} for t in targets_gt]
# if torch.cuda.is_available():
# torch.cuda.synchronize()
# model_time = time.time()
# outputs = model(images)
# outputs = [{k: v.to(cpu_device) for k, v in t.items()} for t in outputs]
# for t in outputs:
# # t["labels"] = (t["labels"] > 0).long()
# assert(torch.allclose(t["labels"] , torch.ones_like(t["labels"])))
# model_time = time.time() - model_time
# # import pdb; pdb.set_trace()
# res = {target["image_id"].item(): output for target, output in zip(targets_gt, outputs)}
# evaluator_time = time.time()
# coco_evaluator.update(res)
# evaluator_time = time.time() - evaluator_time
# metric_logger.update(model_time=model_time, evaluator_time=evaluator_time)
# # gather the stats from all processes
# metric_logger.synchronize_between_processes()
# print("Averaged stats:", metric_logger)
# coco_evaluator.synchronize_between_processes()
# # accumulate predictions from all images
# coco_evaluator.accumulate()
# stats = coco_evaluator.summarize()
# print("#"*10)
# print("Average Recall @100:{:.4f}".format(stats['segm'][8]))
# print("#"*10)
# torch.set_num_threads(n_threads)
# return coco_evaluator