-
Notifications
You must be signed in to change notification settings - Fork 1
/
engine.py
205 lines (167 loc) · 8.48 KB
/
engine.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
"""
Train and eval functions used in main.py
Modified from DETR (https://github.com/facebookresearch/detr)
"""
import math
from models import postprocessors
import os
import sys
from typing import Iterable
import torch
import torch.distributed as dist
import util.misc as utils
from datasets.coco_eval import CocoEvaluator
from datasets.refexp_eval import RefExpEvaluator
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
from datasets.a2d_eval import calculate_precision_at_k_and_iou_metrics
def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module,
data_loader: Iterable, optimizer: torch.optim.Optimizer,
device: torch.device, lr_scheduler, epoch: int, max_norm: float = 0):
model.train()
criterion.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
header = 'Epoch: [{}]'.format(epoch)
print_freq = 10
for samples, targets in metric_logger.log_every(data_loader, print_freq, header):
samples = samples.to(device)
captions = [t["caption"] for t in targets]
targets = utils.targets_to(targets, device)
outputs = model(samples, captions, targets)
loss_dict = criterion(outputs, targets)
weight_dict = criterion.weight_dict
losses = sum(loss_dict[k] * weight_dict[k] for k in loss_dict.keys() if k in weight_dict)
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = utils.reduce_dict(loss_dict)
loss_dict_reduced_unscaled = {f'{k}_unscaled': v
for k, v in loss_dict_reduced.items()}
loss_dict_reduced_scaled = {k: v * weight_dict[k]
for k, v in loss_dict_reduced.items() if k in weight_dict}
losses_reduced_scaled = sum(loss_dict_reduced_scaled.values())
loss_value = losses_reduced_scaled.item()
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
print(loss_dict_reduced)
sys.exit(1)
optimizer.zero_grad()
losses.backward()
if max_norm > 0:
grad_total_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
else:
grad_total_norm = utils.get_total_grad_norm(model.parameters(), max_norm)
optimizer.step()
# lr_scheduler.step()
metric_logger.update(loss=loss_value, **loss_dict_reduced_scaled, **loss_dict_reduced_unscaled)
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
metric_logger.update(grad_norm=grad_total_norm)
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
@torch.no_grad()
def evaluate(model, criterion, postprocessors, data_loader, evaluator_list, device, args):
model.eval()
criterion.eval()
metric_logger = utils.MetricLogger(delimiter=" ")
header = 'Test:'
for samples, targets in metric_logger.log_every(data_loader, 10, header):
samples = samples.to(device)
captions = [t["caption"] for t in targets]
targets = utils.targets_to(targets, device)
outputs = model(samples, captions, targets)
loss_dict = criterion(outputs, targets)
weight_dict = criterion.weight_dict
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = utils.reduce_dict(loss_dict)
loss_dict_reduced_scaled = {k: v * weight_dict[k]
for k, v in loss_dict_reduced.items() if k in weight_dict}
loss_dict_reduced_unscaled = {f'{k}_unscaled': v
for k, v in loss_dict_reduced.items()}
metric_logger.update(loss=sum(loss_dict_reduced_scaled.values()),
**loss_dict_reduced_scaled,
**loss_dict_reduced_unscaled)
orig_target_sizes = torch.stack([t["orig_size"] for t in targets], dim=0)
results = postprocessors['bbox'](outputs, orig_target_sizes)
if 'segm' in postprocessors.keys():
target_sizes = torch.stack([t["size"] for t in targets], dim=0)
results = postprocessors['segm'](results, outputs, orig_target_sizes, target_sizes)
res = {target['image_id']: output for target, output in zip(targets, results)}
for evaluator in evaluator_list:
evaluator.update(res)
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
for evaluator in evaluator_list:
evaluator.synchronize_between_processes()
# accumulate predictions from all images
refexp_res = None
for evaluator in evaluator_list:
if isinstance(evaluator, CocoEvaluator):
evaluator.accumulate()
evaluator.summarize()
elif isinstance(evaluator, RefExpEvaluator):
refexp_res = evaluator.summarize()
stats = {k: meter.global_avg for k, meter in metric_logger.meters.items()}
# update stats
for evaluator in evaluator_list:
if isinstance(evaluator, CocoEvaluator):
if "bbox" in postprocessors.keys():
stats["coco_eval_bbox"] = evaluator.coco_eval["bbox"].stats.tolist()
if "segm" in postprocessors.keys():
stats["coco_eval_masks"] = evaluator.coco_eval["segm"].stats.tolist()
if refexp_res is not None:
stats.update(refexp_res)
return stats
@torch.no_grad()
def evaluate_a2d(model, data_loader, postprocessor, device, args):
model.eval()
predictions = []
metric_logger = utils.MetricLogger(delimiter=" ")
header = 'Test:'
for samples, targets in metric_logger.log_every(data_loader, 10, header):
image_ids = [t['image_id'] for t in targets]
samples = samples.to(device)
captions = [t["caption"] for t in targets]
targets = utils.targets_to(targets, device)
outputs = model(samples, captions, targets)
orig_target_sizes = torch.stack([t["orig_size"] for t in targets], dim=0)
target_sizes = torch.stack([t["size"] for t in targets], dim=0)
processed_outputs = postprocessor(outputs, orig_target_sizes, target_sizes)
for p, image_id in zip(processed_outputs, image_ids):
for s, m in zip(p['scores'], p['rle_masks']):
predictions.append({'image_id': image_id,
'category_id': 1, # dummy label, as categories are not predicted in ref-vos
'segmentation': m,
'score': s.item()})
# gather and merge predictions from all gpus
gathered_pred_lists = utils.all_gather(predictions)
predictions = [p for p_list in gathered_pred_lists for p in p_list]
# evaluation
eval_metrics = {}
if utils.is_main_process():
if args.dataset_file == 'a2d':
coco_gt = COCO(os.path.join(args.a2d_path, 'a2d_sentences_test_annotations_in_coco_format.json'))
elif args.dataset_file == 'jhmdb':
coco_gt = COCO(os.path.join(args.jhmdb_path, 'jhmdb_sentences_gt_annotations_in_coco_format.json'))
elif args.dataset_file == 'refcocoVideo':
coco_gt = COCO(os.path.join(args.coco_path, 'refcocoVideo/finetune_refcoco_val.json'))
else:
raise NotImplementedError
coco_pred = coco_gt.loadRes(predictions)
coco_eval = COCOeval(coco_gt, coco_pred, iouType='segm')
coco_eval.params.useCats = 0 # ignore categories as they are not predicted in ref-vos task
coco_eval.evaluate()
coco_eval.accumulate()
coco_eval.summarize()
ap_labels = ['mAP 0.5:0.95', 'AP 0.5', 'AP 0.75', 'AP 0.5:0.95 S', 'AP 0.5:0.95 M', 'AP 0.5:0.95 L']
ap_metrics = coco_eval.stats[:6]
eval_metrics = {l: m for l, m in zip(ap_labels, ap_metrics)}
# Precision and IOU
precision_at_k, overall_iou, mean_iou = calculate_precision_at_k_and_iou_metrics(coco_gt, coco_pred)
eval_metrics.update({f'P@{k}': m for k, m in zip([0.5, 0.6, 0.7, 0.8, 0.9], precision_at_k)})
eval_metrics.update({'overall_iou': overall_iou, 'mean_iou': mean_iou})
print(eval_metrics)
# sync all processes before starting a new epoch or exiting
dist.barrier()
return eval_metrics