def __init__(self,
video_src: str,
video_output: str,
text_output: str,
obj_list: list,
input_sizes: list,
reid_cpkt: str,
compound_coef: int,
force_input_size=None,
threshold=0.2,
iou_threshold=0.2,
use_cuda=True,
use_float16=False,
cudnn_fastest=True,
cudnn_benchmark=True,
max_dist=0.2,
min_confidence=0.3,
nms_max_overlap=0.5,
max_iou_distance=0.7,
max_age=70,
n_init=3,
nn_budget=100,
selected_target=None):
# I/O
# Video's path
self.video_src = video_src # set int to use webcam, set str to read from a video file
self.video_output = video_output # output to the specific position
# text path
self.text_output = text_output # output to the file with the csv format
# DETECTOR
self.compound_coef = compound_coef
self.force_input_size = force_input_size # set None to use default size
self.threshold = threshold
self.iou_threshold = iou_threshold
self.use_cuda = use_cuda
self.use_float16 = use_float16
cudnn.fastest = cudnn_fastest
cudnn.benchmark = cudnn_benchmark
# coco_name
self.obj_list = obj_list
# input size
self.input_sizes = input_sizes
self.input_size = input_sizes[self.compound_coef] if force_input_size is None else force_input_size
# load detector model
model = EfficientDetBackbone(compound_coef=self.compound_coef, num_classes=len(obj_list))
model.load_state_dict(torch.load(f'weights/efficientdet-d{self.compound_coef}.pth'))
model.requires_grad_(False)
model.eval()
if self.use_cuda and torch.cuda.is_available():
self.detector = model.cuda()
if self.use_float16:
self.detector = model.half()
# TRACKER
self.reid_cpkt = reid_cpkt
self.max_dist = max_dist
self.min_confidence = min_confidence
self.nms_max_overlap = nms_max_overlap
self.max_iou_distance = max_iou_distance
self.max_age = max_age
self.n_init = n_init
self.nn_budget = nn_budget
# load tracker model,
self.trackers = []
self.selected_target = selected_target
for num in range(0, len(self.selected_target)):
self.trackers.append(build_tracker(reid_cpkt,
max_dist,
min_confidence,
nms_max_overlap,
max_iou_distance,
max_age,
n_init,
nn_budget,
use_cuda))
# video frames
self.frame_id = 0
def main(i):
compound_coef = i
force_input_size = None # set None to use default size
# replace this part with your project's anchor config
anchor_ratios = [(1.0, 1.0), (1.4, 0.7), (0.7, 1.4)]
anchor_scales = [2 ** 0, 2 ** (1.0 / 3.0), 2 ** (2.0 / 3.0)]
threshold = 0.2
iou_threshold = 0.2
use_cuda = True
use_float16 = False
cudnn.fastest = True
cudnn.benchmark = True
obj_list = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
'fire hydrant', '', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep',
'cow', 'elephant', 'bear', 'zebra', 'giraffe', '', 'backpack', 'umbrella', '', '', 'handbag', 'tie',
'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove',
'skateboard', 'surfboard', 'tennis racket', 'bottle', '', 'wine glass', 'cup', 'fork', 'knife', 'spoon',
'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut',
'cake', 'chair', 'couch', 'potted plant', 'bed', '', 'dining table', '', '', 'toilet', '', 'tv',
'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink',
'refrigerator', '', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier',
'toothbrush']
out_dict = dict()
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
input_size = input_sizes[compound_coef] if force_input_size is None else force_input_size
model = EfficientDetBackbone(compound_coef=compound_coef, num_classes=len(obj_list),
ratios=anchor_ratios, scales=anchor_scales)
model.load_state_dict(torch.load(f'weights/efficientdet-d{compound_coef}.pth', map_location='cpu'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
base_dir = '/data/jiashenc/jackson/'
print('Processing Det-' + str(i))
for k in range(1000000, 1100000):
if k % 1000 == 0:
print(' Finish {} frames'.format(k + 1))
img_path = os.path.join(base_dir, 'frame{}.jpg'.format(k))
ori_imgs, framed_imgs, framed_metas = preprocess(img_path, max_size=input_size)
if use_cuda:
x = torch.stack([torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32 if not use_float16 else torch.float16).permute(0, 3, 1, 2)
with torch.no_grad():
features, regression, classification, anchors = model(x)
regressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()
out = postprocess(x,
anchors, regression, classification,
regressBoxes, clipBoxes,
threshold, iou_threshold)
out = invert_affine(framed_metas, out)
to_json(out, out_dict)
with open(os.path.join(base_dir, '10', 'res-{:d}.json'.format(i)), 'w') as f:
json.dump(out_dict, f)
out_dict = dict()
# obj_list = params['obj_list']
# obj_list = params['obj_list_fangzhenchui']
# obj_list = params['obj_list_xianjia']
obj_list = params['obj_list_fushusheshi']
# obj_list = params['obj_list_zhongchui']
# obj_list = params['obj_list_daodixian']
# obj_list = params['obj_list_tongdao']
# obj_list = params['obj_list_ganta']
# input_sizes = [512, 640, 768, 896, 1024, 1280, 1408, 1536, 1536]
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
# input_sizes = [512, 640, 768, 896, 1024, 1280, 1408, 768, 1536]
det_save_odgt = f'val_result/fushusheshi_4class_d3_lr_1e-3_only_bad_epoch260_1201.odgt'
det_save_eval_log_txt = f'val_result/fushusheshi_4class_d3_lr_1e-3_only_bad_epoch260_1201.txt'
threshold = 0.05
if __name__ == '__main__':
with torch.cuda.device(0):
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(obj_list),
ratios=eval(params['anchors_ratios']),
scales=eval(params['anchors_scales']))
macs, params = get_model_complexity_info(model, (3, 224, 224),
as_strings=True,
print_per_layer_stat=True,
verbose=True)
print('{:<30} {:<8}'.format('Computational complexity: ', macs))
print('{:<30} {:<8}'.format('Number of parameters: ', params))
def load_apex_model(compound_coef, obj_list, params, weights_path):
opt_level = 'O1'
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(obj_list),
ratios=eval(params['anchors_ratios']),
scales=eval(params['anchors_scales']))
checkpoint = torch.load(weights_path)
model = model.cuda(gpu)
optimizer = torch.optim.AdamW(model.parameters(), lr)
model, optimizer = amp.initialize(model, optimizer, opt_level=opt_level)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
amp.load_state_dict(checkpoint['amp'])
model.requires_grad_(False)
model.cuda(gpu)
model = model.eval()
return model
coco_eval.accumulate()
coco_eval.summarize()
if __name__ == '__main__':
SET_NAME = params['val_set']
VAL_GT = f'datasets/{params["project_name"]}/annotations/instances_{SET_NAME}.json'
VAL_IMGS = f'datasets/{params["project_name"]}/{SET_NAME}/'
MAX_IMAGES = 10000
coco_gt = COCO(VAL_GT)
image_ids = coco_gt.getImgIds()[:MAX_IMAGES]
if override_prev_results or not os.path.exists(
f'{SET_NAME}_bbox_results.json'):
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(obj_list),
ratios=eval(params['anchors_ratios']),
scales=eval(params['anchors_scales']))
model.load_state_dict(
torch.load(weights_path, map_location=torch.device('cpu')))
model.requires_grad_(False)
model.eval()
if use_cuda:
model.cuda(gpu)
if use_float16:
model.half()
evaluate_coco(VAL_IMGS, SET_NAME, image_ids, coco_gt, model)
_eval(coco_gt, image_ids, f'{SET_NAME}_bbox_results.json')
# tf bilinear interpolation is different from any other's, just make do
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
input_size = input_sizes[compound_coef] if force_input_size is None else force_input_size
ori_imgs, framed_imgs, framed_metas = preprocess(img_path, max_size=input_size)
if use_cuda:
x = torch.stack([torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32 if not use_float16 else torch.float16).permute(0, 3, 1, 2)
model = EfficientDetBackbone(compound_coef=compound_coef, num_classes=len(obj_list),
# replace this part with your project's anchor config
ratios=[(1.0, 1.0), (1.4, 0.7), (0.7, 1.4)],
scales=[2 ** 0, 2 ** (1.0 / 3.0), 2 ** (2.0 / 3.0)])
model.load_state_dict(torch.load('logs/wheat/efficientdet-d0_38_7356.pth'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
with torch.no_grad():
features, regression, classification, anchors = model(x)
print('BBox')
coco_eval = COCOeval(coco_gt, coco_pred, 'bbox')
coco_eval.params.imgIds = image_ids
coco_eval.evaluate()
coco_eval.accumulate()
coco_eval.summarize()
if __name__ == '__main__':
SET_NAME = params['val_set']
VAL_GT = f'datasets/{params["project_name"]}/annotations/instances_{SET_NAME}.json'
VAL_IMGS = f'datasets/{params["project_name"]}/{SET_NAME}/'
MAX_IMAGES = 10000
coco_gt = COCO(VAL_GT)
image_ids = coco_gt.getImgIds()[:MAX_IMAGES]
if not os.path.exists(f'{SET_NAME}_bbox_results.json'):
model = EfficientDetBackbone(compound_coef=compound_coef, num_classes=len(obj_list))
model.load_state_dict(torch.load(weights_path))
model.requires_grad_(False)
model.eval()
if use_cuda:
model.cuda()
image_ids = evaluate_coco(VAL_IMGS, SET_NAME, image_ids, coco_gt, model)
eval(coco_gt, image_ids, f'{SET_NAME}_bbox_results.json')
else:
eval(coco_gt, image_ids, f'{SET_NAME}_bbox_results.json')
# tf bilinear interpolation is different from any other's, just make do
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
input_size = input_sizes[
compound_coef] if force_input_size is None else force_input_size
ori_imgs, framed_imgs, framed_metas = preprocess(img_path, max_size=input_size)
if use_cuda:
x = torch.stack([torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32 if not use_float16 else torch.float16).permute(
0, 3, 1, 2)
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(obj_list),
ratios=anchor_ratios,
scales=anchor_scales)
model.load_state_dict(torch.load(weight, map_location='cpu'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
with torch.no_grad():
features, regression, classification, anchors = model(x)
regressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()
def train(args):
assert args.weight_path, 'must indicate the path of initial weight'
if (os.path.exists(f'{args.weight_path}/train_log.txt')):
os.remove(f'{args.weight_path}/train_log.txt')
if (os.path.exists(f'{args.weight_path}/pre_trained_weight.pth')):
os.remove(f'{args.weight_path}/pre_trained_weight.pth')
print("Hi")
present_time = time.strftime("%Y_%m_%d_%H_%M_%S", time.localtime())
params = Params(f'projects/eye.yml')
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
train_params = {'batch_size': args.batch_size,
'shuffle': True,
'drop_last': True,
'collate_fn': collater,
'num_workers': args.num_workers}
val_params = {'batch_size': args.batch_size,
'shuffle': False,
'drop_last': True,
'collate_fn': collater,
'num_workers': args.num_workers}
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
model = EfficientDetBackbone(num_classes=len(params.obj_list), compound_coef=args.compound_coef,
ratios=eval(params.anchors_ratios), scales=eval(params.anchors_scales))
init_weights(model)
# warp the model with loss function, to reduce the memory usage on gpu0 and speedup
model = ModelWithLoss(model)
model = model.cuda()
if args.optim == 'adamw':
optimizer = torch.optim.AdamW(model.parameters(), args.lr)
else:
optimizer = torch.optim.SGD(model.parameters(), args.lr, momentum=0.9, nesterov=True)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=args.patience, verbose=True) # unit is epoch
img_list = glob.glob(f"{args.dataset_path}/train/*")
normal_img_list = []
yellow_img_list = []
for img in img_list:
if (img.find("n_") != -1):
normal_img_list.append(img)
else:
yellow_img_list.append(img)
random.shuffle(normal_img_list)
random.shuffle(yellow_img_list)
normal_val_num = int(len(normal_img_list) / 5)
yellow_val_num = int(len(yellow_img_list) / 5)
train_img_list = normal_img_list[normal_val_num:] + yellow_img_list[yellow_val_num:]
val_img_list = normal_img_list[:normal_val_num] + yellow_img_list[:yellow_val_num]
train_anno_txt_path = f"{args.dataset_path}/train.txt"
val_anno_txt_path = f"{args.dataset_path}/train.txt"
train_transform = transforms.Compose([# Normalizer(mean=params.mean, std=params.std),
Augmenter(),
randomScaleWidth(),
randomBlur(),
# randomBrightness(),
# randomHue(),
# randomSaturation(),
Normalizer(mean=params.mean, std=params.std),
Resizer(input_sizes[args.compound_coef])])
val_transform = transforms.Compose([# Normalizer(mean=params.mean, std=params.std),
Augmenter(),
Normalizer(mean=params.mean, std=params.std),
Resizer(input_sizes[args.compound_coef])])
train_set = EyeDataset(train_img_list, train_anno_txt_path, train_transform)
val_set = EyeDataset(val_img_list, val_anno_txt_path, val_transform)
train_generator = DataLoader(train_set, **train_params)
val_generator = DataLoader(val_set, **val_params)
model.model.load_state_dict(torch.load(f'{args.weight_path}/init_weight.pth')["model_state_dict"])
optimizer.load_state_dict(torch.load(f'{args.weight_path}/init_weight.pth')["optimizer_state_dict"])
scheduler.load_state_dict(torch.load(f'{args.weight_path}/init_weight.pth')["scheduler_state_dict"])
model.train()
best_val_loss = 1e5
for epoch in range(args.epoch):
model.train()
total_loss_ls = []
total_correct = 0
total = 0
for data in train_generator:
imgs = data['img'].cuda()
annot = data['annot'].cuda()
optimizer.zero_grad()
reg_loss, cls_head_loss, cls_correct_num, total_num = model(imgs, annot, obj_list=params.obj_list)
total_correct += cls_correct_num
total += total_num
reg_loss = reg_loss.mean()
loss = cls_head_loss + reg_loss
total_loss_ls.append(loss.item())
if (loss == 0 or not torch.isfinite(loss)):
continue
loss.backward()
optimizer.step()
total_loss = np.mean(total_loss_ls)
scheduler.step(total_loss)
with open(f'{args.weight_path}/train_log.txt', 'a') as fp:
fp.write(f'Epoch: {epoch} loss: {total_loss:.6f} | acc: {total_correct / total * 100:.2f}\n')
model.eval()
with torch.no_grad():
total = 0
total_correct = 0
total_loss_ls = []
for data in val_generator:
imgs = data['img'].cuda()
annot = data['annot'].cuda()
reg_loss, cls_head_loss, cls_correct_num, total_num = model(imgs, annot, obj_list=params.obj_list)
total += total_num
total_correct += cls_correct_num
reg_loss = reg_loss.mean()
loss = cls_head_loss + reg_loss
total_loss_ls.append(loss.item())
total_loss = np.mean(total_loss_ls)
with open(f'{args.weight_path}/train_log.txt', 'a') as fp:
fp.write(f'Epoch: {epoch} loss: {total_loss:.6f} | acc: {total_correct / total * 100:.2f}\n\n')
if (total_loss < best_val_loss):
best_val_loss = total_loss
torch.save({
"model_state_dict": model.model.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"scheduler_state_dict": scheduler.state_dict(),
}, f"{args.weight_path}/pre_trained_weight.pth")
import yaml
from backbone import EfficientDetBackbone
import torch
class Params:
def __init__(self, project_file):
self.params = yaml.safe_load(open(project_file).read())
def __getattr__(self, item):
return self.params.get(item, None)
params = Params('projects/coco.yml')
model = EfficientDetBackbone(num_classes=len(params.obj_list),
compound_coef=0,
ratios=eval(params.anchors_ratios),
scales=eval(params.anchors_scales))
_ = model.load_state_dict(torch.load("./weights/efficientdet-d0.pth"),
strict=False)
model.eval()
_ = model(torch.Tensor(1, 3, 512, 512))
trace_model = torch.jit.trace(model, (torch.Tensor(1, 3, 512, 512), ))
# trace_model.save('./weights/efficientdet-d0_jit.pt')
trace_model.save('./weights/efficientdet.pt')
def toCategory(obj):
mlem = dict()
for k, v in enumerate(obj):
mlem[k] = v
return mlem
color_list = standard_to_bgr(STANDARD_COLORS)
# tf bilinear interpolation is different from any other's, just make do
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
input_size = input_sizes[
compound_coef] if force_input_size is None else force_input_size
model = EfficientDetBackbone(num_classes=len(params.obj_list),
compound_coef=4,
ratios=eval(params.anchors_ratios),
scales=eval(params.anchors_scales))
# print(model)
# weights_path = './weights/efficientdet-d4.pth'
weights_path = 'C:/Users/giang/Desktop/efficientdet-d4_24_3500.pth'
# model.load_state_dict(torch.load(weights_path, map_location = 'cpu'), strict=False)
from efficientdet.model import Classifier
# model.backbone_net.model._conv_stem.conv = nn.Conv2d(4, 48, kernel_size=(3, 3), stride=(2, 2), bias=False)
# model.classifier.header.pointwise_conv.conv = nn.Conv2d(224, 9, kernel_size=(1, 1), stride=(1, 1))
model.classifier = Classifier(in_channels=model.fpn_num_filters[4],
num_anchors=model.num_anchors,
num_classes=1,
num_layers=model.box_class_repeats[4],
pyramid_levels=model.pyramid_levels[4])
model.load_state_dict(torch.load(weights_path), strict=False)
# model.load_state_dict(torch.load(f'weights/efficientdet-d{compound_coef}.pth', map_location='cpu'))
def train_det(opt, cfg):
# # Write history
# if 'backlog' not in opt.config:
# with open(os.path.join(opt.saved_path, f'{opt.project}_backlog.yml'), 'w') as f:
# doc = open(f'projects/{opt.project}.yml', 'r')
# f.write('#History log file')
# f.write(f'\n__backlog__: {now.strftime("%Y/%m/%d %H:%M:%S")}\n')
# f.write(doc.read())
# f.write('\n# Manual seed used')
# f.write(f'\nmanual_seed: {cfg.manual_seed}')
# else:
# with open(os.path.join(opt.saved_path, f'{opt.project}_history.yml'), 'w') as f:
# doc = open(f'projects/{opt.project}.yml', 'r')
# f.write(doc.read())
training_params = {
'batch_size': cfg.batch_size,
'shuffle': True,
'drop_last': True,
'collate_fn': collater,
'num_workers': opt.num_workers
}
val_params = {
'batch_size': cfg.batch_size,
'shuffle': False,
'drop_last': True,
'collate_fn': collater,
'num_workers': opt.num_workers
}
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
training_set = DataGenerator(
data_path=os.path.join(opt.data_path, 'Train'),
class_ids=cfg.dictionary_class_name.keys(),
transform=transforms.Compose([
Augmenter(),
Normalizer(mean=cfg.mean, std=cfg.std),
Resizer(input_sizes[cfg.compound_coef])
]),
pre_augments=['', *[f'{aug}_' for aug in cfg.augment_list]]
if cfg.augment_list else None)
training_generator = DataLoader(training_set, **training_params)
val_set = DataGenerator(
# root_dir=os.path.join(opt.data_path, cfg.project_name),
data_path=os.path.join(opt.data_path, 'Validation'),
class_ids=cfg.dictionary_class_name.keys(),
transform=transforms.Compose([
Normalizer(mean=cfg.mean, std=cfg.std),
Resizer(input_sizes[cfg.compound_coef])
]))
val_generator = DataLoader(val_set, **val_params)
model = EfficientDetBackbone(num_classes=len(cfg.dictionary_class_name),
compound_coef=cfg.compound_coef,
ratios=eval(cfg.anchor_ratios),
scales=eval(cfg.anchor_scales))
# load last weights
if opt.load_weights is not None:
if opt.load_weights.endswith('.pth'):
weights_path = opt.load_weights
else:
weights_path = get_last_weights(opt.saved_path)
try:
last_step = int(
os.path.basename(weights_path).split('_')[-1].split('.')[0])
except:
last_step = 0
try:
ret = model.load_state_dict(torch.load(weights_path), strict=False)
except RuntimeError as e:
print(f'[Warning] Ignoring {e}')
print(
'[Warning] Don\'t panic if you see this, '
'this might be because you load a pretrained weights with different number of classes. '
'The rest of the weights should be loaded already.')
print(
f'[Info] loaded weights: {os.path.basename(weights_path)}, resuming checkpoint from step: {last_step}'
)
else:
last_step = 0
print('[Info] initializing weights...')
init_weights(model)
# freeze backbone if train head_only
if cfg.training_layer.lower() == 'heads':
def freeze_backbone(m):
classname = m.__class__.__name__
for ntl in ['EfficientNet', 'BiFPN']:
if ntl in classname:
for param in m.parameters():
param.requires_grad = False
model.apply(freeze_backbone)
print('[Info] freezed backbone')
# https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
# apply sync_bn when using multiple gpu and batch_size per gpu is lower than 4
# useful when gpu memory is limited.
# because when bn is disable, the training will be very unstable or slow to converge,
# apply sync_bn can solve it,
# by packing all mini-batch across all gpus as one batch and normalize, then send it back to all gpus.
# but it would also slow down the training by a little bit.
if cfg.num_gpus > 1 and cfg.batch_size // cfg.num_gpus < 4:
model.apply(replace_w_sync_bn)
use_sync_bn = True
else:
use_sync_bn = False
# warp the model with loss function, to reduce the memory usage on gpu0 and speedup
model = ModelWithLoss(model, debug=opt.debug)
if cfg.num_gpus > 0:
model = model.cuda()
if cfg.num_gpus > 1:
model = CustomDataParallel(model, cfg.num_gpus)
if use_sync_bn:
patch_replication_callback(model)
if cfg.optimizer.lower() == 'adamw':
optimizer = torch.optim.AdamW(model.parameters(), cfg.learning_rate)
if cfg.optimizer.lower() == 'srsgd':
optimizer = SRSGD(model.parameters(),
lr=cfg.learning_rate,
weight_decay=5e-4,
iter_count=100)
else:
optimizer = torch.optim.SGD(model.parameters(),
cfg.learning_rate,
momentum=0.9,
nesterov=True)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
# Setup complete, then start training
now = datetime.datetime.now()
opt.saved_path = opt.saved_path + f'/trainlogs_{now.strftime("%Y%m%d_%H%M%S")}'
if opt.log_path is None:
opt.log_path = opt.saved_path
os.makedirs(opt.log_path, exist_ok=True)
os.makedirs(opt.saved_path, exist_ok=True)
# Write history
if 'backlog' not in opt.config:
with open(
os.path.join(opt.saved_path,
f'{now.strftime("%Y%m%d%H%M%S")}.backlog.json'),
'w') as f:
backlog = dict(cfg.to_pascal_case())
backlog['__metadata__'] = 'Backlog at ' + now.strftime(
"%Y/%m/%d %H:%M:%S")
json.dump(backlog, f)
else:
with open(
os.path.join(opt.saved_path,
f'{now.strftime("%Y%m%d%H%M%S")}.history.json'),
'w') as f:
history = dict(cfg.to_pascal_case())
history['__metadata__'] = now.strftime("%Y/%m/%d %H:%M:%S")
json.dump(history, f)
writer = SummaryWriter(opt.log_path + f'/tensorboard')
epoch = 0
best_loss = 1e5
best_epoch = 0
step = max(0, last_step)
model.train()
num_iter_per_epoch = len(training_generator)
try:
for epoch in range(cfg.no_epochs):
last_epoch = step // num_iter_per_epoch
if epoch < last_epoch:
continue
epoch_loss = []
progress_bar = tqdm(training_generator)
for iter, data in enumerate(progress_bar):
if iter < step - last_epoch * num_iter_per_epoch:
progress_bar.set_description(
f'Skip {iter} < {step} - {last_epoch} * {num_iter_per_epoch}'
)
progress_bar.update()
continue
try:
imgs = data['img']
annot = data['annot']
if cfg.num_gpus == 1:
# if only one gpu, just send it to cuda:0
# elif multiple gpus, send it to multiple gpus in CustomDataParallel, not here
imgs = imgs.cuda()
annot = annot.cuda()
optimizer.zero_grad()
cls_loss, reg_loss = model(
imgs, annot, obj_list=cfg.dictionary_class_name.keys())
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss = cls_loss + reg_loss
if loss == 0 or not torch.isfinite(loss):
continue
loss.backward()
# torch.nn.utils.clip_grad_norm_(model.parameters(), 0.1)
optimizer.step()
epoch_loss.append(float(loss))
progress_bar.set_description(
'Step: {}. Epoch: {}/{}. Iteration: {}/{}. Cls loss: {:.5f}. Reg loss: {:.5f}. '
'Total loss: {:.5f}'.format(step, epoch, cfg.no_epochs,
iter + 1,
num_iter_per_epoch,
cls_loss.item(),
reg_loss.item(),
loss.item()))
writer.add_scalars('Loss', {'train': loss}, step)
writer.add_scalars('Regression_loss', {'train': reg_loss},
step)
writer.add_scalars('Classification_loss',
{'train': cls_loss}, step)
# log learning_rate
current_lr = optimizer.param_groups[0]['lr']
writer.add_scalar('learning_rate', current_lr, step)
step += 1
except Exception as e:
print('[Error]', traceback.format_exc())
print(e)
continue
scheduler.step(np.mean(epoch_loss))
model.eval()
loss_regression_ls = []
loss_classification_ls = []
for iter, data in enumerate(val_generator):
with torch.no_grad():
imgs = data['img']
annot = data['annot']
if cfg.num_gpus == 1:
imgs = imgs.cuda()
annot = annot.cuda()
cls_loss, reg_loss = model(
imgs, annot, obj_list=cfg.dictionary_class_name.keys())
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss = cls_loss + reg_loss
if loss == 0 or not torch.isfinite(loss):
continue
loss_classification_ls.append(cls_loss.item())
loss_regression_ls.append(reg_loss.item())
cls_loss = np.mean(loss_classification_ls)
reg_loss = np.mean(loss_regression_ls)
loss = cls_loss + reg_loss
progress_bar.set_description(
'Val. Epoch: {}/{}. Classification loss: {:1.5f}. Regression loss: {:1.5f}.'
' Total loss: {:1.5f}'.format(epoch, cfg.no_epochs, cls_loss,
reg_loss, loss))
writer.add_scalars('Loss', {'val': loss}, step)
writer.add_scalars('Regression_loss', {'val': reg_loss}, step)
writer.add_scalars('Classification_loss', {'val': cls_loss}, step)
if cfg.only_best_weights:
if loss + opt.es_min_delta < best_loss:
best_loss = loss
best_epoch = epoch
save_checkpoint(
model,
f"{opt.saved_path}/det_d{cfg.compound_coef}_{epoch}_{step}.pth"
)
else:
if loss + opt.es_min_delta < best_loss:
best_loss = loss
best_epoch = epoch
save_checkpoint(
model,
f"{opt.saved_path}/det_d{cfg.compound_coef}_{epoch}_{step}.pth"
)
model.train()
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
'[Info] Stop training at epoch {}. The lowest loss achieved is {}'
.format(epoch, best_loss))
break
print(
f'[Info] Finished training. Best loss achieved {best_loss} at epoch {best_epoch}.'
)
except KeyboardInterrupt:
save_checkpoint(
model, f"{opt.saved_path}/d{cfg.compound_coef}_{epoch}_{step}.pth")
writer.close()
writer.close()
# tf bilinear interpolation is different from any other's, just make do
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
input_size = input_sizes[
compound_coef] if force_input_size is None else force_input_size
ori_imgs, framed_imgs, framed_metas = preprocess(img_path, max_size=input_size)
if use_cuda:
x = torch.stack([torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32 if not use_float16 else torch.float16).permute(
0, 3, 1, 2)
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(obj_list),
ratios=anchor_ratios,
scales=anchor_scales)
model.load_state_dict(
torch.load(f'firstThinking/bdd100k/efficientdet-d1_40_45500.pth'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
with torch.no_grad():
features, regression, classification, anchors = model(x)
regressBoxes = BBoxTransform()
def effdet_detection(content, effdet):
video_src = 0 # set int to use webcam, set str to read from a video file
compound_coef = 0
force_input_size = None # set None to use default size
threshold = 0.5
iou_threshold = 0.2
use_cuda = True
use_float16 = False
cudnn.fastest = True
cudnn.benchmark = True
obj_list = [
'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train',
'truck', 'boat', 'traffic light', 'fire hydrant', '', 'stop sign',
'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep',
'cow', 'elephant', 'bear', 'zebra', 'giraffe', '', 'backpack',
'umbrella', '', '', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis',
'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove',
'skateboard', 'surfboard', 'tennis racket', 'bottle', '', 'wine glass',
'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich',
'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake',
'chair', 'couch', 'potted plant', 'bed', '', 'dining table', '', '',
'toilet', '', 'tv', 'laptop', 'mouse', 'remote', 'keyboard',
'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator',
'', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier',
'toothbrush'
]
# tf bilinear interpolation is different from any other's, just make do
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
input_size = input_sizes[
compound_coef] if force_input_size is None else force_input_size
# load model
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(obj_list))
model.load_state_dict(
torch.load(f'weights/efficientdet-d{compound_coef}.pth'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
# function for display
def display(preds, imgs, content, effdet):
for i in range(len(imgs)):
if len(preds[i]['rois']) == 0:
return imgs[i]
for j in range(len(preds[i]['rois'])):
(x1, y1, x2, y2) = preds[i]['rois'][j].astype(np.int)
#cv2.rectangle(imgs[i], (x1, y1), (x2, y2), (255, 255, 0), 2)
obj = obj_list[preds[i]['class_ids'][j]]
score = float(preds[i]['scores'][j])
if obj == content:
effdet.send_message_to_scratch(
(x1 + x2) * 0.5 * 0.625 - 200) #发送指定类别的识别框位置到scratch
print((x1 + x2) * 0.5 * 0.625 - 200)
cv2.rectangle(imgs[i], (x1, y1), (x2, y2), (255, 255, 0),
2)
cv2.putText(imgs[i], '{}, {:.3f}'.format(obj, score),
(x1, y1 + 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5,
(255, 255, 0), 1)
return imgs[i]
# Box
regressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()
# Video capture
cap = cv2.VideoCapture(video_src)
while True:
ret, frame = cap.read()
if not ret:
break
# frame preprocessing
ori_imgs, framed_imgs, framed_metas = preprocess_video(
frame, max_size=input_size)
if use_cuda:
x = torch.stack(
[torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32 if not use_float16 else torch.float16).permute(
0, 3, 1, 2)
# model predict
with torch.no_grad():
features, regression, classification, anchors = model(x)
out = postprocess(x, anchors, regression, classification,
regressBoxes, clipBoxes, threshold,
iou_threshold)
# result
out = invert_affine(framed_metas, out)
img_show = display(out, ori_imgs, content, effdet)
# show frame by frame
cv2.imshow('frame', img_show)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cap.release()
cv2.destroyAllWindows()
class ObjectDetectionService(PTServingBaseService):
def __init__(self, model_name, model_path):
# effdet
self.model_name = model_name
self.model_path = os.path.join(os.path.dirname(__file__),
'models_best.pth')
self.input_image_key = 'images'
self.anchor_ratios = [(1.0, 1.0), (1.4, 0.7), (0.7, 1.4)]
self.anchor_scales = [2**0, 2**(1.0 / 3.0), 2**(2.0 / 3.0)]
self.compound_coef = 0
self.threshold = 0.2
self.iou_threshold = 0.2
self.obj_list = [
'一次性快餐盒', '书籍纸张', '充电宝', '剩饭剩菜', '包', '垃圾桶', '塑料器皿', '塑料玩具',
'塑料衣架', '大骨头', '干电池', '快递纸袋', '插头电线', '旧衣服', '易拉罐', '枕头', '果皮果肉',
'毛绒玩具', '污损塑料', '污损用纸', '洗护用品', '烟蒂', '牙签', '玻璃器皿', '砧板', '筷子',
'纸盒纸箱', '花盆', '茶叶渣', '菜帮菜叶', '蛋壳', '调料瓶', '软膏', '过期药物', '酒瓶',
'金属厨具', '金属器皿', '金属食品罐', '锅', '陶瓷器皿', '鞋', '食用油桶', '饮料瓶', '鱼骨'
]
self.input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
self.input_size = self.input_sizes[self.compound_coef]
self.model = EfficientDetBackbone(compound_coef=self.compound_coef,
num_classes=len(self.obj_list),
ratios=self.anchor_ratios,
scales=self.anchor_scales)
self.model.load_state_dict(
torch.load(self.model_path, map_location='cpu'))
self.model.requires_grad_(False)
self.model.eval()
def _preprocess(self, data):
preprocessed_data = {}
for k, v in data.items():
for file_name, file_content in v.items():
ori_imgs, framed_imgs, framed_metas = preprocess(
file_content, max_size=self.input_size)
preprocessed_data[k] = [framed_imgs, framed_metas]
return preprocessed_data
def _inference(self, data):
"""
model inference function
Here are a inference example of resnet, if you use another model, please modify this function
"""
framed_imgs, framed_metas = data[self.input_image_key]
if torch.cuda.is_available():
x = torch.stack(
[torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
self.model = self.model.cuda()
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32).permute(0, 3, 1, 2)
#if use_float16:
# model = model.half()
with torch.no_grad():
features, regression, classification, anchors = self.model(x)
regressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()
out = postprocess(x, anchors, regression, classification,
regressBoxes, clipBoxes, self.threshold,
self.iou_threshold)
out = invert_affine(framed_metas, out)
result = OrderedDict()
result['detection_classes'] = []
result['detection_scores'] = []
result['detection_boxes'] = []
for i in range(len(out)):
if len(out[i]['rois']) == 0:
continue
for j in range(len(out[i]['rois'])):
x1, y1, x2, y2 = out[i]['rois'][j].astype(np.int)
result['detection_boxes'].append([x1, y1, x2, y2])
obj = self.obj_list[out[i]['class_ids'][j]]
result['detection_classes'].append(obj)
score = float(out[i]['scores'][j])
result['detection_scores'].append(score)
return result
def _postprocess(self, data):
return data
def inference(self, data):
'''
Wrapper function to run preprocess, inference and postprocess functions.
Parameters
----------
data : map of object
Raw input from request.
Returns
-------
list of outputs to be sent back to client.
data to be sent back
'''
pre_start_time = time.time()
data = self._preprocess(data)
infer_start_time = time.time()
# Update preprocess latency metric
pre_time_in_ms = (infer_start_time - pre_start_time) * 1000
logger.info('preprocess time: ' + str(pre_time_in_ms) + 'ms')
if self.model_name + '_LatencyPreprocess' in MetricsManager.metrics:
MetricsManager.metrics[self.model_name +
'_LatencyPreprocess'].update(pre_time_in_ms)
data = self._inference(data)
infer_end_time = time.time()
infer_in_ms = (infer_end_time - infer_start_time) * 1000
logger.info('infer time: ' + str(infer_in_ms) + 'ms')
data = self._postprocess(data)
# Update inference latency metric
post_time_in_ms = (time.time() - infer_end_time) * 1000
logger.info('postprocess time: ' + str(post_time_in_ms) + 'ms')
if self.model_name + '_LatencyInference' in MetricsManager.metrics:
MetricsManager.metrics[self.model_name +
'_LatencyInference'].update(post_time_in_ms)
# Update overall latency metric
if self.model_name + '_LatencyOverall' in MetricsManager.metrics:
MetricsManager.metrics[self.model_name +
'_LatencyOverall'].update(pre_time_in_ms +
post_time_in_ms)
logger.info('latency: ' +
str(pre_time_in_ms + infer_in_ms + post_time_in_ms) + 'ms')
data['latency_time'] = str(
round(pre_time_in_ms + infer_in_ms + post_time_in_ms, 1)) + ' ms'
return data
def main(args):
print("Hi")
if (os.path.exists(f"{args.weight_path}/test_log.out")):
os.remove(f"{args.weight_path}/test_log.out")
assert args.weight_path, 'must indicate the path of pre-trained weight'
params = Params(f'projects/eye.yml')
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
test_params = {
'batch_size': args.batch_size,
'shuffle': False,
'drop_last': True,
'collate_fn': collater,
'num_workers': args.num_workers
}
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
model = EfficientDetBackbone(num_classes=len(params.obj_list),
compound_coef=args.compound_coef,
ratios=eval(params.anchors_ratios),
scales=eval(params.anchors_scales))
init_weights(model)
model = ModelWithLoss(model)
model = model.cuda()
if args.optim == 'adamw':
optimizer = torch.optim.AdamW(model.parameters(), args.lr)
else:
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=0.9,
nesterov=True)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=5,
verbose=True)
model.model.load_state_dict(
torch.load(f'{args.weight_path}/pre_trained_weight.pth')
['model_state_dict'])
optimizer.load_state_dict(
torch.load(f'{args.weight_path}/pre_trained_weight.pth')
['optimizer_state_dict'])
scheduler.load_state_dict(
torch.load(f'{args.weight_path}/pre_trained_weight.pth')
['scheduler_state_dict'])
test_img_list = glob.glob(f'{args.dataset_path}/test/*')
test_anno_txt_path = f'{args.dataset_path}/test.txt'
test_transform = transforms.Compose(
[ # Normalizer(mean=params.mean, std=params.std),
Augmenter(),
Normalizer(mean=params.mean, std=params.std),
Resizer(input_sizes[args.compound_coef])
])
test_set = EyeDataset(test_img_list, test_anno_txt_path, test_transform)
test_generator = DataLoader(test_set, **test_params)
model.eval()
with torch.no_grad():
total = 0
total_correct = 0
total_loss_ls = []
for data in test_generator:
imgs = data['img'].cuda()
annot = data['annot'].cuda()
reg_loss, cls_head_loss, cls_correct_num, total_num = model(
imgs, annot, obj_list=params.obj_list)
total_correct += cls_correct_num
total += total_num
reg_loss = reg_loss.mean()
loss = reg_loss + cls_head_loss
total_loss_ls.append(loss.item())
total_loss = np.mean(total_loss_ls)
with open(f"{args.weight_path}/test_log.out", 'a') as fp:
fp.write(
f'Testing loss: {total_loss:.6f} | acc: {total_correct / total * 100:.2f}\n'
)
def train(opt):
params = Params(f'projects/{opt.project}.yml')
params.num_gpus = 4
# opt.log_path = 'C:/Users/giang/Desktop/result_temp/'
if params.num_gpus == 0:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
if torch.cuda.is_available():
torch.cuda.manual_seed(42)
else:
torch.manual_seed(42)
opt.saved_path = opt.saved_path + f'/{params.project_name}/'
opt.log_path = opt.log_path + f'/{params.project_name}/tensorboard/'
os.makedirs(opt.log_path, exist_ok=True)
os.makedirs(opt.saved_path, exist_ok=True)
training_params = {
'batch_size': opt.batch_size,
'shuffle': True,
'drop_last': True,
'collate_fn': collater,
'num_workers': opt.num_workers
}
val_params = {
'batch_size': opt.batch_size * 4,
'shuffle': False,
'drop_last': True,
'collate_fn': collater,
'num_workers': opt.num_workers
}
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
# root_train = 'D:/Etri_tracking_data/Etri_full/train_1024/'
# side_train = 'D:/Etri_tracking_data/Etri_full/train_Sejin_1024/'
# ground_truth_train = 'D:/Etri_tracking_data/Etri_full/train_1024.txt'
root_train = '/home/../../data3/giangData/train_1024/'
side_train = '/home/../../data3/giangData/train_Sejin_1024/'
ground_truth_train = '/home/../../data3/giangData/train_1024.txt'
training_set = TobyCustom(root_dir=root_train, side_dir = side_train, \
annot_path = ground_truth_train, \
transform=ComposeAlb([Flip_X(), \
Flip_Y(), \
Equalize(), \
Brightness(), \
Constrast(), \
Resizer(input_sizes[opt.compound_coef], num_channels=3), \
Normalizer(mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225])]))
training_generator = DataLoader(training_set, **training_params)
# root_val = 'D:/Etri_tracking_data/Etri_full/val_1024/'
# side_val = 'D:/Etri_tracking_data/Etri_full/val_Sejin_1024/'
# ground_truth_val = 'D:/Etri_tracking_data/Etri_full/val_1024.txt'
root_val = '/home/../../data3/giangData/val_1024/'
side_val = '/home/../../data3/giangData/val_Sejin_1024/'
ground_truth_val = '/home/../../data3/giangData/val_1024.txt'
val_set = TobyCustom(root_dir=root_val, side_dir = side_val, \
annot_path = ground_truth_val, \
transform=ComposeAlb([Resizer(input_sizes[opt.compound_coef], num_channels=3),
Normalizer(mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225])]))
val_generator = DataLoader(val_set, **val_params)
model = EfficientDetBackbone(num_classes=len(params.obj_list),
compound_coef=opt.compound_coef,
ratios=eval(params.anchors_ratios),
scales=eval(params.anchors_scales))
from efficientdet.model import Classifier
# model.backbone_net.model._conv_stem.conv = nn.Conv2d(4, 48, kernel_size=(3, 3), stride=(2, 2), bias=False)
# model.classifier.header.pointwise_conv.conv = nn.Conv2d(224, 9, kernel_size=(1, 1), stride=(1, 1))
model.classifier = Classifier(
in_channels=model.fpn_num_filters[opt.compound_coef],
num_anchors=model.num_anchors,
num_classes=1,
num_layers=model.box_class_repeats[opt.compound_coef],
pyramid_levels=model.pyramid_levels[opt.compound_coef])
# opt.load_weights = 'C:/Users/giang/Desktop/efficientdet-d4_107_15228_6.1788892433756875.pth'
opt.load_weights = './../result_3channel_21/save/coco/efficientdet-d4_21_3000.pth'
# block'
# for EfficientNetB5, please test again with B4
# load last weights
if opt.load_weights is not None:
if opt.load_weights.endswith('.pth'):
weights_path = opt.load_weights
else:
weights_path = get_last_weights(opt.saved_path)
try:
last_step = int(
os.path.basename(weights_path).split('_')[-1].split('.')[0])
except:
last_step = 0
try:
ret = model.load_state_dict(torch.load(weights_path), strict=False)
except RuntimeError as e:
print(f'[Warning] Ignoring {e}')
print(
'[Warning] Don\'t panic if you see this, this might be because you load a pretrained weights with different number of classes. The rest of the weights should be loaded already.'
)
print(
f'[Info] loaded weights: {os.path.basename(weights_path)}, resuming checkpoint from step: {last_step}'
)
else:
last_step = 0
print('[Info] initializing weights...')
init_weights(model)
'''
============================================
Modify model
'''
# from efficientdet.model import Classifier
# model.backbone_net.model._conv_stem.conv = nn.Conv2d(4, 48, kernel_size=(3, 3), stride=(2, 2), bias=False)
# model.classifier.header.pointwise_conv.conv = nn.Conv2d(224, 9, kernel_size=(1, 1), stride=(1, 1))
# model.classifier = Classifier(in_channels=model.fpn_num_filters[opt.compound_coef], num_anchors=model.num_anchors,
# num_classes=1,
# num_layers=model.box_class_repeats[opt.compound_coef],
# pyramid_levels=model.pyramid_levels[opt.compound_coef])
'''
=============================================
'''
# freeze backbone if train head_only
if opt.head_only:
def freeze_backbone(m):
classname = m.__class__.__name__
for ntl in ['EfficientNet', 'BiFPN']:
if ntl in classname:
for param in m.parameters():
param.requires_grad = False
model.apply(freeze_backbone)
print('[Info] freezed backbone')
# https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
# apply sync_bn when using multiple gpu and batch_size per gpu is lower than 4
# useful when gpu memory is limited.
# because when bn is disable, the training will be very unstable or slow to converge,
# apply sync_bn can solve it,
# by packing all mini-batch across all gpus as one batch and normalize, then send it back to all gpus.
# but it would also slow down the training by a little bit.
if params.num_gpus > 1 and opt.batch_size // params.num_gpus < 4:
model.apply(replace_w_sync_bn)
use_sync_bn = True
else:
use_sync_bn = False
writer = SummaryWriter(
opt.log_path +
f'/{datetime.datetime.now().strftime("%Y%m%d-%H%M%S")}/')
# warp the model with loss function, to reduce the memory usage on gpu0 and speedup
model = ModelWithLoss(model, debug=opt.debug)
if params.num_gpus > 0:
model = model.cuda()
if params.num_gpus > 1:
model = CustomDataParallel(model, params.num_gpus)
if use_sync_bn:
patch_replication_callback(model)
if opt.optim == 'adamw':
optimizer = torch.optim.AdamW(model.parameters(), opt.lr)
else:
optimizer = torch.optim.SGD(model.parameters(),
opt.lr,
momentum=0.9,
nesterov=True)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
epoch = 0
best_loss = 1e5
best_epoch = 0
step = max(0, last_step)
model.train()
num_iter_per_epoch = len(training_generator)
try:
for epoch in range(opt.num_epochs):
last_epoch = step // num_iter_per_epoch
if epoch < last_epoch:
continue
epoch_loss = []
progress_bar = tqdm(training_generator)
for iter, data in enumerate(progress_bar):
if iter < step - last_epoch * num_iter_per_epoch:
progress_bar.update()
continue
try:
imgs = data['img']
annot = data['annot']
image_path = data['image_path']
# print(image_path)
if params.num_gpus == 1:
# if only one gpu, just send it to cuda:0
# elif multiple gpus, send it to multiple gpus in CustomDataParallel, not here
imgs = imgs.cuda()
annot = annot.cuda()
optimizer.zero_grad()
cls_loss, reg_loss = model(imgs,
annot,
obj_list=params.obj_list)
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss = cls_loss + reg_loss
if loss == 0 or not torch.isfinite(loss):
continue
loss.backward()
# torch.nn.utils.clip_grad_norm_(model.parameters(), 0.1)
optimizer.step()
epoch_loss.append(float(loss))
progress_bar.set_description(
'Step: {}. Epoch: {}/{}. Iteration: {}/{}. Cls loss: {:.5f}. Reg loss: {:.5f}. Total loss: {:.5f}'
.format(step, epoch, opt.num_epochs, iter + 1,
num_iter_per_epoch, cls_loss.item(),
reg_loss.item(), loss.item()))
writer.add_scalars('Loss', {'train': loss}, step)
writer.add_scalars('Regression_loss', {'train': reg_loss},
step)
writer.add_scalars('Classfication_loss',
{'train': cls_loss}, step)
# log learning_rate
current_lr = optimizer.param_groups[0]['lr']
writer.add_scalar('learning_rate', current_lr, step)
step += 1
if step % opt.save_interval == 0 and step > 0:
save_checkpoint(
model,
f'efficientdet-d{opt.compound_coef}_{epoch}_{step}.pth'
)
print('checkpoint...')
except Exception as e:
print('[Error]', traceback.format_exc())
print(e)
continue
scheduler.step(np.mean(epoch_loss))
if epoch % opt.val_interval == 0:
model.eval()
loss_regression_ls = []
loss_classification_ls = []
for iter, data in enumerate(val_generator):
with torch.no_grad():
imgs = data['img']
annot = data['annot']
if params.num_gpus == 1:
imgs = imgs.cuda()
annot = annot.cuda()
cls_loss, reg_loss = model(imgs,
annot,
obj_list=params.obj_list)
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss = cls_loss + reg_loss
if loss == 0 or not torch.isfinite(loss):
continue
loss_classification_ls.append(cls_loss.item())
loss_regression_ls.append(reg_loss.item())
cls_loss = np.mean(loss_classification_ls)
reg_loss = np.mean(loss_regression_ls)
loss = cls_loss + reg_loss
print(
'Val. Epoch: {}/{}. Classification loss: {:1.5f}. Regression loss: {:1.5f}. Total loss: {:1.5f}'
.format(epoch, opt.num_epochs, cls_loss, reg_loss, loss))
writer.add_scalars('Loss', {'val': loss}, step)
writer.add_scalars('Regression_loss', {'val': reg_loss}, step)
writer.add_scalars('Classfication_loss', {'val': cls_loss},
step)
print('\n')
if loss + opt.es_min_delta < best_loss:
best_loss = loss
best_epoch = epoch
save_loss = round(loss, 4)
save_checkpoint(
model,
f'efficientdet-d{opt.compound_coef}_{epoch}_{step}_{loss}.pth'
)
model.train()
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
'[Info] Stop training at epoch {}. The lowest loss achieved is {}'
.format(epoch, best_loss))
break
except KeyboardInterrupt:
save_checkpoint(
model, f'efficientdet-d{opt.compound_coef}_{epoch}_{step}.pth')
writer.close()
writer.close()
def test(opt):
compound_coef = 2
force_input_size = None # set None to use default size
img_id = opt.img_id
img_path = opt.img_path
img_path = img_path + str(img_id) + '.jpg'
# replace this part with your project's anchor config
anchor_ratios = [(1.0, 1.0), (1.4, 0.7), (0.7, 1.4)]
anchor_scales = [2**0, 2**(1.0 / 3.0), 2**(2.0 / 3.0)]
threshold = 0.2
iou_threshold = 0.2
use_cuda = True
use_float16 = False
cudnn.fastest = True
cudnn.benchmark = True
obj_list = ['02010001', '02010002']
color_list = standard_to_bgr(STANDARD_COLORS)
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
input_size = input_sizes[
compound_coef] if force_input_size is None else force_input_size
ori_imgs, framed_imgs, framed_metas = preprocess(img_path,
max_size=input_size)
if use_cuda:
x = torch.stack([torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32 if not use_float16 else torch.float16).permute(
0, 3, 1, 2)
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(obj_list),
ratios=anchor_ratios,
scales=anchor_scales)
model.load_state_dict(torch.load(opt.weights, map_location='cpu'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
with torch.no_grad():
features, regression, classification, anchors = model(x)
regressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()
out = postprocess(x, anchors, regression, classification, regressBoxes,
clipBoxes, threshold, iou_threshold)
def display(preds, imgs, imshow=True, imwrite=False, img_id=1):
for i in range(len(imgs)):
if len(preds[i]['rois']) == 0:
continue
imgs[i] = imgs[i].copy()
imgs[i] = cv2.cvtColor(imgs[i], cv2.COLOR_BGR2RGB)
for j in range(len(preds[i]['rois'])):
x1, y1, x2, y2 = preds[i]['rois'][j].astype(np.int)
obj = obj_list[preds[i]['class_ids'][j]]
score = float(preds[i]['scores'][j])
plot_one_box(imgs[i], [x1, y1, x2, y2],
label=obj,
score=score,
color=color_list[get_index_label(obj, obj_list)])
if imshow:
cv2.imshow('img', imgs[i])
cv2.waitKey(0)
if imwrite:
str1 = 'test/' + str(img_id) + '.jpg'
cv2.imwrite(str1, imgs[i])
out = invert_affine(framed_metas, out)
display(out, ori_imgs, imshow=False, imwrite=True, img_id=img_id)
print('running speed test...')
with torch.no_grad():
print('test1: model inferring and postprocessing')
print('inferring image for 10 times...')
t1 = time.time()
for _ in range(10):
_, regression, classification, anchors = model(x)
out = postprocess(x, anchors, regression, classification,
regressBoxes, clipBoxes, threshold,
iou_threshold)
out = invert_affine(framed_metas, out)
tempList = []
for j in range(len(out[0]['class_ids'])):
tempout = {}
tempout['image_id'] = img_id
if out[0]['class_ids'][j] == 1:
tempout['category_id'] = 2
else:
tempout['category_id'] = 1
tempout['score'] = out[0]['scores'][j].astype(np.float64)
tempout['bbox'] = [
(out[0]['rois'][j][0]).astype(np.float64),
(out[0]['rois'][j][1]).astype(np.float64),
(out[0]['rois'][j][2]).astype(np.float64) -
(out[0]['rois'][j][0]).astype(np.float64),
(out[0]['rois'][j][3]).astype(np.float64) -
(out[0]['rois'][j][1]).astype(np.float64),
]
tempList.append(tempout)
t2 = time.time()
tact_time = (t2 - t1) / 10
print(f'{tact_time} seconds, {1 / tact_time} FPS, @batch_size 1')
with open("test/" + str(img_id) + ".json", "w") as f:
json.dump(tempList, f)
print("生成标注后的图片(" + str(img_id) + ".jpg)和json(" + str(img_id) +
".json)到test文件夹中...")
def efficientDet_video_inference(video_src,compound_coef = 0,force_input_size=None,
frame_skipping = 3,
threshold=0.2,out_path=None,imshow=False,
display_fps=False):
#deep-sort variables
# Definition of the parameters
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
model_filename = '/home/shaheryar/Desktop/Projects/Football-Monitoring/deep_sort/model_weights/mars-small128.pb'
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = nn_matching.NearestNeighborDistanceMetric("cosine", max_cosine_distance, nn_budget)
tracker = Tracker(metric,n_init=5)
# efficientDet-pytorch variables
iou_threshold = 0.4
use_cuda = True
use_float16 = False
cudnn.fastest = True
cudnn.benchmark = True
input_size = input_sizes[compound_coef] if force_input_size is None else force_input_size
# load model
model = EfficientDetBackbone(compound_coef=compound_coef, num_classes=len(obj_list))
model.load_state_dict(torch.load(f'weights/efficientdet-d{compound_coef}.pth'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
regressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()
# Video capture
cap = cv2.VideoCapture(video_src)
frame_width = int(cap.get(3))
frame_height = int(cap.get(4))
fourcc = cv2.VideoWriter_fourcc(*'MPEG')
fps = cap.get(cv2.CAP_PROP_FPS)
print("Video fps",fps)
if(out_path is not None):
outp = cv2.VideoWriter(out_path, fourcc, fps, (frame_width, frame_height))
i=0
start= time.time()
current_frame_fps=0
while True:
ret, frame = cap.read()
if not ret:
break
t1=time.time()
if (frame_skipping==0 or i%frame_skipping==0):
# if(True):
# frame preprocessing (running detections)
ori_imgs, framed_imgs, framed_metas, t1 = preprocess_video(frame, width=input_size, height=input_size)
if use_cuda:
x = torch.stack([fi.cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
# model predict
t1=time.time()
with torch.no_grad():
features, regression, classification, anchors = model(x)
out = postprocess(x,
anchors, regression, classification,
regressBoxes, clipBoxes,
threshold, iou_threshold)
# Post processing
out = invert_affine(framed_metas, out)
# decoding bbox ,object name and scores
boxes,classes,scores =decode_predictions(out[0])
org_boxes = boxes.copy()
t2 = time.time() - t1
# feature extraction for deep sort
boxes = [convert_bbox_to_deep_sort_format(frame.shape, b) for b in boxes]
features = encoder(frame,boxes)
detections = [Detection(bbox, 1.0, feature) for bbox, feature in zip(boxes, features)]
boxes = np.array([d.tlwh for d in detections])
# print(boxes)
scores = np.array([d.confidence for d in detections])
indices = preprocessing.non_max_suppression(boxes, nms_max_overlap, scores)
detections = [detections[i] for i in indices]
tracker.predict()
tracker.update(detections)
i = i + 1
img_show=frame.copy()
for j in range(len(org_boxes)):
img_show =drawBoxes(img_show,org_boxes[j],(255,255,0),str(tracker.tracks[j].track_id))
for track in tracker.tracks:
if not track.is_confirmed() or track.time_since_update > 1:
continue
bbox = track.to_tlbr()
x1=int(bbox[0])
y1 = int(bbox[1])
x2 = int(bbox[2])
y2=int(bbox[3])
roi= frame[y1:y2,x1:x2]
cv2.rectangle(img_show, (x1, y1), (x2, y2), update_color_association(roi, track.track_id), 2)
cv2.putText(img_show, str(track.track_id), (x1, y1), 0, 5e-3 * 100, (255, 255, 0), 1)
if display_fps:
current_frame_fps=1/t2
else:
current_frame_fps=0
cv2.putText(img_show, 'FPS: {0:.2f}'.format(current_frame_fps), (30, 50), cv2.FONT_HERSHEY_SIMPLEX, 1,
(255, 255, 0),
2, cv2.LINE_AA)
if (i % int(fps) == 0):
print("Processed ", str(int(i / fps)), "seconds")
print("Time taken",time.time()-start)
# print(color_dict)
if imshow:
img_show=cv2.resize(img_show,(0,0),fx=0.75,fy=0.75)
cv2.imshow('Frame',img_show)
# Press Q on keyboard to exit
if cv2.waitKey(1) & 0xFF == ord('q'):
break
if out_path is not None:
outp.write(img_show)
cap.release()
outp.release()
def EfficientDetNode():
rospy.init_node('efficient_det_node', anonymous=True)
rospy.Subscriber('input', String, image_callback, queue_size=1)
pub = rospy.Publisher('/image_detections', Detection2DArray, queue_size=10)
rate = rospy.Rate(1) # 10hz
path_list = os.listdir(path)
path_list.sort(key=lambda x: int(x.split('.')[0]))
stamp_file = open(stamp_path)
stamp_lines = stamp_file.readlines()
stamp_i = 0
for filename in path_list:
img_path = filename
cur_frame = img_path[:-4]
img_path = path + "/" + img_path
cur_stamp = ((float)(stamp_lines[stamp_i][-13:].strip('\n')))
# cur_stamp = rospy.Time.from_sec(
# ((float)(stamp_lines[stamp_i][-13:].strip('\n'))))
stamp_i += 1
detection_results = Detection2DArray()
# tf bilinear interpolation is different from any other's, just make do
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
input_size = input_sizes[
compound_coef] if force_input_size is None else force_input_size
ori_imgs, framed_imgs, framed_metas = preprocess(img_path,
max_size=input_size)
if use_cuda:
x = torch.stack(
[torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32 if not use_float16 else torch.float16).permute(
0, 3, 1, 2)
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(obj_list),
ratios=anchor_ratios,
scales=anchor_scales)
model.load_state_dict(
torch.load(f'weights/efficientdet-d{compound_coef}.pth',
map_location='cpu'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
with torch.no_grad():
features, regression, classification, anchors = model(x)
regressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()
out = postprocess(x, anchors, regression, classification,
regressBoxes, clipBoxes, threshold,
iou_threshold)
out = invert_affine(framed_metas, out)
display(cur_frame, out, ori_imgs, imshow=False, imwrite=True)
for i in range(len(out)):
for j in range(len(out[i]['rois'])):
x1, y1, x2, y2 = out[i]['rois'][j].astype(np.int)
obj = obj_list[out[i]['class_ids'][j]]
score = float(out[i]['scores'][j])
result = ObjectHypothesisWithPose()
result.score = score
if (obj == 'car'):
result.id = 0
if (obj == 'person'):
result.id = 1
if (obj == 'cyclist'):
result.id = 2
detection_msg = Detection2D()
detection_msg.bbox.center.x = (x1 + x2) / 2
detection_msg.bbox.center.y = (y1 + y2) / 2
detection_msg.bbox.size_x = x2 - x1
detection_msg.bbox.size_y = y2 - y1
detection_msg.results.append(result)
detection_results.detections.append(detection_msg)
rospy.loginfo("%d: %lf", detection_msg.results[0].id,
detection_msg.results[0].score)
detection_results.header.seq = cur_frame
#detection_results.header.stamp = cur_stamp
rospy.loginfo(detection_results.header.stamp)
pub.publish(detection_results)
if not os.path.exists(txt_path):
os.makedirs(txt_path)
#with open(f'txt/{cur_frame}.txt', 'w') as f:
with open(f'{txt_path}/{cur_frame}.txt', 'w') as f:
#f.write(str((float)(stamp_lines[stamp_i][-13:].strip('\n'))) + "\n")
f.write(str(cur_stamp) + "\n")
for detection in detection_results.detections:
f.write(str(detection.bbox.center.x) + " ")
f.write(str(detection.bbox.center.y) + " ")
f.write(str(detection.bbox.size_x) + " ")
f.write(str(detection.bbox.size_y) + " ")
f.write(str(detection.results[0].id) + " ")
f.write(str(detection.results[0].score) + "\n")
f.close()
rate.sleep()
print('running speed test...')
with torch.no_grad():
print('test1: model inferring and postprocessing')
print('inferring image for 10 times...')
t1 = time.time()
for _ in range(10):
_, regression, classification, anchors = model(x)
out = postprocess(x, anchors, regression, classification,
regressBoxes, clipBoxes, threshold,
iou_threshold)
out = invert_affine(framed_metas, out)
t2 = time.time()
tact_time = (t2 - t1) / 10
print(f'{tact_time} seconds, {1 / tact_time} FPS, @batch_size 1')
def getImageDetections(imagePath, weights, nms_threshold, confidenceParam, coefficient):
"""
Runs the detections and returns all detection into a single structure.
Parameters
----------
imagePath : str
Path to all images.
weights : str
path to the weights.
nms_threshold : float
non-maximum supression threshold.
confidenceParam : float
confidence score for the detections (everything above this threshold is considered a valid detection).
coefficient : int
coefficient of the current efficientdet model (from d1 to d7).
Returns
-------
detectionsList : List
return a list with all predicted bounding-boxes.
"""
compound_coef = coefficient
force_input_size = None # set None to use default size
img_path = imagePath
threshold = confidenceParam
iou_threshold = nms_threshold
use_cuda = True
use_float16 = False
cudnn.fastest = True
cudnn.benchmark = True
obj_list = ['class_name']
# tf bilinear interpolation is different from any other's, just make do
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
input_size = input_sizes[compound_coef] if force_input_size is None else force_input_size
ori_imgs, framed_imgs, framed_metas = preprocess(img_path, max_size=input_size)
if use_cuda:
x = torch.stack([torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32 if not use_float16 else torch.float16).permute(0, 3, 1, 2)
model = EfficientDetBackbone(compound_coef=compound_coef, num_classes=len(obj_list),
# replace this part with your project's anchor config
ratios=[(1.0, 1.0), (1.4, 0.7), (0.7, 1.4)],
scales=[2 ** 0, 2 ** (1.0 / 3.0), 2 ** (2.0 / 3.0)])
model.load_state_dict(torch.load(rootDir+'logs/' + project + '/' + weights))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
with torch.no_grad():
features, regression, classification, anchors = model(x)
regressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()
out = postprocess(x,
anchors, regression, classification,
regressBoxes, clipBoxes,
threshold, iou_threshold)
out = invert_affine(framed_metas, out)
for i in range(len(ori_imgs)):
if len(out[i]['rois']) == 0:
continue
detectionsList = []
for j in range(len(out[i]['rois'])):
(x1, y1, x2, y2) = out[i]['rois'][j].astype(np.int)
detectionsList.append((float(out[i]['scores'][j]), x1, y1, x2, y2))
return detectionsList
return self.params.get(item, None)
params = Params(f'projects/{project}_crop.yml')
obj_list = params.obj_list
obj_list.sort()
print(obj_list)
color_list = standard_to_bgr(STANDARD_COLORS)
# tf bilinear interpolation is different from any other's, just make do
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
input_size = input_sizes[
compound_coef] if force_input_size is None else force_input_size
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(obj_list),
ratios=anchor_ratios,
scales=anchor_scales)
# model.load_state_dict(torch.load(f'logs/{project}/efficientdet-d{compound_coef}_{number}.pth', map_location='cpu'))
# model.load_state_dict(torch.load(f'/data/efdet/logs/{project}/efficientdet-d{compound_coef}_{number}.pth', map_location='cpu'))
model.load_state_dict(
torch.load(
f'/data/efdet/logs/{project}/crop/weights/{save_time}/efficientdet-d{compound_coef}_{number}.pth',
map_location='cpu'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
def train(opt):
params = Params(f'projects/{opt.project}.yml')
if params.num_gpus == 0:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
if torch.cuda.is_available():
torch.cuda.manual_seed(42)
else:
torch.manual_seed(42)
opt.saved_path = opt.saved_path + f'/{params.project_name}/'
opt.log_path = opt.log_path + f'/{params.project_name}/tensorboard/'
os.makedirs(opt.log_path, exist_ok=True)
os.makedirs(opt.saved_path, exist_ok=True)
training_params = {'batch_size': opt.batch_size,
'shuffle': True,
'drop_last': True,
'collate_fn': collater,
'num_workers': opt.num_workers}
val_params = {'batch_size': opt.batch_size,
'shuffle': False,
'drop_last': True,
'collate_fn': collater,
'num_workers': opt.num_workers}
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
training_set = CocoDataset(root_dir=os.path.join(opt.data_path, params.project_name), set=params.train_set,
transform=transforms.Compose([Normalizer(mean=params.mean, std=params.std),
Augmenter(),
Resizer(input_sizes[opt.compound_coef])]))
training_generator = DataLoader(training_set, **training_params)
val_set = CocoDataset(root_dir=os.path.join(opt.data_path, params.project_name), set=params.val_set,
transform=transforms.Compose([Normalizer(mean=params.mean, std=params.std),
Resizer(input_sizes[opt.compound_coef])]))
val_generator = DataLoader(val_set, **val_params)
model = EfficientDetBackbone(num_classes=len(params.obj_list), compound_coef=opt.compound_coef,
ratios=eval(params.anchors_ratios), scales=eval(params.anchors_scales))
# load last weights
if opt.load_weights is not None:
if opt.load_weights.endswith('.pth'):
weights_path = opt.load_weights
else:
weights_path = get_last_weights(opt.saved_path)
try:
last_step = int(os.path.basename(weights_path).split('_')[-1].split('.')[0])
except:
last_step = 0
try:
ret = model.load_state_dict(torch.load(weights_path), strict=False)
except RuntimeError as e:
print(f'[Warning] Ignoring {e}')
print(
'[Warning] Don\'t panic if you see this, this might be because you load a pretrained weights with different number of classes. The rest of the weights should be loaded already.')
print(f'[Info] loaded weights: {os.path.basename(weights_path)}, resuming checkpoint from step: {last_step}')
else:
last_step = 0
print('[Info] initializing weights...')
init_weights(model)
# freeze backbone if train head_only
if opt.head_only:
def freeze_backbone(m):
classname = m.__class__.__name__
for ntl in ['EfficientNet', 'BiFPN']:
if ntl in classname:
for param in m.parameters():
param.requires_grad = False
model.apply(freeze_backbone)
print('[Info] freezed backbone')
# https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
# apply sync_bn when using multiple gpu and batch_size per gpu is lower than 4
# useful when gpu memory is limited.
# because when bn is disable, the training will be very unstable or slow to converge,
# apply sync_bn can solve it,
# by packing all mini-batch across all gpus as one batch and normalize, then send it back to all gpus.
# but it would also slow down the training by a little bit.
if params.num_gpus > 1 and opt.batch_size // params.num_gpus < 4:
model.apply(replace_w_sync_bn)
use_sync_bn = True
else:
use_sync_bn = False
writer = SummaryWriter(opt.log_path + f'/{datetime.datetime.now().strftime("%Y%m%d-%H%M%S")}/')
# warp the model with loss function, to reduce the memory usage on gpu0 and speedup
model = ModelWithLoss(model, debug=opt.debug)
if params.num_gpus > 0:
model = model.cuda()
if params.num_gpus > 1:
model = CustomDataParallel(model, params.num_gpus)
if use_sync_bn:
patch_replication_callback(model)
if opt.optim == 'adamw':
optimizer = torch.optim.AdamW(model.parameters(), opt.lr)
else:
optimizer = torch.optim.SGD(model.parameters(), opt.lr, momentum=0.9, nesterov=True)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=3, verbose=True)
epoch = 0
best_loss = 1e5
best_epoch = 0
step = max(0, last_step)
model.train()
num_iter_per_epoch = len(training_generator)
try:
for epoch in range(opt.num_epochs):
last_epoch = step // num_iter_per_epoch
if epoch < last_epoch:
continue
epoch_loss = []
progress_bar = tqdm(training_generator)
for iter, data in enumerate(progress_bar):
if iter < step - last_epoch * num_iter_per_epoch:
progress_bar.update()
continue
try:
imgs = data['img']
annot = data['annot']
if params.num_gpus == 1:
# if only one gpu, just send it to cuda:0
# elif multiple gpus, send it to multiple gpus in CustomDataParallel, not here
imgs = imgs.cuda()
annot = annot.cuda()
optimizer.zero_grad()
cls_loss, reg_loss = model(imgs, annot, obj_list=params.obj_list)
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss = cls_loss + reg_loss
if loss == 0 or not torch.isfinite(loss):
continue
loss.backward()
# torch.nn.utils.clip_grad_norm_(model.parameters(), 0.1)
optimizer.step()
epoch_loss.append(float(loss))
progress_bar.set_description(
'Step: {}. Epoch: {}/{}. Iteration: {}/{}. Cls loss: {:.5f}. Reg loss: {:.5f}. Total loss: {:.5f}'.format(
step, epoch, opt.num_epochs, iter + 1, num_iter_per_epoch, cls_loss.item(),
reg_loss.item(), loss.item()))
writer.add_scalars('Loss', {'train': loss}, step)
writer.add_scalars('Regression_loss', {'train': reg_loss}, step)
writer.add_scalars('Classfication_loss', {'train': cls_loss}, step)
# log learning_rate
current_lr = optimizer.param_groups[0]['lr']
writer.add_scalar('learning_rate', current_lr, step)
step += 1
#if step % opt.save_interval == 0 and step > 0:
if epoch % opt.save_interval == 0 :
#save_checkpoint(model, f'efficientdet-d{opt.compound_coef}_{epoch}_{step}.pth')
save_checkpoint(model, f'efficientdet-d{opt.compound_coef}_{epoch}.pth')
print('checkpoint...')
except Exception as e:
print('[Error]', traceback.format_exc())
print(e)
continue
scheduler.step(np.mean(epoch_loss))
if epoch % opt.val_interval == 0:
model.eval()
loss_regression_ls = []
loss_classification_ls = []
for iter, data in enumerate(val_generator):
with torch.no_grad():
imgs = data['img']
annot = data['annot']
if params.num_gpus == 1:
imgs = imgs.cuda()
annot = annot.cuda()
cls_loss, reg_loss = model(imgs, annot, obj_list=params.obj_list)
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss = cls_loss + reg_loss
if loss == 0 or not torch.isfinite(loss):
continue
loss_classification_ls.append(cls_loss.item())
loss_regression_ls.append(reg_loss.item())
cls_loss = np.mean(loss_classification_ls)
reg_loss = np.mean(loss_regression_ls)
loss = cls_loss + reg_loss
print(
'Val. Epoch: {}/{}. Classification loss: {:1.5f}. Regression loss: {:1.5f}. Total loss: {:1.5f}'.format(
epoch, opt.num_epochs, cls_loss, reg_loss, loss))
writer.add_scalars('Loss', {'val': loss}, step)
writer.add_scalars('Regression_loss', {'val': reg_loss}, step)
writer.add_scalars('Classfication_loss', {'val': cls_loss}, step)
if loss + opt.es_min_delta < best_loss:
best_loss = loss
best_epoch = epoch
save_checkpoint(model, f'efficientdet-d{opt.compound_coef}_best_{epoch}.pth')
model.train()
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print('[Info] Stop training at epoch {}. The lowest loss achieved is {}'.format(epoch, best_loss))
break
except KeyboardInterrupt:
save_checkpoint(model, f'efficientdet-d{opt.compound_coef}_interrupted_{epoch}_{step}.pth')
writer.close()
writer.close()
def train(opt):
params = Params(f'projects/{opt.project}.yml')
if opt.project == "vcoco":
num_obj_class = 90
num_union_action = 25
num_inst_action = 51
else:
assert opt.project == "hico-det"
num_obj_class = 90
num_union_action = 117
num_inst_action = 234
if params.num_gpus == 0:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
opt.saved_path = opt.saved_path + f'/{params.project_name}/'
opt.log_path = opt.log_path + f'/{params.project_name}/tensorboard/'
os.makedirs(opt.log_path, exist_ok=True)
os.makedirs(opt.saved_path, exist_ok=True)
training_params = {
'batch_size': opt.batch_size,
'shuffle': True,
'drop_last': True,
'collate_fn': collater,
'num_workers': opt.num_workers,
'pin_memory': False
}
val_params = {
'batch_size': opt.batch_size * 2,
'shuffle': False,
'drop_last': True,
'collate_fn': collater,
'num_workers': opt.num_workers,
'pin_memory': False
}
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
train_transform = transforms.Compose([
Normalizer(mean=params.mean, std=params.std),
Augmenter(),
Resizer(input_sizes[opt.compound_coef])
])
val_transform = transforms.Compose([
Normalizer(mean=params.mean, std=params.std),
Resizer(input_sizes[opt.compound_coef])
])
if opt.project == "vcoco":
# training_set = VCOCO_Dataset(root_dir="./datasets/vcoco", set=params.train_set, color_prob=1,
# transform=train_transform)
# val_set = VCOCO_Dataset(root_dir="./datasets/vcoco", set=params.val_set,
# transform=val_transform)
exit(-999)
else:
training_set = HICO_DET_Dataset(root_dir="data/hico_20160224_det",
set="train",
color_prob=1,
transform=train_transform)
val_set = HICO_DET_Dataset(root_dir="data/hico_20160224_det",
set="test",
transform=val_transform)
training_generator = DataLoader(training_set, **training_params)
val_generator = DataLoader(val_set, **val_params)
model = EfficientDetBackbone(num_classes=num_obj_class,
num_union_classes=num_union_action,
num_inst_classes=num_inst_action,
compound_coef=opt.compound_coef,
ratios=eval(params.anchors_ratios),
scales=eval(params.anchors_scales))
model.train()
print("num_classes:", num_obj_class)
print("num_union_classes:", num_union_action)
print("instance_action_list", num_inst_action)
# load last weights
if opt.load_weights is not None:
if opt.load_weights.endswith('.pth'):
weights_path = opt.load_weights
else:
weights_path = get_last_weights(opt.saved_path)
try:
# last_step = int(os.path.basename(weights_path).split('_')[-1].split('.')[0])
# last_epoch = int(os.path.basename(weights_path).split('_')[-2].split('.')[0]) + 1
last_epoch = int(
os.path.basename(weights_path).split('_')[-1].split('.')[0])
last_step = last_epoch * len(training_generator)
except:
last_step = 0
try:
init_weights(model)
print(weights_path)
model_dict = model.state_dict()
pretrained_dict = torch.load(weights_path,
map_location=torch.device('cpu'))
new_pretrained_dict = {}
for k, v in pretrained_dict.items():
if k in model_dict:
new_pretrained_dict[k] = v
elif ("instance_branch.object_" + k) in model_dict:
new_pretrained_dict["instance_branch.object_" + k] = v
# print("instance_branch.object_"+k)
ret = model.load_state_dict(new_pretrained_dict, strict=False)
except RuntimeError as e:
print(f'[Warning] Ignoring {e}')
print(
'[Warning] Don\'t panic if you see this, this might be because you load a pretrained weights with different number of classes. The rest of the weights should be loaded already.'
)
print(
f'[Info] loaded weights: {os.path.basename(weights_path)}, resuming checkpoint from step: {last_step}'
)
else:
last_step = 0
print('[Info] initializing weights...')
init_weights(model)
# freeze backbone if train head_only
if opt.head_only:
model.apply(freeze_backbone)
freeze_bn_backbone(model)
print('[Info] freezed backbone')
if opt.freeze_object_detection:
freeze_object_detection(model)
freeze_bn_object_detection(model)
# model.apply(freeze_object_detection)
print('[Info] freezed object detection branch')
# https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
# apply sync_bn when using multiple gpu and batch_size per gpu is lower than 4
# useful when gpu memory is limited.
# because when bn is disable, the training will be very unstable or slow to converge,
# apply sync_bn can solve it,
# by packing all mini-batch across all gpus as one batch and normalize, then send it back to all gpus.
# but it would also slow down the training by a little bit.
if params.num_gpus > 1 and opt.batch_size // params.num_gpus < 8:
model.apply(replace_w_sync_bn)
use_sync_bn = True
else:
use_sync_bn = False
# if os.path.exists('nohup.out'):
# os.remove('nohup.out')
# f = open('nohup.out', 'w')
# f.close()
if os.path.exists(opt.log_path):
import shutil
shutil.rmtree(opt.log_path)
os.makedirs(opt.log_path)
writer = SummaryWriter(
opt.log_path +
f'/{datetime.datetime.now().strftime("%Y%m%d-%H%M%S")}/')
# warp the model with loss function, to reduce the memory usage on gpu0 and speedup
model = ModelWithLoss(model, dataset=opt.project, debug=opt.debug)
if params.num_gpus > 0:
model = model.cuda()
if params.num_gpus > 1:
model = CustomDataParallel(model, params.num_gpus)
if use_sync_bn:
patch_replication_callback(model)
if opt.head_only:
print('[Info] freezed SyncBN backbone')
freeze_bn_backbone(model.module.model)
if opt.freeze_object_detection:
print('[Info] freezed SyncBN object detection')
freeze_bn_object_detection(model.module.model)
if opt.optim == 'adamw':
# optimizer = torch.optim.AdamW(model.parameters(), opt.lr)
optimizer = torch.optim.AdamW(
filter(lambda p: p.requires_grad, model.parameters()), opt.lr)
elif opt.optim == "adam":
# optimizer = torch.optim.Adam(model.parameters(), opt.lr)
optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, model.parameters()), opt.lr)
else:
# optimizer = torch.optim.SGD(model.parameters(), opt.lr, momentum=0.9, nesterov=True)
optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad,
model.parameters()),
opt.lr,
momentum=0.9,
nesterov=True)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=2,
verbose=True,
min_lr=1e-7)
epoch = 0
best_loss = 1e5
best_epoch = 0
step = max(0, last_step)
num_iter_per_epoch = (len(training_generator) + opt.accumulate_batch -
1) // opt.accumulate_batch
start_time = time.time()
try:
for epoch in range(opt.num_epochs):
last_epoch = step // num_iter_per_epoch + 1
if epoch < last_epoch:
continue
if epoch in [12, 16]:
optimizer.param_groups[0][
'lr'] = optimizer.param_groups[0]['lr'] / 10
epoch_loss = []
for iter, data in enumerate(training_generator):
try:
imgs = data['img']
annot = data['annot']
# torch.cuda.empty_cache()
if params.num_gpus == 1:
# if only one gpu, just send it to cuda:0
# elif multiple gpus, send it to multiple gpus in CustomDataParallel, not here
imgs = imgs.cuda()
for key in annot:
annot[key] = annot[key].cuda()
union_act_cls_loss, union_sub_reg_loss, union_obj_reg_loss, union_diff_reg_loss, \
inst_act_cls_loss, inst_obj_cls_loss, inst_obj_reg_loss = model(imgs, annot["instance"], annot["interaction"])
union_act_cls_loss = union_act_cls_loss.mean()
union_sub_reg_loss = union_sub_reg_loss.mean()
union_obj_reg_loss = union_obj_reg_loss.mean()
union_diff_reg_loss = union_diff_reg_loss.mean()
inst_act_cls_loss = inst_act_cls_loss.mean()
inst_obj_cls_loss = inst_obj_cls_loss.mean()
inst_obj_reg_loss = inst_obj_reg_loss.mean()
union_loss = union_act_cls_loss + union_sub_reg_loss + union_obj_reg_loss + union_diff_reg_loss
instance_loss = inst_act_cls_loss + inst_obj_cls_loss + inst_obj_reg_loss
loss = union_loss + inst_act_cls_loss
if loss == 0 or not torch.isfinite(loss):
continue
batch_loss = loss / opt.accumulate_batch
batch_loss.backward()
if (iter + 1) % opt.accumulate_batch == 0 or iter == len(
training_generator) - 1:
optimizer.step()
optimizer.zero_grad()
step += 1
loss = loss.item()
union_loss = union_loss.item()
instance_loss = instance_loss.item()
epoch_loss.append(float(loss))
current_lr = optimizer.param_groups[0]['lr']
if step % opt.log_interval == 0:
writer.add_scalars('Union Action Classification Loss',
{'train': union_act_cls_loss}, step)
writer.add_scalars('Union Subject Regression Loss',
{'train': union_sub_reg_loss}, step)
writer.add_scalars('Union Object Regression Loss',
{'train': union_obj_reg_loss}, step)
writer.add_scalars('Union Diff Regression Loss',
{'train': union_diff_reg_loss},
step)
writer.add_scalars(
'Instance Action Classification Loss',
{'train': inst_act_cls_loss}, step)
writer.add_scalars(
'Instance Object Classification Loss',
{'train': inst_obj_cls_loss}, step)
writer.add_scalars('Instance Regression Loss',
{'train': inst_obj_reg_loss}, step)
writer.add_scalars('Total Loss', {'train': loss}, step)
writer.add_scalars('Union Loss', {'train': union_loss},
step)
writer.add_scalars('Instance Loss',
{'train': instance_loss}, step)
# log learning_rate
writer.add_scalar('learning_rate', current_lr, step)
if iter % 20 == 0:
end_time = time.time()
print(
'Step: {}. Epoch: {}/{}. Iteration: {}/{}. Union loss: {:.5f}. Instance loss: {:.5f}. '
' Total loss: {:.5f}. Learning rate: {:.5f} Time: {:.2f}s'
.format(step, epoch, opt.num_epochs,
(iter + 1) // opt.accumulate_batch,
num_iter_per_epoch, union_loss,
instance_loss, loss, current_lr,
end_time - start_time))
start_time = time.time()
except Exception as e:
print('[Error]', traceback.format_exc())
print(e)
continue
# scheduler.step(np.mean(epoch_loss))
save_checkpoint(model,
f'efficientdet-d{opt.compound_coef}_{epoch}.pth')
print('checkpoint...')
if epoch % opt.val_interval == 0:
# model.eval()
union_loss_ls = []
instance_loss_ls = []
union_act_cls_loss_ls = []
union_obj_cls_loss_ls = []
union_act_reg_loss_ls = []
union_sub_reg_loss_ls = []
union_obj_reg_loss_ls = []
union_diff_reg_loss_ls = []
inst_act_cls_loss_ls = []
inst_obj_cls_loss_ls = []
inst_obj_reg_loss_ls = []
val_loss = []
for iter, data in enumerate(val_generator):
if (iter + 1) % 50 == 0:
print("%d/%d" % (iter + 1, len(val_generator)))
with torch.no_grad():
imgs = data['img']
annot = data['annot']
if params.num_gpus == 1:
imgs = imgs.cuda()
for key in annot:
annot[key] = annot[key].cuda()
union_act_cls_loss, union_sub_reg_loss, union_obj_reg_loss, union_diff_reg_loss, \
inst_act_cls_loss, inst_obj_cls_loss, inst_obj_reg_loss = model(imgs, annot["instance"], annot["interaction"])
union_act_cls_loss = union_act_cls_loss.mean()
union_sub_reg_loss = union_sub_reg_loss.mean()
union_obj_reg_loss = union_obj_reg_loss.mean()
union_diff_reg_loss = union_diff_reg_loss.mean()
inst_act_cls_loss = inst_act_cls_loss.mean()
inst_obj_cls_loss = inst_obj_cls_loss.mean()
inst_obj_reg_loss = inst_obj_reg_loss.mean()
union_loss = union_act_cls_loss + union_sub_reg_loss + union_obj_reg_loss + union_diff_reg_loss
instance_loss = inst_act_cls_loss + inst_obj_cls_loss + inst_obj_reg_loss
loss = union_loss + inst_act_cls_loss
if loss == 0 or not torch.isfinite(loss):
continue
val_loss.append(loss.item())
union_act_cls_loss_ls.append(union_act_cls_loss.item())
union_sub_reg_loss_ls.append(union_sub_reg_loss.item())
union_obj_reg_loss_ls.append(union_obj_reg_loss.item())
union_diff_reg_loss_ls.append(
union_diff_reg_loss.item())
# union_obj_cls_loss_ls.append(union_obj_cls_loss.item())
# union_act_reg_loss_ls.append(union_act_reg_loss.item())
inst_act_cls_loss_ls.append(inst_act_cls_loss.item())
inst_obj_cls_loss_ls.append(inst_obj_cls_loss.item())
inst_obj_reg_loss_ls.append(inst_obj_reg_loss.item())
union_loss_ls.append(union_loss.item())
instance_loss_ls.append(instance_loss.item())
union_loss = np.mean(union_loss_ls)
instance_loss = np.mean(instance_loss_ls)
union_act_cls_loss = np.mean(union_act_cls_loss_ls)
union_sub_reg_loss = np.mean(union_sub_reg_loss_ls)
union_obj_reg_loss = np.mean(union_obj_reg_loss_ls)
union_diff_reg_loss = np.mean(union_diff_reg_loss_ls)
inst_act_cls_loss = np.mean(inst_act_cls_loss_ls)
inst_obj_cls_loss = np.mean(inst_obj_cls_loss_ls)
inst_obj_reg_loss = np.mean(inst_obj_reg_loss_ls)
loss = union_loss + inst_act_cls_loss
print(
'Val. Epoch: {}/{}. Union loss: {:1.5f}. Instance loss: {:1.5f}. '
'Total loss: {:1.5f}'.format(epoch, opt.num_epochs,
union_loss, instance_loss,
loss))
writer.add_scalars('Union Action Classification Loss',
{'val': union_act_cls_loss}, step)
writer.add_scalars('Union Subject Regression Loss',
{'val': union_sub_reg_loss}, step)
writer.add_scalars('Union Object Regression Loss',
{'val': union_obj_reg_loss}, step)
writer.add_scalars('Union Diff Regression Loss',
{'val': union_diff_reg_loss}, step)
writer.add_scalars('Instance Action Classification Loss',
{'val': inst_act_cls_loss}, step)
writer.add_scalars('Instance Object Classification Loss',
{'val': inst_obj_cls_loss}, step)
writer.add_scalars('Instance Regression Loss',
{'val': inst_obj_reg_loss}, step)
writer.add_scalars('Total Loss', {'val': loss}, step)
writer.add_scalars('Union Loss', {'val': union_loss}, step)
writer.add_scalars('Instance Loss', {'val': instance_loss},
step)
if loss + opt.es_min_delta < best_loss:
best_loss = loss
best_epoch = epoch
save_checkpoint(
model,
f'efficientdet-d{opt.compound_coef}_{best_epoch}_best.pth'
)
# model.train()
# scheduler.step()
scheduler.step(np.mean(val_loss))
if optimizer.param_groups[0]['lr'] < opt.lr / 100:
break
# Early stopping
# if epoch - best_epoch > opt.es_patience > 0:
# print('[Info] Stop training at epoch {}. The lowest loss achieved is {}'.format(epoch, loss))
# break
except KeyboardInterrupt:
# save_checkpoint(model, f'efficientdet-d{opt.compound_coef}_{epoch}_{step}.pth')
writer.close()
writer.close()
def model_fn(model_dir, compound_coef=0, use_cuda=False, use_float16=False):
# based entirely off of
# https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch/blob/master/coco_eval.py
print(f'building and loading efficientdet d{compound_coef}')
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(PARAMS['obj_list']),
ratios=eval(PARAMS['anchors_ratios']),
scales=eval(PARAMS['anchors_scales']))
state_dict = torch.hub.load_state_dict_from_url(
url=get_weights_url(c=compound_coef),
model_dir=model_dir,
map_location=torch.device('cpu'))
model.load_state_dict(state_dict)
model.requires_grad_(False)
model.eval()
if use_cuda:
model.cuda(0)
if use_float16:
model.half()
return model
color_list = standard_to_bgr(STANDARD_COLORS)
# tf bilinear interpolation is different from any other's, just make do
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536, 1536]
input_size = input_sizes[compound_coef] if force_input_size is None else force_input_size
ori_imgs, framed_imgs, framed_metas = preprocess(img_path, mean=[0.538, 0.474, 0.442], std=[0.301, 0.294, 0.291], max_size=input_size)
if use_cuda:
x = torch.stack([torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32 if not use_float16 else torch.float16).permute(0, 3, 1, 2)
model = EfficientDetBackbone(compound_coef=compound_coef, num_classes=len(obj_list),
ratios=anchor_ratios, scales=anchor_scales)
# model.load_state_dict(torch.load(f'weights/efficientdet-d{compound_coef}.pth', map_location='cpu'))
model.load_state_dict(torch.load(f'weights/efficientdet-d4_460_0.3313012893001238.pth', map_location='cpu'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
with torch.no_grad():
features, regression, classification, anchors = model(x)
regressBoxes = BBoxTransform()
def train(opt):
'''
Input: get_args()
Function: Train the model.
'''
params = Params(f'projects/{opt.project}.yml')
if params.num_gpus == 0:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
if torch.cuda.is_available():
torch.cuda.manual_seed(42)
else:
torch.manual_seed(42)
opt.saved_path = opt.saved_path + f'/{params.project_name}/'
opt.log_path = opt.log_path + f'/{params.project_name}/tensorboard/'
os.makedirs(opt.log_path, exist_ok=True)
os.makedirs(opt.saved_path, exist_ok=True)
# evaluation json file
pred_folder = f'{OPT.data_path}/{OPT.project}/predictions'
os.makedirs(pred_folder, exist_ok=True)
evaluation_pred_file = f'{pred_folder}/instances_bbox_results.json'
training_params = {
'batch_size': opt.batch_size,
'shuffle': True,
'drop_last': True,
'collate_fn': collater,
'num_workers': opt.num_workers
}
val_params = {
'batch_size': opt.batch_size,
'shuffle': True,
'drop_last': True,
'collate_fn': collater,
'num_workers': opt.num_workers
}
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
training_set = CocoDataset(root_dir=os.path.join(opt.data_path,
params.project_name),
set=params.train_set,
transform=torchvision.transforms.Compose([
Normalizer(mean=params.mean,
std=params.std),
Augmenter(),
Resizer(input_sizes[opt.compound_coef])
]))
training_generator = DataLoader(training_set, **training_params)
val_set = CocoDataset(root_dir=os.path.join(opt.data_path,
params.project_name),
set=params.val_set,
transform=torchvision.transforms.Compose([
Normalizer(mean=params.mean, std=params.std),
Resizer(input_sizes[opt.compound_coef])
]))
val_generator = DataLoader(val_set, **val_params)
model = EfficientDetBackbone(num_classes=len(params.obj_list),
compound_coef=opt.compound_coef,
ratios=eval(params.anchors_ratios),
scales=eval(params.anchors_scales))
# load last weights
if opt.load_weights is not None:
if opt.load_weights.endswith('.pth'):
weights_path = opt.load_weights
else:
weights_path = get_last_weights(opt.saved_path)
try:
last_step = int(
os.path.basename(weights_path).split('_')[-1].split('.')[0])
except Exception as exception:
last_step = 0
try:
_ = model.load_state_dict(torch.load(weights_path), strict=False)
except RuntimeError as rerror:
print(f'[Warning] Ignoring {rerror}')
print('[Warning] Don\'t panic if you see this, '\
'this might be because you load a pretrained weights with different number of classes.'\
' The rest of the weights should be loaded already.')
print(
f'[Info] loaded weights: {os.path.basename(weights_path)}, resuming checkpoint from step: {last_step}'
)
else:
last_step = 0
print('[Info] initializing weights...')
init_weights(model)
# freeze backbone if train head_only
if opt.head_only:
def freeze_backbone(mdl):
classname = mdl.__class__.__name__
for ntl in ['EfficientNet', 'BiFPN']:
if ntl in classname:
for param in mdl.parameters():
param.requires_grad = False
model.apply(freeze_backbone)
print('[Info] freezed backbone')
# https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
# apply sync_bn when using multiple gpu and batch_size per gpu is lower than 4
# useful when gpu memory is limited.
# because when bn is disable, the training will be very unstable or slow to converge,
# apply sync_bn can solve it,
# by packing all mini-batch across all gpus as one batch and normalize, then send it back to all gpus.
# but it would also slow down the training by a little bit.
if params.num_gpus > 1 and opt.batch_size // params.num_gpus < 4:
model.apply(replace_w_sync_bn)
use_sync_bn = True
else:
use_sync_bn = False
writer = SummaryWriter(
opt.log_path +
f'/{datetime.datetime.now().strftime("%Y%m%d-%H%M%S")}/')
# warp the model with loss function, to reduce the memory usage on gpu0 and speedup
model = ModelWithLoss(model, debug=opt.debug)
if params.num_gpus > 0:
model = model.cuda()
if params.num_gpus > 1:
model = CustomDataParallel(model, params.num_gpus)
if use_sync_bn:
patch_replication_callback(model)
if opt.optim == 'adamw':
optimizer = torch.optim.AdamW(model.parameters(), opt.lr)
else:
optimizer = torch.optim.SGD(model.parameters(),
opt.lr,
momentum=0.9,
nesterov=True)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
epoch = 0
best_loss = 1e5
best_epoch = 0
step = max(0, last_step)
model.train()
num_iter_per_epoch = len(training_generator)
num_val_iter_per_epoch = len(val_generator)
# Limit the no.of preds to #images in val.
# Here, I averaged the #obj to 5 for computational efficacy
if opt.max_preds_toeval > 0:
opt.max_preds_toeval = len(val_generator) * opt.batch_size * 5
try:
for epoch in range(opt.num_epochs):
last_epoch = step // num_iter_per_epoch
if epoch < last_epoch:
continue
epoch_loss = []
progress_bar = tqdm(training_generator)
for iternum, data in enumerate(progress_bar):
if iternum < step - last_epoch * num_iter_per_epoch:
progress_bar.update()
continue
try:
imgs = data['img']
annot = data['annot']
if params.num_gpus == 1:
# if only one gpu, just send it to cuda:0
# elif multiple gpus, send it to multiple gpus in CustomDataParallel, not here
imgs = imgs.cuda()
annot = annot.cuda()
optimizer.zero_grad()
if iternum % int(num_iter_per_epoch *
(opt.eval_percent_epoch / 100)) != 0:
model.debug = False
cls_loss, reg_loss, _ = model(imgs,
annot,
obj_list=params.obj_list)
else:
model.debug = True
cls_loss, reg_loss, imgs_labelled = model(
imgs, annot, obj_list=params.obj_list)
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss = cls_loss + reg_loss
if loss == 0 or not torch.isfinite(loss):
continue
loss.backward()
# torch.nn.utils.clip_grad_norm_(model.parameters(), 0.1)
optimizer.step()
epoch_loss.append(float(loss))
progress_bar.set_description(
'Step: {}. Epoch: {}/{}. Iteration: {}/{}. Cls loss: {:.5f}. Reg loss: {:.5f}. Total loss: {:.5f}'
.format(step, epoch, opt.num_epochs, iternum + 1,
num_iter_per_epoch, cls_loss.item(),
reg_loss.item(), loss.item()))
writer.add_scalars('Loss', {'train': loss}, step)
writer.add_scalars('Regression_loss', {'train': reg_loss},
step)
writer.add_scalars('Classfication_loss',
{'train': cls_loss}, step)
if iternum % int(
num_iter_per_epoch *
(opt.eval_percent_epoch / 100)) == 0 and step > 0:
# create grid of images
imgs_labelled = np.asarray(imgs_labelled)
imgs_labelled = torch.from_numpy(
imgs_labelled) # (N, H, W, C)
imgs_labelled.transpose_(1, 3) # (N, C, H, W)
imgs_labelled.transpose_(2, 3)
img_grid = torchvision.utils.make_grid(imgs_labelled)
# write to tensorboard
writer.add_image('Training_images',
img_grid,
global_step=step)
#########################################################start EVAL#####################################################
model.eval()
model.debug = False # Don't print images in tensorboard now.
# remove json
if os.path.exists(evaluation_pred_file):
os.remove(evaluation_pred_file)
loss_regression_ls = []
loss_classification_ls = []
model.evalresults = [
] # Empty the results for next evaluation.
imgs_to_viz = []
num_validation_steps = int(
num_val_iter_per_epoch *
(opt.eval_sampling_percent / 100))
for valiternum, valdata in enumerate(val_generator):
with torch.no_grad():
imgs = valdata['img']
annot = valdata['annot']
resizing_imgs_scales = valdata['scale']
new_ws = valdata['new_w']
new_hs = valdata['new_h']
imgs_ids = valdata['img_id']
if params.num_gpus >= 1:
imgs = imgs.cuda()
annot = annot.cuda()
if valiternum % (num_validation_steps //
(opt.num_visualize_images //
opt.batch_size)) != 0:
model.debug = False
cls_loss, reg_loss, _ = model(
imgs,
annot,
obj_list=params.obj_list,
resizing_imgs_scales=
resizing_imgs_scales,
new_ws=new_ws,
new_hs=new_hs,
imgs_ids=imgs_ids)
else:
model.debug = True
cls_loss, reg_loss, val_imgs_labelled = model(
imgs,
annot,
obj_list=params.obj_list,
resizing_imgs_scales=
resizing_imgs_scales,
new_ws=new_ws,
new_hs=new_hs,
imgs_ids=imgs_ids)
imgs_to_viz += list(val_imgs_labelled)
loss_classification_ls.append(cls_loss.item())
loss_regression_ls.append(reg_loss.item())
if valiternum > (num_validation_steps):
break
cls_loss = np.mean(loss_classification_ls)
reg_loss = np.mean(loss_regression_ls)
loss = cls_loss + reg_loss
print(
'Val. Epoch: {}/{}. Classification loss: {:1.5f}. Regression loss: {:1.5f}. Total loss: {:1.5f}'
.format(epoch, opt.num_epochs, cls_loss, reg_loss,
loss))
writer.add_scalars('Loss', {'val': loss}, step)
writer.add_scalars('Regression_loss',
{'val': reg_loss}, step)
writer.add_scalars('Classfication_loss',
{'val': cls_loss}, step)
# create grid of images
val_imgs_labelled = np.asarray(imgs_to_viz)
val_imgs_labelled = torch.from_numpy(
val_imgs_labelled) # (N, H, W, C)
val_imgs_labelled.transpose_(1, 3) # (N, C, H, W)
val_imgs_labelled.transpose_(2, 3)
val_img_grid = torchvision.utils.make_grid(
val_imgs_labelled, nrow=2)
# write to tensorboard
writer.add_image('Eval_Images', val_img_grid, \
global_step=(step))
if opt.max_preds_toeval > 0:
json.dump(model.evalresults,
open(evaluation_pred_file, 'w'),
indent=4)
try:
val_results = calc_mAP_fin(params.project_name,\
params.val_set, evaluation_pred_file, \
val_gt=f'{OPT.data_path}/{OPT.project}/annotations/instances_{params.val_set}.json')
for catgname in val_results:
metricname = 'Average Precision (AP) @[ IoU = 0.50 | area = all | maxDets = 100 ]'
evalscore = val_results[catgname][
metricname]
writer.add_scalars(
f'mAP@IoU=0.5 and area=all',
{f'{catgname}': evalscore}, step)
except Exception as exption:
print("Unable to perform evaluation", exption)
if loss + opt.es_min_delta < best_loss:
best_loss = loss
best_epoch = epoch
save_checkpoint(
model,
f'efficientdet-d{opt.compound_coef}_{epoch}_{step}.pth'
)
model.train()
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
'[Info] Stop training at epoch {}. The lowest loss achieved is {}'
.format(epoch, best_loss))
break
#########################################################EVAL#####################################################
# log learning_rate
current_lr = optimizer.param_groups[0]['lr']
writer.add_scalar('learning_rate', current_lr, step)
step += 1
if step % opt.save_interval == 0 and step > 0:
save_checkpoint(
model,
f'efficientdet-d{opt.compound_coef}_{epoch}_{step}.pth'
)
print('checkpoint...')
except Exception as exception:
print('[Error]', traceback.format_exc())
print(exception)
continue
scheduler.step(np.mean(epoch_loss))
except KeyboardInterrupt:
save_checkpoint(
model, f'efficientdet-d{opt.compound_coef}_{epoch}_{step}.pth')
writer.close()
writer.close()
def infer(self, image):
img = np.array(image)
img = img[:, :, ::-1] #rgb 2 bgr
anchor_ratios = [(1.0, 1.0), (1.4, 0.7), (0.7, 1.4)]
anchor_scales = [2**0, 2**(1.0 / 3.0), 2**(2.0 / 3.0)]
threshold = 0.25
iou_threshold = 0.25
force_input_size = None
use_cuda = False
use_float16 = False
cudnn.fastest = False
cudnn.benchmark = False
input_size = 512
ori_imgs, framed_imgs, framed_metas = preprocess(img,
max_size=input_size)
if use_cuda:
x = torch.stack(
[torch.from_numpy(fi).cuda() for fi in framed_imgs], 0)
else:
x = torch.stack([torch.from_numpy(fi) for fi in framed_imgs], 0)
x = x.to(torch.float32 if not use_float16 else torch.float16).permute(
0, 3, 1, 2)
model = EfficientDetBackbone(compound_coef=0,
num_classes=len(self.labels),
ratios=anchor_ratios,
scales=anchor_scales)
model.load_state_dict(torch.load(self.path, map_location='cpu'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
with torch.no_grad():
features, regression, classification, anchors = model(x)
regressBoxes = BBoxTransform()
clipBoxes = ClipBoxes()
out = postprocess(x, anchors, regression, classification,
regressBoxes, clipBoxes, threshold,
iou_threshold)
pred = invert_affine(framed_metas, out)
results = []
for i in range(len(ori_imgs)):
if len(pred[i]['rois']) == 0:
continue
ori_imgs[i] = ori_imgs[i].copy()
for j in range(len(pred[i]['rois'])):
xt1, yt1, xbr, ybr = pred[i]['rois'][j].astype(np.float64)
xt1 = float(xt1)
yt1 = float(yt1)
xbr = float(xbr)
yb4 = float(ybr)
obj = str(pred[i]['class_ids'][j])
obj_label = self.labels.get(obj)
obj_score = str(pred[i]['scores'][j])
results.append({
"confidence": str(obj_score),
"label": obj_label,
"points": [xt1, yt1, xbr, ybr],
"type": "rectangle",
})
return results
'surfboard', 'tennis racket', 'bottle', '', 'wine glass', 'cup', 'fork',
'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange',
'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair',
'couch', 'potted plant', 'bed', '', 'dining table', '', '', 'toilet', '',
'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave',
'oven', 'toaster', 'sink', 'refrigerator', '', 'book', 'clock', 'vase',
'scissors', 'teddy bear', 'hair drier', 'toothbrush'
]
# tf bilinear interpolation is different from any other's, just make do
input_sizes = [512, 640, 768, 896, 1024, 1280, 1280, 1536]
input_size = input_sizes[
compound_coef] if force_input_size is None else force_input_size
model = EfficientDetBackbone(compound_coef=compound_coef,
num_classes=len(obj_list),
ratios=anchor_ratios,
scales=anchor_scales)
model.load_state_dict(torch.load(f'weights/efficientdet-d{compound_coef}.pth'))
model.requires_grad_(False)
model.eval()
if use_cuda:
model = model.cuda()
if use_float16:
model = model.half()
def detect(image):
# convert image to array
frame = np.array(image)
import torch
import yaml
from torch import nn
import sys
sys.path.append("..")
from backbone import EfficientDetBackbone
import numpy as np
class Params:
def __init__(self, project_file):
self.params = yaml.safe_load(open(project_file).read())
def __getattr__(self, item):
return self.params.get(item, None)
device = torch.device('cuda')
params = Params(f'../projects/voc2coco0712.yml')
model = EfficientDetBackbone(num_classes=len(params.obj_list), compound_coef=2, onnx_export=True,
ratios=eval(params.anchors_ratios), scales=eval(params.anchors_scales)).to(device)
model.backbone_net.model.set_swish(memory_efficient=False)
dummy_input = torch.randn((1,768,768,3), dtype=torch.float32).to(device)
model.load_state_dict(torch.load(f'../weights/efficientdet-d2_onnx.pth'))
# opset_version can be changed to 10 or other number, based on your need
torch.onnx.export(model, dummy_input, '../convert/efficientdet-d2_onnx.onnx', verbose=True, input_names=["input"], output_names=["regression", "classification"])
