def get_dataset(datasetnames):
names = datasetnames.split('+')
dataset = RoiDataset(get_imdb(names[0]))
print('load dataset {}'.format(names[0]))
for name in names[1:]:
tmp = RoiDataset(get_imdb(name))
dataset += tmp
print('load and add dataset {}'.format(name))
return dataset
def get_dataset(datasetnames):
names = datasetnames.split(
'+') # ['voc_2007_trainval', 'voc_2012_trainval']
dataset = RoiDataset(get_imdb(names[0])) # 'voc_2007_trainval'
print('load dataset {}'.format(names[0]))
for name in names[1:]:
tmp = RoiDataset(get_imdb(name))
dataset += tmp
print('load and add dataset {}'.format(name))
return dataset
def combined_roidb(imdb_names, training=True):
"""
Combine multiple roidbs
"""
def get_training_roidb(imdb):
"""Returns a roidb (Region of Interest database) for use in training."""
if cfg.TRAIN.USE_FLIPPED:
print('Appending horizontally-flipped training examples...')
imdb.append_flipped_images()
print('done')
print('Preparing training data...')
prepare_roidb(imdb)
# ratio_index = rank_roidb_ratio(imdb)
print('done')
return imdb.roidb
def get_roidb(imdb_name):
imdb = get_imdb(imdb_name)
print('Loaded dataset `{:s}` for training'.format(imdb.name))
roidb = get_training_roidb(imdb)
return roidb
roidbs = [get_roidb(s) for s in imdb_names.split('+')]
roidb = roidbs[0]
if len(roidbs) > 1:
for r in roidbs[1:]:
roidb.extend(r)
imdb = get_imdb(imdb_names.split('+')[0])
# imdb = dataset.imdb.imdb(imdb_names, tmp.classes)
else:
imdb = get_imdb(imdb_names)
if training:
roidb = filter_roidb(roidb)
return imdb, roidb
def train_model(dataset, trainset, num_classes, net, pad, cachepath):
cachefolder = osp.join(cachepath, dataset+'_'+trainset, net)
if not osp.isdir(cachefolder):
os.makedirs(cachefolder)
ptag = 'S' + ('%d_'*len(options.seg.sizes)) % tuple(options.seg.sizes) \
+ 'IB%d_B%d_E%d-uniform' % (options.seg.imbatch,options.seg.batchsize,options.seg.epoch)
if options.seg.trainflip:
ptag += '_F'
tag = 'S' + ('%d_'*len(options.seg.sizes)) % tuple(options.seg.sizes) \
+ 'IB%d_B%d_E%d-uniform' % (options.seg.imbatch,options.seg.batchsize,options.seg.ftepochall)
if options.seg.trainflip:
tag += '_F'
trainfolder = osp.join(cachefolder,'TRAIN')
finetunefolder = osp.join(cachefolder,'FT-TR')
if not osp.isdir(finetunefolder):
os.makedirs(finetunefolder)
prefile = osp.join(trainfolder,ptag+'.caffemodel')
targetfile = osp.join(finetunefolder,tag+'.caffemodel')
targetlock = osp.join(finetunefolder,tag+'.lock')
if osp.exists(targetfile) or not osp.exists(prefile):
return
try:
os.mkdir(targetlock)
except Exception as e:
return
# looks like it is a tricky task to redirect the stdout/stderr
# leave it as of now
# logfolder = osp.join(trainfolder,'logs')
print '%s_%s<-%s: %s' % (dataset,trainset,net,tag)
time.sleep(5)
datasetname = '%s_%s' % (dataset,trainset)
segdb = get_imdb(datasetname)
# create the solver and train file
solverpath, _ = dump_prototxts(dataset, trainset, num_classes, segdb.num_images, net, pad)
# solverpath = osp.join(options.netpath,net,'pysol-seg-'+dataset+'-'+trainset+'.prototxt')
# start training
# caffe.set_random_seed(options.seed)
np.random.seed(options.seed)
sw = SolverWrapper(solverpath, segdb, finetunefolder, tag, prefile)
sw.train_model()
del sw
# after training
os.rmdir(targetlock)
def train():
pretrained_model = os.path.join(cfg.PRETRAINED_DIR, 'npy', 'yolov2.npy')
assert os.path.exists(pretrained_model), \
'Model path {} does not exist!'.format(pretrained_model)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)
name = DATASET + '_train'
snapshot_infix = DATASET
output_dir = os.path.join(cfg.TRAIN.TRAINED_DIR, DATASET)
imdb = get_imdb(name)
sw = SolverWrapper(imdb, snapshot_infix, output_dir, pretrained_model)
sw.train_net()
def train():
# define the hyper parameters first
args = parse_args()
args.steplr_epoch = cfg.steplr_epoch
args.steplr_factor = cfg.steplr_factor
args.weight_decay = cfg.weight_decay
args.momentum = cfg.momentum
print('Called with args:')
print(args)
lr = args.lr
# initial tensorboardX writer
if args.use_tfboard:
if args.exp_name == 'default':
writer = SummaryWriter()
else:
writer = SummaryWriter('runs/' + args.exp_name)
args.imdb_name = 'trainval'
args.imdbval_name = 'trainval'
output_dir = args.output_dir
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# load dataset
print('loading dataset....')
train_dataset = RoiDataset(get_imdb(args.imdb_name))
print('dataset loaded.')
print('training rois number: {}'.format(len(train_dataset)))
train_dataloader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=detection_collate,
drop_last=True)
# initialize the model
print('initialize the model')
tic = time.time()
model = Yolov2(pretrained=True, arch=args.arch)
toc = time.time()
print('model loaded: cost time {:.2f}s'.format(toc - tic))
# initialize the optimizer
optimizer = optim.SGD([{
"params": model.trunk.parameters(),
"lr": args.lr * cfg.former_lr_decay
}, {
"params": model.conv3.parameters()
}, {
"params": model.conv4.parameters()
}],
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
scheduler = StepLR(optimizer,
step_size=args.steplr_epoch,
gamma=args.steplr_factor)
if args.resume:
print('resume training enable')
resume_checkpoint_name = 'yolov2_epoch_{}.pth'.format(
args.checkpoint_epoch)
resume_checkpoint_path = os.path.join(output_dir,
resume_checkpoint_name)
print('resume from {}'.format(resume_checkpoint_path))
checkpoint = torch.load(resume_checkpoint_path)
model.load_state_dict(checkpoint['model'])
args.start_epoch = checkpoint['epoch'] + 1
lr = checkpoint['lr']
print('learning rate is {}'.format(lr))
adjust_learning_rate(optimizer, lr)
if args.use_cuda:
model.cuda()
if args.mGPUs:
model = nn.DataParallel(model)
# set the model mode to train because we have some layer whose behaviors are different when in training and testing.
# such as Batch Normalization Layer.
model.train()
iters_per_epoch = int(len(train_dataset) / args.batch_size)
# start training
for epoch in range(args.start_epoch, args.max_epochs + 1):
loss_temp = 0
tic = time.time()
train_data_iter = iter(train_dataloader)
scheduler.step()
lr = get_lr(optimizer)
if cfg.multi_scale and epoch in cfg.epoch_scale:
cfg.scale_range = cfg.epoch_scale[epoch]
print('change scale range to {}'.format(cfg.scale_range))
for step in range(iters_per_epoch):
if cfg.multi_scale and (step + 1) % cfg.scale_step == 0:
scale_index = np.random.randint(*cfg.scale_range)
cfg.input_size = cfg.input_sizes[scale_index]
##print('change input size {}'.format(cfg.input_size))
im_data, boxes, gt_classes, num_obj = next(train_data_iter)
if args.use_cuda:
im_data = im_data.cuda()
boxes = boxes.cuda()
gt_classes = gt_classes.cuda()
num_obj = num_obj.cuda()
im_data_variable = Variable(im_data)
box_loss, iou_loss, class_loss = model(im_data_variable,
boxes,
gt_classes,
num_obj,
training=True)
loss = box_loss.mean()+ iou_loss.mean() \
+ class_loss.mean()
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_temp += loss.item()
if (step + 1) % args.display_interval == 0:
toc = time.time()
loss_temp /= args.display_interval
iou_loss_v = iou_loss.mean().item()
box_loss_v = box_loss.mean().item()
class_loss_v = class_loss.mean().item()
log = "[epoch %2d][step %4d/%4d] loss: %.4f, lr: %.2e, iou_loss: %.4f, box_loss: %.4f, cls_loss: %.4f" \
% (epoch, step+1, iters_per_epoch, loss_temp, lr, iou_loss_v, box_loss_v, class_loss_v)
print(log)
logfile = os.path.join(output_dir, 'training_log.txt')
with open(logfile, 'a') as f:
print(log, file=f)
if args.use_tfboard:
n_iter = (epoch - 1) * iters_per_epoch + step + 1
writer.add_scalar('lr', lr, n_iter)
writer.add_scalar('losses/loss', loss_temp, n_iter)
writer.add_scalar('losses/iou_loss', iou_loss_v, n_iter)
writer.add_scalar('losses/box_loss', box_loss_v, n_iter)
writer.add_scalar('losses/cls_loss', class_loss_v, n_iter)
loss_temp = 0
tic = time.time()
if epoch % args.save_interval == 0:
save_name = os.path.join(output_dir,
'yolov2_epoch_{}.pth'.format(epoch))
torch.save(
{
'model':
model.module.state_dict()
if args.mGPUs else model.state_dict(),
'epoch':
epoch,
'lr':
lr
}, save_name)
def test():
args = parse_args()
args.conf_thresh = 0.005
args.nms_thresh = 0.45
if args.vis:
args.conf_thresh = 0.5
print('Called with args:')
print(args)
# prepare dataset
val_imdb = get_imdb(args.dataset)
val_dataset = RoiDataset(val_imdb, train=False)
val_dataloader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False)
# load model
model = Yolov2(arch=args.arch)
# weight_loader = WeightLoader()
# weight_loader.load(model, 'yolo-voc.weights')
# print('loaded')
model_path = os.path.join(args.output_dir, args.model_name + '.pth')
print('loading model from {}'.format(model_path))
if torch.cuda.is_available():
checkpoint = torch.load(model_path)
else:
checkpoint = torch.load(model_path, map_location='cpu')
model.load_state_dict(checkpoint['model'])
if args.use_cuda:
model.cuda()
model.eval()
print('model loaded')
dataset_size = len(val_imdb.image_index)
print('classes: ', val_imdb.num_classes)
all_boxes = [[[] for _ in range(dataset_size)]
for _ in range(val_imdb.num_classes)]
det_file = os.path.join(args.output_dir, 'detections.pkl')
results = []
img_id = -1
with torch.no_grad():
for batch, (im_data, im_infos) in enumerate(val_dataloader):
if args.use_cuda:
im_data_variable = Variable(im_data).cuda()
else:
im_data_variable = Variable(im_data)
yolo_outputs = model(im_data_variable)
for i in range(im_data.size(0)):
img_id += 1
output = [item[i].data for item in yolo_outputs]
im_info = {'width': im_infos[i][0], 'height': im_infos[i][1]}
detections = yolo_eval(output,
im_info,
conf_threshold=args.conf_thresh,
nms_threshold=args.nms_thresh)
if img_id % 100 == 0:
print('im detect [{}/{}]'.format(img_id + 1,
len(val_dataset)))
if len(detections) > 0:
for cls in range(val_imdb.num_classes):
inds = torch.nonzero(detections[:, -1] == cls).view(-1)
if inds.numel() > 0:
cls_det = torch.zeros((inds.numel(), 5))
cls_det[:, :4] = detections[inds, :4]
cls_det[:, 4] = detections[inds,
4] * detections[inds, 5]
all_boxes[cls][img_id] = cls_det.cpu().numpy()
img = Image.open(val_imdb.image_path_at(img_id))
if len(detections) > 0:
detect_result = {}
boxes = detections[:, :5].cpu().numpy()
classes = detections[:, -1].long().cpu().numpy()
class_names = val_imdb.classes
num_boxes = boxes.shape[0]
labels = []
for i in range(num_boxes):
det_bbox = tuple(
np.round(boxes[i, :4]).astype(np.int64))
score = boxes[i, 4]
gt_class_ind = classes[i]
class_name = class_names[gt_class_ind]
disp_str = '{}: {:.2f}'.format(class_name, score)
bbox = tuple(np.round(boxes[i, :4]).astype(np.int64))
xmin = bbox[0]
ymin = bbox[1]
xmax = bbox[2]
ymax = bbox[3]
box2d = {}
box2d["x1"] = str(xmin)
box2d["y1"] = str(ymin)
box2d["x2"] = str(xmax)
box2d["y2"] = str(ymax)
bbox = {}
bbox["box2d"] = box2d
bbox["category"] = class_name
labels.append(bbox)
detect_result["ImageID"] = os.path.basename(
val_imdb.image_path_at(img_id))
detect_result["labels"] = labels
results.append(detect_result)
if args.vis:
img = Image.open(val_imdb.image_path_at(img_id))
if len(detections) == 0:
continue
det_boxes = detections[:, :5].cpu().numpy()
det_classes = detections[:, -1].long().cpu().numpy()
im2show = draw_detection_boxes(
img,
det_boxes,
det_classes,
class_names=val_imdb.classes)
plt.figure()
plt.imshow(im2show)
plt.show()
#if img_id > 10:
# break
print(results)
results_file = os.path.join(args.output_dir, 'detections.json')
with open(results_file, 'w') as f:
json.dump(results,
f,
ensure_ascii=False,
indent=4,
sort_keys=True,
separators=(',', ': '))
with open(det_file, 'wb') as f:
pickle.dump(all_boxes, f, pickle.HIGHEST_PROTOCOL)
val_imdb.evaluate_detections(all_boxes, output_dir=args.output_dir)
import _init_paths
from config.options import options
from dataset.factory import get_imdb
import pdb
for split in ['train', 'trainval']:
name = 'aug-voc2012_{}'.format(split)
print 'DS: ' + name
imdb = get_imdb(name)
imdb.compute_stats()
for split in ['train']:
name = 'context_{}'.format(split)
print 'DS: ' + name
imdb = get_imdb(name)
imdb.compute_stats()
name = 'context33_{}'.format(split)
print 'DS: ' + name
imdb = get_imdb(name)
imdb.compute_stats()
name = 'context20_{}'.format(split)
print 'DS: ' + name
imdb = get_imdb(name)
imdb.compute_stats()
def test():
args = parse_args()
args.conf_thresh = 0.005
args.nms_thresh = 0.45
if args.vis:
args.conf_thresh = 0.5
print('Called with args:')
print(args)
# prepare dataset
if args.dataset == 'voc07trainval':
args.imdbval_name = 'voc_2007_trainval'
elif args.dataset == 'voc07test':
args.imdbval_name = 'voc_2007_test'
else:
raise NotImplementedError
val_imdb = get_imdb(args.imdbval_name)
val_dataset = RoiDataset(val_imdb, train=False)
val_dataloader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False)
# load model
model = Yolov2()
# weight_loader = WeightLoader()
# weight_loader.load(model, 'yolo-voc.weights')
# print('loaded')
model_path = os.path.join(args.output_dir, args.model_name + '.pth')
print('loading model from {}'.format(model_path))
if torch.cuda.is_available():
checkpoint = torch.load(model_path)
else:
checkpoint = torch.load(model_path, map_location='cpu')
model.load_state_dict(checkpoint['model'])
if args.use_cuda:
model.cuda()
model.eval()
print('model loaded')
dataset_size = len(val_imdb.image_index)
all_boxes = [[[] for _ in range(dataset_size)]
for _ in range(val_imdb.num_classes)]
det_file = os.path.join(args.output_dir, 'detections.pkl')
img_id = -1
with torch.no_grad():
for batch, (im_data, im_infos) in enumerate(val_dataloader):
if args.use_cuda:
im_data_variable = Variable(im_data).cuda()
else:
im_data_variable = Variable(im_data)
yolo_outputs = model(im_data_variable)
for i in range(im_data.size(0)):
img_id += 1
output = [item[i].data for item in yolo_outputs]
im_info = {'width': im_infos[i][0], 'height': im_infos[i][1]}
detections = yolo_eval(output,
im_info,
conf_threshold=args.conf_thresh,
nms_threshold=args.nms_thresh)
print('im detect [{}/{}]'.format(img_id + 1, len(val_dataset)))
if len(detections) > 0:
for cls in range(val_imdb.num_classes):
inds = torch.nonzero(detections[:, -1] == cls).view(-1)
if inds.numel() > 0:
cls_det = torch.zeros((inds.numel(), 5))
cls_det[:, :4] = detections[inds, :4]
cls_det[:, 4] = detections[inds,
4] * detections[inds, 5]
all_boxes[cls][img_id] = cls_det.cpu().numpy()
if args.vis:
img = Image.open(val_imdb.image_path_at(img_id))
if len(detections) == 0:
continue
det_boxes = detections[:, :5].cpu().numpy()
det_classes = detections[:, -1].long().cpu().numpy()
im2show = draw_detection_boxes(
img,
det_boxes,
det_classes,
class_names=val_imdb.classes)
plt.figure()
plt.imshow(im2show)
plt.show()
with open(det_file, 'wb') as f:
pickle.dump(all_boxes, f, pickle.HIGHEST_PROTOCOL)
val_imdb.evaluate_detections(all_boxes, output_dir=args.output_dir)
def get_roidb(imdb_name):
imdb = get_imdb(imdb_name)
print('Loaded dataset `{:s}` for training'.format(imdb.name))
roidb = get_training_roidb(imdb)
return roidb
def get_dataset(datasetnames):
names = datasetnames
dataset = RoiDataset(get_imdb(names))
return dataset
def test():
args = parse_args()
if args.vis:
args.conf_thresh = 0.5
# load test data
if args.dataset == 'voc07test':
dataset_name = 'voc_2007_test'
elif args.dataset == 'voc12test':
dataset_name = 'voc_2012_test'
else:
raise NotImplementedError
test_imdb = get_imdb(dataset_name)
test_dataset = RoiDataset(test_imdb, train=False)
test_dataloader = DataLoader(test_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=False)
# load model
model = YOLOv2()
if not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
weight_file_path = os.path.join(
args.output_dir, 'yolov2_epoch_{}.pth'.format(args.check_epoch))
if torch.cuda.is_available:
state_dict = torch.load(weight_file_path)
else:
state_dict = torch.load(weight_file_path, map_location='cpu')
model.load_state_dict(state_dict['model'])
if args.use_cuda:
model = model.cuda()
model.eval()
num_data = len(test_dataset)
all_boxes = [[[] for _ in range(num_data)]
for _ in range(test_imdb.num_classes)]
img_id = -1
det_file = os.path.join(args.output_dir, 'detections.pkl')
with torch.no_grad():
for batch_size, (im_data, im_infos) in enumerate(test_dataloader):
if args.use_cuda:
im_data = im_data.cuda()
im_infos = im_infos.cuda()
im_data_variable = Variable(im_data)
outputs = model(im_data_variable)
for i in range(im_data.size(0)):
img_id += 1
output = [item[i].data for item in outputs]
im_info = im_infos[i]
detections = eval(output, im_info, args.conf_thresh,
args.nms_thresh)
if len(detections) > 0:
for i in range(cfg.CLASS_NUM):
idxs = torch.nonzero(detections[:, -1] == i).view(-1)
if idxs.numel() > 0:
cls_det = torch.zeros((idxs.numel(), 5))
cls_det[:, :4] = detections[idxs, :4]
cls_det[:, 4] = detections[idxs,
4] * detections[idxs, 5]
all_boxes[i][img_id] = cls_det.cpu().numpy()
if args.vis:
img = Image.open(test_imdb.image_path_at(img_id))
if len(detections) == 0:
continue
det_boxes = detections[:, :5].cpu().numpy()
det_classes = detections[:, -1].long().cpu().numpy()
imshow = draw_detection_boxes(
img,
det_boxes,
det_classes,
class_names=test_imdb.classes)
plt.figure()
plt.imshow(imshow)
plt.show()
with open(det_file, 'wb') as f:
pickle.dump(all_boxes, f, pickle.HIGHEST_PROTOCOL)
test_imdb.evaluate_detections(all_boxes, output_dir=args.output_dir)
