def get_dataset(dataset, args):
if dataset.lower() == 'coco':
train_dataset = gdata.COCOInstance(root='/home/xcq/PycharmProjects/datasets/coco/',splits='instances_train2017')
val_dataset = gdata.COCOInstance(root='/home/xcq/PycharmProjects/datasets/coco/',splits='instances_val2017', skip_empty=False)
val_metric = COCOInstanceMetric(val_dataset, args.save_prefix + '_eval', cleanup=True)
else:
raise NotImplementedError('Dataset: {} not implemented.'.format(dataset))
return train_dataset, val_dataset, val_metric
def get_dataset(dataset, args):
if dataset.lower() == 'coco':
val_dataset = gdata.COCOInstance(splits='instances_val2017', skip_empty=False)
val_metric = COCOInstanceMetric(val_dataset, args.save_prefix + '_eval',
cleanup=not args.save_json)
else:
raise NotImplementedError('Dataset: {} not implemented.'.format(dataset))
return val_dataset, val_metric
def get_dataset(dataset, args):
if dataset.lower() == 'coco':
train_dataset = gdata.COCOInstance(splits='instances_train2017')
val_dataset = gdata.COCOInstance(splits='instances_val2017', skip_empty=False)
starting_id = 0
if args.horovod and MPI:
length = len(val_dataset)
shard_len = length // hvd.size()
rest = length % hvd.size()
# Compute the start index for this partition
starting_id = shard_len * hvd.rank() + min(hvd.rank(), rest)
val_metric = COCOInstanceMetric(val_dataset, args.save_prefix + '_eval',
use_ext=args.use_ext, starting_id=starting_id)
else:
raise NotImplementedError('Dataset: {} not implemented.'.format(dataset))
if args.horovod and MPI:
val_dataset = val_dataset.shard(hvd.size(), hvd.rank())
return train_dataset, val_dataset, val_metric
def get_dataset(args):
train_dataset = gdata.COCOInstance(root=args.datasetloc,
splits='instances_train2017')
val_dataset = gdata.COCOInstance(root=args.datasetloc,
splits='instances_val2017',
skip_empty=False)
starting_id = 0
if args.horovod and MPI:
length = len(val_dataset)
shard_len = length // hvd.size()
rest = length % hvd.size()
# Compute the start index for this partition
starting_id = shard_len * hvd.rank() + min(hvd.rank(), rest)
val_metric = COCOInstanceMetric(val_dataset,
args.save_prefix + '_eval',
use_ext=False,
starting_id=starting_id)
if args.horovod and MPI:
val_dataset = val_dataset.shard(hvd.size(), hvd.rank())
return (train_dataset, val_dataset, val_metric)
def test_coco_instance():
if not osp.isdir(osp.expanduser('~/.mxnet/datasets/coco')):
return
# use valid only, loading training split is very slow
val = data.COCOInstance(splits=('instances_val2017',))
name = str(val)
assert len(val.classes) > 0
for _ in range(10):
index = np.random.randint(0, len(val))
_ = val[index]
def get_dataset(dataset, args):
if dataset.lower() == 'coco':
train_dataset = gdata.COCOInstance(splits='instances_train2017')
val_dataset = gdata.COCOInstance(splits='instances_val2017',
skip_empty=False)
val_metric = COCOInstanceMetric(val_dataset,
args.save_prefix + '_eval',
cleanup=True)
elif dataset.lower() == 'imaterialist':
from utils.iMaterialistDataset import COCOiMaterialist
train_dataset = COCOiMaterialist(root='datasets/imaterialist/',
splits='rle_instances_train')
val_dataset = COCOiMaterialist(root='datasets/imaterialist/',
splits='resize_rle_instances_val',
skip_empty=False)
val_metric = COCOInstanceMetric(val_dataset,
args.save_prefix + '_eval',
cleanup=True)
else:
raise NotImplementedError(
'Dataset: {} not implemented.'.format(dataset))
return train_dataset, val_dataset, val_metric
def get_dataset(dataset, args):
if dataset.lower() == 'coco':
train_dataset = gdata.COCOInstance(splits='instances_train2017')
val_dataset = gdata.COCOInstance(splits='instances_val2017',
skip_empty=False)
val_metric = COCOInstanceMetric(val_dataset,
args.save_prefix + '_eval',
cleanup=True)
elif dataset.lower == "voc":
input_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([.485, .456, .406], [.229, .224, .225]),
])
train_dataset = gdata.VOCSegmentation(split="train",
root=args.data_dir,
transform=input_transform)
val_dataset = gdata.VOCSegmentation(split="val",
root=args.data_dir,
transform=input_transform)
val_metric = SegmentationMetric(train_dataset.num_class)
else:
raise NotImplementedError(
'Dataset: {} not implemented.'.format(dataset))
return train_dataset, val_dataset, val_metric
def get_dataset(dataset, args):
if dataset.lower() == 'blbalbla':
train_dataset = gdata.COCOInstance(splits='instances_train2017')
val_dataset = gdata.COCOInstance(splits='instances_val2017',
skip_empty=False)
val_metric = COCOInstanceMetric(val_dataset,
args.save_prefix + '_eval',
cleanup=True)
print(val_metric)
elif dataset.lower() == 'coco':
# normal loading
annotations_train = json.load(
open('/home/nexus/primates/annotations/annot_train.json'))
annotations_val = json.load(
open('/home/nexus/primates/annotations/annot_test.json'))
img_path_train = '/home/nexus/primates/train/'
img_path_val = '/home/nexus/primates/test/'
train_set = []
def jsonToSet(annotations_train_1, img_path):
for key in tqdm(annotations_train_1.keys()):
el = annotations_train_1[key]
img = mx.nd.array(
imread(img_path + key.split('.')[0] +
'.png').astype('float32'))
img_annot = []
masks = []
for reg in el['regions']:
x = reg['shape_attributes']['all_points_x']
y = reg['shape_attributes']['all_points_y']
_mask = [
np.swapaxes(np.asarray([x, y]).astype('float32'),
axis1=0,
axis2=1)
]
masks.append(_mask)
# print(masks[-1])
mask = load_mask(img.asnumpy(),
np.expand_dims(_mask, axis=-1))
# print(mask[0].shape)
bboxes = extract_bboxes(mask)
# print(bboxes)
bboxes = [[
bboxes[0][1], bboxes[0][0], bboxes[0][3], bboxes[0][2]
]]
# print(bboxes)
annot = np.append(bboxes, [0.0])
img_annot.append(annot.astype('float32'))
annot = np.asarray(img_annot).astype('float32')
# mask = load_mask(img, masks)
# bboxes = extract_bboxes(mask)
train_set.append((img, annot, masks))
return train_set
train_dataset = gdata.COCOInstance(splits='instances_train2017')
val_dataset = gdata.COCOInstance(splits='instances_val2017',
skip_empty=False)
val_metric = COCOInstanceMetric(val_dataset,
args.save_prefix + '_eval',
cleanup=True)
train_dataset = jsonToSet(annotations_train, img_path_train)
val_dataset = jsonToSet(annotations_val, img_path_train)
from mxnet.gluon.data.dataset import ArrayDataset
train_dataset = ArrayDataset(train_dataset)
val_dataset = ArrayDataset(val_dataset)
#from gluoncv.data.batchify import Tuple, Append, MaskRCNNTrainBatchify
#train_transform = presets.rcnn.MaskRCNNDefaultTrainTransform(short, max_size, net)
#train_dataset = train_dataset.transform(train_transform)
#val_dataset = val_dataset.transform(train_transform)
else:
raise NotImplementedError(
'Dataset: {} not implemented.'.format(dataset))
return train_dataset, val_dataset, val_metric
# init params
ctx = mx.cpu()
model_path = "/Users/rensike/.mxnet/models/mask_rcnn_resnet50_v1b_coco-a3527fdc.params"
num_workers = 0
# init model
net = get_model("mask_rcnn_resnet50_v1b_coco",pretrained=False,pretrained_base=False)
net.load_parameters(model_path)
net.collect_params().reset_ctx(ctx)
# load val dataset
val_dataset = gdata.COCOInstance(splits='instances_val',root="/Users/rensike/Files/temp/coco_mini", skip_empty=False)
# val_dataset = gdata.VOCSegmentation(root="/Users/rensike/Files/temp/voc_mini",split="val")
# eval_metric = SegmentationMetric(nclass=val_dataset.num_class)
eval_metric = COCOInstanceMetric(val_dataset,"coco_eval")
# load val dataloader
val_bfn = batchify.Tuple(*[batchify.Append() for _ in range(2)])
val_data_loader = mx.gluon.data.DataLoader(
val_dataset.transform(MaskRCNNDefaultValTransform(net.short, net.max_size)),
1, False, batchify_fn=val_bfn, last_batch='keep', num_workers=num_workers)
# do evaluate
eval_metric.reset()
net.hybridize(static_alloc=True)
names, values = validate(net, val_data_loader, [ctx], eval_metric,len(val_dataset))
