Skip to content
This repository has been archived by the owner on Jul 2, 2021. It is now read-only.

Commit

Permalink
add YOLOv2Tiny
Browse files Browse the repository at this point in the history
  • Loading branch information
@Hakuyume
Hakuyume committed Oct 18, 2018
1 parent 5f0c195 commit a2e8cb7
Show file tree
Hide file tree
Showing 3 changed files with 197 additions and 74 deletions.
1 change: 1 addition & 0 deletions chainercv/links/__init__.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -20,4 +20,5 @@
from chainercv.links.model.ssd import SSD512 # NOQA
from chainercv.links.model.vgg import VGG16 # NOQA
from chainercv.links.model.yolo import YOLOv2 # NOQA
from chainercv.links.model.yolo import YOLOv2Tiny # NOQA
from chainercv.links.model.yolo import YOLOv3 # NOQA
2 changes: 2 additions & 0 deletions chainercv/links/model/yolo/__init__.py
View file
Original file line number Diff line number Diff line change
@@ -1,6 +1,8 @@
from chainercv.links.model.yolo.yolo_base import YOLOBase # NOQA
from chainercv.links.model.yolo.yolo_v2 import DarknetExtractor # NOQA
from chainercv.links.model.yolo.yolo_v2 import Darknet19Extractor # NOQA
from chainercv.links.model.yolo.yolo_v2 import YOLOv2 # NOQA
from chainercv.links.model.yolo.yolo_v2 import YOLOv2Tiny # NOQA
from chainercv.links.model.yolo.yolo_v3 import Darknet53Extractor # NOQA
from chainercv.links.model.yolo.yolo_v3 import ResidualBlock # NOQA
from chainercv.links.model.yolo.yolo_v3 import YOLOv3 # NOQA
268 changes: 194 additions & 74 deletions chainercv/links/model/yolo/yolo_v2.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -29,60 +29,7 @@ def _reorg(x):
return F.reshape(x, (n, c * 4, h // 2, w // 2))


class Darknet19Extractor(chainer.ChainList):
"""A Darknet19 based feature extractor for YOLOv2.
This is a feature extractor for :class:`~chainercv.links.model.yolo.YOLOv2`
"""

insize = 416
grid = 13

def __init__(self):
super(Darknet19Extractor, self).__init__()

# Darknet19
for k, n_conv in enumerate((1, 1, 3, 3, 5, 5)):
for i in range(n_conv):
if i % 2 == 0:
self.append(
Conv2DBNActiv(32 << k, 3, pad=1, activ=_leaky_relu))
else:
self.append(
Conv2DBNActiv(32 << (k - 1), 1, activ=_leaky_relu))

# additional links
self.append(Conv2DBNActiv(1024, 3, pad=1, activ=_leaky_relu))
self.append(Conv2DBNActiv(1024, 3, pad=1, activ=_leaky_relu))
self.append(Conv2DBNActiv(64, 1, activ=_leaky_relu))
self.append(Conv2DBNActiv(1024, 3, pad=1, activ=_leaky_relu))

def __call__(self, x):
"""Compute a feature map from a batch of images.
Args:
x (ndarray): An array holding a batch of images.
The images should be resized to :math:`416\\times 416`.
Returns:
Variable:
"""

h = x
for i, link in enumerate(self):
h = link(h)
if i == 12:
tmp = h
elif i == 19:
h, tmp = tmp, h
elif i == 20:
h = F.concat((_reorg(h), tmp))
if i in {0, 1, 4, 7, 12}:
h = _maxpool(h)
return h


class YOLOv2(YOLOBase):
class YOLOv2Base(YOLOBase):
"""YOLOv2.
This is a model of YOLOv2 [#]_.
Expand Down Expand Up @@ -113,24 +60,8 @@ class YOLOv2(YOLOBase):
"""

_models = {
'voc0712': {
'param': {'n_fg_class': 20},
'url': 'https://chainercv-models.preferred.jp/'
'yolo_v2_voc0712_converted_2018_05_03.npz',
'cv2': True
},
}

_anchors = (
(1.73145, 1.3221),
(4.00944, 3.19275),
(8.09892, 5.05587),
(4.84053, 9.47112),
(10.0071, 11.2364))

def __init__(self, n_fg_class=None, pretrained_model=None):
super(YOLOv2, self).__init__()
super(YOLOv2Base, self).__init__()

param, path = utils.prepare_pretrained_model(
{'n_fg_class': n_fg_class}, pretrained_model, self._models)
Expand All @@ -139,7 +70,7 @@ def __init__(self, n_fg_class=None, pretrained_model=None):
self.use_preset('visualize')

with self.init_scope():
self.extractor = Darknet19Extractor()
self.extractor = self._extractorDarknet19Extractor()
self.subnet = Convolution2D(
len(self._anchors) * (4 + 1 + self.n_fg_class), 1)

Expand All @@ -153,11 +84,11 @@ def __init__(self, n_fg_class=None, pretrained_model=None):
chainer.serializers.load_npz(path, self, strict=False)

def to_cpu(self):
super(YOLOv2, self).to_cpu()
super(YOLOv2Base, self).to_cpu()
self._default_bbox = cuda.to_cpu(self._default_bbox)

def to_gpu(self, device=None):
super(YOLOv2, self).to_gpu(device)
super(YOLOv2Base, self).to_gpu(device)
self._default_bbox = cuda.to_gpu(self._default_bbox, device)

def __call__(self, x):
Expand Down Expand Up @@ -233,3 +164,192 @@ def _decode(self, loc, obj, conf):
score = self.xp.hstack(score).astype(np.float32)

return bbox, label, score


class Darknet19Extractor(chainer.ChainList):
"""A Darknet19 based feature extractor for YOLOv2.
This is a feature extractor for :class:`~chainercv.links.model.yolo.YOLOv2`
"""

insize = 416
grid = 13

def __init__(self):
super(Darknet19Extractor, self).__init__()

# Darknet19
for k, n_conv in enumerate((1, 1, 3, 3, 5, 5)):
for i in range(n_conv):
if i % 2 == 0:
self.append(
Conv2DBNActiv(32 << k, 3, pad=1, activ=_leaky_relu))
else:
self.append(
Conv2DBNActiv(32 << (k - 1), 1, activ=_leaky_relu))

# additional links
self.append(Conv2DBNActiv(1024, 3, pad=1, activ=_leaky_relu))
self.append(Conv2DBNActiv(1024, 3, pad=1, activ=_leaky_relu))
self.append(Conv2DBNActiv(64, 1, activ=_leaky_relu))
self.append(Conv2DBNActiv(1024, 3, pad=1, activ=_leaky_relu))

def __call__(self, x):
"""Compute a feature map from a batch of images.
Args:
x (ndarray): An array holding a batch of images.
The images should be resized to :math:`416\\times 416`.
Returns:
Variable:
"""

h = x
for i, link in enumerate(self):
h = link(h)
if i == 12:
tmp = h
elif i == 19:
h, tmp = tmp, h
elif i == 20:
h = F.concat((_reorg(h), tmp))
if i in {0, 1, 4, 7, 12}:
h = _maxpool(h)
return h


class YOLOv2(YOLOv2Base):
"""YOLOv2.
This is a model of YOLOv2 [#]_.
This model uses :class:`~chainercv.links.model.yolo.Darknet19Extractor` as
its feature extractor.
.. [#] Joseph Redmon, Ali Farhadi.
YOLO9000: Better, Faster, Stronger. CVPR 2017.
Args:
n_fg_class (int): The number of classes excluding the background.
pretrained_model (string): The weight file to be loaded.
This can take :obj:`'voc0712'`, `filepath` or :obj:`None`.
The default value is :obj:`None`.
* :obj:`'voc0712'`: Load weights trained on trainval split of \
PASCAL VOC 2007 and 2012. \
The weight file is downloaded and cached automatically. \
:obj:`n_fg_class` must be :obj:`20` or :obj:`None`. \
These weights were converted from the darknet model \
provided by `the original implementation \
<https://pjreddie.com/darknet/yolov2/>`_. \
The conversion code is \
`chainercv/examples/yolo/darknet2npz.py`.
* `filepath`: A path of npz file. In this case, :obj:`n_fg_class` \
must be specified properly.
* :obj:`None`: Do not load weights.
"""

_extractor = Darknet19Extractor

_models = {
'voc0712': {
'param': {'n_fg_class': 20},
'url': 'https://chainercv-models.preferred.jp/'
'yolo_v2_voc0712_converted_2018_05_03.npz',
'cv2': True
},
}

_anchors = (
(1.73145, 1.3221),
(4.00944, 3.19275),
(8.09892, 5.05587),
(4.84053, 9.47112),
(10.0071, 11.2364))


class DarknetExtractor(chainer.ChainList):
"""A Darknet based feature extractor for YOLOv2Tiny.
This is a feature extractor for
:class:`~chainercv.links.model.yolo.YOLOv2Tiny`
"""

insize = 416
grid = 13

def __init__(self):
super(DarknetExtractor, self).__init__()

# Darknet
for k in range(6):
self.append(Conv2DBNActiv(16 << k, 3, pad=1, activ=_leaky_relu))

# additional link
self.append(Conv2DBNActiv(1024, 3, pad=1, activ=_leaky_relu))

def __call__(self, x):
"""Compute a feature map from a batch of images.
Args:
x (ndarray): An array holding a batch of images.
The images should be resized to :math:`416\\times 416`.
Returns:
Variable:
"""

h = x
for i, link in enumerate(self):
h = link(h)
if i < 5:
h = _maxpool(h)
return h


class YOLOv2Tiny(YOLOv2Base):
"""YOLOv2 tiny.
This is a model of YOLOv2 tiny a.k.a. Tiny YOLO.
This model uses :class:`~chainercv.links.model.yolo.DarknetExtractor` as
its feature extractor.
Args:
n_fg_class (int): The number of classes excluding the background.
pretrained_model (string): The weight file to be loaded.
This can take :obj:`'voc0712'`, `filepath` or :obj:`None`.
The default value is :obj:`None`.
* :obj:`'voc0712'`: Load weights trained on trainval split of \
PASCAL VOC 2007 and 2012. \
The weight file is downloaded and cached automatically. \
:obj:`n_fg_class` must be :obj:`20` or :obj:`None`. \
These weights were converted from the darknet model \
provided by `the original implementation \
<https://pjreddie.com/darknet/yolov2/>`_. \
The conversion code is \
`chainercv/examples/yolo/darknet2npz.py`.
* `filepath`: A path of npz file. In this case, :obj:`n_fg_class` \
must be specified properly.
* :obj:`None`: Do not load weights.
"""

_extractor = DarknetExtractor

_models = {
'voc0712': {
'param': {'n_fg_class': 20},
'url': 'https://chainercv-models.preferred.jp/'
'yolo_v2_tiny_voc0712_converted_2018_10_19.npz',
'cv2': True
},
}

_anchors = (
(1.19, 1.08),
(4.41, 3.42),
(11.38, 6.63),
(5.11, 9.42),
(10.52, 16.62))

0 comments on commit a2e8cb7

Please sign in to comment.