def get_body(self, resnet_layer_id):
'''
Create the feature extraction network of the SSD based on resnet34.
The first layer of the res-net is converted into grayscale by averaging the weights of the 3 channels
of the original resnet.
Returns
-------
network: gluon.nn.HybridSequential
The body network for feature extraction based on resnet
'''
pretrained = resnet34_v1(pretrained=True, ctx=ctx)
pretrained_2 = resnet34_v1(pretrained=True, ctx=mx.cpu(0))
first_weights = pretrained_2.features[0].weight.data().mean(
axis=1).expand_dims(axis=1)
# First weights could be replaced with individual channels.
body = gluon.nn.HybridSequential()
with body.name_scope():
first_layer = gluon.nn.Conv2D(channels=64,
kernel_size=(7, 7),
padding=(3, 3),
strides=(2, 2),
in_channels=1,
use_bias=False)
first_layer.initialize(mx.init.Normal(), ctx=ctx)
first_layer.weight.set_data(first_weights)
body.add(first_layer)
body.add(*pretrained.features[1:-resnet_layer_id])
return body
def make_cnn(ctx=mx.gpu()):
p_dropout = 0.5
pretrained = resnet34_v1(pretrained=True, ctx=ctx)
pretrained_2 = resnet34_v1(pretrained=True, ctx=mx.cpu(0))
first_weights = pretrained_2.features[0].weight.data().mean(axis=1).expand_dims(axis=1)
# First weights could be replaced with individual channels.
body = gluon.nn.HybridSequential()
with body.name_scope():
first_layer = gluon.nn.Conv2D(channels=64, kernel_size=(7, 7), padding=(3, 3), strides=(2, 2), in_channels=1, use_bias=False)
first_layer.initialize(mx.init.Normal(), ctx=ctx)
first_layer.weight.set_data(first_weights)
body.add(first_layer)
body.add(*pretrained.features[1:6])
output = gluon.nn.HybridSequential()
output.add(gluon.nn.Flatten())
output.add(gluon.nn.Dense(64, activation='relu'))
output.add(gluon.nn.Dropout(p_dropout))
output.add(gluon.nn.Dense(64, activation='relu'))
output.add(gluon.nn.Dropout(p_dropout))
output.add(gluon.nn.Dense(4, activation='sigmoid'))
output.collect_params().initialize(mx.init.Normal(), ctx=ctx)
body.add(output)
return body
def get_body(self,resnet_layer_id):
pretrained = resnet34_v1(pretrained=True,ctx=self.ctx)
pretrained_2 = resnet34_v1(pretrained=True,ctx=mx.cpu(0))
first_weights = pretrained_2.features[0].weight.data().mean(axis=1).expand_dims(axis=1)
body = gluon.nn.HybridSequential()
with body.name_scope():
first_layer = gluon.nn.Conv2D(channels=64,kernel_size=(7,7),padding=(3,3),strides=(2,2),in_channels=1,use_bias=False)
first_layer.initialize(mx.init.Xavier(),ctx=self.ctx)
first_layer.weight.set_data(first_weights)
body.add(first_layer)
body.add(*pretrained.features[1:-resnet_layer_id])
return body
def __init__(self):
super(BackboneResnet34, self).__init__()
self.info = "resnet34"
self.features = vision.resnet34_v1(pretrained=True).features[0:-2]
self.append = nn.Sequential()
self.append.add(
nn.Conv2D(1024,
kernel_size=3,
padding=6,
dilation=6,
weight_initializer=mx.init.Xavier()), nn.BatchNorm(),
nn.Activation("relu"),
nn.Conv2D(1024,
kernel_size=1,
padding=0,
dilation=1,
weight_initializer=mx.init.Xavier()), nn.BatchNorm(),
nn.Activation("relu"))
self.extra = nn.Sequential()
self.extra.add(
ssd_extra_one(channels=512, stride=2, padding=1),
ssd_extra_one(channels=256, stride=2, padding=1),
ssd_extra_one(channels=256, stride=1,
padding=0), #smaller feature map
ssd_extra_one(channels=256, stride=1, padding=0),
)
self.features.collect_params().setattr("lr_mult", 0.1)
self.extra.initialize()
self.append.initialize()
return
if autograd.is_training(
): #for hybridBlock, must call hybridize() to make this take effect!!
return anchors, cls_preds, bbox_pred
return self.postprocess(anchors, cls_preds, bbox_pred)
if 0:
net = SSD_CUSTOM(100)
ctx = mx.gpu()
net.collect_params().reset_ctx(ctx)
x = nd.zeros((32, 3, 256 * 2, 256 * 2), ctx=ctx)
y = net(x)
if 0:
#net = get_resnet_34()
net = vision.resnet34_v1(pretrained=True).features
ctx = mx.gpu()
net.collect_params().reset_ctx(ctx)
x = nd.zeros((32, 3, 256 * 2, 256 * 2), ctx=ctx)
y = net(x)
for layer in net:
x = layer(x)
print(layer.name, x.shape)
data = mx.sym.var("data")
for sym in net(data).get_internals():
print(sym)
if 0:
# net = model_zoo.get_model('ssd_300_vgg16_atrous_voc', pretrained=False, pretrained_base=False)
#net = vgg16_atrous_300(pretrained=True)
net = BackboneResnet34()
