def get_featuremap_model(config, num_classes):
inputs = Input(shape=(None, None, 3))
base_layers = ResNet50(inputs)
featuremap_model = Model(inputs, base_layers)
return featuremap_model
def get_model(config, num_classes):
# 设定输入
inputs = Input(shape=(None, None, 3))
# 设定roi的个数以及他们的位置坐标，None一般为32，可以在config.py中的num.rois中设置
roi_input = Input(shape=(None, 4))
# 获取特征提取层的输出
base_layers = ResNet50(inputs)
# 一般采用9个anchor，可以在config.py中设置
num_anchors = len(config.anchor_box_scales) * len(config.anchor_box_ratios)
# 获取RPN
rpn = get_rpn(base_layers, num_anchors)
# 建立rpn模型
model_rpn = Model(inputs, rpn[:2])
# 获取分类模型的输出
classifier = get_classifier(base_layers,
roi_input,
config.num_rois,
nb_classes=num_classes,
trainable=True)
# 建立分类模型
model_classifier = Model([inputs, roi_input], classifier)
# 同时包含RPN和分类网络的model，用来保存和读取权重信息
model_all = Model([inputs, roi_input], rpn[:2] + classifier)
# 返回模型
return model_rpn, model_classifier, model_all
def get_predict_model(config, num_classes):
inputs = Input(shape=(None, None, 3))
roi_input = Input(shape=(None, 4))
feature_map_input = Input(shape=(None, None, 1024))
#----------------------------------------------------#
# 假设输入为600,600,3
# 获得一个38,38,1024的共享特征层base_layers
#----------------------------------------------------#
base_layers = ResNet50(inputs)
#----------------------------------------------------#
# 每个特征点9个先验框
#----------------------------------------------------#
num_anchors = len(config.anchor_box_scales) * len(config.anchor_box_ratios)
#----------------------------------------------------#
# 将共享特征层传入建议框网络
# 该网络结果会对先验框进行调整获得建议框
#----------------------------------------------------#
rpn = get_rpn(base_layers, num_anchors)
model_rpn = Model(inputs, rpn + [base_layers])
#----------------------------------------------------#
# 将共享特征层和建议框传入classifier网络
# 该网络结果会对建议框进行调整获得预测框
#----------------------------------------------------#
classifier = get_classifier(feature_map_input, roi_input, num_classes,
config.pooling_regions)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
return model_rpn, model_classifier_only
def get_predict_model(config, num_classes):
# 设定输入
inputs = Input(shape=(None, None, 3))
# 设定roi的个数以及他们的位置坐标，None一般为32，可以在config.py中的num.rois中设置
roi_input = Input(shape=(None, 4))
# 输入特征层规格
feature_map_input = Input(shape=(None, None, 1024))
# 获取特征提取层的输出
base_layers = ResNet50(inputs)
# 生成多少个anchor box
num_anchors = len(config.anchor_box_scales) * len(config.anchor_box_ratios)
# 获取RPN
rpn = get_rpn(base_layers, num_anchors)
# 建立rpn模型
model_rpn = Model(inputs, rpn)
# 获取分类模型的输出
classifier = get_classifier(feature_map_input,
roi_input,
config.num_rois,
nb_classes=num_classes,
trainable=True)
# 分类模型
model_classifier_only = Model([feature_map_input, roi_input], classifier)
# 返回模型
return model_rpn, model_classifier_only
def resnet_retinanet(num_classes, inputs=None, num_anchors=None, submodels=None, name='retinanet'):
if inputs==None:
inputs = keras.layers.Input(shape=(600, 600, 3))
else:
inputs = inputs
resnet = ResNet50(inputs)
if num_anchors is None:
num_anchors = AnchorParameters.default.num_anchors()
C3, C4, C5 = resnet.outputs[1:]
P5 = keras.layers.Conv2D(256, kernel_size=1, strides=1, padding='same', name='C5_reduced')(C5)
# 38x38x256
P5_upsampled = layers.UpsampleLike(name='P5_upsampled')([P5, C4])
P5 = keras.layers.Conv2D(256, kernel_size=3, strides=1, padding='same', name='P5')(P5)
# 38x38x256
P4 = keras.layers.Conv2D(256, kernel_size=1, strides=1, padding='same', name='C4_reduced')(C4)
P4 = keras.layers.Add(name='P4_merged')([P5_upsampled, P4])
P4_upsampled = layers.UpsampleLike(name='P4_upsampled')([P4, C3])
P4 = keras.layers.Conv2D(256, kernel_size=3, strides=1, padding='same', name='P4')(P4)
# 75x75x256
P3 = keras.layers.Conv2D(256, kernel_size=1, strides=1, padding='same', name='C3_reduced')(C3)
P3 = keras.layers.Add(name='P3_merged')([P4_upsampled, P3])
P3 = keras.layers.Conv2D(256, kernel_size=3, strides=1, padding='same', name='P3')(P3)
P6 = keras.layers.Conv2D(256, kernel_size=3, strides=2, padding='same', name='P6')(C5)
P7 = keras.layers.Activation('relu', name='C6_relu')(P6)
P7 = keras.layers.Conv2D(256, kernel_size=3, strides=2, padding='same', name='P7')(P7)
features = [P3, P4, P5, P6, P7]
regression_model = make_last_layer_loc(num_classes,num_anchors)
classification_model = make_last_layer_cls(num_classes,num_anchors)
regressions = []
classifications = []
for feature in features:
regression = regression_model(feature)
classification = classification_model(feature)
regressions.append(regression)
classifications.append(classification)
regressions = keras.layers.Concatenate(axis=1, name="regression")(regressions)
classifications = keras.layers.Concatenate(axis=1, name="classification")(classifications)
pyramids = [regressions,classifications]
model = keras.models.Model(inputs=inputs, outputs=pyramids, name=name)
return model
def centernet(input_shape,
num_classes,
backbone='resnet50',
max_objects=100,
mode="train",
num_stacks=2):
assert backbone in ['resnet50', 'hourglass']
image_input = Input(shape=input_shape)
if backbone == 'resnet50':
#-----------------------------------#
# 对输入图片进行特征提取
# 512, 512, 3 -> 16, 16, 2048
#-----------------------------------#
C5 = ResNet50(image_input)
#--------------------------------------------------------------------------------------------------------#
# 对获取到的特征进行上采样，进行分类预测和回归预测
# 16, 16, 2048 -> 32, 32, 256 -> 64, 64, 128 -> 128, 128, 64 -> 128, 128, 64 -> 128, 128, num_classes
# -> 128, 128, 64 -> 128, 128, 2
# -> 128, 128, 64 -> 128, 128, 2
#--------------------------------------------------------------------------------------------------------#
y1, y2, y3 = centernet_head(C5, num_classes)
if mode == "train":
model = Model(inputs=image_input, outputs=[y1, y2, y3])
return model
elif mode == "predict":
detections = Lambda(lambda x: decode(*x, max_objects=max_objects))(
[y1, y2, y3])
prediction_model = Model(inputs=image_input, outputs=detections)
return prediction_model
elif mode == "heatmap":
prediction_model = Model(inputs=image_input, outputs=y1)
return prediction_model
else:
outs = HourglassNetwork(image_input, num_stacks, num_classes)
if mode == "train":
temp_outs = []
for out in outs:
temp_outs += out
model = Model(inputs=image_input, outputs=out)
return model
elif mode == "predict":
y1, y2, y3 = outs[-1]
detections = Lambda(lambda x: decode(*x, max_objects=max_objects))(
[y1, y2, y3])
prediction_model = Model(inputs=image_input, outputs=[detections])
return prediction_model
elif mode == "heatmap":
y1, y2, y3 = outs[-1]
prediction_model = Model(inputs=image_input, outputs=y1)
return prediction_model
def __init__(self, args):
self.args = args
kwargs = {"num_classes": self.args.num_class}
if args.net == "resnet18":
from nets.resnet import ResNet18
model = ResNet18(pretrained=(args.load == 'imagenet'), **kwargs)
elif args.net == "resnet50":
from nets.resnet import ResNet50
model = ResNet50(pretrained=(args.load == 'imagenet'), **kwargs)
elif args.net == "wideresnet282":
from nets.wideresnet import WRN28_2
model = WRN28_2(**kwargs)
else:
raise NotImplementedError
print("Number of parameters",
sum(p.numel() for p in model.parameters() if p.requires_grad))
self.model = nn.DataParallel(model).cuda()
self.optimizer = torch.optim.SGD(self.model.parameters(),
lr=self.args.lr,
momentum=0.9,
nesterov=True,
weight_decay=5e-4)
self.criterion = nn.CrossEntropyLoss(ignore_index=-1)
self.criterion_nored = nn.CrossEntropyLoss(reduction="none")
self.kwargs = {"num_workers": 12, "pin_memory": False}
self.train_loader, self.val_loader = make_data_loader(
args, **self.kwargs)
self.best = 0
self.best_epoch = 0
self.acc = []
self.train_acc = []
self.med_clean = []
self.med_noisy = []
self.perc_clean = []
self.perc_noisy = []
self.reductor_plot = umap.UMAP(n_components=2)
self.toPIL = torchvision.transforms.ToPILImage()
self.unorm = UnNormalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
def get_model(config, num_classes):
inputs = Input(shape=(None, None, 3))
roi_input = Input(shape=(None, 4))
base_layers = ResNet50(inputs)
num_anchors = len(config.anchor_box_scales) * len(config.anchor_box_ratios)
rpn = get_rpn(base_layers, num_anchors)
model_rpn = Model(inputs, rpn[:2])
classifier = get_classifier(base_layers,
roi_input,
config.num_rois,
nb_classes=num_classes,
trainable=True)
model_classifier = Model([inputs, roi_input], classifier)
model_all = Model([inputs, roi_input], rpn[:2] + classifier)
return model_rpn, model_classifier, model_all
def get_predict_model(config, num_classes):
inputs = Input(shape=(None, None, 3))
roi_input = Input(shape=(None, 4))
feature_map_input = Input(shape=(None, None, 1024))
base_layers = ResNet50(inputs)
num_anchors = len(config.anchor_box_scales) * len(config.anchor_box_ratios)
rpn = get_rpn(base_layers, num_anchors)
model_rpn = Model(inputs, rpn)
classifier = get_classifier(feature_map_input,
roi_input,
config.num_rois,
nb_classes=num_classes,
trainable=True)
model_classifier_only = Model([feature_map_input, roi_input], classifier)
return model_rpn, model_classifier_only
def get_predict_model(num_classes, backbone, num_anchors=9):
inputs = Input(shape=(None, None, 3))
roi_input = Input(shape=(None, 4))
if backbone == 'vgg':
feature_map_input = Input(shape=(None, None, 512))
#----------------------------------------------------#
# 假设输入为600,600,3
# 获得一个37,37,512的共享特征层base_layers
#----------------------------------------------------#
base_layers = VGG16(inputs)
#----------------------------------------------------#
# 将共享特征层传入建议框网络
# 该网络结果会对先验框进行调整获得建议框
#----------------------------------------------------#
rpn = get_rpn(base_layers, num_anchors)
#----------------------------------------------------#
# 将共享特征层和建议框传入classifier网络
# 该网络结果会对建议框进行调整获得预测框
#----------------------------------------------------#
classifier = get_vgg_classifier(feature_map_input, roi_input, 7,
num_classes)
else:
feature_map_input = Input(shape=(None, None, 1024))
#----------------------------------------------------#
# 假设输入为600,600,3
# 获得一个38,38,1024的共享特征层base_layers
#----------------------------------------------------#
base_layers = ResNet50(inputs)
#----------------------------------------------------#
# 将共享特征层传入建议框网络
# 该网络结果会对先验框进行调整获得建议框
#----------------------------------------------------#
rpn = get_rpn(base_layers, num_anchors)
#----------------------------------------------------#
# 将共享特征层和建议框传入classifier网络
# 该网络结果会对建议框进行调整获得预测框
#----------------------------------------------------#
classifier = get_resnet50_classifier(feature_map_input, roi_input, 14,
num_classes)
model_rpn = Model(inputs, rpn + [base_layers])
model_classifier_only = Model([feature_map_input, roi_input], classifier)
return model_rpn, model_classifier_only
def get_model(num_classes,
backbone,
num_anchors=9,
input_shape=[None, None, 3]):
inputs = Input(shape=input_shape)
roi_input = Input(shape=(None, 4))
if backbone == 'vgg':
#----------------------------------------------------#
# 假设输入为600,600,3
# 获得一个37,37,512的共享特征层base_layers
#----------------------------------------------------#
base_layers = VGG16(inputs)
#----------------------------------------------------#
# 将共享特征层传入建议框网络
# 该网络结果会对先验框进行调整获得建议框
#----------------------------------------------------#
rpn = get_rpn(base_layers, num_anchors)
#----------------------------------------------------#
# 将共享特征层和建议框传入classifier网络
# 该网络结果会对建议框进行调整获得预测框
#----------------------------------------------------#
classifier = get_vgg_classifier(base_layers, roi_input, 7, num_classes)
else:
#----------------------------------------------------#
# 假设输入为600,600,3
# 获得一个38,38,1024的共享特征层base_layers
#----------------------------------------------------#
base_layers = ResNet50(inputs)
#----------------------------------------------------#
# 将共享特征层传入建议框网络
# 该网络结果会对先验框进行调整获得建议框
#----------------------------------------------------#
rpn = get_rpn(base_layers, num_anchors)
#----------------------------------------------------#
# 将共享特征层和建议框传入classifier网络
# 该网络结果会对建议框进行调整获得预测框
#----------------------------------------------------#
classifier = get_resnet50_classifier(base_layers, roi_input, 14,
num_classes)
model_rpn = Model(inputs, rpn)
model_all = Model([inputs, roi_input], rpn + classifier)
return model_rpn, model_all
def get_model(config, num_classes):
inputs = Input(shape=(None, None, 3))
roi_input = Input(shape=(None, 4))
base_layers = ResNet50(inputs) # base_layer是ResNet提取的特征图
num_anchors = len(config.anchor_box_scales) * len(
config.anchor_box_ratios) #3*3
rpn = get_rpn(base_layers, num_anchors) #RPN的输出是粗分类和粗定位
model_rpn = Model(inputs, rpn[:2])
#classifier是网络的最后输出 包括精分类和精定位 RoiPooling
classifier = get_classifier(base_layers,
roi_input,
config.num_rois,
nb_classes=num_classes,
trainable=True)
model_classifier = Model([inputs, roi_input], classifier)
model_all = Model([inputs, roi_input], rpn[:2] + classifier)
return model_rpn, model_classifier, model_all
def get_model(config, num_classes):
''' RPN网络输出 置信度+ 框回归 '''
inputs = Input(shape=(None, None, 3))
roi_input = Input(shape=(None, 4))
base_layers = ResNet50(inputs) # 共享特征层FeatureMap
# 1 rpn网络
num_anchors = len(config.anchor_box_scales) * len(
config.anchor_box_ratios) # 9=3*3
rpn = get_rpn(base_layers,
num_anchors) # rpn网络输出[x_class, x_regr, base_layers]
model_rpn = Model(inputs, rpn[:2]) # ？？？
# 2 fast rcnn网络
classifier = get_classifier(base_layers,
roi_input,
config.num_rois,
nb_classes=num_classes,
trainable=True)
model_classifier = Model([inputs, roi_input], classifier)
# 3 综合
model_all = Model([inputs, roi_input], rpn[:2] + classifier)
return model_rpn, model_classifier, model_all
def get_model(config, num_classes):
'''
创建训练网络模型
Parameters
----------
config
num_classes
Returns
-------
'''
# 输入主干提取的图片
inputs = Input(shape=(None, None, 3))
# roi-pooling层的输入，从rpn层获得的最后建议框，None:一张图片中建议框数量
roi_input = Input(shape=(None, 4))
# 假设输入为600,600,3
# 获得一个38,38,1024的共享特征层base_layers
base_layers = ResNet50(inputs)
# 每个特征点9个先验框，先验框边长数量*先验框宽高比例数
num_anchors = len(config.anchor_box_scales) * len(config.anchor_box_ratios)
# 将共享特征层传入建议框网络
# 该网络结果会对先验框进行调整获得建议框
rpn = get_rpn(base_layers, num_anchors)
# 这是包含在下面moel_all中的子网络，不必单独训练model_rpn,因为下面model_all训练时会同步更新
# model_rpn的权重。model_rpn的作用是每次训练和预测时生成用于截取特征图的建议框。训练时每个batch_size
# 都会更新权重，向好的方向调整，获得的建议框也会越来越精确。额~，我怎么突然想起了GAN网络...
model_rpn = Model(inputs, rpn)
# 将共享特征层和建议框传入classifier网络
# 该网络结果会对建议框进行调整获得预测框
classifier = get_classifier(base_layers, roi_input, num_classes,
config.pooling_regions)
# 构建包含rpn和分类的两个网络，一起训练，使得两个网络的损失函数整体最小
model_all = Model([inputs, roi_input], rpn + classifier)
return model_rpn, model_all
def RetinaFace(cfg, backbone="mobilenet"):
inputs = Input(shape=(None, None, 3))
if backbone == "mobilenet":
C3, C4, C5 = MobileNet(inputs)
elif backbone == "resnet50":
C3, C4, C5 = ResNet50(inputs)
else:
raise ValueError('Unsupported backbone - `{}`, Use mobilenet, resnet50.'.format(backbone))
leaky = 0
if (cfg['out_channel'] <= 64):
leaky = 0.1
P3 = Conv2D_BN_Leaky(cfg['out_channel'], kernel_size=1, strides=1, padding='same', name='C3_reduced', leaky=leaky)(C3)
P4 = Conv2D_BN_Leaky(cfg['out_channel'], kernel_size=1, strides=1, padding='same', name='C4_reduced', leaky=leaky)(C4)
P5 = Conv2D_BN_Leaky(cfg['out_channel'], kernel_size=1, strides=1, padding='same', name='C5_reduced', leaky=leaky)(C5)
P5_upsampled = UpsampleLike(name='P5_upsampled')([P5, P4])
P4 = Add(name='P4_merged')([P5_upsampled, P4])
P4 = Conv2D_BN_Leaky(cfg['out_channel'], kernel_size=3, strides=1, padding='same', name='Conv_P4_merged', leaky=leaky)(P4)
P4_upsampled = UpsampleLike(name='P4_upsampled')([P4, P3])
P3 = Add(name='P3_merged')([P4_upsampled, P3])
P3 = Conv2D_BN_Leaky(cfg['out_channel'], kernel_size=3, strides=1, padding='same', name='Conv_P3_merged', leaky=leaky)(P3)
SSH1 = SSH(P3, cfg['out_channel'], leaky=leaky)
SSH2 = SSH(P4, cfg['out_channel'], leaky=leaky)
SSH3 = SSH(P5, cfg['out_channel'], leaky=leaky)
SSH_all = [SSH1,SSH2,SSH3]
bbox_regressions = Concatenate(axis=1,name="bbox_reg")([BboxHead(feature) for feature in SSH_all])
classifications = Concatenate(axis=1,name="cls")([ClassHead(feature) for feature in SSH_all])
ldm_regressions = Concatenate(axis=1,name="ldm_reg")([LandmarkHead(feature) for feature in SSH_all])
output = [bbox_regressions, classifications, ldm_regressions]
model = Model(inputs=inputs, outputs=output)
return model
def get_predict_model(config, num_classes):
'''
训练时两步一起训练,使得loss全部最小,预测时分开,加快预测速度
Parameters
----------
config
num_classes
Returns
-------
'''
# 输入主干提取的图片
inputs = Input(shape=(None, None, 3))
# roi-pooling层的输入，从rpn层获得的最后建议框，None:一张图片中建议框数量
roi_input = Input(shape=(None, 4))
# 主干网络输出的特征层，预测时作为分类网络的输入
feature_map_input = Input(shape=(None, None, 1024))
# 假设输入为600,600,3, 获得一个38,38,1024的共享特征层base_layers
base_layers = ResNet50(inputs)
# 每个特征点9个先验框
num_anchors = len(config.anchor_box_scales) * len(config.anchor_box_ratios)
# 将共享特征层传入建议框网络
# 该网络结果会对先验框进行调整获得建议框
rpn = get_rpn(base_layers, num_anchors)
model_rpn = Model(inputs, rpn + [base_layers])
# 将共享特征层和建议框传入classifier网络
# 该网络结果会对建议框进行调整获得预测框
classifier = get_classifier(feature_map_input, roi_input, num_classes,
config.pooling_regions)
# 此处仅构建分类模型，与训练时的模型不同
model_classifier_only = Model([feature_map_input, roi_input], classifier)
return model_rpn, model_classifier_only
def centernet(input_shape,
num_classes,
backbone='resnet50',
max_objects=100,
mode="train",
num_stacks=2):
assert backbone in ['resnet50', 'hourglass']
output_size = input_shape[0] // 4
image_input = Input(shape=input_shape)
hm_input = Input(shape=(output_size, output_size, num_classes))
wh_input = Input(shape=(max_objects, 2))
reg_input = Input(shape=(max_objects, 2))
reg_mask_input = Input(shape=(max_objects, ))
index_input = Input(shape=(max_objects, ))
if backbone == 'resnet50':
#-----------------------------------#
# 对输入图片进行特征提取
# 512, 512, 3 -> 16, 16, 2048
#-----------------------------------#
C5 = ResNet50(image_input)
#--------------------------------------------------------------------------------------------------------#
# 对获取到的特征进行上采样，进行分类预测和回归预测
# 16, 16, 2048 -> 32, 32, 256 -> 64, 64, 128 -> 128, 128, 64 -> 128, 128, 64 -> 128, 128, num_classes
# -> 128, 128, 64 -> 128, 128, 2
# -> 128, 128, 64 -> 128, 128, 2
#--------------------------------------------------------------------------------------------------------#
y1, y2, y3 = centernet_head(C5, num_classes)
if mode == "train":
loss_ = Lambda(loss, name='centernet_loss')([
y1, y2, y3, hm_input, wh_input, reg_input, reg_mask_input,
index_input
])
model = Model(inputs=[
image_input, hm_input, wh_input, reg_input, reg_mask_input,
index_input
],
outputs=[loss_])
return model
elif mode == "predict":
detections = Lambda(lambda x: decode(*x, max_objects=max_objects))(
[y1, y2, y3])
prediction_model = Model(inputs=image_input, outputs=detections)
return prediction_model
elif mode == "heatmap":
prediction_model = Model(inputs=image_input, outputs=y1)
return prediction_model
else:
outs = HourglassNetwork(image_input, num_stacks, num_classes)
if mode == "train":
loss_all = []
for out in outs:
y1, y2, y3 = out
loss_ = Lambda(loss)([
y1, y2, y3, hm_input, wh_input, reg_input, reg_mask_input,
index_input
])
loss_all.append(loss_)
loss_all = Lambda(tf.reduce_mean, name='centernet_loss')(loss_all)
model = Model(inputs=[
image_input, hm_input, wh_input, reg_input, reg_mask_input,
index_input
],
outputs=loss_all)
return model
elif mode == "predict":
y1, y2, y3 = outs[-1]
detections = Lambda(lambda x: decode(*x, max_objects=max_objects))(
[y1, y2, y3])
prediction_model = Model(inputs=image_input, outputs=[detections])
return prediction_model
elif mode == "heatmap":
y1, y2, y3 = outs[-1]
prediction_model = Model(inputs=image_input, outputs=y1)
return prediction_model
def RetinaFace(cfg, backbone="mobilenet"):
inputs = Input(shape=(None, None, 3))
#-------------------------------------------#
# 获得三个shape的有效特征层
# 分别是C3 80, 80, 64
# C4 40, 40, 128
# C5 20, 20, 256
#-------------------------------------------#
if backbone == "mobilenet":
C3, C4, C5 = MobileNet(inputs)
elif backbone == "resnet50":
C3, C4, C5 = ResNet50(inputs)
else:
raise ValueError(
'Unsupported backbone - `{}`, Use mobilenet, resnet50.'.format(
backbone))
leaky = 0
if (cfg['out_channel'] <= 64):
leaky = 0.1
#-------------------------------------------#
# 获得三个shape的有效特征层
# 分别是P3 80, 80, 64
# P4 40, 40, 64
# P5 20, 20, 64
#-------------------------------------------#
P3 = Conv2D_BN_Leaky(cfg['out_channel'],
kernel_size=1,
strides=1,
padding='same',
kernel_initializer=random_normal(stddev=0.02),
name='C3_reduced',
leaky=leaky)(C3)
P4 = Conv2D_BN_Leaky(cfg['out_channel'],
kernel_size=1,
strides=1,
padding='same',
kernel_initializer=random_normal(stddev=0.02),
name='C4_reduced',
leaky=leaky)(C4)
P5 = Conv2D_BN_Leaky(cfg['out_channel'],
kernel_size=1,
strides=1,
padding='same',
kernel_initializer=random_normal(stddev=0.02),
name='C5_reduced',
leaky=leaky)(C5)
#-------------------------------------------#
# P5上采样和P4特征融合
# P4 40, 40, 64
#-------------------------------------------#
P5_upsampled = UpsampleLike(name='P5_upsampled')([P5, P4])
P4 = Add(name='P4_merged')([P5_upsampled, P4])
P4 = Conv2D_BN_Leaky(cfg['out_channel'],
kernel_size=3,
strides=1,
padding='same',
kernel_initializer=random_normal(stddev=0.02),
name='Conv_P4_merged',
leaky=leaky)(P4)
#-------------------------------------------#
# P4上采样和P3特征融合
# P3 80, 80, 64
#-------------------------------------------#
P4_upsampled = UpsampleLike(name='P4_upsampled')([P4, P3])
P3 = Add(name='P3_merged')([P4_upsampled, P3])
P3 = Conv2D_BN_Leaky(cfg['out_channel'],
kernel_size=3,
strides=1,
padding='same',
kernel_initializer=random_normal(stddev=0.02),
name='Conv_P3_merged',
leaky=leaky)(P3)
SSH1 = SSH(P3, cfg['out_channel'], leaky=leaky)
SSH2 = SSH(P4, cfg['out_channel'], leaky=leaky)
SSH3 = SSH(P5, cfg['out_channel'], leaky=leaky)
SSH_all = [SSH1, SSH2, SSH3]
#-------------------------------------------#
# 将所有结果进行堆叠
#-------------------------------------------#
bbox_regressions = Concatenate(
axis=1, name="bbox_reg")([BboxHead(feature) for feature in SSH_all])
classifications = Concatenate(
axis=1, name="cls")([ClassHead(feature) for feature in SSH_all])
ldm_regressions = Concatenate(
axis=1, name="ldm_reg")([LandmarkHead(feature) for feature in SSH_all])
output = [bbox_regressions, classifications, ldm_regressions]
model = Model(inputs=inputs, outputs=output)
return model
def centernet(input_shape,
num_classes,
backbone='resnet50',
max_objects=100,
mode="train",
num_stacks=2):
assert backbone in ['resnet50', 'hourglass']
output_size = input_shape[0] // 4
image_input = Input(shape=input_shape)
hm_input = Input(shape=(output_size, output_size, num_classes))
wh_input = Input(shape=(max_objects, 2))
reg_input = Input(shape=(max_objects, 2))
reg_mask_input = Input(shape=(max_objects, ))
index_input = Input(shape=(max_objects, ))
if backbone == 'resnet50':
#-------------------------------#
# 编码器
#-------------------------------#
C5 = ResNet50(image_input)
y1, y2, y3 = centernet_head(C5, num_classes)
if mode == "train":
loss_ = Lambda(loss, name='centernet_loss')([
y1, y2, y3, hm_input, wh_input, reg_input, reg_mask_input,
index_input
])
model = Model(inputs=[
image_input, hm_input, wh_input, reg_input, reg_mask_input,
index_input
],
outputs=[loss_])
return model
else:
detections = Lambda(lambda x: decode(
*x, max_objects=max_objects, num_classes=num_classes))(
[y1, y2, y3])
prediction_model = Model(inputs=image_input, outputs=detections)
return prediction_model
else:
outs = HourglassNetwork(image_input, num_stacks, num_classes)
if mode == "train":
loss_all = []
for out in outs:
y1, y2, y3 = out
loss_ = Lambda(loss)([
y1, y2, y3, hm_input, wh_input, reg_input, reg_mask_input,
index_input
])
loss_all.append(loss_)
loss_all = Lambda(tf.reduce_mean, name='centernet_loss')(loss_all)
model = Model(inputs=[
image_input, hm_input, wh_input, reg_input, reg_mask_input,
index_input
],
outputs=loss_all)
return model
else:
y1, y2, y3 = outs[-1]
detections = Lambda(lambda x: decode(
*x, max_objects=max_objects, num_classes=num_classes))(
[y1, y2, y3])
prediction_model = Model(inputs=image_input, outputs=[detections])
return prediction_model
num_params += np.prod(data[key][name].shape)
return num_params
def check_shape(net):
for var in tf.global_variables():
key, name = net.find_key_name(var)
var_shape = tuple(var.get_shape().as_list())
np_shape = data[key][name].shape
if var_shape != np_shape:
print('incorrect var shape')
if __name__ == '__main__':
from os.path import join, abspath, dirname
pretrain_root = abspath(join(dirname(__file__), '../../model/pretrained_model/ori_resnet/'))
resnet50 = ResNet50()
resnet101 = ResNet101()
resnet152 = ResNet152()
X = tf.placeholder(tf.float32, shape=(None, 224, 224, 3))
y = resnet50(X)
data = np.load(join(pretrain_root, 'resnet50.npy'), encoding='bytes').item()
print('our resnet50 has parameters: %d'%count_total_variables())
print('pretrained resnet50 parameters: %d'%count_total_npvars(data))
check_shape(resnet50)
tf.reset_default_graph()
X = tf.placeholder(tf.float32, shape=(None, 224, 224, 3))
y = resnet101(X)
data = np.load(join(pretrain_root, 'resnet101.npy'), encoding='bytes').item()
print('our resnet101 has parameters: %d'%count_total_variables())
from utils import load_image
from tensorflow.keras.callbacks import TensorBoard, ModelCheckpoint, ReduceLROnPlateau, EarlyStopping
from tensorflow.keras import optimizers
import tensorflow as tf
if __name__ == "__main__":
gpus = tf.config.list_physical_devices(device_type="GPU")
for gpu in gpus:
tf.config.experimental.set_memory_growth(device=gpu, enable=True)
train_dir = "datasets/train"
val_dir = "datasets/val"
batch_size = 2048
image_size = 224
model = ResNet50(input_shape=(image_size, image_size, 3), classes=2)
# model.load_weights("ResNet50.h5")
model.summary()
opt = optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
model.compile(optimizer=opt,
loss=tf.keras.losses.sparse_categorical_crossentropy,
metrics=[tf.keras.metrics.sparse_categorical_accuracy])
plot_model(model, to_file='models/ResNet50.png',show_shapes=True, show_layer_names=True)
logging = TensorBoard(log_dir="logs")
checkpoint = ModelCheckpoint("models" + "/ResNet_{epoch:04d}-{val_loss:.4f}.h5", monitor="val_loss", mode="min",
verbose=1, save_weights_only=True, save_best_only=True, period=3)
early_stopping = EarlyStopping(monitor='val_loss', min_delta=0, patience=15, verbose=1)
def resnet_retinanet(num_classes, input_shape=(600, 600, 3), name='retinanet'):
inputs = keras.layers.Input(shape=input_shape)
resnet = ResNet50(inputs)
#-----------------------------------------#
# 取出三个有效特征层，分别是C3、C4、C5
# C3 75,75,512
# C4 38,38,1024
# C5 19,19,2048
#-----------------------------------------#
C3, C4, C5 = resnet.outputs[1:]
# 75,75,512 -> 75,75,256
P3 = keras.layers.Conv2D(256,
kernel_size=1,
strides=1,
padding='same',
name='C3_reduced')(C3)
# 38,38,1024 -> 38,38,256
P4 = keras.layers.Conv2D(256,
kernel_size=1,
strides=1,
padding='same',
name='C4_reduced')(C4)
# 19,19,2048 -> 19,19,256
P5 = keras.layers.Conv2D(256,
kernel_size=1,
strides=1,
padding='same',
name='C5_reduced')(C5)
# 19,19,256 -> 38,38,256
P5_upsampled = layers.UpsampleLike(name='P5_upsampled')([P5, P4])
# 38,38,256 + 38,38,256 -> 38,38,256
P4 = keras.layers.Add(name='P4_merged')([P5_upsampled, P4])
# 38,38,256 -> 75,75,256
P4_upsampled = layers.UpsampleLike(name='P4_upsampled')([P4, P3])
# 75,75,256 + 75,75,256 -> 75,75,256
P3 = keras.layers.Add(name='P3_merged')([P4_upsampled, P3])
# 75,75,256 -> 75,75,256
P3 = keras.layers.Conv2D(256,
kernel_size=3,
strides=1,
padding='same',
name='P3')(P3)
# 38,38,256 -> 38,38,256
P4 = keras.layers.Conv2D(256,
kernel_size=3,
strides=1,
padding='same',
name='P4')(P4)
# 19,19,256 -> 19,19,256
P5 = keras.layers.Conv2D(256,
kernel_size=3,
strides=1,
padding='same',
name='P5')(P5)
# 19,19,2048 -> 10,10,256
P6 = keras.layers.Conv2D(256,
kernel_size=3,
strides=2,
padding='same',
name='P6')(C5)
P7 = keras.layers.Activation('relu', name='C6_relu')(P6)
# 10,10,256 -> 5,5,256
P7 = keras.layers.Conv2D(256,
kernel_size=3,
strides=2,
padding='same',
name='P7')(P7)
features = [P3, P4, P5, P6, P7]
num_anchors = AnchorParameters.default.num_anchors()
regression_model = make_last_layer_loc(num_classes, num_anchors)
classification_model = make_last_layer_cls(num_classes, num_anchors)
regressions = []
classifications = []
#----------------------------------------------------------#
# 将获取到的P3, P4, P5, P6, P7传入到
# Retinahead里面进行预测，获得回归预测结果和分类预测结果
# 将所有特征层的预测结果进行堆叠
#----------------------------------------------------------#
for feature in features:
regression = regression_model(feature)
classification = classification_model(feature)
regressions.append(regression)
classifications.append(classification)
regressions = keras.layers.Concatenate(axis=1,
name="regression")(regressions)
classifications = keras.layers.Concatenate(
axis=1, name="classification")(classifications)
model = keras.models.Model(inputs, [regressions, classifications],
name=name)
return model