def classifier(base_layers,
input_rois,
num_rois,
nb_classes=21,
trainable=False):
# compile times on theano tend to be very high, so we use smaller ROI pooling regions to workaround
if K.backend() == 'tensorflow':
pooling_regions = 14
input_shape = (num_rois, 14, 14, 1024)
elif K.backend() == 'theano':
pooling_regions = 7
input_shape = (num_rois, 1024, 7, 7)
out_roi_pool = RoiPoolingConv(pooling_regions,
num_rois)([base_layers, input_rois])
out = classifier_layers(out_roi_pool,
input_shape=input_shape,
trainable=True)
out = TimeDistributed(Flatten())(out)
out_class = TimeDistributed(Dense(nb_classes,
activation='softmax',
kernel_initializer='zero'),
name='dense_class_{}'.format(nb_classes))(out)
# note: no regression target for bg class
out_regr = TimeDistributed(Dense(4 * (nb_classes - 1),
activation='linear',
kernel_initializer='zero'),
name='dense_regress_{}'.format(nb_classes))(out)
return [out_class, out_regr]
def classifier(base_layers, input_rois, num_rois, nb_classes=21, trainable=False, alpha=1.0, depth_mult=1):
# compile times on theano tend to be very high, so we use smaller ROI pooling regions to workaround
if K.backend() == 'tensorflow':
pooling_regions = 14
input_shape = (num_rois, 14, 14, 1024)
elif K.backend() == 'theano':
pooling_regions = 7
input_shape = (num_rois, 1024, 7, 7)
out_roi_pool = RoiPoolingConv(pooling_regions, num_rois)([base_layers, input_rois])
out = classifier_layers(out_roi_pool, trainable=True, alpha=1.0, depth_multiplier=1)
out = TimeDistributed(AveragePooling2D(name='Global_average_Pooling_classifier_layer'), name='TimeDistributed_AVG')(out)
out = TimeDistributed(Flatten(name='flatten'), name='TimeDistributed_flatten')(out)
#out = TimeDistributed(Dense(4096, activation='relu', name='fc1'))(out)
#out = TimeDistributed(Dense(4096, activation='relu', name='fc2'))(out)
out_class = TimeDistributed(Dense(nb_classes, activation='softmax', kernel_initializer='zero', name='dense_class'),
name='dense_class_{}'.format(nb_classes))(out)
# note: no regression target for bg class
out_regr = TimeDistributed(Dense(4 * (nb_classes - 1), activation='linear', kernel_initializer='zero', name='dense_regr'),
name='dense_regress_{}'.format(nb_classes))(out)
return [out_class, out_regr]
def classifier(base_layers,
input_rois,
num_rois,
nb_classes=21,
trainable=False):
if K.backend() == 'tensorflow':
pooling_regions = 7
input_shape = (num_rois, 7, 7, 512)
out_roi_pool = RoiPoolingConv(pooling_regions,
num_rois)([base_layers, input_rois])
out = TimeDistributed(Flatten(name='flatten'))(out_roi_pool)
out = TimeDistributed(Dense(4096, activation='relu', name='fc1'))(out)
out = TimeDistributed(Dense(4096, activation='relu', name='fc2'))(out)
out_class = TimeDistributed(Dense(nb_classes,
activation='softmax',
kernel_initializer='zero'),
name='dense_class_{}'.format(nb_classes))(out)
# note: no regression target for bg class
out_regr = TimeDistributed(Dense(4 * (nb_classes - 1),
activation='linear',
kernel_initializer='zero'),
name='dense_regress_{}'.format(nb_classes))(out)
return [out_class, out_regr]
def classifier(base_layers, input_rois, num_rois, nb_classes=21, trainable=False):
'''
Construct a ROI classifier from base layers and input roi. Only ROI classifier uses
the TimeDistributed Layer.
# Args
| base_layers: base layers
| input_rois: a placeholder for input rois
| num_rois: the number of ROI to be processed each time (? to be verified)
| num_classes: number of classes including background
# Return
| out_class: (num_rois, nb_classes) softmax probailities
| out_regr: (num_rois, 4*(nb_classes-1)) regression result
'''
# compile times on theano tend to be very high, so we use smaller ROI pooling regions to workaround
# The pooling_region defines the fixed feature map used in ROI pooling layer
# if K.backend() == 'tensorflow':
# pooling_regions = 7
# input_shape = (num_rois, 7, 7, 512)
# elif K.backend() == 'theano':
# pooling_regions = 7
# input_shape = (num_rois, 512, 7, 7)
pooling_regions = 7
out_roi_pool = RoiPoolingConv(pooling_regions, num_rois)([base_layers, input_rois])
out = TimeDistributed(Flatten(name='flatten'), name='TimeDistributed_flatten')(out_roi_pool)
out = TimeDistributed(Dense(4096, activation='relu', name='fc1'))(out)
out = TimeDistributed(Dense(4096, activation='relu', name='fc2'))(out)
# https://keras.io/layers/core/#dense
# Input nD tensor with shape: (batch_size, ..., input_dim)
# Output has shape (batch_size, ..., units).
out_class = TimeDistributed(Dense(nb_classes, activation='softmax', kernel_initializer='zero'),
name='dense_class_{}'.format(nb_classes))(out)
# note: no regression target for bg class
out_regr = TimeDistributed(Dense(4 * (nb_classes - 1), activation='linear', kernel_initializer='zero'),
name='dense_regress_{}'.format(nb_classes))(out)
return [out_class, out_regr]
def classifier(base_layers,
input_rois,
num_rois,
nb_classes=11,
trainable=False):
# compile times on theano tend to be very high, so we use smaller ROI pooling regions to workaround
if K.backend() == 'tensorflow':
pooling_regions = 7
input_shape = (num_rois, 7, 7, 512) # return x的通道数
elif K.backend() == 'theano':
pooling_regions = 7
input_shape = (num_rois, 512, 7, 7)
# RoiPoolingConv：返回的shape为(1, 32, 7, 7, 512)
# 含义是batch_size，预选框的个数，特征图宽，特征图高度，特征图深度
out_roi_pool = RoiPoolingConv(pooling_regions,
num_rois)([base_layers, input_rois])
# TimeDistributed：输入至少为3D张量，下标为1的维度将被认为是时间维。即对以一个维度下的变量当作一个完整变量来看待。本文是32。
# 你要实现的目的就是对32个预选宽提出的32个图片做出判断。
out = TimeDistributed(Flatten(name='flatten'))(out_roi_pool)
# 4096=32*
out = TimeDistributed(Dense(4096, activation='relu', name='fc1'))(out)
out = TimeDistributed(Dense(4096, activation='relu', name='fc2'))(out)
# out_class的shape:(?, 32, 21);
out_class = TimeDistributed(Dense(nb_classes,
activation='softmax',
kernel_initializer='zero'),
name='dense_class_{}'.format(nb_classes))(out)
# note: no regression target for bg class
# out_regr的shape: (?, 32, 80)
out_regr = TimeDistributed(Dense(4 * (nb_classes - 1),
activation='linear',
kernel_initializer='zero'),
name='dense_regress_{}'.format(nb_classes))(out)
# 产生num_rois个out_class和out_reg
# 四个损失函数中的：Fast R-CNN classification和Fast R-CNN regression(proposal ->box)。
return [out_class, out_regr]
def classifier(base_layers,
input_rois,
num_rois,
nb_classes=21,
trainable=False):
# compile times on theano tend to be very high, so we use smaller ROI pooling regions to workaround
if K.backend() == 'tensorflow':
pooling_regions = 14
input_shape = (num_rois, 14, 14, 1024)
elif K.backend() == 'theano':
pooling_regions = 7
input_shape = (num_rois, 1024, 7, 7)
# RoiPoolingConv的作用是分别将每一个ROI对应的原图区域resize为指定大小(14,14),
# 然后，把所有的resize后的矩阵在0轴上串联起来。
# 其作用就是将尺寸不一而同的ROI都调整为相同的大小，输入给后续的层，因为后续的层有全连接层，所以需要规范输入图的大小
out_roi_pool = RoiPoolingConv(pooling_regions, num_rois)(
[base_layers,
input_rois]) #shape=(1, num_rois, 14, 14, 1024) dtype=float32
out = classifier_layers(
out_roi_pool, input_shape=input_shape,
trainable=True) #9个卷积层#shape=(1, num_rois, 1, 1, 2048) dtype=float32>
out = TimeDistributed(Flatten())(
out) #shape=(?, num_rois, 2048) dtype=float32
out_class = TimeDistributed(
Dense(nb_classes, activation='softmax', kernel_initializer='zero'),
name='dense_class_{}'.format(nb_classes))(
out) #shape=(?, num_rois, nb_classes) dtype=float32
# note: no regression target for bg class
out_regr = TimeDistributed(
Dense(4 * (nb_classes - 1),
activation='linear',
kernel_initializer='zero'),
name='dense_regress_{}'.format(nb_classes))(
out) #shape=(?, num_rois, 4 * (nb_classes - 1)) dtype=float32
return [out_class, out_regr]
def classifier(base_layers,
input_rois,
num_rois,
nb_classes=11,
trainable=False):
# compile times on theano tend to be very high, so we use smaller ROI pooling regions to workaround
if K.backend() == 'tensorflow':
pooling_regions = 14
input_shape = (num_rois, 14, 14, 1024) # return x的通道数
elif K.backend() == 'theano':
pooling_regions = 7
input_shape = (num_rois, 1024, 7, 7)
# # 该层的输入为feature maps和roi的坐标信息
out_roi_pool = RoiPoolingConv(pooling_regions,
num_rois)([base_layers, input_rois])
# 输出的是（None, num_riois, 2048)的feature map
out = classifier_layers(out_roi_pool,
input_shape=input_shape,
trainable=True)
# 因为是对num_rois个feature maps分别处理的，所以需要使用timedistributed进行包装
out = TimeDistributed(Flatten())(out)
out_class = TimeDistributed(Dense(nb_classes,
activation='softmax',
kernel_initializer='zero'),
name='dense_class_{}'.format(nb_classes))(out)
# note: no regression target for bg class
# 我们可以使用包装器TimeDistributed包装Dense，以产生针对各个时间步信号的独立全连接
out_regr = TimeDistributed(Dense(4 * (nb_classes - 1),
activation='linear',
kernel_initializer='zero'),
name='dense_regress_{}'.format(nb_classes))(out)
return [out_class, out_regr] # 一共有num_riois个out_class和out_regr
base_layers)
x_class = Conv2D(num_anchors, (1, 1), activation='sigmoid', kernel_initializer='uniform', name='rpn_out_class')(x)
x_regr = Conv2D(num_anchors * 4, (1, 1), activation='linear', kernel_initializer='zero', name='rpn_out_regress')(x)
return [x_class, x_regr, base_layers]
def classifier(base_layers, input_rois, num_rois, nb_classes=21, trainable=False):
# compile times on theano tend to be very high, so we use smaller ROI pooling regions to workaround
# -if K.backend() == 'tensorflow':
if K.image_data_format == 'channels_last':
pooling_regions = 7
input_shape = (num_rois, 7, 7, 512)
# -elif K.backend() == 'theano':
elif K.image_data_format = 'channels_first':
pooling_regions = 7
input_shape = (num_rois, 512, 7, 7)
out_roi_pool = RoiPoolingConv(pooling_regions, num_rois)([base_layers, input_rois])
out = TimeDistributed(Flatten(name='flatten'))(out_roi_pool)
out = TimeDistributed(Dense(4096, activation='relu', name='fc1'))(out)
out = TimeDistributed(Dense(4096, activation='relu', name='fc2'))(out)
out_class = TimeDistributed(Dense(nb_classes, activation='softmax', kernel_initializer='zero'),
name='dense_class_{}'.format(nb_classes))(out)
# note: no regression target for bg class
out_regr = TimeDistributed(Dense(4 * (nb_classes - 1), activation='linear', kernel_initializer='zero'),
name='dense_regress_{}'.format(nb_classes))(out)
return [out_class, out_regr]
