def __define_grouped_convolutions(self, input_shape, n_groups,
n_channels_per_branch, is_dilated,
layer_num):
"""
Define layers inside grouped convolutional block.
"""
n_channels_in = input_shape[1]
n_branches = 5
n_channels_per_group_in = int(n_channels_in /
n_groups) # 每一个group的输入通道数
n_channels_out = int(n_groups * n_branches *
n_channels_per_branch) # timeception的输出总通道数
n_channels_per_group_out = int(n_channels_out /
n_groups) # 每一个group的输出通道数
# 通道数要能被group整除
assert n_channels_in % n_groups == 0
assert n_channels_out % n_groups == 0
# type of multi-scale kernels to use: either multi_kernel_sizes or multi_dilation_rates
if is_dilated: # 固定尺度的卷积核，多尺度的空洞卷积
kernel_sizes = (3, 3, 3)
dilation_rates = (1, 2, 3)
else:
kernel_sizes = (3, 5, 7) # 多尺度的卷积核
dilation_rates = (1, 1, 1)
input_shape_per_group = list(input_shape) # (N,C,T,H,W)
input_shape_per_group[1] = n_channels_per_group_in
# loop on groups, and define convolutions in each group
for idx_group in range(n_groups):
group_num = idx_group + 1
self.__define_temporal_convolutional_block(input_shape_per_group,
n_channels_per_branch,
kernel_sizes,
dilation_rates,
layer_num, group_num)
# activation
layer_name = 'relu_tc%d' % (layer_num)
layer = ReLU()
layer._name = layer_name
setattr(self, layer_name, layer)
# shuffle channels
layer_name = 'shuffle_tc%d' % (layer_num)
layer = ChannelShuffleLayer(n_channels_out, n_groups)
layer._name = layer_name
setattr(self, layer_name, layer)
def __define_grouped_convolutions(self, input_shape, n_groups,
n_channels_per_branch, is_dilated,
layer_num):
'''
Define layers inside grouped convolutional block. 定义timeception中的操作
:param input_shape: (32, 1024, 128, 7, 7)：1024输入的通道数
:param n_groups: 8
:param n_channels_per_branch: [32,40,50,62]
:param is_dilated:
:param layer_num: 当前的层数
:return:
'''
n_channels_in = input_shape[1]
n_branches = 5
n_channels_per_group_in = int(n_channels_in /
n_groups) #每个group输入通道的个数 = 128
n_channels_out = int(n_groups * n_branches *
n_channels_per_branch) #整个temporal层的输出通道数
n_channels_per_group_out = int(n_channels_out /
n_groups) #每个group的输出通道的个数
assert n_channels_in % n_groups == 0 #断言表达式，否的话抛出异常
assert n_channels_out % n_groups == 0
# type of multi-scale kernels to use: either multi_kernel_sizes or multi_dilation_rates
# 设置膨胀与核大小已应对多变的复杂动作时长
if is_dilated:
kernel_sizes = (3, 3, 3)
dilation_rates = (1, 2, 3)
else:
kernel_sizes = (3, 5, 7)
dilation_rates = (1, 1, 1)
input_shape_per_group = list(input_shape)
input_shape_per_group[
1] = n_channels_per_group_in #(32, 128, 128, 7, 7)
# loop on groups, and define convolutions in each group
# 对每一层的每一个group进行定义具体的操作
for idx_group in range(n_groups):
group_num = idx_group + 1 #当前的group数
#定义每个group的操作
self.__define_temporal_convolutional_block(input_shape_per_group,
n_channels_per_branch,
kernel_sizes,
dilation_rates,
layer_num, group_num)
# activation 定义激活操作
layer_name = 'relu_tc%d' % (layer_num)
layer = ReLU()
layer._name = layer_name
setattr(self, layer_name, layer)
# shuffle channels 定义混洗操作
layer_name = 'shuffle_tc%d' % (layer_num)
layer = ChannelShuffleLayer(n_channels_out, n_groups)
layer._name = layer_name
setattr(self, layer_name, layer)
