The Beautiful Future
caffe conv layer 본문
### LayerSetUp
channel_axis_ = 1 // n c h w
first_spatial_axis = channel_axis_ + 1 // h , w
num_spatial_axes_ = 4 // n c h w
num_output_ = num of filters
conv_out_channels_ = num_output_
weight_shape0 = conv_out_channels_
weight_shape1 = conv_in_channels_ / group_
weight_shape2 = kernel_h
weight_shape3 = kernel_w
conv_out_spatial_dim_ = top_h * top_w
kernel_dim_ = kernel_c *kernel_h * kernel_w
weight_offset_ = conv_out_channels_ * kernel_dim_ / group_
col_offset_ = conv_out_spatial_dim_ * kernel_dim_
output_offset_ = conv_out_channels_ * conv_out_spatial_dim_ / group_
col_buffer_shape_ = kernel_dim_ x output_h x output_w
top_dim = top_c * top_h * top_w
bottom_dim = bottom_c * bottom_h * bottom_w
num_kernels_im2col_ = conv_in_channels_ * conv_out_spatial_dim_
num_kernels_col2im_ = bottom_dim_
out_spatial_dim_ = out_h * out_w
## offset
weight_offset_ = conv_out_channels_ * kernel_dim_ / group_
col_offset_ = conv_out_spatial_dim_ * kernel_dim_ // for column buffer
output_offset_ = conv_out_channels_ * conv_out_spatial_dim_ / group_
'DNN' 카테고리의 다른 글
| TensorFlow CIFAR10 Example 분석1, Write TFRecord (0) | 2018.04.13 |
|---|---|
| transpose convolution visualization (0) | 2018.02.05 |
| minimal filtering algorithm, Shmuel Winograd (0) | 2017.10.12 |
| Adam (0) | 2017.05.24 |
| Installing Tensorflow on Windows7 from Sources (0) | 2017.03.16 |