The dimensionality of the last conv's output is 6*6*30，but how can we reshape it to 1470。

[sysuzyq]

layer {
name: "conv_reg"
type: "Convolution"
bottom: "add_conv2"
top: "conv_reg"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 30 ################output :6 * 6 * 30
kernel_size: 3
stride: 1
pad: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0.0
}
}
}
layer {
name: "reg_reshape"
type: "Reshape"
bottom: "conv_reg"
top: "regression"
reshape_param {
axis: 1
shape {
dim: 1470 #############but here is 7 * 7 * 30, whish is (classes+num_object*5) * side *side
}
}
}
Your input is 448 * 448, but the feature map of last conv layer is 6 * 6, then the output should be 6 * 6 * 30 = 1080. So should we reshape to 1080 instead of 1470?
look forward to your reply, thanks

[ICTwangbiao]

total stride=64, then the last feature map size is 448/64=7, so 7730=1470.
side = weight (or height) / stride

[sysuzyq]

@ICTwangbiao thanks for your reply, but it confuses me. when i use the equation " h_o = (h_i + 2 * pad_h - kernel_h) / stride_h +1 " to calculate the output's side layer by layer, finally i get 6 . what's wrong with me? thanks

[quhezheng]

@ICTwangbiao (classes+num_object*5) * side *side comes from the paper.
But this code use 7 instead of 30 to represent one cell. https://github.com/yeahkun/caffe-yolo/blob/master/src/caffe/layers/box_data_layer.cpp#L142 make it very clear.:

CHECK_EQ(count, locations * 7)

Here locations is 7*7=49. So dim: 1470 is wrong here. I don't know why the code can run. The right should be dim: 343

[ICTwangbiao]

@quhezheng "CHECK_EQ(count, locations * 7)" is used to check your ground truth (of course, the number of your labels should be locations * 7 {class_LABEL , difficult, isobj, x, y, w ,h}), so 343 your mentioned is the number of your ground truth.
BUT 1470 is the number of your regression layer outputs: locations * (class_NUMBER + score_confidence + coordinate infos)
In @sysuzyq 's case, I guess he/she want detect 25 classes if num_object=1, so outputs = locations * (25+5)