Hi,
I got the different output between PC and latform in TDA4. This model can also work normally in TDA2.
The architecture is look like the figure below.
The difference is begin with third layer. This layer is merged by convolution and relu.
layer {
name: "down2/bd"
type: "Convolution"
bottom: "down2/bn"
top: "down2/bd"
param {
lr_mult: 0.0
decay_mult: 0.0
}
convolution_param {
num_output: 24
bias_term: false
pad: 1
kernel_size: 3
group: 24
stride: 2
}
}
layer {
name: "down2/bd_relu"
type: "ReLU"
bottom: "down2/bd"
top: "down2/bd"
}
There are some files about the net and param in the last.
I ask you a favor for making output become the same between PC and TDA4 platform.
Thanks,
