问题
Assuming we have a layer like this:
layer {
name: "fully-connected"
type: "InnerProduct"
bottom: "bottom"
top: "top"
inner_product_param {
num_output: 1
}
}
The output is batch_size x 1. In several papers (for exmaple link1 page 3 picture on the top, or link2 page 4 on top)I have seen that they used such a layer in the end to come up with a 2D image for pixel-wise prediction. How is it possible to transform this into a 2D image? I was thinking of reshape or deconvolution, but I cannot figure out how that would work. A simple example would be helpful
UPDATE: My input images are 304x228 and my ground_truth (depth images) are 75x55.
################# Main net ##################
layer {
name: "conv1"
type: "Convolution"
bottom: "data"
top: "conv1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 96
kernel_size: 11
stride: 4
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu1"
type: "ReLU"
bottom: "conv1"
top: "conv1"
}
layer {
name: "norm1"
type: "LRN"
bottom: "conv1"
top: "norm1"
lrn_param {
local_size: 5
alpha: 0.0001
beta: 0.75
}
}
layer {
name: "pool1"
type: "Pooling"
bottom: "norm1"
top: "pool1"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
layer {
name: "conv2"
type: "Convolution"
bottom: "pool1"
top: "conv2"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 256
pad: 2
kernel_size: 5
group: 2
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu2"
type: "ReLU"
bottom: "conv2"
top: "conv2"
}
layer {
name: "norm2"
type: "LRN"
bottom: "conv2"
top: "norm2"
lrn_param {
local_size: 5
alpha: 0.0001
beta: 0.75
}
}
layer {
name: "pool2"
type: "Pooling"
bottom: "norm2"
top: "pool2"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
layer {
name: "conv3"
type: "Convolution"
bottom: "pool2"
top: "conv3"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 384
pad: 1
kernel_size: 3
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu3"
type: "ReLU"
bottom: "conv3"
top: "conv3"
}
layer {
name: "conv4"
type: "Convolution"
bottom: "conv3"
top: "conv4"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 384
pad: 1
kernel_size: 3
group: 2
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu4"
type: "ReLU"
bottom: "conv4"
top: "conv4"
}
layer {
name: "conv5"
type: "Convolution"
bottom: "conv4"
top: "conv5"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 256
pad: 1
kernel_size: 3
group: 2
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relu5"
type: "ReLU"
bottom: "conv5"
top: "conv5"
}
layer {
name: "pool5"
type: "Pooling"
bottom: "conv5"
top: "pool5"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
layer {
name: "fc6"
type: "InnerProduct"
bottom: "pool5"
top: "fc6"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
name: "relufc6"
type: "ReLU"
bottom: "fc6"
top: "fc6"
}
layer {
name: "drop6"
type: "Dropout"
bottom: "fc6"
top: "fc6"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7"
type: "InnerProduct"
bottom: "fc6"
top: "fc7"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 4070
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 0.1
}
}
}
layer {
type: "Reshape"
name: "reshape"
bottom: "fc7"
top: "fc7_reshaped"
reshape_param {
shape { dim: 1 dim: 1 dim: 55 dim: 74 }
}
}
layer {
name: "deconv1"
type: "Deconvolution"
bottom: "fc7_reshaped"
top: "deconv1"
convolution_param {
num_output: 64
kernel_size: 5
pad: 2
stride: 1
#group: 256
weight_filler {
type: "bilinear"
}
bias_term: false
}
}
#########################
layer {
name: "conv6"
type: "Convolution"
bottom: "data"
top: "conv6"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 63
kernel_size: 9
stride: 2
pad: 1
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu6"
type: "ReLU"
bottom: "conv6"
top: "conv6"
}
layer {
name: "pool6"
type: "Pooling"
bottom: "conv6"
top: "pool6"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
########################
layer {
name: "concat"
type: "Concat"
bottom: "deconv1"
bottom: "pool6"
top: "concat"
concat_param {
concat_dim: 1
}
}
layer {
name: "conv7"
type: "Convolution"
bottom: "concat"
top: "conv7"
convolution_param {
num_output: 64
kernel_size: 5
pad: 2
stride: 1
weight_filler {
type: "gaussian"
std: 0.011
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu7"
type: "ReLU"
bottom: "conv7"
top: "conv7"
relu_param{
negative_slope: 0.01
engine: CUDNN
}
}
layer {
name: "conv8"
type: "Convolution"
bottom: "conv7"
top: "conv8"
convolution_param {
num_output: 64
kernel_size: 5
pad: 2
stride: 1
weight_filler {
type: "gaussian"
std: 0.011
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu8"
type: "ReLU"
bottom: "conv8"
top: "conv8"
relu_param{
negative_slope: 0.01
engine: CUDNN
}
}
layer {
name: "conv9"
type: "Convolution"
bottom: "conv8"
top: "conv9"
convolution_param {
num_output: 1
kernel_size: 5
pad: 2
stride: 1
weight_filler {
type: "gaussian"
std: 0.011
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu9"
type: "ReLU"
bottom: "conv9"
top: "result"
relu_param{
negative_slope: 0.01
engine: CUDNN
}
}
log:
I1108 19:34:57.239722 4277 data_layer.cpp:41] output data size: 1,1,228,304
I1108 19:34:57.243340 4277 data_layer.cpp:41] output data size: 1,1,55,74
I1108 19:34:57.247392 4277 net.cpp:150] Setting up conv1
I1108 19:34:57.247407 4277 net.cpp:157] Top shape: 1 96 55 74 (390720)
I1108 19:34:57.248191 4277 net.cpp:150] Setting up pool1
I1108 19:34:57.248196 4277 net.cpp:157] Top shape: 1 96 27 37 (95904)
I1108 19:34:57.253263 4277 net.cpp:150] Setting up conv2
I1108 19:34:57.253276 4277 net.cpp:157] Top shape: 1 256 27 37 (255744)
I1108 19:34:57.254202 4277 net.cpp:150] Setting up pool2
I1108 19:34:57.254220 4277 net.cpp:157] Top shape: 1 256 13 18 (59904)
I1108 19:34:57.269943 4277 net.cpp:150] Setting up conv3
I1108 19:34:57.269961 4277 net.cpp:157] Top shape: 1 384 13 18 (89856)
I1108 19:34:57.285303 4277 net.cpp:150] Setting up conv4
I1108 19:34:57.285338 4277 net.cpp:157] Top shape: 1 384 13 18 (89856)
I1108 19:34:57.294801 4277 net.cpp:150] Setting up conv5
I1108 19:34:57.294841 4277 net.cpp:157] Top shape: 1 256 13 18 (59904)
I1108 19:34:57.295207 4277 net.cpp:150] Setting up pool5
I1108 19:34:57.295210 4277 net.cpp:157] Top shape: 1 256 6 9 (13824)
I1108 19:34:57.743222 4277 net.cpp:150] Setting up fc6
I1108 19:34:57.743259 4277 net.cpp:157] Top shape: 1 4096 (4096)
I1108 19:34:57.881680 4277 net.cpp:150] Setting up fc7
I1108 19:34:57.881718 4277 net.cpp:157] Top shape: 1 4070 (4070)
I1108 19:34:57.881826 4277 net.cpp:150] Setting up reshape
I1108 19:34:57.881846 4277 net.cpp:157] Top shape: 1 1 55 74 (4070)
I1108 19:34:57.884768 4277 net.cpp:150] Setting up conv6
I1108 19:34:57.885309 4277 net.cpp:150] Setting up pool6
I1108 19:34:57.885327 4277 net.cpp:157] Top shape: 1 63 55 74 (256410)
I1108 19:34:57.885395 4277 net.cpp:150] Setting up concat
I1108 19:34:57.885412 4277 net.cpp:157] Top shape: 1 64 55 74 (260480)
I1108 19:34:57.886759 4277 net.cpp:150] Setting up conv7
I1108 19:34:57.886786 4277 net.cpp:157] Top shape: 1 64 55 74 (260480)
I1108 19:34:57.897269 4277 net.cpp:150] Setting up conv8
I1108 19:34:57.897303 4277 net.cpp:157] Top shape: 1 64 55 74 (260480)
I1108 19:34:57.899129 4277 net.cpp:150] Setting up conv9
I1108 19:34:57.899138 4277 net.cpp:157] Top shape: 1 1 55 74 (4070)
回答1:
The value of num_output of the last fully connected layer will not be 1 for pixel wise prediction. It will be equal to w*h of the input image.
What made you feel that the value will be 1?
Edit 1:
Below are the dimensions of each layer mentioned in link1 page 3 figure:
LAYER OUTPUT DIM [c*h*w]
course1 96*h1*w1 conv layer
course2 256*h2*w2 conv layer
course3 384*h3*w3 conv layer
course4 384*h4*w4 conv layer
course5 256*h5*w5 conv layer
course6 4096*1*1 fc layer
course7 X*1*1 fc layer where 'X' could be interpreted as w*h
To understand this further, lets assume we have a network to predict the pixels of the image. The images are of size 10*10. Thus, the final output of the fc layer too will be having the dimension 100*1*1(like in course7). This could be interpreted as 10*10.
Now the question will be, how can the 1d array predict a 2d image correctly. For this, you have to note that the loss is calculated for this output, using the labels which could be corresponding to the pixel data. Thus during training, the weights will learn to predict the pixel data.
EDIT 2:
Trying to draw the net using draw_net.py in caffe, gives you this:
The relu layer connected with conv6 and fc6 has the same name, leading to a complicated connectivity in the drawn image. I am not sure on whether this will cause some issues during training, but I would suggest you to rename one of the relu layers to a unique name to avoid some unforseen issues.
Coming back to your question, there doesn't seem to be an upsampling happening after fully connected layers. As seen in the log:
I1108 19:34:57.881680 4277 net.cpp:150] Setting up fc7
I1108 19:34:57.881718 4277 net.cpp:157] Top shape: 1 4070 (4070)
I1108 19:34:57.881826 4277 net.cpp:150] Setting up reshape
I1108 19:34:57.881846 4277 net.cpp:157] Top shape: 1 1 55 74 (4070)
I1108 19:34:57.884768 4277 net.cpp:150] Setting up conv6
I1108 19:34:57.885309 4277 net.cpp:150] Setting up pool6
I1108 19:34:57.885327 4277 net.cpp:157] Top shape: 1 63 55 74 (256410)
fc7 has output dimension of 4070*1*1. This is being reshaped to 1*55*74 to be passed as an input to conv6 layer.
The output of the whole network is produced in conv9, which has an output dimension of 1*55*74, which is exactly similar to the dimension of the labels (depth data).
Please do pinpoint on where you feel the upsample is happening, if my answer is still not clear.
回答2:
if you simply need fully-connected networks like the conventional multi-layer perceptron, use 2D blobs (shape (N, D)) and call the InnerProductLayer.
来源：https://stackoverflow.com/questions/40483458/caffe-reshape-upsample-fully-connected-layer