the code in pytorch is here, I only used the vgg19 arch.
In order to make the cifar10 dataset preprocessed same in caffe and pytorch, I remove all the transforms in main.py but toTensor(). I found the cifar10 dataset's range is [0,1] in pytorch, but caffe is [0,255], so I scale 1/255 below. Any extra preprocessing is canceled to make things easier.
here is my caffe net definition prototxt:
layer {
name: "data"
type: "Data"
top: "data"
top: "label"
include {
phase: TRAIN
}
transform_param {
scale: 0.00392156862745
}
data_param {
source: "/home/snk/Documents/digits-jobs/20180106-135745-4cac/train_db"
batch_size: 128
backend: LMDB
}
}
layer {
name: "data"
type: "Data"
top: "data"
top: "label"
include {
phase: TEST
}
transform_param {
scale: 0.00392156862745
}
data_param {
source: "/home/snk/Documents/digits-jobs/20180106-135745-4cac/val_db"
batch_size: 128
backend: LMDB
}
}
layer {
name: "conv1_1"
type: "Convolution"
bottom: "data"
top: "conv1_1"
convolution_param {
num_output: 64
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu1_1"
type: "ReLU"
bottom: "conv1_1"
top: "conv1_1"
}
layer {
name: "conv1_2"
type: "Convolution"
bottom: "conv1_1"
top: "conv1_2"
convolution_param {
num_output: 64
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu1_2"
type: "ReLU"
bottom: "conv1_2"
top: "conv1_2"
}
layer {
name: "pool1"
type: "Pooling"
bottom: "conv1_2"
top: "pool1"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "conv2_1"
type: "Convolution"
bottom: "pool1"
top: "conv2_1"
convolution_param {
num_output: 128
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu2_1"
type: "ReLU"
bottom: "conv2_1"
top: "conv2_1"
}
layer {
name: "conv2_2"
type: "Convolution"
bottom: "conv2_1"
top: "conv2_2"
convolution_param {
num_output: 128
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu2_2"
type: "ReLU"
bottom: "conv2_2"
top: "conv2_2"
}
layer {
name: "pool2"
type: "Pooling"
bottom: "conv2_2"
top: "pool2"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "conv3_1"
type: "Convolution"
bottom: "pool2"
top: "conv3_1"
convolution_param {
num_output: 256
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu3_1"
type: "ReLU"
bottom: "conv3_1"
top: "conv3_1"
}
layer {
name: "conv3_2"
type: "Convolution"
bottom: "conv3_1"
top: "conv3_2"
convolution_param {
num_output: 256
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu3_2"
type: "ReLU"
bottom: "conv3_2"
top: "conv3_2"
}
layer {
name: "conv3_3"
type: "Convolution"
bottom: "conv3_2"
top: "conv3_3"
convolution_param {
num_output: 256
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu3_3"
type: "ReLU"
bottom: "conv3_3"
top: "conv3_3"
}
layer {
name: "conv3_4"
type: "Convolution"
bottom: "conv3_3"
top: "conv3_4"
convolution_param {
num_output: 256
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu3_4"
type: "ReLU"
bottom: "conv3_4"
top: "conv3_4"
}
layer {
name: "pool3"
type: "Pooling"
bottom: "conv3_4"
top: "pool3"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "conv4_1"
type: "Convolution"
bottom: "pool3"
top: "conv4_1"
convolution_param {
num_output: 512
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu4_1"
type: "ReLU"
bottom: "conv4_1"
top: "conv4_1"
}
layer {
name: "conv4_2"
type: "Convolution"
bottom: "conv4_1"
top: "conv4_2"
convolution_param {
num_output: 512
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu4_2"
type: "ReLU"
bottom: "conv4_2"
top: "conv4_2"
}
layer {
name: "conv4_3"
type: "Convolution"
bottom: "conv4_2"
top: "conv4_3"
convolution_param {
num_output: 512
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu4_3"
type: "ReLU"
bottom: "conv4_3"
top: "conv4_3"
}
layer {
name: "conv4_4"
type: "Convolution"
bottom: "conv4_3"
top: "conv4_4"
convolution_param {
num_output: 512
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu4_4"
type: "ReLU"
bottom: "conv4_4"
top: "conv4_4"
}
layer {
name: "pool4"
type: "Pooling"
bottom: "conv4_4"
top: "pool4"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "conv5_1"
type: "Convolution"
bottom: "pool4"
top: "conv5_1"
convolution_param {
num_output: 512
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu5_1"
type: "ReLU"
bottom: "conv5_1"
top: "conv5_1"
}
layer {
name: "conv5_2"
type: "Convolution"
bottom: "conv5_1"
top: "conv5_2"
convolution_param {
num_output: 512
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu5_2"
type: "ReLU"
bottom: "conv5_2"
top: "conv5_2"
}
layer {
name: "conv5_3"
type: "Convolution"
bottom: "conv5_2"
top: "conv5_3"
convolution_param {
num_output: 512
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu5_3"
type: "ReLU"
bottom: "conv5_3"
top: "conv5_3"
}
layer {
name: "conv5_4"
type: "Convolution"
bottom: "conv5_3"
top: "conv5_4"
convolution_param {
num_output: 512
pad: 1
kernel_size: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu5_4"
type: "ReLU"
bottom: "conv5_4"
top: "conv5_4"
}
layer {
name: "pool5"
type: "Pooling"
bottom: "conv5_4"
top: "pool5"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "drop6"
type: "Dropout"
bottom: "pool5"
top: "drop_pool5"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc6"
type: "InnerProduct"
bottom: "drop_pool5"
top: "fc6"
inner_product_param {
num_output: 512
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu6"
type: "ReLU"
bottom: "fc6"
top: "fc6"
}
layer {
name: "drop7"
type: "Dropout"
bottom: "fc6"
top: "fc6"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7"
type: "InnerProduct"
bottom: "fc6"
top: "fc7"
inner_product_param {
num_output: 512
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "relu7"
type: "ReLU"
bottom: "fc7"
top: "fc7"
}
layer {
name: "fc8"
type: "InnerProduct"
bottom: "fc7"
top: "fc8"
inner_product_param {
num_output: 10
weight_filler {
type: "xavier"
}
bias_filler {
type: "xavier"
}
}
}
layer {
name: "acc"
type: "Accuracy"
bottom: "fc8"
bottom: "label"
top: "accuracy"
include {
phase: TEST
}
}
layer {
name: "loss"
type: "SoftmaxWithLoss"
bottom: "fc8"
bottom: "label"
top: "loss"
}
Here is my solver.prototxt
test_iter: 79
test_interval: 391
base_lr: 0.05
display: 40
max_iter: 117300
lr_policy: "step"
gamma: 0.5
momentum: 0.9
weight_decay: 0.0005
stepsize: 11730
snapshot: 3910
snapshot_prefix: "snapshot"
solver_mode: GPU
net: "train_val.prototxt"
solver_type: SGD
The fact is pytorch can train the model at a very quick speed, the acc after one epoch is about 20%, but in caffe the loss is always around 2.3XXX (about -log(0.1), random guess loss), I suspect it was because the different weight initialization, so I changed caffe's filler.hpp's xavier to 'efficient backprop'(i.e U(-std, std), std=1/(sqrt(fan_in))), but it didn't work.
Now, the only difference is the bias initialization method, in pytorch it uses the weight's fan_in, but in caffe I think it uses the output_num as fan_in (because its shape is [1,N], N is the number of output neurons, and in filler.hpp it use blob.count() / blob.num() as fan_in for xavier ).
Who can help me ? I think if all the configuration is the same, the training process is almost the same ,but it breaked my opinion.