本部分剖析Caffe中Net::Backward()函數,即反向傳播計算過程。從LeNet網絡角度出發,且調試網絡爲訓練網絡,共9層網絡。
入口信息
Net::Backward()函數中調用BackwardFromTo函數,從網絡最後一層到網絡第一層反向調用每個網絡層的Backward。
void Net<Dtype>::BackwardFromTo(int start, int end) {
for (int i = start; i >= end; --i) {
if (layer_need_backward_[i]) {
layers_[i]->Backward(
top_vecs_[i], bottom_need_backward_[i], bottom_vecs_[i]);
if (debug_info_) { BackwardDebugInfo(i); }
}
}
}
第九層 SoftmaxWithLossLayer
代碼實現如下:
void SoftmaxWithLossLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
// bottom_diff shape:64*10
Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
// prob_data shape:64*10
const Dtype* prob_data = prob_.gpu_data();
// top_data shape:(1)
const Dtype* top_data = top[0]->gpu_data();
// 將Softmax層預測的結果prob複製到bottom_diff中
caffe_gpu_memcpy(prob_.count() * sizeof(Dtype), prob_data, bottom_diff);
// label shape:64*1
const Dtype* label = bottom[1]->gpu_data();
// dim = 640 / 64 = 10
const int dim = prob_.count() / outer_num_;
// nthreads = 64 / 1 = 64
const int nthreads = outer_num_ * inner_num_;
// Since this memory is never used for anything else,
// we use to to avoid allocating new GPU memory.
Dtype* counts = prob_.mutable_gpu_diff();
// 該函數將bottom_diff(此時爲每個類的預測概率)對應的正確類別(label)的概率值-1,其他數據沒變。見公式推導。
SoftmaxLossBackwardGPU<Dtype><<<CAFFE_GET_BLOCKS(nthreads),
CAFFE_CUDA_NUM_THREADS>>>(nthreads, top_data, label, bottom_diff,
outer_num_, dim, inner_num_, has_ignore_label_, ignore_label_, counts);
// 代碼展開開始,代碼有修改
__global__ void SoftmaxLossBackwardGPU(...) {
CUDA_KERNEL_LOOP(index, nthreads) {
const int label_value = static_cast<int>(label[index]);
bottom_diff[index * dim + label_value] -= 1;
counts[index] = 1;
}
}
// 代碼展開結束
Dtype valid_count = -1;
// 注意爲loss的權值,對該權值(一般爲1或者0)歸一化(除以64)
// Scale gradient
const Dtype loss_weight = top[0]->cpu_diff()[0];
if (normalize_) {
caffe_scal(prob_.count(), loss_weight / count, bottom_diff);
} else {
caffe_scal(prob_.count(), loss_weight / outer_num_, bottom_diff);
}
}
說明:
- SoftmaxWithLossLayer是沒有學習參數的,因此不需要對該層的參數做調整,只需要計算bottom_diff(理解反向傳播算法的鏈式求導,求bottom_diff對上一層的輸出求導,是爲了進一步計算調整上一層權值)
- 以上代碼核心部分在SoftmaxLossBackwardGPU。該函數將bottom_diff(此時爲每個類的預測概率)對應的正確類別(label)的概率值-1,其他數據沒變。
第八層 InnerProduct
template <typename Dtype>
void InnerProductLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down,
const vector<Blob<Dtype>*>& bottom) {
//對參數求偏導,top_diff*bottom_data=blobs_diff
// 注意,此處(Dtype)1., this->blobs_[0]->mutable_gpu_diff()
// 中的(Dtype)1.:使得在一個solver的iteration中的多個iter_size
// 的梯度沒有清零,而得以累加
if (this->param_propagate_down_[0]) {
const Dtype* top_diff = top[0]->gpu_diff();
const Dtype* bottom_data = bottom[0]->gpu_data();
// Gradient with respect to weight
caffe_gpu_gemm<Dtype>(CblasTrans, CblasNoTrans, N_, K_, M_, (Dtype)1.,
top_diff, bottom_data, (Dtype)1., this->blobs_[0]->mutable_gpu_diff());
}
// 對偏置求偏導top_diff*bias=blobs_diff
if (bias_term_ && this->param_propagate_down_[1]) {
const Dtype* top_diff = top[0]->gpu_diff();
// Gradient with respect to bias
caffe_gpu_gemv<Dtype>(CblasTrans, M_, N_, (Dtype)1., top_diff,
bias_multiplier_.gpu_data(), (Dtype)1.,
this->blobs_[1]->mutable_gpu_diff());
}
//對上一層輸出求偏導top_diff*blobs_data=bottom_diff
if (propagate_down[0]) {
const Dtype* top_diff = top[0]->gpu_diff();
// Gradient with respect to bottom data
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, M_, K_, N_, (Dtype)1.,
top_diff, this->blobs_[0]->gpu_data(), (Dtype)0.,
bottom[0]->mutable_gpu_diff());
}
}
第七層 ReLU
cpu代碼分析如下,注,該層沒有參數,只需對輸入求導
void ReLULayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down,
const vector<Blob<Dtype>*>& bottom) {
if (propagate_down[0]) {
const Dtype* bottom_data = bottom[0]->cpu_data();
const Dtype* top_diff = top[0]->cpu_diff();
Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
const int count = bottom[0]->count();
//見公式推導
Dtype negative_slope = this->layer_param_.relu_param().negative_slope();
for (int i = 0; i < count; ++i) {
bottom_diff[i] = top_diff[i] * ((bottom_data[i] > 0)
+ negative_slope * (bottom_data[i] <= 0));
}
}
}
公式推導
第五層 Pooling
Maxpooling的cpu代碼分析如下,注,該層沒有參數,只需對輸入求導
void PoolingLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
const Dtype* top_diff = top[0]->cpu_diff();
Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
// bottom_diff初始化置0
caffe_set(bottom[0]->count(), Dtype(0), bottom_diff);
const int* mask = NULL; // suppress warnings about uninitialized variables
...
// 在前向計算時max_idx中保存了top_data中的點是有bottom_data中的點得來的在該feature map中的座標
mask = max_idx_.cpu_data();
// 主循環,按(N,C,H,W)方式便利top_data中每個點
for (int n = 0; n < top[0]->num(); ++n) {
for (int c = 0; c < channels_; ++c) {
for (int ph = 0; ph < pooled_height_; ++ph) {
for (int pw = 0; pw < pooled_width_; ++pw) {
const int index = ph * pooled_width_ + pw;
const int bottom_index = mask[index];
// 見公式推導
bottom_diff[bottom_index] += top_diff[index];
}
}
bottom_diff += bottom[0]->offset(0, 1);
top_diff += top[0]->offset(0, 1);
mask += top[0]->offset(0, 1);
}
}
}
第四層 Convolution
void ConvolutionLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
const Dtype* weight = this->blobs_[0]->cpu_data();
Dtype* weight_diff = this->blobs_[0]->mutable_cpu_diff();
for (int i = 0; i < top.size(); ++i) {
const Dtype* top_diff = top[i]->cpu_diff();
const Dtype* bottom_data = bottom[i]->cpu_data();
Dtype* bottom_diff = bottom[i]->mutable_cpu_diff();
// Bias gradient, if necessary.
if (this->bias_term_ && this->param_propagate_down_[1]) {
Dtype* bias_diff = this->blobs_[1]->mutable_cpu_diff();
// 對於每個Batch中的樣本,計算偏置的偏導
for (int n = 0; n < this->num_; ++n) {
this->backward_cpu_bias(bias_diff, top_diff + n * this->top_dim_);
}
}
if (this->param_propagate_down_[0] || propagate_down[i]) {
// 對於每個Batch中的樣本,關於權值及輸入求導部分代碼展開了函數(非可運行代碼)
for (int n = 0; n < this->num_; ++n) {
// gradient w.r.t. weight. Note that we will accumulate diffs.
//top_diff(50*64) * bottom_data(500*64,Transpose) = weight_diff(50*500)
// 注意,此處(Dtype)1., this->blobs_[0]->mutable_gpu_diff()
// 中的(Dtype)1.:使得在一個solver的iteration中的多個iter_size
// 的梯度沒有清零,而得以累加
caffe_cpu_gemm<Dtype>(CblasNoTrans, CblasTrans, conv_out_channels_ / group_,
kernel_dim_, conv_out_spatial_dim_,
(Dtype)1., top_diff + n * this->top_dim_, bottom_data + n * this->bottom_dim_,
(Dtype)1., weight_diff);
// gradient w.r.t. bottom data, if necessary.
// weight(50*500,Transpose) * top_diff(50*64) = bottom_diff(500*64)
caffe_cpu_gemm<Dtype>(CblasTrans, CblasNoTrans, kernel_dim_,
conv_out_spatial_dim_, conv_out_channels_ ,
(Dtype)1., weight, top_diff + n * this->top_dim_,
(Dtype)0., bottom_diff + n * this->bottom_dim_);
}
}
}
}
- 第四層的bottom維度(N,C,H,W)=(64,20,12,12),top的維度bottom維度(N,C,H,W)=(64,50,8,8),由於每個樣本單獨處理,所以只需要關注(C,H,W)的維度,分別爲(20,12,12)和(50,8,8)
- 根據(Caffe)卷積的實現,該層可以寫成矩陣相乘的形式Weight_data×Bottom_dataT=Top_data
- Weight_data的維度爲Cout×(C∗K∗K)=50×500
- Bottom_data的維度爲(H∗W)×(C∗K∗K)=64×500,64爲8∗8個卷積核的位置,500=C∗K∗K=20∗5∗5
Top_data的維度爲64×50 - 寫成矩陣表示後,從某種角度上與全連接從(也是表示成矩陣相乘)相同,因此,可以借鑑全連接層的推導。