OrgLion-ML
/
darknet
mirror of https://github.com/AlexeyAB/darknet.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Another one fix of gradient for BiFPN

This commit is contained in:
AlexeyAB 2020-02-24 20:01:53 +03:00
parent cfa40fe890
commit 3cb9125b95
1 changed files with 10 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
src/blas_kernels.cu
View File

@ -943,11 +943,15 @@ __global__ void backward_shortcut_multilayer_kernel(int size, int src_outputs, i
if (weights_normalizion == RELU_NORMALIZATION) w = relu(w) / sum;
else if (weights_normalizion == SOFTMAX_NORMALIZATION) w = expf(w - max_val) / sum;
if (weights_normalizion == RELU_NORMALIZATION) grad = w;
else if (weights_normalizion == SOFTMAX_NORMALIZATION) grad = w*(1-w);
delta_out[id] += delta_in[id] * w; // [0 or c or (c, h ,w)]
float weights_update_tmp = delta_in[id] * in[id] * grad;
if (!isnan(weights_update_tmp) && !isinf(weights_update_tmp))
weight_updates_gpu[src_i / step] += weights_update_tmp;
atomicAdd(&weight_updates_gpu[src_i / step], weights_update_tmp);
//weight_updates_gpu[src_i / step] += weights_update_tmp;
}
else delta_out[id] += delta_in[id];
@ -967,11 +971,15 @@ __global__ void backward_shortcut_multilayer_kernel(int size, int src_outputs, i
if (weights_normalizion == RELU_NORMALIZATION) w = relu(w) / sum;
else if (weights_normalizion == SOFTMAX_NORMALIZATION) w = expf(w - max_val) / sum;
if (weights_normalizion == RELU_NORMALIZATION) grad = w;
else if (weights_normalizion == SOFTMAX_NORMALIZATION) grad = w*(1 - w);
layer_delta[add_index] += delta_in[id] * w;
float weights_update_tmp = delta_in[id] * add[add_index] * grad;
if (!isnan(weights_update_tmp) && !isinf(weights_update_tmp))
weight_updates_gpu[weights_index] += weights_update_tmp;
atomicAdd(&weight_updates_gpu[weights_index], weights_update_tmp);
//weight_updates_gpu[weights_index] += weights_update_tmp;
}
else layer_delta[add_index] += delta_in[id];
}