lightsalt2011 / bp-cpp Goto Github PK

// 前向部分
auto x           = choice_rows(train_norm_images,   image_indexs, ibatch * batch_size, batch_size);
auto y           = choice_rows(train_onehot_labels, image_indexs, ibatch * batch_size, batch_size);
auto hidden      = gemm_mul(x,          false, input_to_hidden,  false) + hidden_bias;
auto hidden_act  = hidden.relu();
auto output      = gemm_mul(hidden_act, false, hidden_to_output, false) + output_bias;
auto probability = output.sigmoid();
float loss       = compute_loss(probability, y);

// 反向部分
auto doutput           = (probability - y) / batch_size;
auto doutput_bias      = row_sum(doutput);
auto dhidden_to_output = gemm_mul(hidden_act, true, doutput, false);
auto dhidden_act       = gemm_mul(doutput, false, hidden_to_output, true);
auto dhidden           = delta_relu(dhidden_act, hidden);
auto dinput_to_hidden  = gemm_mul(x, true, dhidden, false);
auto dhidden_bias      = row_sum(dhidden);
optim.update_params(...);