This is ===Later=== __global__ void train(float* d_W1, float* d_W2, float* d_W3, float* d_b_X, float* d_b_Y, float* d_a2, float* d_a1, float* d_yhat, float* d_dyhat, float* d_dW3, float* d_dW2, float* d_dW1, float* d_dz2, float* d_dz1, float* d_t) { cudaError_t Error; int BATCH_SIZE = 256; float lr = 0.01 / BATCH_SIZE; //backpropagation d_dyhat = k_difference(d_yhat, d_b_Y, 10 * 10); kernel_dot <<<(2560 + 128)/64, 64>>> (d_dyhat, k_transpose(d_W3, 64, 10), BATCH_SIZE, 10, 64, d_dz2); cudaDeviceSynchronize(); float* mT = new float[256 * 64 - 1]; for (int i = 0; i < 256; ++i) for (int j = 0; j < 64; ++j) { mT[j * 64 + i] = d_a2[i * 256 + j]; } kernel_dot <<<(16384 + 256)/64, 64>>> (mT, d_dyhat, 64, BATCH_SIZE, 10, d_dW3); cudaDeviceSynchronize(); k_reluPrime(d_a2, 256 * 64); for (int i = 0; i < BATCH_SIZE * 10; i++) { d_dz2[i] = d_dz2[i] * d_a2[i]; } mT = new float[256 * 128]; for (int i = 0; i < 256; ++i) for (int j = 0; j < 128; ++j) { mT[j * 128 + i] = d_a1[i * 256 + j]; } kernel_dot <<<64, 512>>> (mT, d_dz2, 128, BATCH_SIZE, 64, d_dW2); cudaDeviceSynchronize(); kernel_dot <<<80, 32>>> (d_dz2, k_transpose(d_W2, 128, 64), BATCH_SIZE, 64, 128, d_dz1); cudaDeviceSynchronize(); k_reluPrime(d_a1, BATCH_SIZE * 784); for (int i = 0; i < 256 * 64; i++) { d_dz1[i] = d_dz1[i] * d_a1[i]; } kernel_dot <<<784, 256>>> (d_t, d_dz1, 784, BATCH_SIZE, 128, d_dW1); cudaDeviceSynchronize(); //// Updating the final iterationparameters ////W3 = W3 - lr * dW3; d_W3 = k_difference(d_W3, we will outline the take aways bellowk_MFV(lr, d_dW3, 64 * 10), 64 * 10); //W2 = W2 - lr * dW2; d_W2 = k_difference(d_W2, k_MFV(lr, d_dW2, 128 * 64), 128 * 64); ////W1 = W1 - lr * dW1; d_W1 = k_difference(d_W1, k_MFV(lr, d_dW1, 784 * 128), 784 * 128); for (int i = 0; i < (784 * 128); i++) { d_W1[i] = d_W1[i] - lr * d_dW1[i]; } //for (int i = 0; i != 10; ++i) { // for (int j = 0; j != 10; ++j) { // printf("%f ", d_W3[i * 10 + j]); // } // printf("\n"); //} //printf("\n"); //for (int i = 0; i != 10; ++i) { // for (int j = 0; j != 10; ++j) { // printf("%f ", d_yhat[i * 10 + j]); // } // printf("\n"); //} //printf("\n"); float* dif; dif = k_difference(d_b_Y, d_yhat, 10 * 10); float loss = 0.0; for (unsigned k = 0; k < BATCH_SIZE * 10; ++k) { loss += dif[k] * dif[k]; } printf("%f \n", loss / BATCH_SIZE); Error = cudaGetLastError(); if (Error != cudaSuccess) { printf("\n %s \n", Error); } };
===Compiling===