From 23cb35e6c8eae8b59fab161036ae3f417a55c8db Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Fri, 30 Mar 2018 11:46:51 +0000
Subject: [PATCH] Changed small_object
---
src/convolutional_kernels.cu | 154 +++++++++++++++++++++++++++++++++++++++++----------
1 files changed, 123 insertions(+), 31 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index ee9b534..603d531 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -37,7 +37,7 @@
int i = 0;
float mean = 0;
for(i = 0; i < n; ++i){
- mean += abs(input[i*size + s]);
+ mean += fabs(input[i*size + s]);
}
mean = mean / n;
for(i = 0; i < n; ++i){
@@ -59,7 +59,7 @@
int i = 0;
float mean = 0;
for(i = 0; i < size; ++i){
- mean += abs(weights[f*size + i]);
+ mean += fabs(weights[f*size + i]);
}
mean = mean / size;
for(i = 0; i < size; ++i){
@@ -78,22 +78,22 @@
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < size) output_f16[idx] = __float2half(input_f32[idx]);
- //if (idx < size) *((unsigned int *)output_f16 + idx) = __float2half(input_f32[idx]);
+ //if (idx < size) *((unsigned short *)output_f16 + idx) = __float2half(input_f32[idx]);
}
-void cuda_convert_f32_to_f16(float* input_f32, size_t size, half *output_f16) {
- cuda_f32_to_f16 <<< size / BLOCK + 1, BLOCK, 0, get_cuda_stream() >>> (input_f32, size, output_f16);
+void cuda_convert_f32_to_f16(float* input_f32, size_t size, float *output_f16) {
+ cuda_f32_to_f16 <<< size / BLOCK + 1, BLOCK, 0, get_cuda_stream() >>> (input_f32, size, (half *)output_f16);
}
__global__ void cuda_f16_to_f32(half* input_f16, size_t size, float *output_f32)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < size) output_f32[idx] = __half2float(input_f16[idx]);
- //if (idx < size) output_f32[idx] = __half2float(*((unsigned int *)input_f16 + idx));
+ //if (idx < size) output_f32[idx] = __half2float(*((unsigned short *)input_f16 + idx));
}
-void cuda_convert_f16_to_f32(half* input_f16, size_t size, float *output_f32) {
- cuda_f16_to_f32 <<< size / BLOCK + 1, BLOCK, 0, get_cuda_stream() >>> (input_f16, size, output_f32);
+void cuda_convert_f16_to_f32(float* input_f16, size_t size, float *output_f32) {
+ cuda_f16_to_f32 <<< size / BLOCK + 1, BLOCK, 0, get_cuda_stream() >>> ((half *)input_f16, size, output_f32);
}
half *cuda_make_f16_from_f32_array(float *src, size_t n)
@@ -102,7 +102,7 @@
size_t size = sizeof(half)*n;
check_error(cudaMalloc((void **)&dst16, size));
if (src) {
- cuda_convert_f32_to_f16(src, n, dst16);
+ cuda_convert_f32_to_f16(src, n, (float *)dst16);
}
if (!dst16) error("Cuda malloc failed\n");
return dst16;
@@ -124,8 +124,8 @@
}
#ifdef CUDNN
- //float one = 1; // alpha[0], beta[0] is float for HALF and FLOAT
- float alpha = 1, beta = 0;
+ float one = 1; // alpha[0], beta[0] is float for HALF and FLOAT
+ float alpha = 1, beta = 0;
#ifdef CUDNN_HALF
// Note: For improved performance it is advised to use beta[0] = 0.0.
@@ -135,27 +135,26 @@
// More: http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#tensor_ops
const size_t input16_size = l.batch*l.c*l.w*l.h;
- static size_t max_input16_size = input16_size;
- static half* input16 = cuda_make_f16_from_f32_array(NULL, max_input16_size);
-
const size_t output16_size = l.batch*l.out_c*l.out_h*l.out_w;
- static size_t max_output16_size = output16_size;
- static half* output16 = cuda_make_f16_from_f32_array(NULL, max_output16_size);
- if (max_input16_size < input16_size) {
- max_input16_size = input16_size;
- cuda_free((float *)input16);
- input16 = cuda_make_f16_from_f32_array(state.input, max_input16_size);
+ if (*state.net.max_input16_size < input16_size) {
+ //printf("\n input16_size: cur = %zu \t max = %zu \n", input16_size, *state.net.max_input16_size);
+ *state.net.max_input16_size = input16_size;
+ if (*state.net.input16_gpu) cuda_free(*state.net.input16_gpu);
+ *state.net.input16_gpu = (float *)cuda_make_f16_from_f32_array(NULL, *state.net.max_input16_size);
}
+ float *input16 = *state.net.input16_gpu;
- if (max_output16_size < output16_size) {
- max_output16_size = output16_size;
- cuda_free((float *)output16);
- output16 = cuda_make_f16_from_f32_array(NULL, max_output16_size);
+ if (*state.net.max_output16_size < output16_size) {
+ *state.net.max_output16_size = output16_size;
+ if (*state.net.output16_gpu) cuda_free(*state.net.output16_gpu);
+ *state.net.output16_gpu = (float *)cuda_make_f16_from_f32_array(NULL, *state.net.max_output16_size);
}
+ float *output16 = *state.net.output16_gpu;
cuda_convert_f32_to_f16(state.input, input16_size, input16);
+ //fill_ongpu(output16_size / 2, 0, (float *)output16, 1);
cudnnConvolutionForward(cudnn_handle(),
&alpha,
l.srcTensorDesc,
@@ -171,10 +170,11 @@
output16);
cuda_convert_f16_to_f32(output16, output16_size, l.output_gpu);
+
#else
cudnnConvolutionForward(cudnn_handle(),
- &alpha,
+ &one,
l.srcTensorDesc,
state.input,
l.weightDesc,
@@ -183,7 +183,7 @@
l.fw_algo,
state.workspace,
l.workspace_size,
- &beta,
+ &one,
l.dstTensorDesc,
l.output_gpu);
#endif
@@ -205,8 +205,10 @@
if (l.batch_normalize) {
forward_batchnorm_layer_gpu(l, state);
- }
- add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.n, l.out_w*l.out_h);
+ }
+ else {
+ add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.n, l.out_w*l.out_h);
+ }
activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
//if(l.dot > 0) dot_error_gpu(l);
@@ -230,7 +232,85 @@
if(l.xnor) state.input = l.binary_input_gpu;
#ifdef CUDNN
- float one = 1;
+ float one = 1;
+ float alpha = 1, beta = 0;
+
+#ifdef CUDNN_HALF
+
+ const size_t input16_size = l.batch*l.c*l.w*l.h;
+ const size_t delta16_size = l.batch*l.n*l.out_w*l.out_h;
+
+ if (*state.net.max_input16_size < input16_size) {
+ *state.net.max_input16_size = input16_size;
+ if(*state.net.input16_gpu) cuda_free(*state.net.input16_gpu);
+ *state.net.input16_gpu = (float *)cuda_make_f16_from_f32_array(NULL, *state.net.max_input16_size);
+ }
+ float *input16 = *state.net.input16_gpu;
+
+ if (*state.net.max_output16_size < delta16_size) {
+ *state.net.max_output16_size = delta16_size;
+ if(*state.net.output16_gpu) cuda_free(*state.net.output16_gpu);
+ *state.net.output16_gpu = (float *)cuda_make_f16_from_f32_array(NULL, *state.net.max_output16_size);
+ }
+ float *delta16 = *state.net.output16_gpu;
+
+ cuda_convert_f32_to_f16(state.input, input16_size, input16);
+ cuda_convert_f32_to_f16(l.delta_gpu, delta16_size, delta16);
+
+ // convert input: state.input (x), l.delta_gpu (y) from fp32 to fp16
+ // get output: l.weight_updates_gpu (dw) and convert it to fp32 (ONLY if it is fp16)
+
+ // calculate conv weight updates
+ // Already: l.weight_updates_gpu = (l.weight_updates_gpu - l.weight*decay*batch*subdivision)*momentum
+ // so we should copy f32 to f16, or compute: f16=(w_up - w*d*b*s)*m
+ cuda_convert_f32_to_f16(l.weight_updates_gpu, l.c*l.n*l.size*l.size, l.weight_updates_gpu16);
+
+ cudnnConvolutionBackwardFilter(cudnn_handle(),
+ &one,
+ l.srcTensorDesc,
+ input16, //state.input,
+ l.ddstTensorDesc,
+ delta16, //l.delta_gpu,
+ l.convDesc,
+ l.bf_algo,
+ state.workspace,
+ l.workspace_size,
+ &one,
+ l.dweightDesc,
+ l.weight_updates_gpu16); // l.weight_updates_gpu);
+
+ cuda_convert_f16_to_f32(l.weight_updates_gpu16, l.c*l.n*l.size*l.size, l.weight_updates_gpu);
+
+ if (state.delta) {
+ if (l.binary || l.xnor) swap_binary(&l);
+
+ // http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#cudnnConvolutionBackwardData
+ // calculate delta for the next layer
+ // convert input: l.weights_gpu (w), l.delta_gpu (dy) from fp32 to fp16
+ // get output: state.delta (dx) and convert it to fp32 (ONLY if it is fp16)
+ cudnnConvolutionBackwardData(cudnn_handle(),
+ &alpha,
+ l.weightDesc,
+ l.weights_gpu16, //l.weights_gpu,
+ l.ddstTensorDesc,
+ delta16, //l.delta_gpu,
+ l.convDesc,
+ l.bd_algo,
+ state.workspace,
+ l.workspace_size,
+ &beta,
+ l.dsrcTensorDesc,
+ input16); // state.delta);
+
+ cuda_convert_f16_to_f32(input16, input16_size, state.delta);
+
+ if (l.binary || l.xnor) swap_binary(&l);
+ if (l.xnor) gradient_array_ongpu(original_input, l.batch*l.c*l.h*l.w, HARDTAN, state.delta);
+ }
+#else // CUDNN_HALF
+
+ // calculate conv weight updates
+ // if used: beta=1 then loss decreases faster
cudnnConvolutionBackwardFilter(cudnn_handle(),
&one,
l.srcTensorDesc,
@@ -247,6 +327,8 @@
if(state.delta){
if(l.binary || l.xnor) swap_binary(&l);
+ // http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#cudnnConvolutionBackwardData
+ // calculate delta for the next layer
cudnnConvolutionBackwardData(cudnn_handle(),
&one,
l.weightDesc,
@@ -264,7 +346,9 @@
if(l.xnor) gradient_array_ongpu(original_input, l.batch*l.c*l.h*l.w, HARDTAN, state.delta);
}
-#else
+#endif // CUDNN_HALF
+
+#else // CUDNN
int m = l.n;
int n = l.size*l.size*l.c;
int k = l.out_w*l.out_h;
@@ -317,7 +401,7 @@
{
cuda_push_array(layer.weights_gpu, layer.weights, layer.c*layer.n*layer.size*layer.size);
#ifdef CUDNN_HALF
- cuda_convert_f32_to_f16(layer.weights_gpu, layer.c*layer.n*layer.size*layer.size, (half *)layer.weights_gpu16);
+ cuda_convert_f32_to_f16(layer.weights_gpu, layer.c*layer.n*layer.size*layer.size, layer.weights_gpu16);
#endif
cuda_push_array(layer.biases_gpu, layer.biases, layer.n);
cuda_push_array(layer.weight_updates_gpu, layer.weight_updates, layer.c*layer.n*layer.size*layer.size);
@@ -357,6 +441,14 @@
adam_gpu(size, layer.weights_gpu, layer.m_gpu, layer.v_gpu, layer.B1, layer.B2, learning_rate/batch, layer.eps, layer.t+1);
fill_ongpu(size, 0, layer.weight_updates_gpu, 1);
}else{
+ // update weights:
+ // weights_gpu = weights_gpu*(1 - decay*lr) + weight_updates_gpu*lr / (batch*subdivision) =
+ // weights_gpu*(1 - 0.0005*0.001) + weight_updates_gpu*0.001/(64*8) =
+ // weights_gpu * 0.999 999 5 + weight_updates_gpu * 0.000 001 953125
+ //
+ // weight_updates_gpu = (weight_updates_gpu - weights_gpu*decay*batch*subdivision)*momentum =
+ // (weight_updates_gpu - weights_gpu * 0.0005 * 64 * 8) * 0.9 =
+ // weight_updates_gpu*0.9 - weights_gpu*0.2304
axpy_ongpu(size, -decay*batch, layer.weights_gpu, 1, layer.weight_updates_gpu, 1);
axpy_ongpu(size, learning_rate/batch, layer.weight_updates_gpu, 1, layer.weights_gpu, 1);
scal_ongpu(size, momentum, layer.weight_updates_gpu, 1);
--
Gitblit v1.10.0