crop layer scaling and trans on cpu

Joseph Redmon

2015-04-15 47528e37cf29e0f9da6885213e5aee044bed84ef

 src/connected_layer.c

@@ -1,22 +1,19 @@
#include "connected_layer.h"
#include "utils.h"
#include "mini_blas.h"
#include "cuda.h"
#include "blas.h"
#include "gemm.h"

#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

connected_layer *make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, float learning_rate, float momentum, float decay)
connected_layer *make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation)
{
    fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
    int i;
    connected_layer *layer = calloc(1, sizeof(connected_layer));

    layer->learning_rate = learning_rate;
    layer->momentum = momentum;
    layer->decay = decay;

    layer->inputs = inputs;
    layer->outputs = outputs;
    layer->batch=batch;
@@ -25,46 +22,50 @@
    layer->delta = calloc(batch*outputs, sizeof(float*));

    layer->weight_updates = calloc(inputs*outputs, sizeof(float));
    //layer->weight_adapt = calloc(inputs*outputs, sizeof(float));
    layer->weights = calloc(inputs*outputs, sizeof(float));
    float scale = 1./inputs;
    scale = .05;
    for(i = 0; i < inputs*outputs; ++i)
        layer->weights[i] = scale*2*(rand_uniform()-.5);

    layer->bias_updates = calloc(outputs, sizeof(float));
    //layer->bias_adapt = calloc(outputs, sizeof(float));

    layer->weight_prev = calloc(inputs*outputs, sizeof(float));
    layer->bias_prev = calloc(outputs, sizeof(float));

    layer->weights = calloc(inputs*outputs, sizeof(float));
    layer->biases = calloc(outputs, sizeof(float));
    for(i = 0; i < outputs; ++i){
        //layer->biases[i] = rand_normal()*scale + scale;
        layer->biases[i] = 1;


    float scale = 1./sqrt(inputs);
    for(i = 0; i < inputs*outputs; ++i){
        layer->weights[i] = 2*scale*rand_uniform() - scale;
    }

    #ifdef GPU
    layer->weights_cl = cl_make_array(layer->weights, inputs*outputs);
    layer->biases_cl = cl_make_array(layer->biases, outputs);
    for(i = 0; i < outputs; ++i){
        layer->biases[i] = scale;
    }

    layer->weight_updates_cl = cl_make_array(layer->weight_updates, inputs*outputs);
    layer->bias_updates_cl = cl_make_array(layer->bias_updates, outputs);
#ifdef GPU
    layer->weights_gpu = cuda_make_array(layer->weights, inputs*outputs);
    layer->biases_gpu = cuda_make_array(layer->biases, outputs);

    layer->output_cl = cl_make_array(layer->output, outputs*batch);
    layer->delta_cl = cl_make_array(layer->delta, outputs*batch);
    #endif
    layer->weight_updates_gpu = cuda_make_array(layer->weight_updates, inputs*outputs);
    layer->bias_updates_gpu = cuda_make_array(layer->bias_updates, outputs);

    layer->output_gpu = cuda_make_array(layer->output, outputs*batch);
    layer->delta_gpu = cuda_make_array(layer->delta, outputs*batch);
#endif
    layer->activation = activation;
    fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
    return layer;
}

void update_connected_layer(connected_layer layer)
void update_connected_layer(connected_layer layer, int batch, float learning_rate, float momentum, float decay)
{
    axpy_cpu(layer.outputs, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
    scal_cpu(layer.outputs, layer.momentum, layer.bias_updates, 1);
    axpy_cpu(layer.outputs, learning_rate/batch, layer.bias_updates, 1, layer.biases, 1);
    scal_cpu(layer.outputs, momentum, layer.bias_updates, 1);

    scal_cpu(layer.inputs*layer.outputs, 1.-layer.learning_rate*layer.decay, layer.weights, 1);
    axpy_cpu(layer.inputs*layer.outputs, layer.learning_rate, layer.weight_updates, 1, layer.weights, 1);
    scal_cpu(layer.inputs*layer.outputs, layer.momentum, layer.weight_updates, 1);
    axpy_cpu(layer.inputs*layer.outputs, -decay*batch, layer.weights, 1, layer.weight_updates, 1);
    axpy_cpu(layer.inputs*layer.outputs, learning_rate/batch, layer.weight_updates, 1, layer.weights, 1);
    scal_cpu(layer.inputs*layer.outputs, momentum, layer.weight_updates, 1);
}

void forward_connected_layer(connected_layer layer, float *input)
void forward_connected_layer(connected_layer layer, network_state state)
{
    int i;
    for(i = 0; i < layer.batch; ++i){
@@ -73,14 +74,14 @@
    int m = layer.batch;
    int k = layer.inputs;
    int n = layer.outputs;
    float *a = input;
    float *a = state.input;
    float *b = layer.weights;
    float *c = layer.output;
    gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
    activate_array(layer.output, layer.outputs*layer.batch, layer.activation);
}

void backward_connected_layer(connected_layer layer, float *input, float *delta)
void backward_connected_layer(connected_layer layer, network_state state)
{
    int i;
    gradient_array(layer.output, layer.outputs*layer.batch, layer.activation, layer.delta);
@@ -90,7 +91,7 @@
    int m = layer.inputs;
    int k = layer.batch;
    int n = layer.outputs;
    float *a = input;
    float *a = state.input;
    float *b = layer.delta;
    float *c = layer.weight_updates;
    gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);
@@ -101,7 +102,7 @@

    a = layer.delta;
    b = layer.weights;
    c = delta;
    c = state.delta;

    if(c) gemm(0,1,m,n,k,1,a,k,b,k,0,c,n);
}
@@ -110,64 +111,69 @@

void pull_connected_layer(connected_layer layer)
{
    cl_read_array(layer.weights_cl, layer.weights, layer.inputs*layer.outputs);
    cl_read_array(layer.biases_cl, layer.biases, layer.outputs);
    cuda_pull_array(layer.weights_gpu, layer.weights, layer.inputs*layer.outputs);
    cuda_pull_array(layer.biases_gpu, layer.biases, layer.outputs);
    cuda_pull_array(layer.weight_updates_gpu, layer.weight_updates, layer.inputs*layer.outputs);
    cuda_pull_array(layer.bias_updates_gpu, layer.bias_updates, layer.outputs);
}

void update_connected_layer_gpu(connected_layer layer)
void push_connected_layer(connected_layer layer)
{
    axpy_ongpu(layer.outputs, layer.learning_rate, layer.bias_updates_cl, 1, layer.biases_cl, 1);
    scal_ongpu(layer.outputs, layer.momentum, layer.bias_updates_cl, 1);

    scal_ongpu(layer.inputs*layer.outputs, 1.-layer.learning_rate*layer.decay, layer.weights_cl, 1);
    axpy_ongpu(layer.inputs*layer.outputs, layer.learning_rate, layer.weight_updates_cl, 1, layer.weights_cl, 1);
    scal_ongpu(layer.inputs*layer.outputs, layer.momentum, layer.weight_updates_cl, 1);
    pull_connected_layer(layer);
    cuda_push_array(layer.weights_gpu, layer.weights, layer.inputs*layer.outputs);
    cuda_push_array(layer.biases_gpu, layer.biases, layer.outputs);
    cuda_push_array(layer.weight_updates_gpu, layer.weight_updates, layer.inputs*layer.outputs);
    cuda_push_array(layer.bias_updates_gpu, layer.bias_updates, layer.outputs);
}

void forward_connected_layer_gpu(connected_layer layer, cl_mem input)
void update_connected_layer_gpu(connected_layer layer, int batch, float learning_rate, float momentum, float decay)
{
    axpy_ongpu(layer.outputs, learning_rate/batch, layer.bias_updates_gpu, 1, layer.biases_gpu, 1);
    scal_ongpu(layer.outputs, momentum, layer.bias_updates_gpu, 1);

    axpy_ongpu(layer.inputs*layer.outputs, -decay*batch, layer.weights_gpu, 1, layer.weight_updates_gpu, 1);
    axpy_ongpu(layer.inputs*layer.outputs, learning_rate/batch, layer.weight_updates_gpu, 1, layer.weights_gpu, 1);
    scal_ongpu(layer.inputs*layer.outputs, momentum, layer.weight_updates_gpu, 1);
}

void forward_connected_layer_gpu(connected_layer layer, network_state state)
{
    int i;
    for(i = 0; i < layer.batch; ++i){
        cl_mem sub = cl_sub_array(layer.output_cl, i*layer.outputs, layer.outputs);
        copy_ongpu(layer.outputs, layer.biases_cl, 1, sub, 1);
        clReleaseMemObject(sub);
        copy_ongpu_offset(layer.outputs, layer.biases_gpu, 0, 1, layer.output_gpu, i*layer.outputs, 1);
    }
    int m = layer.batch;
    int k = layer.inputs;
    int n = layer.outputs;
    cl_mem a = input;
    cl_mem b = layer.weights_cl;
    cl_mem c = layer.output_cl;
    float * a = state.input;
    float * b = layer.weights_gpu;
    float * c = layer.output_gpu;
    gemm_ongpu(0,0,m,n,k,1,a,k,b,n,1,c,n);
    activate_array_ongpu(layer.output_cl, layer.outputs*layer.batch, layer.activation);
    activate_array_ongpu(layer.output_gpu, layer.outputs*layer.batch, layer.activation);
}

void backward_connected_layer_gpu(connected_layer layer, cl_mem input, cl_mem delta)
void backward_connected_layer_gpu(connected_layer layer, network_state state)
{
    int i;
    gradient_array_ongpu(layer.output_cl, layer.outputs*layer.batch, layer.activation, layer.delta_cl);
    gradient_array_ongpu(layer.output_gpu, layer.outputs*layer.batch, layer.activation, layer.delta_gpu);
    for(i = 0; i < layer.batch; ++i){
        cl_mem sub = cl_sub_array(layer.delta_cl, i*layer.outputs, layer.outputs);
        axpy_ongpu(layer.outputs, 1, sub, 1, layer.bias_updates_cl, 1);
        clReleaseMemObject(sub);
        axpy_ongpu_offset(layer.outputs, 1, layer.delta_gpu, i*layer.outputs, 1, layer.bias_updates_gpu, 0, 1);
    }
    int m = layer.inputs;
    int k = layer.batch;
    int n = layer.outputs;
    cl_mem a = input;
    cl_mem b = layer.delta_cl;
    cl_mem c = layer.weight_updates_cl;
    float * a = state.input;
    float * b = layer.delta_gpu;
    float * c = layer.weight_updates_gpu;
    gemm_ongpu(1,0,m,n,k,1,a,m,b,n,1,c,n);

    m = layer.batch;
    k = layer.outputs;
    n = layer.inputs;

    a = layer.delta_cl;
    b = layer.weights_cl;
    c = delta;
    a = layer.delta_gpu;
    b = layer.weights_gpu;
    c = state.delta;

    if(c) gemm_ongpu(0,1,m,n,k,1,a,k,b,k,0,c,n);
}
    #endif
#endif