From 71a9929af6c3d3ffb9527bb921c5cc4a20971ff6 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Fri, 17 Mar 2017 22:47:21 +0000
Subject: [PATCH] Fixed x & y coords less than 0
---
src/network_kernels.cu | 531 +++++++++++++++++++++++++++++++++++++---------------------
1 files changed, 334 insertions(+), 197 deletions(-)
diff --git a/src/network_kernels.cu b/src/network_kernels.cu
index 1f3f2e0..313cd6d 100644
--- a/src/network_kernels.cu
+++ b/src/network_kernels.cu
@@ -1,207 +1,104 @@
+#include "cuda_runtime.h"
+#include "curand.h"
+#include "cublas_v2.h"
+
extern "C" {
#include <stdio.h>
#include <time.h>
+#include <assert.h>
#include "network.h"
#include "image.h"
#include "data.h"
#include "utils.h"
+#include "parser.h"
#include "crop_layer.h"
#include "connected_layer.h"
+#include "rnn_layer.h"
+#include "gru_layer.h"
+#include "crnn_layer.h"
+#include "detection_layer.h"
+#include "region_layer.h"
#include "convolutional_layer.h"
-#include "deconvolutional_layer.h"
+#include "activation_layer.h"
#include "maxpool_layer.h"
-#include "cost_layer.h"
+#include "reorg_layer.h"
+#include "avgpool_layer.h"
#include "normalization_layer.h"
-#include "freeweight_layer.h"
+#include "batchnorm_layer.h"
+#include "cost_layer.h"
+#include "local_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
+#include "route_layer.h"
+#include "shortcut_layer.h"
+#include "blas.h"
}
-extern "C" float * get_network_output_gpu_layer(network net, int i);
-extern "C" float * get_network_delta_gpu_layer(network net, int i);
+float * get_network_output_gpu_layer(network net, int i);
+float * get_network_delta_gpu_layer(network net, int i);
+float * get_network_output_gpu(network net);
-void forward_network_gpu(network net, float * input, float * truth, int train)
+void forward_network_gpu(network net, network_state state)
{
+ state.workspace = net.workspace;
int i;
for(i = 0; i < net.n; ++i){
- //clock_t time = clock();
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- forward_convolutional_layer_gpu(layer, input);
- input = layer.output_gpu;
+ state.index = i;
+ layer l = net.layers[i];
+ if(l.delta_gpu){
+ fill_ongpu(l.outputs * l.batch, 0, l.delta_gpu, 1);
}
- else if(net.types[i] == DECONVOLUTIONAL){
- deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
- forward_deconvolutional_layer_gpu(layer, input);
- input = layer.output_gpu;
- }
- else if(net.types[i] == COST){
- cost_layer layer = *(cost_layer *)net.layers[i];
- forward_cost_layer_gpu(layer, input, truth);
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- forward_connected_layer_gpu(layer, input);
- input = layer.output_gpu;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- forward_maxpool_layer_gpu(layer, input);
- input = layer.output_gpu;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- forward_softmax_layer_gpu(layer, input);
- input = layer.output_gpu;
- }
- else if(net.types[i] == DROPOUT){
- if(!train) continue;
- dropout_layer layer = *(dropout_layer *)net.layers[i];
- forward_dropout_layer_gpu(layer, input);
- input = layer.output_gpu;
- }
- else if(net.types[i] == CROP){
- crop_layer layer = *(crop_layer *)net.layers[i];
- forward_crop_layer_gpu(layer, train, input);
- input = layer.output_gpu;
- }
- //cudaDeviceSynchronize();
- //printf("Forward %d %s %f\n", i, get_layer_string(net.types[i]), sec(clock() - time));
+ l.forward_gpu(l, state);
+ state.input = l.output_gpu;
}
}
-void backward_network_gpu(network net, float * input)
+void backward_network_gpu(network net, network_state state)
{
+ state.workspace = net.workspace;
int i;
- float * prev_input;
- float * prev_delta;
+ float * original_input = state.input;
+ float * original_delta = state.delta;
for(i = net.n-1; i >= 0; --i){
- //clock_t time = clock();
+ state.index = i;
+ layer l = net.layers[i];
if(i == 0){
- prev_input = input;
- prev_delta = 0;
+ state.input = original_input;
+ state.delta = original_delta;
}else{
- prev_input = get_network_output_gpu_layer(net, i-1);
- prev_delta = get_network_delta_gpu_layer(net, i-1);
+ layer prev = net.layers[i-1];
+ state.input = prev.output_gpu;
+ state.delta = prev.delta_gpu;
}
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- backward_convolutional_layer_gpu(layer, prev_input, prev_delta);
- }
- else if(net.types[i] == DECONVOLUTIONAL){
- deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
- backward_deconvolutional_layer_gpu(layer, prev_input, prev_delta);
- }
- else if(net.types[i] == COST){
- cost_layer layer = *(cost_layer *)net.layers[i];
- backward_cost_layer_gpu(layer, prev_input, prev_delta);
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- backward_connected_layer_gpu(layer, prev_input, prev_delta);
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- backward_maxpool_layer_gpu(layer, prev_delta);
- }
- else if(net.types[i] == DROPOUT){
- dropout_layer layer = *(dropout_layer *)net.layers[i];
- backward_dropout_layer_gpu(layer, prev_delta);
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- backward_softmax_layer_gpu(layer, prev_delta);
- }
- //printf("Backward %d %s %f\n", i, get_layer_string(net.types[i]), sec(clock() - time));
+ l.backward_gpu(l, state);
}
}
void update_network_gpu(network net)
{
+ cuda_set_device(net.gpu_index);
int i;
+ int update_batch = net.batch*net.subdivisions;
+ float rate = get_current_rate(net);
for(i = 0; i < net.n; ++i){
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- update_convolutional_layer_gpu(layer);
- }
- else if(net.types[i] == DECONVOLUTIONAL){
- deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
- update_deconvolutional_layer_gpu(layer);
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- update_connected_layer_gpu(layer);
+ layer l = net.layers[i];
+ l.t = get_current_batch(net);
+ if(l.update_gpu){
+ l.update_gpu(l, update_batch, rate, net.momentum, net.decay);
}
}
}
-float * get_network_output_gpu_layer(network net, int i)
+void forward_backward_network_gpu(network net, float *x, float *y)
{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.output_gpu;
- }
- else if(net.types[i] == DECONVOLUTIONAL){
- deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
- return layer.output_gpu;
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- return layer.output_gpu;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.output_gpu;
- }
- else if(net.types[i] == CROP){
- crop_layer layer = *(crop_layer *)net.layers[i];
- return layer.output_gpu;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- return layer.output_gpu;
- } else if(net.types[i] == DROPOUT){
- dropout_layer layer = *(dropout_layer *)net.layers[i];
- return layer.output_gpu;
- }
- return 0;
-}
-
-float * get_network_delta_gpu_layer(network net, int i)
-{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.delta_gpu;
- }
- else if(net.types[i] == DECONVOLUTIONAL){
- deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
- return layer.delta_gpu;
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- return layer.delta_gpu;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.delta_gpu;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- return layer.delta_gpu;
- } else if(net.types[i] == DROPOUT){
- if(i == 0) return 0;
- return get_network_delta_gpu_layer(net, i-1);
- }
- return 0;
-}
-
-float train_network_datum_gpu(network net, float *x, float *y)
-{
- //clock_t time = clock();
+ network_state state;
+ state.index = 0;
+ state.net = net;
int x_size = get_network_input_size(net)*net.batch;
int y_size = get_network_output_size(net)*net.batch;
+ if(net.layers[net.n-1].truths) y_size = net.layers[net.n-1].truths*net.batch;
if(!*net.input_gpu){
*net.input_gpu = cuda_make_array(x, x_size);
*net.truth_gpu = cuda_make_array(y, y_size);
@@ -209,63 +106,303 @@
cuda_push_array(*net.input_gpu, x, x_size);
cuda_push_array(*net.truth_gpu, y, y_size);
}
- //printf("trans %f\n", sec(clock() - time));
- //time = clock();
- forward_network_gpu(net, *net.input_gpu, *net.truth_gpu, 1);
- //printf("forw %f\n", sec(clock() - time));
- //time = clock();
- backward_network_gpu(net, *net.input_gpu);
- //printf("back %f\n", sec(clock() - time));
- //time = clock();
- update_network_gpu(net);
+ state.input = *net.input_gpu;
+ state.delta = 0;
+ state.truth = *net.truth_gpu;
+ state.train = 1;
+ forward_network_gpu(net, state);
+ backward_network_gpu(net, state);
+}
+
+float train_network_datum_gpu(network net, float *x, float *y)
+{
+ *net.seen += net.batch;
+ forward_backward_network_gpu(net, x, y);
float error = get_network_cost(net);
- //printf("updt %f\n", sec(clock() - time));
- //time = clock();
+ if (((*net.seen) / net.batch) % net.subdivisions == 0) update_network_gpu(net);
+
return error;
}
+typedef struct {
+ network net;
+ data d;
+ float *err;
+} train_args;
+
+void *train_thread(void *ptr)
+{
+ train_args args = *(train_args*)ptr;
+ free(ptr);
+ cuda_set_device(args.net.gpu_index);
+ *args.err = train_network(args.net, args.d);
+ return 0;
+}
+
+pthread_t train_network_in_thread(network net, data d, float *err)
+{
+ pthread_t thread;
+ train_args *ptr = (train_args *)calloc(1, sizeof(train_args));
+ ptr->net = net;
+ ptr->d = d;
+ ptr->err = err;
+ if(pthread_create(&thread, 0, train_thread, ptr)) error("Thread creation failed");
+ return thread;
+}
+
+void pull_updates(layer l)
+{
+ if(l.type == CONVOLUTIONAL){
+ cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.n);
+ cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.n*l.size*l.size*l.c);
+ if(l.scale_updates) cuda_pull_array(l.scale_updates_gpu, l.scale_updates, l.n);
+ } else if(l.type == CONNECTED){
+ cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
+ cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.outputs*l.inputs);
+ }
+}
+
+void push_updates(layer l)
+{
+ if(l.type == CONVOLUTIONAL){
+ cuda_push_array(l.bias_updates_gpu, l.bias_updates, l.n);
+ cuda_push_array(l.weight_updates_gpu, l.weight_updates, l.n*l.size*l.size*l.c);
+ if(l.scale_updates) cuda_push_array(l.scale_updates_gpu, l.scale_updates, l.n);
+ } else if(l.type == CONNECTED){
+ cuda_push_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
+ cuda_push_array(l.weight_updates_gpu, l.weight_updates, l.outputs*l.inputs);
+ }
+}
+
+void update_layer(layer l, network net)
+{
+ int update_batch = net.batch*net.subdivisions;
+ float rate = get_current_rate(net);
+ l.t = get_current_batch(net);
+ if(l.update_gpu){
+ l.update_gpu(l, update_batch, rate, net.momentum, net.decay);
+ }
+}
+
+void merge_weights(layer l, layer base)
+{
+ if (l.type == CONVOLUTIONAL) {
+ axpy_cpu(l.n, 1, l.biases, 1, base.biases, 1);
+ axpy_cpu(l.n*l.size*l.size*l.c, 1, l.weights, 1, base.weights, 1);
+ if (l.scales) {
+ axpy_cpu(l.n, 1, l.scales, 1, base.scales, 1);
+ }
+ } else if(l.type == CONNECTED) {
+ axpy_cpu(l.outputs, 1, l.biases, 1, base.biases, 1);
+ axpy_cpu(l.outputs*l.inputs, 1, l.weights, 1, base.weights, 1);
+ }
+}
+
+void scale_weights(layer l, float s)
+{
+ if (l.type == CONVOLUTIONAL) {
+ scal_cpu(l.n, s, l.biases, 1);
+ scal_cpu(l.n*l.size*l.size*l.c, s, l.weights, 1);
+ if (l.scales) {
+ scal_cpu(l.n, s, l.scales, 1);
+ }
+ } else if(l.type == CONNECTED) {
+ scal_cpu(l.outputs, s, l.biases, 1);
+ scal_cpu(l.outputs*l.inputs, s, l.weights, 1);
+ }
+}
+
+
+void pull_weights(layer l)
+{
+ if(l.type == CONVOLUTIONAL){
+ cuda_pull_array(l.biases_gpu, l.biases, l.n);
+ cuda_pull_array(l.weights_gpu, l.weights, l.n*l.size*l.size*l.c);
+ if(l.scales) cuda_pull_array(l.scales_gpu, l.scales, l.n);
+ } else if(l.type == CONNECTED){
+ cuda_pull_array(l.biases_gpu, l.biases, l.outputs);
+ cuda_pull_array(l.weights_gpu, l.weights, l.outputs*l.inputs);
+ }
+}
+
+void push_weights(layer l)
+{
+ if(l.type == CONVOLUTIONAL){
+ cuda_push_array(l.biases_gpu, l.biases, l.n);
+ cuda_push_array(l.weights_gpu, l.weights, l.n*l.size*l.size*l.c);
+ if(l.scales) cuda_push_array(l.scales_gpu, l.scales, l.n);
+ } else if(l.type == CONNECTED){
+ cuda_push_array(l.biases_gpu, l.biases, l.outputs);
+ cuda_push_array(l.weights_gpu, l.weights, l.outputs*l.inputs);
+ }
+}
+
+void distribute_weights(layer l, layer base)
+{
+ if(l.type == CONVOLUTIONAL){
+ cuda_push_array(l.biases_gpu, base.biases, l.n);
+ cuda_push_array(l.weights_gpu, base.weights, l.n*l.size*l.size*l.c);
+ if(base.scales) cuda_push_array(l.scales_gpu, base.scales, l.n);
+ } else if(l.type == CONNECTED){
+ cuda_push_array(l.biases_gpu, base.biases, l.outputs);
+ cuda_push_array(l.weights_gpu, base.weights, l.outputs*l.inputs);
+ }
+}
+
+
+void merge_updates(layer l, layer base)
+{
+ if (l.type == CONVOLUTIONAL) {
+ axpy_cpu(l.n, 1, l.bias_updates, 1, base.bias_updates, 1);
+ axpy_cpu(l.n*l.size*l.size*l.c, 1, l.weight_updates, 1, base.weight_updates, 1);
+ if (l.scale_updates) {
+ axpy_cpu(l.n, 1, l.scale_updates, 1, base.scale_updates, 1);
+ }
+ } else if(l.type == CONNECTED) {
+ axpy_cpu(l.outputs, 1, l.bias_updates, 1, base.bias_updates, 1);
+ axpy_cpu(l.outputs*l.inputs, 1, l.weight_updates, 1, base.weight_updates, 1);
+ }
+}
+
+void distribute_updates(layer l, layer base)
+{
+ if(l.type == CONVOLUTIONAL){
+ cuda_push_array(l.bias_updates_gpu, base.bias_updates, l.n);
+ cuda_push_array(l.weight_updates_gpu, base.weight_updates, l.n*l.size*l.size*l.c);
+ if(base.scale_updates) cuda_push_array(l.scale_updates_gpu, base.scale_updates, l.n);
+ } else if(l.type == CONNECTED){
+ cuda_push_array(l.bias_updates_gpu, base.bias_updates, l.outputs);
+ cuda_push_array(l.weight_updates_gpu, base.weight_updates, l.outputs*l.inputs);
+ }
+}
+
+void sync_layer(network *nets, int n, int j)
+{
+ //printf("Syncing layer %d\n", j);
+ int i;
+ network net = nets[0];
+ layer base = net.layers[j];
+ cuda_set_device(net.gpu_index);
+ pull_weights(base);
+ for (i = 1; i < n; ++i) {
+ cuda_set_device(nets[i].gpu_index);
+ layer l = nets[i].layers[j];
+ pull_weights(l);
+ merge_weights(l, base);
+ }
+ scale_weights(base, 1./n);
+ for (i = 0; i < n; ++i) {
+ cuda_set_device(nets[i].gpu_index);
+ layer l = nets[i].layers[j];
+ distribute_weights(l, base);
+ }
+ //printf("Done syncing layer %d\n", j);
+}
+
+typedef struct{
+ network *nets;
+ int n;
+ int j;
+} sync_args;
+
+void *sync_layer_thread(void *ptr)
+{
+ sync_args args = *(sync_args*)ptr;
+ sync_layer(args.nets, args.n, args.j);
+ free(ptr);
+ return 0;
+}
+
+pthread_t sync_layer_in_thread(network *nets, int n, int j)
+{
+ pthread_t thread;
+ sync_args *ptr = (sync_args *)calloc(1, sizeof(sync_args));
+ ptr->nets = nets;
+ ptr->n = n;
+ ptr->j = j;
+ if(pthread_create(&thread, 0, sync_layer_thread, ptr)) error("Thread creation failed");
+ return thread;
+}
+
+void sync_nets(network *nets, int n, int interval)
+{
+ int j;
+ int layers = nets[0].n;
+ pthread_t *threads = (pthread_t *) calloc(layers, sizeof(pthread_t));
+
+ *nets[0].seen += interval * (n-1) * nets[0].batch * nets[0].subdivisions;
+ for (j = 0; j < n; ++j){
+ *nets[j].seen = *nets[0].seen;
+ }
+ for (j = 0; j < layers; ++j) {
+ threads[j] = sync_layer_in_thread(nets, n, j);
+ }
+ for (j = 0; j < layers; ++j) {
+ pthread_join(threads[j], 0);
+ }
+ free(threads);
+}
+
+float train_networks(network *nets, int n, data d, int interval)
+{
+ int i;
+ int batch = nets[0].batch;
+ int subdivisions = nets[0].subdivisions;
+ assert(batch * subdivisions * n == d.X.rows);
+ pthread_t *threads = (pthread_t *) calloc(n, sizeof(pthread_t));
+ float *errors = (float *) calloc(n, sizeof(float));
+
+ float sum = 0;
+ for(i = 0; i < n; ++i){
+ data p = get_data_part(d, i, n);
+ threads[i] = train_network_in_thread(nets[i], p, errors + i);
+ }
+ for(i = 0; i < n; ++i){
+ pthread_join(threads[i], 0);
+ //printf("%f\n", errors[i]);
+ sum += errors[i];
+ }
+ //cudaDeviceSynchronize();
+ if (get_current_batch(nets[0]) % interval == 0) {
+ printf("Syncing... ");
+ fflush(stdout);
+ sync_nets(nets, n, interval);
+ printf("Done!\n");
+ }
+ //cudaDeviceSynchronize();
+ free(threads);
+ free(errors);
+ return (float)sum/(n);
+}
+
float *get_network_output_layer_gpu(network net, int i)
{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.output;
- }
- else if(net.types[i] == DECONVOLUTIONAL){
- deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
- return layer.output;
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- cuda_pull_array(layer.output_gpu, layer.output, layer.outputs*layer.batch);
- return layer.output;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.output;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- pull_softmax_layer_output(layer);
- return layer.output;
- }
- return 0;
+ layer l = net.layers[i];
+ if(l.type != REGION) cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
+ return l.output;
}
float *get_network_output_gpu(network net)
{
int i;
- for(i = net.n-1; i > 0; --i) if(net.types[i] != COST) break;
+ for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
return get_network_output_layer_gpu(net, i);
}
float *network_predict_gpu(network net, float *input)
{
-
+ cuda_set_device(net.gpu_index);
int size = get_network_input_size(net) * net.batch;
- float * input_gpu = cuda_make_array(input, size);
- forward_network_gpu(net, input_gpu, 0, 0);
+ network_state state;
+ state.index = 0;
+ state.net = net;
+ state.input = cuda_make_array(input, size);
+ state.truth = 0;
+ state.train = 0;
+ state.delta = 0;
+ forward_network_gpu(net, state);
float *out = get_network_output_gpu(net);
- cuda_free(input_gpu);
+ cuda_free(state.input);
return out;
}
--
Gitblit v1.10.0