From aebe937710ced03d03f73ab23f410f29685655c1 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 11 Aug 2016 18:54:24 +0000
Subject: [PATCH] what do you even write here?
---
src/network.c | 132 ++++++++++++++++++++++++++++++++++++++++----
1 files changed, 120 insertions(+), 12 deletions(-)
diff --git a/src/network.c b/src/network.c
index 063a1bb..91baafe 100644
--- a/src/network.c
+++ b/src/network.c
@@ -8,17 +8,25 @@
#include "crop_layer.h"
#include "connected_layer.h"
+#include "gru_layer.h"
+#include "rnn_layer.h"
+#include "crnn_layer.h"
+#include "local_layer.h"
#include "convolutional_layer.h"
+#include "activation_layer.h"
#include "deconvolutional_layer.h"
#include "detection_layer.h"
#include "region_layer.h"
#include "normalization_layer.h"
+#include "batchnorm_layer.h"
#include "maxpool_layer.h"
+#include "reorg_layer.h"
#include "avgpool_layer.h"
#include "cost_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
#include "route_layer.h"
+#include "shortcut_layer.h"
int get_current_batch(network net)
{
@@ -26,6 +34,17 @@
return batch_num;
}
+void reset_momentum(network net)
+{
+ if (net.momentum == 0) return;
+ net.learning_rate = 0;
+ net.momentum = 0;
+ net.decay = 0;
+ #ifdef GPU
+ if(gpu_index >= 0) update_network_gpu(net);
+ #endif
+}
+
float get_current_rate(network net)
{
int batch_num = get_current_batch(net);
@@ -41,12 +60,16 @@
for(i = 0; i < net.num_steps; ++i){
if(net.steps[i] > batch_num) return rate;
rate *= net.scales[i];
+ if(net.steps[i] > batch_num - 1) reset_momentum(net);
}
return rate;
case EXP:
return net.learning_rate * pow(net.gamma, batch_num);
case POLY:
+ if (batch_num < net.burn_in) return net.learning_rate * pow((float)batch_num / net.burn_in, net.power);
return net.learning_rate * pow(1 - (float)batch_num / net.max_batches, net.power);
+ case RANDOM:
+ return net.learning_rate * pow(rand_uniform(0,1), net.power);
case SIG:
return net.learning_rate * (1./(1.+exp(net.gamma*(batch_num - net.step))));
default:
@@ -60,12 +83,24 @@
switch(a){
case CONVOLUTIONAL:
return "convolutional";
+ case ACTIVE:
+ return "activation";
+ case LOCAL:
+ return "local";
case DECONVOLUTIONAL:
return "deconvolutional";
case CONNECTED:
return "connected";
+ case RNN:
+ return "rnn";
+ case GRU:
+ return "gru";
+ case CRNN:
+ return "crnn";
case MAXPOOL:
return "maxpool";
+ case REORG:
+ return "reorg";
case AVGPOOL:
return "avgpool";
case SOFTMAX:
@@ -82,8 +117,12 @@
return "cost";
case ROUTE:
return "route";
+ case SHORTCUT:
+ return "shortcut";
case NORMALIZATION:
return "normalization";
+ case BATCHNORM:
+ return "batchnorm";
default:
break;
}
@@ -105,8 +144,10 @@
void forward_network(network net, network_state state)
{
+ state.workspace = net.workspace;
int i;
for(i = 0; i < net.n; ++i){
+ state.index = i;
layer l = net.layers[i];
if(l.delta){
scal_cpu(l.outputs * l.batch, 0, l.delta, 1);
@@ -115,14 +156,26 @@
forward_convolutional_layer(l, state);
} else if(l.type == DECONVOLUTIONAL){
forward_deconvolutional_layer(l, state);
+ } else if(l.type == ACTIVE){
+ forward_activation_layer(l, state);
+ } else if(l.type == LOCAL){
+ forward_local_layer(l, state);
} else if(l.type == NORMALIZATION){
forward_normalization_layer(l, state);
+ } else if(l.type == BATCHNORM){
+ forward_batchnorm_layer(l, state);
} else if(l.type == DETECTION){
forward_detection_layer(l, state);
} else if(l.type == REGION){
forward_region_layer(l, state);
} else if(l.type == CONNECTED){
forward_connected_layer(l, state);
+ } else if(l.type == RNN){
+ forward_rnn_layer(l, state);
+ } else if(l.type == GRU){
+ forward_gru_layer(l, state);
+ } else if(l.type == CRNN){
+ forward_crnn_layer(l, state);
} else if(l.type == CROP){
forward_crop_layer(l, state);
} else if(l.type == COST){
@@ -131,12 +184,16 @@
forward_softmax_layer(l, state);
} else if(l.type == MAXPOOL){
forward_maxpool_layer(l, state);
+ } else if(l.type == REORG){
+ forward_reorg_layer(l, state);
} else if(l.type == AVGPOOL){
forward_avgpool_layer(l, state);
} else if(l.type == DROPOUT){
forward_dropout_layer(l, state);
} else if(l.type == ROUTE){
forward_route_layer(l, net);
+ } else if(l.type == SHORTCUT){
+ forward_shortcut_layer(l, state);
}
state.input = l.output;
}
@@ -155,12 +212,23 @@
update_deconvolutional_layer(l, rate, net.momentum, net.decay);
} else if(l.type == CONNECTED){
update_connected_layer(l, update_batch, rate, net.momentum, net.decay);
+ } else if(l.type == RNN){
+ update_rnn_layer(l, update_batch, rate, net.momentum, net.decay);
+ } else if(l.type == GRU){
+ update_gru_layer(l, update_batch, rate, net.momentum, net.decay);
+ } else if(l.type == CRNN){
+ update_crnn_layer(l, update_batch, rate, net.momentum, net.decay);
+ } else if(l.type == LOCAL){
+ update_local_layer(l, update_batch, rate, net.momentum, net.decay);
}
}
}
float *get_network_output(network net)
{
+ #ifdef GPU
+ if (gpu_index >= 0) return get_network_output_gpu(net);
+ #endif
int i;
for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
return net.layers[i].output;
@@ -172,15 +240,7 @@
float sum = 0;
int count = 0;
for(i = 0; i < net.n; ++i){
- if(net.layers[i].type == COST){
- sum += net.layers[i].output[0];
- ++count;
- }
- if(net.layers[i].type == DETECTION){
- sum += net.layers[i].cost[0];
- ++count;
- }
- if(net.layers[i].type == REGION){
+ if(net.layers[i].cost){
sum += net.layers[i].cost[0];
++count;
}
@@ -200,7 +260,9 @@
int i;
float *original_input = state.input;
float *original_delta = state.delta;
+ state.workspace = net.workspace;
for(i = net.n-1; i >= 0; --i){
+ state.index = i;
if(i == 0){
state.input = original_input;
state.delta = original_delta;
@@ -214,10 +276,16 @@
backward_convolutional_layer(l, state);
} else if(l.type == DECONVOLUTIONAL){
backward_deconvolutional_layer(l, state);
+ } else if(l.type == ACTIVE){
+ backward_activation_layer(l, state);
} else if(l.type == NORMALIZATION){
backward_normalization_layer(l, state);
+ } else if(l.type == BATCHNORM){
+ backward_batchnorm_layer(l, state);
} else if(l.type == MAXPOOL){
if(i != 0) backward_maxpool_layer(l, state);
+ } else if(l.type == REORG){
+ backward_reorg_layer(l, state);
} else if(l.type == AVGPOOL){
backward_avgpool_layer(l, state);
} else if(l.type == DROPOUT){
@@ -230,10 +298,20 @@
if(i != 0) backward_softmax_layer(l, state);
} else if(l.type == CONNECTED){
backward_connected_layer(l, state);
+ } else if(l.type == RNN){
+ backward_rnn_layer(l, state);
+ } else if(l.type == GRU){
+ backward_gru_layer(l, state);
+ } else if(l.type == CRNN){
+ backward_crnn_layer(l, state);
+ } else if(l.type == LOCAL){
+ backward_local_layer(l, state);
} else if(l.type == COST){
backward_cost_layer(l, state);
} else if(l.type == ROUTE){
backward_route_layer(l, net);
+ } else if(l.type == SHORTCUT){
+ backward_shortcut_layer(l, state);
}
}
}
@@ -245,6 +323,8 @@
if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
#endif
network_state state;
+ state.index = 0;
+ state.net = net;
state.input = x;
state.delta = 0;
state.truth = y;
@@ -293,10 +373,13 @@
return (float)sum/(n*batch);
}
+
float train_network_batch(network net, data d, int n)
{
int i,j;
network_state state;
+ state.index = 0;
+ state.net = net;
state.train = 1;
state.delta = 0;
float sum = 0;
@@ -321,6 +404,11 @@
int i;
for(i = 0; i < net->n; ++i){
net->layers[i].batch = b;
+ #ifdef CUDNN
+ if(net->layers[i].type == CONVOLUTIONAL){
+ cudnn_convolutional_setup(net->layers + i);
+ }
+ #endif
}
}
@@ -331,17 +419,21 @@
net->w = w;
net->h = h;
int inputs = 0;
+ size_t workspace_size = 0;
//fprintf(stderr, "Resizing to %d x %d...", w, h);
//fflush(stderr);
for (i = 0; i < net->n; ++i){
layer l = net->layers[i];
if(l.type == CONVOLUTIONAL){
resize_convolutional_layer(&l, w, h);
+ }else if(l.type == CROP){
+ resize_crop_layer(&l, w, h);
}else if(l.type == MAXPOOL){
resize_maxpool_layer(&l, w, h);
+ }else if(l.type == REORG){
+ resize_reorg_layer(&l, w, h);
}else if(l.type == AVGPOOL){
resize_avgpool_layer(&l, w, h);
- break;
}else if(l.type == NORMALIZATION){
resize_normalization_layer(&l, w, h);
}else if(l.type == COST){
@@ -349,11 +441,25 @@
}else{
error("Cannot resize this type of layer");
}
+ if(l.workspace_size > workspace_size) workspace_size = l.workspace_size;
inputs = l.outputs;
net->layers[i] = l;
w = l.out_w;
h = l.out_h;
+ if(l.type == AVGPOOL) break;
}
+#ifdef GPU
+ if(gpu_index >= 0){
+ cuda_free(net->workspace);
+ net->workspace = cuda_make_array(0, (workspace_size-1)/sizeof(float)+1);
+ }else {
+ free(net->workspace);
+ net->workspace = calloc(1, workspace_size);
+ }
+#else
+ free(net->workspace);
+ net->workspace = calloc(1, workspace_size);
+#endif
//fprintf(stderr, " Done!\n");
return 0;
}
@@ -433,6 +539,8 @@
#endif
network_state state;
+ state.net = net;
+ state.index = 0;
state.input = input;
state.truth = 0;
state.train = 0;
@@ -566,10 +674,10 @@
free_layer(net.layers[i]);
}
free(net.layers);
- #ifdef GPU
+#ifdef GPU
if(*net.input_gpu) cuda_free(*net.input_gpu);
if(*net.truth_gpu) cuda_free(*net.truth_gpu);
if(net.input_gpu) free(net.input_gpu);
if(net.truth_gpu) free(net.truth_gpu);
- #endif
+#endif
}
--
Gitblit v1.10.0