From aa5996d58e68edfbefe51061856aecd549dd09c4 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 13 Jan 2015 01:27:08 +0000
Subject: [PATCH] Faster
---
src/network.c | 393 +++++++++++++++++++++----------------------------------
1 files changed, 149 insertions(+), 244 deletions(-)
diff --git a/src/network.c b/src/network.c
index 69942e8..5c5ce9d 100644
--- a/src/network.c
+++ b/src/network.c
@@ -31,148 +31,6 @@
return net;
}
-#ifdef GPU
-
-void forward_network_gpu(network net, cl_mem input, cl_mem truth, int train)
-{
- //printf("start\n");
- 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_cl;
- }
- 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_cl;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- forward_maxpool_layer_gpu(layer, input);
- input = layer.output_cl;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- forward_softmax_layer_gpu(layer, input);
- input = layer.output_cl;
- }
- //printf("%d %f\n", i, sec(clock()-time));
- /*
- else if(net.types[i] == CROP){
- crop_layer layer = *(crop_layer *)net.layers[i];
- forward_crop_layer(layer, input);
- input = layer.output;
- }
- else if(net.types[i] == NORMALIZATION){
- normalization_layer layer = *(normalization_layer *)net.layers[i];
- forward_normalization_layer(layer, input);
- input = layer.output;
- }
- */
- }
-}
-
-void backward_network_gpu(network net, cl_mem input)
-{
- int i;
- cl_mem prev_input;
- cl_mem prev_delta;
- for(i = net.n-1; i >= 0; --i){
- if(i == 0){
- prev_input = input;
- prev_delta = 0;
- }else{
- prev_input = get_network_output_cl_layer(net, i-1);
- prev_delta = get_network_delta_cl_layer(net, i-1);
- }
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- backward_convolutional_layer_gpu(layer, 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] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- backward_softmax_layer_gpu(layer, prev_delta);
- }
- }
-}
-
-void update_network_gpu(network net)
-{
- int i;
- 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] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- update_connected_layer_gpu(layer);
- }
- }
-}
-
-cl_mem get_network_output_cl_layer(network net, int i)
-{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.output_cl;
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- return layer.output_cl;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.output_cl;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- return layer.output_cl;
- }
- return 0;
-}
-
-cl_mem get_network_delta_cl_layer(network net, int i)
-{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.delta_cl;
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- return layer.delta_cl;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.delta_cl;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- return layer.delta_cl;
- }
- return 0;
-}
-
-#endif
void forward_network(network net, float *input, float *truth, int train)
{
@@ -216,6 +74,7 @@
if(!train) continue;
dropout_layer layer = *(dropout_layer *)net.layers[i];
forward_dropout_layer(layer, input);
+ input = layer.output;
}
else if(net.types[i] == FREEWEIGHT){
if(!train) continue;
@@ -244,6 +103,7 @@
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
+ //secret_update_connected_layer((connected_layer *)net.layers[i]);
update_connected_layer(layer);
}
}
@@ -261,12 +121,16 @@
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.output;
} else if(net.types[i] == DROPOUT){
- return get_network_output_layer(net, i-1);
+ dropout_layer layer = *(dropout_layer *)net.layers[i];
+ return layer.output;
} else if(net.types[i] == FREEWEIGHT){
return get_network_output_layer(net, i-1);
} else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.output;
+ } else if(net.types[i] == CROP){
+ crop_layer layer = *(crop_layer *)net.layers[i];
+ return layer.output;
} else if(net.types[i] == NORMALIZATION){
normalization_layer layer = *(normalization_layer *)net.layers[i];
return layer.output;
@@ -292,6 +156,7 @@
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.delta;
} else if(net.types[i] == DROPOUT){
+ if(i == 0) return 0;
return get_network_delta_layer(net, i-1);
} else if(net.types[i] == FREEWEIGHT){
return get_network_delta_layer(net, i-1);
@@ -355,12 +220,16 @@
}
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- backward_convolutional_layer(layer, prev_delta);
+ backward_convolutional_layer(layer, prev_input, prev_delta);
}
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
if(i != 0) backward_maxpool_layer(layer, prev_delta);
}
+ else if(net.types[i] == DROPOUT){
+ dropout_layer layer = *(dropout_layer *)net.layers[i];
+ backward_dropout_layer(layer, prev_delta);
+ }
else if(net.types[i] == NORMALIZATION){
normalization_layer layer = *(normalization_layer *)net.layers[i];
if(i != 0) backward_normalization_layer(layer, prev_input, prev_delta);
@@ -380,81 +249,15 @@
}
}
-
-#ifdef GPU
-float train_network_datum_gpu(network net, float *x, float *y)
-{
- int x_size = get_network_input_size(net)*net.batch;
- int y_size = get_network_output_size(net)*net.batch;
- clock_t time = clock();
- if(!*net.input_cl){
- *net.input_cl = cl_make_array(x, x_size);
- *net.truth_cl = cl_make_array(y, y_size);
- }else{
- cl_write_array(*net.input_cl, x, x_size);
- cl_write_array(*net.truth_cl, y, y_size);
- }
- //printf("trans %f\n", sec(clock()-time));
- time = clock();
- forward_network_gpu(net, *net.input_cl, *net.truth_cl, 1);
- //printf("forw %f\n", sec(clock()-time));
- time = clock();
- backward_network_gpu(net, *net.input_cl);
- //printf("back %f\n", sec(clock()-time));
- time = clock();
- float error = get_network_cost(net);
- update_network_gpu(net);
- //printf("updt %f\n", sec(clock()-time));
- time = clock();
- return error;
-}
-
-float train_network_sgd_gpu(network net, data d, int n)
-{
- int batch = net.batch;
- float *X = calloc(batch*d.X.cols, sizeof(float));
- float *y = calloc(batch*d.y.cols, sizeof(float));
-
- int i;
- float sum = 0;
- for(i = 0; i < n; ++i){
- get_random_batch(d, batch, X, y);
- float err = train_network_datum_gpu(net, X, y);
- sum += err;
- }
- free(X);
- free(y);
- return (float)sum/(n*batch);
-}
-
-float train_network_data_gpu(network net, data d, int n)
-{
- int batch = net.batch;
- float *X = calloc(batch*d.X.cols, sizeof(float));
- float *y = calloc(batch*d.y.cols, sizeof(float));
-
- int i;
- float sum = 0;
- for(i = 0; i < n; ++i){
- get_next_batch(d, batch, i*batch, X, y);
- float err = train_network_datum_gpu(net, X, y);
- sum += err;
- }
- free(X);
- free(y);
- return (float)sum/(n*batch);
-}
-#endif
-
-
float train_network_datum(network net, float *x, float *y)
{
+ #ifdef GPU
+ if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
+ #endif
forward_network(net, x, y, 1);
- //int class = get_predicted_class_network(net);
backward_network(net, x);
float error = get_network_cost(net);
update_network(net);
- //return (y[class]?1:0);
return error;
}
@@ -475,6 +278,26 @@
free(y);
return (float)sum/(n*batch);
}
+
+float train_network(network net, data d)
+{
+ int batch = net.batch;
+ int n = d.X.rows / batch;
+ float *X = calloc(batch*d.X.cols, sizeof(float));
+ float *y = calloc(batch*d.y.cols, sizeof(float));
+
+ int i;
+ float sum = 0;
+ for(i = 0; i < n; ++i){
+ get_next_batch(d, batch, i*batch, X, y);
+ float err = train_network_datum(net, X, y);
+ sum += err;
+ }
+ free(X);
+ free(y);
+ return (float)sum/(n*batch);
+}
+
float train_network_batch(network net, data d, int n)
{
int i,j;
@@ -494,23 +317,69 @@
return (float)sum/(n*batch);
}
-
-void train_network(network net, data d)
+void set_learning_network(network *net, float rate, float momentum, float decay)
{
int i;
- int correct = 0;
- for(i = 0; i < d.X.rows; ++i){
- correct += train_network_datum(net, d.X.vals[i], d.y.vals[i]);
- if(i%100 == 0){
- visualize_network(net);
- cvWaitKey(10);
+ net->learning_rate=rate;
+ net->momentum = momentum;
+ net->decay = decay;
+ for(i = 0; i < net->n; ++i){
+ if(net->types[i] == CONVOLUTIONAL){
+ convolutional_layer *layer = (convolutional_layer *)net->layers[i];
+ layer->learning_rate=rate;
+ layer->momentum = momentum;
+ layer->decay = decay;
+ }
+ else if(net->types[i] == CONNECTED){
+ connected_layer *layer = (connected_layer *)net->layers[i];
+ layer->learning_rate=rate;
+ layer->momentum = momentum;
+ layer->decay = decay;
}
}
- visualize_network(net);
- cvWaitKey(100);
- fprintf(stderr, "Accuracy: %f\n", (float)correct/d.X.rows);
}
+
+void set_batch_network(network *net, int b)
+{
+ net->batch = b;
+ int i;
+ for(i = 0; i < net->n; ++i){
+ if(net->types[i] == CONVOLUTIONAL){
+ convolutional_layer *layer = (convolutional_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == MAXPOOL){
+ maxpool_layer *layer = (maxpool_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == CONNECTED){
+ connected_layer *layer = (connected_layer *)net->layers[i];
+ layer->batch = b;
+ } else if(net->types[i] == DROPOUT){
+ dropout_layer *layer = (dropout_layer *) net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == FREEWEIGHT){
+ freeweight_layer *layer = (freeweight_layer *) net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == SOFTMAX){
+ softmax_layer *layer = (softmax_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == COST){
+ cost_layer *layer = (cost_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == CROP){
+ crop_layer *layer = (crop_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ }
+}
+
+
int get_network_input_size_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
@@ -527,6 +396,9 @@
} else if(net.types[i] == DROPOUT){
dropout_layer layer = *(dropout_layer *) net.layers[i];
return layer.inputs;
+ } else if(net.types[i] == CROP){
+ crop_layer layer = *(crop_layer *) net.layers[i];
+ return layer.c*layer.h*layer.w;
}
else if(net.types[i] == FREEWEIGHT){
freeweight_layer layer = *(freeweight_layer *) net.layers[i];
@@ -536,6 +408,7 @@
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.inputs;
}
+ printf("Can't find input size\n");
return 0;
}
@@ -551,6 +424,10 @@
image output = get_maxpool_image(layer);
return output.h*output.w*output.c;
}
+ else if(net.types[i] == CROP){
+ crop_layer layer = *(crop_layer *) net.layers[i];
+ return layer.c*layer.crop_height*layer.crop_width;
+ }
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.outputs;
@@ -567,6 +444,7 @@
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.inputs;
}
+ printf("Can't find output size\n");
return 0;
}
@@ -664,29 +542,20 @@
}
}
-void top_predictions(network net, int n, int *index)
+void top_predictions(network net, int k, int *index)
{
- int i,j;
- int k = get_network_output_size(net);
+ int size = get_network_output_size(net);
float *out = get_network_output(net);
- float thresh = FLT_MAX;
- for(i = 0; i < n; ++i){
- float max = -FLT_MAX;
- int max_i = -1;
- for(j = 0; j < k; ++j){
- float val = out[j];
- if(val > max && val < thresh){
- max = val;
- max_i = j;
- }
- }
- index[i] = max_i;
- thresh = max;
- }
+ top_k(out, size, k, index);
}
+
float *network_predict(network net, float *input)
{
+ #ifdef GPU
+ if(gpu_index >= 0) return network_predict_gpu(net, input);
+ #endif
+
forward_network(net, input, 0, 0);
float *out = get_network_output(net);
return out;
@@ -722,7 +591,7 @@
int i,j,b;
int k = get_network_output_size(net);
matrix pred = make_matrix(test.X.rows, k);
- float *X = calloc(net.batch*test.X.rows, sizeof(float));
+ float *X = calloc(net.batch*test.X.cols, sizeof(float));
for(i = 0; i < test.X.rows; i += net.batch){
for(b = 0; b < net.batch; ++b){
if(i+b == test.X.rows) break;
@@ -784,18 +653,54 @@
}
}
+void compare_networks(network n1, network n2, data test)
+{
+ matrix g1 = network_predict_data(n1, test);
+ matrix g2 = network_predict_data(n2, test);
+ int i;
+ int a,b,c,d;
+ a = b = c = d = 0;
+ for(i = 0; i < g1.rows; ++i){
+ int truth = max_index(test.y.vals[i], test.y.cols);
+ int p1 = max_index(g1.vals[i], g1.cols);
+ int p2 = max_index(g2.vals[i], g2.cols);
+ if(p1 == truth){
+ if(p2 == truth) ++d;
+ else ++c;
+ }else{
+ if(p2 == truth) ++b;
+ else ++a;
+ }
+ }
+ printf("%5d %5d\n%5d %5d\n", a, b, c, d);
+ float num = pow((abs(b - c) - 1.), 2.);
+ float den = b + c;
+ printf("%f\n", num/den);
+}
+
float network_accuracy(network net, data d)
{
matrix guess = network_predict_data(net, d);
- float acc = matrix_accuracy(d.y, guess);
+ float acc = matrix_topk_accuracy(d.y, guess,1);
free_matrix(guess);
return acc;
}
+float *network_accuracies(network net, data d)
+{
+ static float acc[2];
+ matrix guess = network_predict_data(net, d);
+ acc[0] = matrix_topk_accuracy(d.y, guess,1);
+ acc[1] = matrix_topk_accuracy(d.y, guess,5);
+ free_matrix(guess);
+ return acc;
+}
+
+
float network_accuracy_multi(network net, data d, int n)
{
matrix guess = network_predict_data_multi(net, d, n);
- float acc = matrix_accuracy(d.y, guess);
+ float acc = matrix_topk_accuracy(d.y, guess,1);
free_matrix(guess);
return acc;
}
--
Gitblit v1.10.0