mem management:avoid potential infinite allocation

src/cnn/include/models.h

@@ -22,6 +22,13 @@ Network* create_network_alexnet(float learning_rate, int dropout, int activation
 */
 Network* create_network_VGG16(float learning_rate, int dropout, int activation, int initialisation, int size_output);
 
+
+/*
+* Renvoie un réseau suivant l'architecture VGG16 originel pour prendre en entrée 3x227x227
+* et une sortie de taille 1 000
+*/
+Network* create_network_VGG16_227(float learning_rate, int dropout, int activation, int initialisation);
+
 /*
 * Renvoie un réseau sans convolution, similaire à celui utilisé dans src/dense
 */

src/cnn/models.c

@@ -36,37 +36,67 @@ Network* create_network_alexnet(float learning_rate, int dropout, int activation
 }
 
 Network* create_network_VGG16(float learning_rate, int dropout, int activation, int initialisation, int size_output) {
-    Network* network = create_network(23, learning_rate, dropout, initialisation, 256, 3);
+    Network* network = create_network(22, learning_rate, dropout, initialisation, 256, 3);
-    add_convolution(network, 3, 64, 1, 0, activation); // Conv3-64
+    add_convolution(network, 3, 64, 1, 1, activation); // Conv3-64
-    add_convolution(network, 3, 64, 1, 0, activation); // Conv3-64
+    add_convolution(network, 3, 64, 1, 1, activation); // Conv3-64
     add_average_pooling(network, 2, 2, 0); // Max Pool
 
-    add_convolution(network, 3, 128, 1, 0, activation); // Conv3-128
+    add_convolution(network, 3, 128, 1, 1, activation); // Conv3-128
     add_convolution(network, 1, 128, 1, 0, activation); // Conv1-128
     add_average_pooling(network, 2, 2, 0); // Max Pool
 
-    add_convolution(network, 3, 256, 1, 0, activation); // Conv3-256
+    add_convolution(network, 3, 256, 1, 1, activation); // Conv3-256
-    add_convolution(network, 3, 256, 1, 0, activation); // Conv3-256
+    add_convolution(network, 3, 256, 1, 1, activation); // Conv3-256
     add_convolution(network, 1, 256, 1, 0, activation); // Conv1-256
     add_average_pooling(network, 2, 2, 0); // Max Pool
 
-    add_convolution(network, 3, 512, 1, 0, activation); // Conv3-512
+    add_convolution(network, 3, 512, 1, 1, activation); // Conv3-512
-    add_convolution(network, 3, 512, 1, 0, activation); // Conv3-512
+    add_convolution(network, 3, 512, 1, 1, activation); // Conv3-512
     add_convolution(network, 1, 512, 1, 0, activation); // Conv1-512
     add_average_pooling(network, 2, 2, 0); // Max Pool
 
-    add_convolution(network, 3, 512, 1, 0, activation); // Conv3-512
+    add_convolution(network, 3, 512, 1, 1, activation); // Conv3-512
-    add_convolution(network, 3, 512, 1, 0, activation); // Conv3-512
+    add_convolution(network, 3, 512, 1, 1, activation); // Conv3-512
     add_convolution(network, 1, 512, 1, 0, activation); // Conv1-512
     add_average_pooling(network, 2, 2, 0); // Max Pool
 
-    add_dense_linearisation(network, 2048, activation);
+    add_dense_linearisation(network, 4096, activation);
-    add_dense(network, 2048, activation);
+    add_dense(network, 4096, activation);
-    add_dense(network, 256, activation);
     add_dense(network, size_output, SOFTMAX);
     return network;
 }
 
+Network* create_network_VGG16_227(float learning_rate, int dropout, int activation, int initialisation) {
+    Network* network = create_network(22, learning_rate, dropout, initialisation, 227, 3);
+    add_convolution(network, 3, 64, 1, 1, activation); // Conv3-64
+    add_convolution(network, 3, 64, 1, 1, activation); // Conv3-64
+    add_average_pooling(network, 2, 2, 0); // Max Pool
+
+    add_convolution(network, 3, 128, 1, 1, activation); // Conv3-128
+    add_convolution(network, 1, 128, 1, 0, activation); // Conv1-128
+    add_average_pooling(network, 2, 2, 0); // Max Pool
+
+    add_convolution(network, 3, 256, 1, 1, activation); // Conv3-256
+    add_convolution(network, 3, 256, 1, 1, activation); // Conv3-256
+    add_convolution(network, 1, 256, 1, 0, activation); // Conv1-256
+    add_average_pooling(network, 2, 2, 0); // Max Pool
+
+    add_convolution(network, 3, 512, 1, 1, activation); // Conv3-512
+    add_convolution(network, 3, 512, 1, 1, activation); // Conv3-512
+    add_convolution(network, 1, 512, 1, 0, activation); // Conv1-512
+    add_average_pooling(network, 2, 2, 0); // Max Pool
+
+    add_convolution(network, 3, 512, 1, 1, activation); // Conv3-512
+    add_convolution(network, 3, 512, 1, 1, activation); // Conv3-512
+    add_convolution(network, 1, 512, 1, 0, activation); // Conv1-512
+    add_average_pooling(network, 2, 2, 0); // Max Pool
+
+    add_dense_linearisation(network, 4096, activation);
+    add_dense(network, 4096, activation);
+    add_dense(network, 1000, SOFTMAX);
+    return network;
+}
+
 Network* create_simple_one(float learning_rate, int dropout, int activation, int initialisation, int input_width, int input_depth) {
     Network* network = create_network(3, learning_rate, dropout, initialisation, input_width, input_depth);
     add_dense_linearisation(network, 80, activation);

src/common/memory_management.c

@@ -136,8 +136,7 @@ void* allocate_memory(int nb_elements, size_t size, Memory* mem) {
         //printf("Mémoire disponible: %ld. Nécessaire: %ld\n", mem->size - ((intptr_t)mem->cursor - (intptr_t)mem->start), nb_elements*size);
         // Sinon on continue sur l'élément suivant de la liste
         if (!mem->next) {
-            //! WARNING: May cause Infinite allocations when trying to allocate more than MEMORY_BLOCK size at once that is not naturally aligned (CUDA only)
+            mem->next = create_memory_block(MEMORY_BLOCK < (nb_elements+1)*size ? (nb_elements+1)*size : MEMORY_BLOCK);
-            mem->next = create_memory_block(MEMORY_BLOCK < nb_elements*size ? nb_elements*size : MEMORY_BLOCK);
         }
         return allocate_memory(nb_elements, size, mem->next);
     }

src/common/memory_management.cu

@@ -136,8 +136,7 @@ void* allocate_memory(int nb_elements, size_t size, Memory* mem) {
         //printf("Mémoire disponible: %ld. Nécessaire: %ld\n", mem->size - ((intptr_t)mem->cursor - (intptr_t)mem->start), nb_elements*size);
         // Sinon on continue sur l'élément suivant de la liste
         if (!mem->next) {
-            //! WARNING: May cause Infinite allocations when trying to allocate more than MEMORY_BLOCK size at once that is not naturally aligned (CUDA only)
+            mem->next = create_memory_block(MEMORY_BLOCK < (nb_elements+1)*size ? (nb_elements+1)*size : MEMORY_BLOCK);
-            mem->next = create_memory_block(MEMORY_BLOCK < nb_elements*size ? nb_elements*size : MEMORY_BLOCK);
         }
         return allocate_memory(nb_elements, size, mem->next);
     }