Merge pull request #205 from deephealthproject/develop

RParedesPalacios · web-flow · commit 8c9878af32e8 · 2020-10-12T12:14:21.000+02:00
Develop
diff --git a/examples/nn/4_NLP/4_nlp_text_generation.cpp b/examples/nn/4_NLP/4_nlp_text_generation.cpp
@@ -63,6 +63,8 @@ Tensor *onehot(Tensor *in, int vocs)
 
 int main(int argc, char **argv) {
 
+    download_flickr();
+
     // Settings
     int epochs = 10;
     int batch_size = 8;
@@ -112,7 +114,7 @@ int main(int argc, char **argv) {
     // Build model
     build(net,
           opt, // Optimizer
-          {"cross_entropy"}, // Losses
+          {"soft_cross_entropy"}, // Losses
           {"accuracy"}, // Metrics
           //CS_GPU({1}) // one GPU
           //CS_GPU({1,1},100) // two GPU with weight sync every 100 batches
@@ -126,16 +128,17 @@ int main(int argc, char **argv) {
     // Load dataset
     Tensor *x_train=Tensor::load("flickr_trX.bin","bin");
     x_train->info();
+
     Tensor *y_train=Tensor::load("flickr_trY.bin","bin");
     y_train->info();
-    //y_train->print();
+
     y_train=onehot(y_train,outvs);
     y_train->reshape_({y_train->shape[0],olength,outvs}); //batch x timesteps x input_dim
+    y_train->info();
 
     // Train model
     for(int i=0;i<epochs;i++) {
       fit(net, {x_train}, {y_train}, batch_size, 1);
-      //evaluate(net,{x_test},{y_test});
     }
 
 }
diff --git a/include/eddl/apis/eddl.h b/include/eddl/apis/eddl.h
@@ -466,7 +466,7 @@ typedef NetLoss * metric;
       *  @return    (void)
     */
     void eval_batch(model net, vector<Tensor *> in, vector<Tensor *> out, vector<int> indices);
-    
+
     /**
       *  @brief Loads the next batch of random samples from the input vector to the output vector
       *
@@ -1829,6 +1829,15 @@ typedef NetLoss * metric;
     */
     void download_eutrans();
 
+    /**
+      *  @brief Downloads Flickr Dataset (small partition)
+      *
+      *  @see
+      *
+      *  @return     (void) The binary files of Flickr
+    */
+    void download_flickr();
+
 
 }
 #endif
diff --git a/include/eddl/tensor/tensor.h b/include/eddl/tensor/tensor.h
@@ -159,7 +159,7 @@ class Tensor {
     void updateShape(const vector<int> &new_shape);
     void updateSize();
     void updateStrides();
-    void updateData(float* ptr, void *ptr2=NULL);
+    void updateData(float* ptr, void *ptr2=NULL,bool setshared=true);
     void deleteData();
 
     /**
diff --git a/src/apis/eddl.cpp b/src/apis/eddl.cpp
@@ -1263,6 +1263,10 @@ namespace eddl {
       download_dataset("eutrans","bin",{"2w0p7f4un6ci94v","g4k1bc6p4bow9tf","egcfin16gl9t92y","n8ks3lyqyhxx1e8"});
     }
 
+    void download_flickr(){
+      download_dataset("flickr","bin",{"452pyxe9x5jpnwb","24c2d5bm6pug8gg"});
+    }
+    
     void download_drive(){
       download_dataset("drive","npy",{"sbd8eu32adcf5oi","qp0j8oiqzf6tc1a"});
     }
diff --git a/src/descriptors/descriptor_conv.cpp b/src/descriptors/descriptor_conv.cpp
@@ -130,8 +130,8 @@ void ConvolDescriptor::build(Tensor *A) {
         // mem for ptr, lowering im2col
         ptrI=get_fmem(A->shape[0] * r * c * kr * kc * kz,"ConvolDescriptor::build");
 	 _profile_add_tensor(A->shape[0] * r * c * kr * kc * kz);
-        new(&matK) Eigen::Map<Eigen::MatrixXf>(K->ptr, kr * kc * kz, nk);
-        new(&matgK) Eigen::Map<Eigen::MatrixXf>(gK->ptr, kr * kc * kz, nk);
+        //new(&matK) Eigen::Map<Eigen::MatrixXf>(K->ptr, kr * kc * kz, nk);
+        //new(&matgK) Eigen::Map<Eigen::MatrixXf>(gK->ptr, kr * kc * kz, nk);
         // convolution: matC=matA*matK
     }
 #ifdef cGPU
diff --git a/src/layers/conv/layer_conv.cpp b/src/layers/conv/layer_conv.cpp
@@ -128,22 +128,21 @@ void LConv::apply_accumulated_gradients() {
 }
 
 Layer *LConv::share(int c, int bs, vector<Layer *> p) {
-    // TODO: share ComvDescriptor
-    LConv *n = new LConv(p[0], cd->ksize, cd->stride, cd->pad,  name, dev,mem_level);
+    LConv *n = new LConv(p[0], cd->ksize, cd->stride, cd->pad,  "share_"+name, dev,mem_level);
     n->orig = this;
     n->isshared=true;
     n->trainable = trainable;
 
     n->cd->use_bias=cd->use_bias;
 
     //share params
+
     for (int i = 0; i < n->params.size(); i++) delete n->params[i];
     n->params.clear();
 
 
     n->cd->K = cd->K;
     n->cd->bias = cd->bias;
-    n->cd->matK = cd->matK;
 
     n->params.push_back(n->cd->K);
     n->params.push_back(n->cd->bias);
diff --git a/src/net/net.cpp b/src/net/net.cpp
@@ -15,9 +15,7 @@
 #include <fstream>
 #include <string>
 #include <chrono>
-#include <thread>
 #include "eddl/net/net.h"
-#include <pthread.h>
 #include "eddl/utils.h"
 #include "eddl/random.h"
 
diff --git a/src/net/net_api.cpp b/src/net/net_api.cpp
@@ -805,7 +805,12 @@ void Net::prepare_recurrent(vtensor tin, vtensor tout, int &inl, int &outl, vten
       for(j=1;j<xtd[i]->ndim;j++)
         shape.push_back(xtd[i]->shape[j]);
 
-      tinr.push_back(Z);
+      vector<int>zero_shape;
+      for(j=0;j<tout[i]->ndim;j++)
+        if (j!=1) zero_shape.push_back(tout[i]->shape[j]);
+
+      if (!isencoder) tinr.push_back(new Tensor(tin[0]->shape,tin[0]->ptr,tin[0]->device));
+      tinr.push_back(Tensor::zeros(zero_shape,tout[i]->device));
       for(j=0;j<outl-1;j++)
         tinr.push_back(new Tensor(shape,xtd[i]->ptr+(j*offset),xtd[i]->device));
     }
@@ -837,12 +842,7 @@ void Net::fit_recurrent(vtensor tin, vtensor tout, int batch, int epochs) {
   int inl;
   int outl;
 
-  vector<int>shape;
-  for(j=0;j<tout[0]->ndim;j++)
-    if (j!=1) shape.push_back(tout[0]->shape[j]);
-  Tensor *Z=Tensor::zeros(shape,tout[0]->device);
-
-  prepare_recurrent(tin,tout,inl,outl,xt,xtd,yt,tinr,toutr,Z);
+  prepare_recurrent(tin,tout,inl,outl,xt,xtd,yt,tinr,toutr);
 
   if (rnet==nullptr) build_rnet(inl,outl);
 
@@ -851,10 +851,13 @@ void Net::fit_recurrent(vtensor tin, vtensor tout, int batch, int epochs) {
   else if (isencoder)
     rnet->fit(tinr,tout,batch,epochs);
   else if (isdecoder)
-    rnet->fit(tin,toutr,batch,epochs);
+    rnet->fit(tinr,toutr,batch,epochs);
 
   if (snets[0]->dev!=DEV_CPU) rnet->sync_weights();
 
+  for(i=0;i<tinr.size();i++) delete(tinr[i]);
+  for(i=0;i<toutr.size();i++) delete(toutr[i]);
+
   if (isencoder) {
     for(i=0;i<xt.size();i++)
       delete xt[i];
@@ -870,8 +873,6 @@ void Net::fit_recurrent(vtensor tin, vtensor tout, int batch, int epochs) {
     yt.clear();
   }
 
-  delete Z;
-
 }
 
 // TODO:  train_batch_recurrent
diff --git a/src/net/net_build.cpp b/src/net/net_build.cpp
@@ -14,9 +14,7 @@
 #include <fstream>
 #include <string>
 #include <chrono>
-#include <thread>
 #include "eddl/net/net.h"
-#include <pthread.h>
 #include "eddl/utils.h"
 #include "eddl/random.h"
 
diff --git a/src/net/net_func.cpp b/src/net/net_func.cpp
@@ -14,9 +14,7 @@
 #include <fstream>
 #include <string>
 #include <chrono>
-#include <thread>
 #include "eddl/net/net.h"
-#include <pthread.h>
 #include "eddl/utils.h"
 #include "eddl/random.h"
 #include "eddl/layers/core/layer_core.h"
diff --git a/src/net/net_unroll.cpp b/src/net/net_unroll.cpp
@@ -14,9 +14,7 @@
 #include <fstream>
 #include <string>
 #include <chrono>
-#include <thread>
 #include "eddl/net/net.h"
-#include <pthread.h>
 #include "eddl/utils.h"
 #include "eddl/random.h"
 
@@ -379,7 +377,7 @@ Net* Net::unroll_dec(int inl, int outl) {
   */
 
   // unroll inputs
-  nin=new vlayer[inl];
+  nin=new vlayer[inl+outl];
   nlayers=new vlayer[outl];
   nout=new vlayer[outl];
 
@@ -411,8 +409,10 @@ Net* Net::unroll_dec(int inl, int outl) {
               Layer *n;
               n=layers[j]->share(i, batch_size, par);
               nlayers[i].push_back(n);
-              if (layers[j]->lin==0)
+              if (layers[j]->lin==0) {
+                nin[i].push_back(n);
                 din.push_back(n);
+              }
             }
           }
           else msg("Unexpected error","unroll");
@@ -433,7 +433,7 @@ Net* Net::unroll_dec(int inl, int outl) {
 /////
 vlayer ninl;
 vlayer noutl;
-for (i = 0; i < inl; i++)
+for (i = 0; i < inl+outl; i++)
   for (j = 0; j < nin[i].size(); j++)
     ninl.push_back(nin[i][j]);
 
diff --git a/src/tensor/tensor.cpp b/src/tensor/tensor.cpp
@@ -141,7 +141,7 @@ void Tensor::updateStrides() {
 
 void Tensor::deleteData(){
     // Carefpdal, you can't know is a pointer is allocated
-    //if (isshared) return;
+    if (isshared) return;
 
     if(this->ptr != nullptr){
         if (this->isCPU()) {
@@ -176,14 +176,15 @@ void Tensor::deleteData(){
     }
 }
 
-void Tensor::updateData(float *fptr, void *fptr2){
+void Tensor::updateData(float *fptr, void *fptr2,bool setshared){
     // TODO: What if the new_pointer is the same?
+    // Solved with setshared for reshape_
     isshared=false;
     if (this->isCPU()) {
         // If null => Reserve memory
         // else => point to data
         if (fptr==nullptr) { this->ptr = get_fmem(this->size,"Tensor::updateData"); }
-        else { this->ptr = fptr; isshared=true;};
+        else { this->ptr = fptr; isshared=setshared;};
 
         // For 2 dimensions, map to data to Eigen for efficiency
         // Efficient operations will be done over ptr2, which also points to ptr
@@ -203,7 +204,7 @@ void Tensor::updateData(float *fptr, void *fptr2){
         // If null => Reserve memory
         // else => point to data  | CAREFUL! This pointer MUST be a GPU pointer. We cannot check it.
         if (fptr == nullptr) { this->ptr = gpu_create_tensor(this->gpu_device, this->size); }
-        else { this->ptr = fptr; isshared=true;}
+        else { this->ptr = fptr; isshared=setshared;}
 
     }
 #endif
@@ -397,6 +398,7 @@ void Tensor::info() {
     cout << setw(cols) << left << "contiguous: "   << true << endl; // for future
     cout << setw(cols) << left << "order: "        << 'C' << endl;  // C=>C order, F=>Fortran order
     cout << setw(cols) << left << "data pointer: " << &this->ptr << endl;
+    cout << setw(cols) << left << "is shared: " << isshared << endl;
     cout << setw(cols) << left << "type: "         << "float" << " (" << sizeof(float) << " bytes)" << endl;
     cout << setw(cols) << left << "device: " << this->getDeviceName() << " (code = " << this->device << ")" << endl;
     cout << "-------------------------------" << endl;
diff --git a/src/tensor/tensor_core.cpp b/src/tensor/tensor_core.cpp
@@ -167,7 +167,7 @@ void Tensor::reshape_(const vector<int> &new_shape){
     updateShape(final_shape);
     updateSize();
     updateStrides();
-    updateData(this->ptr);  // Due to the Eigen mapping
+    updateData(this->ptr,nullptr, false);  // Due to potential the Eigen mapping
 
 }
 
@@ -786,7 +786,7 @@ void Tensor::deselect(Tensor *A, Tensor *B, vector<int> sind, int ini, int end,i
         fpga_deselect(A, B, sind, ini, end, inc,mask_zeros);
       }
     #endif
-	
+
     else {
         msg("unsuppoted select", "Tensor::select");
     }

Original file line number	Diff line number	Diff line change
`@@ -1263,6 +1263,10 @@ namespace eddl {`
`1263`	`1263`	`download_dataset("eutrans","bin",{"2w0p7f4un6ci94v","g4k1bc6p4bow9tf","egcfin16gl9t92y","n8ks3lyqyhxx1e8"});`
`1264`	`1264`	`}`
`1265`	`1265`
	`1266`	`+ void download_flickr(){`
	`1267`	`+ download_dataset("flickr","bin",{"452pyxe9x5jpnwb","24c2d5bm6pug8gg"});`
	`1268`	`+ }`
	`1269`	`+`
`1266`	`1270`	`void download_drive(){`
`1267`	`1271`	`download_dataset("drive","npy",{"sbd8eu32adcf5oi","qp0j8oiqzf6tc1a"});`
`1268`	`1272`	`}`