Fixes 1.0

cmake/eigen.CMakeLists.txt.in

@@ -5,13 +5,7 @@ project(eigen-download NONE)
 # Set base dir
 SET(EP_BASE_DIR @EP_BASE_DIR@)
 
-if(CMAKE_GENERATOR MATCHES "Visual Studio")
-    # Visual Studio needs a more recent version of Eigen3
-    # Version 3.3.90
-	SET(TAG ba9d18b9388acdf27a3900a4f981fab587e59b0c) 
-else()
-	SET(TAG "3.3.7") # Eigen3 must be >= 3.3.7
-endif()
+SET(TAG "3.4.0")
 
 include(ExternalProject)
 ExternalProject_Add(eigen

docs/markdown/eddl_progress.md

@@ -31,10 +31,13 @@
 | Expand        |  🟢️️| 🟢️️ |🟢️️ | 🟢️ | Returns a layer with singleton dimensions expanded to a larger size |
 | Where | 🟢️️ | 🟢️️ | 🟢️️ | ️🔴️ | Return elements chosen from x or y depending on condition |
 | Resize          | 🟢️️ | 🟢️️ |🟢️️ | 🟢️️ | Resize the input image to the given size. `[height, width]` |
-| Clamp / Clip          | 🟢️️ | 🟢️️ |🟢️️ | 🔴️️ | Clamps all elements in input into the range `[min, max]`.  |
-| Repeat          | 🟢️️ | 🟢️️ |🟢️️ | 🔴️️ | Repeats the elements of a tensor along the specified dimension.  |
-
-
+| Clamp / Clip          | 🟢️️ | 🟢️️ |🟢️️ | 🟢️ | Clamps all elements in input into the range `[min, max]`.  |
+| Repeat          | 🟢️️ | 🟢️️ |🟢️️ | 🔴️️ | Repeats the elements of a tensor along the specified dimension. (Elements in an axis can be repeated independently)  |
+| Tile          | 🔴️️ | 🔴 |🔴️ | 🔴️️ | Repeats the elements of a tensor along the specified dimensions.  |
+| Round          | 🔴️️ | 🔴️️ | 🔴️️ | 🔴️️ | Round of the elements of input |
+| Ceil          | 🔴️ | 🔴️️ | 🔴️ | 🔴️️ | Ceil of the elements of input  |
+| Floor          | 🔴️️ | 🔴 |🔴️ | 🔴️️ | Floor of the elements of input  |
+| Broadcast          | 🔴️️ | 🔴 |🔴️ | 🔴️️ | Produce an object that mimics broadcasting. |
 
 
 ## Activations

docs/markdown/eddl_progress_tensor.md

@@ -127,8 +127,9 @@ Numpy-like operations over a raw-tensor object
 
 | Functionality | CPU  | GPU  | Comments |
 | ------------- | ---- | ---- | -------- |
-| tile          | ⚫️   | ⚫️   |  	Construct an array by repeating A the number of times given by reps |
 | repeat        | 🟢️    | 🟢️️️ ️   |  Repeats the elements of a tensor along the specified dimension |
+| tile        | 🔴️️    | 🔴️️️ ️   |  Repeats the elements of a tensor along the specified dimensions |
+| broadcast        | 🔴️️    | 🔴️️️ ️   |  Produce an object that mimics broadcasting. |
 
 
 ### Adding and removing elements

docs/sphinx/source/conf.py

@@ -73,4 +73,7 @@
 
 # html_sidebars = {
 #     '**': ['searchbox.html'],
-# }
+# }
+
+
+github_url = "https://github.com/deephealthproject/eddl"

docs/sphinx/source/intro/build-options.rst

@@ -56,6 +56,8 @@ If you decide to manually install these dependencies in your system (make sure t
         - sphinx_rtd_theme==1.0.*
         - sphinx-tabs==3.2.*
         - breathe==4.31.*
+        - onnx-simplifier==0.3.*
+        - protobuf==3.19.*
 
 .. note::
 

docs/sphinx/source/intro/installation.rst

@@ -123,6 +123,7 @@ You can also install ``EDDL`` from source with cmake.
             # -DCMAKE_C_COMPILER=$(which gcc-7) \
             # -DCMAKE_CXX_COMPILER=$(which g++-7) \
 
+            make -j$(nproc)  # The -j flag is to parallelize the compilation
             make install
 
     .. tab:: MacOS
@@ -142,6 +143,7 @@ You can also install ``EDDL`` from source with cmake.
             cd build
             cmake .. -DCMAKE_PREFIX_PATH=$CONDA_PREFIX -DCMAKE_INSTALL_PREFIX=$CONDA_PREFIX
 
+            make -j$(sysctl -n hw.logicalcpu)  # The -j flag is to parallelize the compilation
             make install
 
     .. tab:: Windows
@@ -186,6 +188,16 @@ See the :doc:`build-options` section for more details about cmake options.
 
         cmake .. -DBUILD_SUPERBUILD=ON -DBUILD_TARGET=CUDNN -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc -DCMAKE_CXX_COMPILER=/usr/bin/g++-7
 
+    6. To update/remote the eddl environment type:
+
+    .. code:: bash
+
+        # Update existing environment
+        conda env update --file environment.yml --prune
+
+        # Remove environment
+        conda remove --name eddl --all
+
 
 Including EDDL in your project
 ---------------------------------

docs/sphinx/source/model/onnx.rst

@@ -46,6 +46,8 @@ you try to "standardize" or "simplifying" using *onnx_simplifier_*.
 
 **Installation**
 
+This is only needed is you are not using the conda environment for the EDDL, as we ship ONNX simplifier with it.
+
 .. code-block:: bash
 
     pip3 install -U pip && pip3 install onnx-simplifier

environment.yml

@@ -21,4 +21,6 @@ dependencies:  # https://docs.conda.io/projects/conda-build/en/latest/resources/
     - sphinx==4.3.*
     - sphinx_rtd_theme==1.0.*
     - sphinx-tabs==3.2.*
-    - breathe==4.31.*
+    - breathe==4.31.*
+    - onnx-simplifier==0.3.*
+    - protobuf==3.19.*

examples/CMakeLists.txt

@@ -125,6 +125,9 @@ if(BUILD_PROTOBUF)
 	add_executable(app_image_classification_resnet34 "applications/1_image_classification_resnet34.cpp")
 	target_link_libraries(app_image_classification_resnet34 eddl)
 
+	add_executable(app_test_model "applications/app_test_model.cpp")
+	target_link_libraries(app_test_model eddl)
+
 
 	# EXAMPLES: RAW ONNX **************************************************************
 	add_executable(onnx_pointer "onnx/1_onnx_pointer.cpp")

examples/applications/app_test_model.cpp

@@ -0,0 +1,181 @@
+
+#include <iostream>
+#include "eddl/apis/eddl.h"
+#include "eddl/serialization/onnx/eddl_onnx.h"
+
+using namespace eddl;
+
+
+Tensor* preprocess_input(Tensor* input, const vector<int> &target_size, bool normalize=true, bool standarize=true){
+    // Define preprocessing constants
+    auto* mean_vec = new Tensor( {0.485, 0.456, 0.406}, {3, 1}, input->device);
+    auto* std_vec = new Tensor( {0.229, 0.224, 0.225}, {3, 1}, input->device);
+
+    // ==========================================================================
+    // ====== SANITY CHECKS =====================================================
+    // ==========================================================================
+    // Check dimension. Input must be a 3D or 4D tensor
+    if(!(input->ndim == 3 || input->ndim == 4)){
+        throw std::runtime_error("A 3D or 4D tensor is expected. " + std::to_string(input->ndim) + "D tensor received.");
+    }
+
+    // Convert from 3D to 4D (if needed)
+    if(input->ndim == 3){
+        input->unsqueeze_(0);
+    }
+    // ==========================================================================
+
+
+    // ==========================================================================
+    // ====== NORMALIZATION =====================================================
+    // ==========================================================================
+
+    // Resize tensor (creates a new instance)
+    Tensor* new_input = input->scale(target_size);  // (height, width)
+
+    // Normalization [0..1]
+    if(normalize){
+        new_input->mult_(1/255.0f);
+    }
+
+    // Standarization: (X-mean)/std
+    if(standarize){
+//    Tensor* mean = Tensor::broadcast(mean_vec, new_input);
+//    Tensor* std = Tensor::broadcast(std_vec, new_input);
+        // 1) [There is no broadcasting...] Repeat dimensions  => Temp!
+        Tensor* mean = Tensor::repeat(mean_vec, target_size[0]*target_size[1], 1);  mean->reshape_(new_input->shape);
+        Tensor* std =  Tensor::repeat(std_vec, target_size[0]*target_size[1], 1); std->reshape_(new_input->shape);
+        new_input->sub_(mean);
+        new_input->div_(std);
+
+        // Free memory
+        delete mean;
+        delete std;
+    }
+    // ==========================================================================
+
+    // Free memory
+    delete mean_vec;
+    delete std_vec;
+
+    return new_input;
+}
+
+
+int main(int argc, char **argv) {
+    // ==========================================================================
+    // ====== SET DEFAULT VARIABLES =============================================
+    // ==========================================================================
+
+    // Step 0: Download the model, the classes and the image we want to classify
+    string image_fname = "../../examples/data/elephant.jpg";
+    string class_names_file = "../../examples/data/imagenet_class_names.txt";
+
+    // Image Classification
+    string model_path = "models/resnet34-v1-7.onnx";  // 3x224x224  // okay
+//    string model_path = "models/mobilenetv2-7.onnx"; // 3x224x224 // Signal: SIGSEGV (Segmentation fault)
+//    string model_path = "models/vgg16-7.onnx";  // 3xHxW  // okay
+//    string model_path = "models/bvlcalexnet-3.onnx";  // 3x224x224  // The onnx node 'LRN' is not supported yet
+//    string model_path = "models/bvlcalexnet-12.onnx";  // 3x224x224  // The onnx node 'LRN' is not supported yet
+//    string model_path = "models/googlenet-3.onnx";  // 3x224x224  // The onnx node 'LRN' is not supported yet
+//    string model_path = "models/densenet-3.onnx";  // 3x224x224  // okay
+//    string model_path = "models/inception-v1-3.onnx";  // 3x224x224  // The onnx node 'LRN' is not supported yet
+//    string model_path = "models/efficientnet-lite4-11.onnx";  // 224x224x3  // The onnx node 'LRN' is not supported yet
+
+    // Object Detection & Image Segmentation
+//    string model_path = "models/tinyyolov2-7.onnx";  // 3x416x416 //Error in Add node Add. The first dimension (batch) of the constant operator must be 1, got 3 (ONNX::ImportNet) ⚠️
+//    string model_path = "models/ssd-10_simp.onnx";  // 3x1200x1200 // Signal: SIGSEGV (Segmentation fault)
+//    string model_path = "models/FasterRCNN-10_simp.onnx";  // 3xHxW // Signal: SIGSEGV (Segmentation fault)
+//    string model_path = "models/yolov4_simp.onnx";  // 416x416x3 // [ONNX::LOG] Detected a padding asymmetry + The onnx node 'Shape' is not supported yet
+//    string model_path = "models/MaskRCNN-10_simp.onnx";  // 3xHxW // Signal: SIGSEGV (Segmentation fault)
+//    string model_path = "models/retinanet-9.onnx";  // 3xHxW // okay
+//    string model_path = "models/yolov3-10_simp.onnx";  // 3x416x416 // The imported model has more than 1 input layer and the shape provided to reshape the model can only be appliedif the model has only one input layer. (ONNX::ImportNet)
+//    string model_path = "models/fcn-resnet50-11_simp.onnx";  // 3xHxW // The onnx node 'Shape' is not supported yet
+//    string model_path = "models/ResNet101-DUC-7.onnx";  // 3xHxW // okay. Error with plot
+
+    // Image Manipulation
+//    string model_path = "models/super-resolution-10.onnx";  // 3x224x224 // Tensors with different size (Tensor::copy)
+//    string model_path = "models/mosaic-9.onnx";  // 3xHxW //  Error importing layer . Only "constant" mode is supported (passed "reflect"). (ONNX::ImportNet)
+
+    // Step 1: Specify input
+    int in_channels = 3;
+    int in_height = 224;
+    int in_width = 224;
+    vector<int> input_shape = {in_channels, in_height, in_width};
+    vector<int> dimensions_order = {0, 1, 2, 3};
+    // ==========================================================================
+
+
+    // ==========================================================================
+    // ====== LOAD ONNX MODEL ===================================================
+    // ==========================================================================
+
+    // Import ONNX model
+    std::cout << "Importing ONNX..." << std::endl;
+    Net *net = import_net_from_onnx_file(model_path, input_shape);
+
+    // ==========================================================================
+    // Print and plot our model
+    net->summary();
+    net->plot("mymodel.pdf");
+
+    // ==========================================================================
+    // ====== APPLY POSTPROCESSING ==============================================
+    // ==========================================================================
+
+    // Step 4 (optional): Add a softmax layer to get probabilities directly from the model, since it
+    // does not include the softmax layer.
+//    layer input = net->lin[0];   // getLayer(net,"input_layer_name");
+//    layer output = net->lout[0];   // getLayer(net,"output_layer_name");
+//    layer new_output = Softmax(output);
+//
+//    // Create model
+//    net = Model({input},{new_output});
+    // ==========================================================================
+
+    // ==========================================================================
+    // ====== COMPILE MODEL =====================================================
+    // ==========================================================================
+
+    // Build model
+    build(net,
+          adam(0.001f), // Optimizer (not used for prediction)
+          {"softmax_cross_entropy"},// Losses (not used for prediction)
+          {"categorical_accuracy"}, // Metrics (not used for prediction)
+          CS_GPU({1}), // Use one GPU
+          false       // Disable model initialization, since we want to use the onnx weights
+    );
+    // ==========================================================================
+
+
+    // ==========================================================================
+    // ====== INFERENCE =========================================================
+    // ==========================================================================
+    // Load test image
+    Tensor *image = Tensor::load(image_fname);
+
+    // Step 3: Preprocess input. (Look up the preprocessing required at the model's page)
+    Tensor* image_preprocessed = preprocess_input(image, {in_height, in_width});
+    image_preprocessed->permute_(dimensions_order);
+
+    // Predict image. Returns a vector of tensors (here one).
+    vector<Tensor*> outputs = net->predict({image_preprocessed});
+    // ==========================================================================
+
+
+    // ==========================================================================
+    // ====== PRINT TOP K CLASSES ===============================================
+    // ==========================================================================
+    // Read imagenet class names from txt file
+    std::cout << "Reading imagenet class names..." << std::endl;
+    vector<string> class_names = eddl::read_txt_file(class_names_file);
+
+    // Print top K predictions
+    int top_k = 5;
+    std::cout << "Top " << top_k << " predictions:" << std::endl;
+    std::cout << eddl::get_topk_predictions(outputs[0], class_names, top_k)  << std::endl;
+    // ==========================================================================
+
+    return 0;
+}
+

examples/test_internals/onnx/test_onnx_auxiliar_layers.cpp

@@ -69,6 +69,9 @@ int main(int argc, char **argv) {
         layer l = in;  // Aux var
 
         l = Reshape(l, {1, 28, 28});
+        l = Repeat(l, 3, 0);
+        l = Repeat(l, 2, 1);
+        l = Repeat(l, 2, 2);
         l = Pad(l, {1, 1, 1, 1}, 0.0);
         l = MaxPool2D(ReLu(Conv2D(l, 30, {3, 3}, {1, 1}, "valid")), {3, 3}, {1, 1}, "same");
 
@@ -194,4 +197,4 @@ int main(int argc, char **argv) {
         return EXIT_SUCCESS;
     else
         return EXIT_FAILURE;
-}
+}

include/eddl/serialization/onnx/layers/core/repeat_onnx.h

@@ -0,0 +1,29 @@
+#if defined(cPROTO)
+#ifndef EDDL_REPEAT_ONNX_H
+#define EDDL_REPEAT_ONNX_H
+#include "eddl/serialization/onnx/onnx.pb.h"
+#include "eddl/serialization/onnx/utils_onnx.h"
+#include "eddl/layers/core/layer_core.h"
+
+/*
+ * ONNX IMPORT
+ */
+
+// OPSET: 13, 6
+Layer* build_repeat_layer(onnx::NodeProto *node,
+                          map<string, onnx::NodeProto *> &constant_node_map, 
+                          map<string, vector<float>> &map_init_values,
+                          map<string, Layer *> &output_node_map,
+                          LOG_LEVEL log_level,
+                          int dev,
+                          int mem);
+
+/*
+ * ONNX EXPORT
+ */
+
+// OPSET: 13, 6
+void build_tile_node(LRepeat *layer, onnx::GraphProto *graph);
+
+#endif // EDDL_REPEAT_ONNX_H
+#endif // cPROTO

include/eddl/serialization/onnx/layers/layers_onnx.h

@@ -68,7 +68,8 @@ enum ONNX_LAYERS {
   SLICE,            // OPSET: 13, 11, 10
   SPLIT,            // OPSET: 13, 11, 2
   EXPAND,           // OPSET: 13, 8
-  CONSTANT          // OPSET: 13, 12, 11, 9, 1
+  CONSTANT,         // OPSET: 13, 12, 11, 9, 1
+  REPEAT            // OPSET: 13, 6
 };
 
 map<string, ONNX_LAYERS> create_enum_map();

include/eddl/utils.h

@@ -70,23 +70,27 @@ bool pathExists(const std::string &s);
 
 string get_parent_dir(const string& fname);
 
+string replace_str(const string& value, const string& oldvalue, const string& newvalue);
+
+string normalize_layer_name(const string& value);
+
 vector<int> compute_squeeze(vector<int> shape, int axis, bool ignore_batch=false);
 vector<int> compute_unsqueeze(vector<int> shape, int axis, bool ignore_batch=false);
 
-vector<int> address2indices(unsigned int address, const vector<int>& shape, const vector<int>& strides);
+vector<int> address2indices(int address, const vector<int>& shape, const vector<int>& strides);
 unsigned int indices2address(const vector<int>& indices, const vector<int>& strides);
 
 // https://isocpp.org/wiki/faq/inline-functions#inline-member-fns
-inline int fast_indices2address(const unsigned int* indices, const unsigned int* strides, unsigned int ndim){
-    unsigned int address = 0;
+inline int fast_indices2address(const int* indices, const int* strides, int ndim){
+    int address = 0;
     for (int i=0; i< ndim; i++){
         address += indices[i] * strides[i];
     }
     return address;
 }
 
 
-inline void fast_address2indices( int address, unsigned int* indices, const unsigned int* shape, const unsigned int* strides, unsigned int ndim){
+inline void fast_address2indices( int address, int* indices, const int* shape, const int* strides, int ndim){
     for(int i=0; i<ndim; i++) {
         indices[i] = address / strides[i] % shape[i];
     }