feat(torch): nbeats

feat(nbeats): make nbeats able to handle signals that are more than 1D

feat(nbeats): expose nbeats net definition in api

Author: fantes

CMakeLists.txt

@@ -773,7 +773,7 @@ if (USE_TORCH)
   if (NOT USE_CPU_ONLY AND CUDA_FOUND)
     list(APPEND TORCH_LIB_DEPS ${TORCH_LOCATION}/lib/libc10_cuda.so ${TORCH_LOCATION}/lib/libtorch_cuda.so)
   else()
-    list(APPEND TORCH_LIB_DEPS iomp5)
+    list(APPEND TORCH_LIB_DEPS ${TORCH_LOCATION}/lib/libtorch_cpu.so iomp5)
   endif()
 
   set(TORCH_INC_DIR ${TORCH_LOCATION}/include/ ${TORCH_LOCATION}/include/torch/csrc/api/include/ ${CMAKE_BINARY_DIR}/pytorch/src/pytorch/torch/include/torch/csrc/api/include ${TORCH_LOCATION}/.. ${CMAKE_BINARY_DIR}/src)

examples/all/sinus/gen.py

@@ -31,7 +31,6 @@
         os.remove("predict/"+f)
     os.rmdir("predict")
 
-
 os.mkdir("train")
 os.mkdir("test")
 os.mkdir("predict")

src/CMakeLists.txt

@@ -81,6 +81,7 @@ if (USE_TORCH)
     backends/torch/torchinputconns.cc
     backends/torch/db.cpp
     backends/torch/db_lmdb.cpp
+    backends/torch/native/templates/nbeats.cc
 	basegraph.cc
 	caffegraphinput.cc
 	backends/torch/torchgraphbackend.cc

src/backends/torch/native/native.h

@@ -0,0 +1,7 @@
+#ifndef NATIVE_H
+#define NATIVE_H
+
+#include "native_net.h"
+#include "native_factory.h"
+
+#endif

src/backends/torch/native/native_factory.h

@@ -0,0 +1,55 @@
+#ifndef NATIVE_FACTORY_H
+#define NATIVE_FACTORY_H
+
+#include "native_net.h"
+#include "./templates/nbeats.h"
+#include "../torchinputconns.h"
+#include "apidata.h"
+
+namespace dd
+{
+  class NativeFactory
+  {
+  public:
+    template <class TInputConnectorStrategy>
+    static NativeModule *from_template(const std::string tdef,
+                                       const APIData template_params,
+                                       const TInputConnectorStrategy &inputc)
+    {
+      (void)(tdef);
+      (void)(template_params);
+      (void)(inputc);
+      return nullptr;
+    }
+
+    static bool valid_template_def(std::string tdef)
+    {
+      if (tdef.find("nbeats") != std::string::npos)
+        return true;
+      return false;
+    }
+
+    static bool is_timeserie(std::string tdef)
+    {
+      if (tdef.find("nbeats") != std::string::npos)
+        return true;
+      return false;
+    }
+  };
+
+  template <>
+  NativeModule *NativeFactory::from_template<CSVTSTorchInputFileConn>(
+      const std::string tdef, const APIData template_params,
+      const CSVTSTorchInputFileConn &inputc)
+  {
+    if (tdef.find("nbeats") != std::string::npos)
+      {
+        std::vector<std::string> p = template_params.get("template_params")
+                                         .get<std::vector<std::string>>();
+        return new NBeats(inputc, p);
+      }
+    else
+      return nullptr;
+  }
+}
+#endif

src/backends/torch/native/native_net.h

@@ -0,0 +1,74 @@
+#ifndef NATIVE_NET_H
+#define NATIVE_NET_H
+
+#include "torch/torch.h"
+
+namespace dd
+{
+
+  class NativeModule : public torch::nn::Module
+  {
+  public:
+    virtual torch::Tensor forward(torch::Tensor x) = 0;
+    virtual ~NativeModule()
+    {
+    }
+    /**
+     * \brief see torch::module::to
+     * @param device cpu / gpu
+     * @param non_blocking
+     */
+    virtual void to(torch::Device device, bool non_blocking = false)
+    {
+      torch::nn::Module::to(device, non_blocking);
+      _device = device;
+    }
+
+    /**
+     * \brief see torch::module::to
+     * @param dtype : torch::kFloat32 or torch::kFloat64
+     * @param non_blocking
+     */
+    virtual void to(torch::Dtype dtype, bool non_blocking = false)
+    {
+      torch::nn::Module::to(dtype, non_blocking);
+      _dtype = dtype;
+    }
+
+    /**
+     * \brief see torch::module::to
+     * @param device cpu / gpu
+     * @param dtype : torch::kFloat32 or torch::kFloat64
+     * @param non_blocking
+     */
+    virtual void to(torch::Device device, torch::Dtype dtype,
+                    bool non_blocking = false)
+    {
+      torch::nn::Module::to(device, dtype, non_blocking);
+      _device = device;
+      _dtype = dtype;
+    }
+
+    virtual torch::Tensor cleanup_output(torch::Tensor output)
+    {
+      return output;
+    }
+
+    virtual torch::Tensor loss(std::string loss, torch::Tensor input,
+                               torch::Tensor output, torch::Tensor target)
+        = 0;
+
+    virtual void update_input_connector(TorchInputInterface &inputc)
+    {
+      (void)(inputc);
+    }
+
+  protected:
+    torch::Dtype _dtype
+        = torch::kFloat32; /**< type of data stored in tensors */
+    torch::Device _device
+        = torch::DeviceType::CPU; /**< device to compute on */
+  };
+}
+
+#endif