Strange silent compiler exit + nim-lang/Nim#16653

flambeau/cpp/std_cpp.nim

@@ -33,3 +33,32 @@ func `$`*(s: CppString): string =
   copyMem(result[0].addr, s.data.unsafeAddr, s.len)
 
 {.pop.}
+
+# std::shared_ptr<T>
+# -----------------------------------------------------------------------
+
+{.push header: "<memory>".}
+
+type
+  CppSharedPtr* {.importcpp: "std::shared_ptr", bycopy.} [T] = object
+
+func make_shared*(T: typedesc): CppSharedPtr[T] {.importcpp: "std::make_shared<'*0>()".}
+
+{.pop.}
+
+# std::vector<T>
+# -----------------------------------------------------------------------
+
+{.push header: "<memory>".}
+
+type
+  CppVector* {.importcpp"std::vector", header: "<vector>", bycopy.} [T] = object
+
+proc init*(V: type CppVector): V {.importcpp: "std::vector<'*0>()", header: "<vector>", constructor.}
+proc init*(V: type CppVector, size: int): V {.importcpp: "std::vector<'*0>(#)", header: "<vector>", constructor.}
+proc len*(v: CppVector): int {.importcpp: "#.size()", header: "<vector>".}
+proc add*[T](v: var CppVector[T], elem: T){.importcpp: "#.push_back(#)", header: "<vector>".}
+proc `[]`*[T](v: CppVector[T], idx: int): T{.importcpp: "#[#]", header: "<vector>".}
+proc `[]`*[T](v: var CppVector[T], idx: int): var T{.importcpp: "#[#]", header: "<vector>".}
+
+{.pop.}

flambeau/raw_bindings/data_api.nim

@@ -0,0 +1,101 @@
+# Flambeau
+# Copyright (c) 2020 Mamy André-Ratsimbazafy
+# Licensed and distributed under either of
+#   * MIT license (license terms in the root directory or at http://opensource.org/licenses/MIT).
+#   * Apache v2 license (license terms in the root directory or at http://www.apache.org/licenses/LICENSE-2.0).
+# at your option. This file may not be copied, modified, or distributed except according to those terms.
+
+import
+  ./tensors,
+  ../cpp/std_cpp
+
+# (Almost) raw bindings to PyTorch Data API
+# -----------------------------------------------------------------------
+#
+# This provides almost raw bindings to PyTorch data API.
+#
+# "Nimification" (camelCase), ergonomic indexing and interoperability with Nim types is left to the "high-level" bindings.
+# This should ease searching PyTorch and libtorch documentation,
+# and make C++ tutorials easily applicable.
+
+# #######################################################################
+#
+#                         Datasets
+#
+# #######################################################################
+#
+# Custom Dataset example: https://github.com/mhubii/libtorch_custom_dataset
+# libtorch/include/torch/csrc/api/include/torch/data/datasets/base.h
+
+type
+  Example*{.bycopy, importcpp: "torch::data::Example".}
+      [Data, Target] = object
+    data*: Data
+    target*: Target
+
+  # TODO: https://github.com/nim-lang/Nim/issues/16653
+  #   generics + {.inheritable.} doesn't work
+  BatchDataset*
+        {.bycopy, pure, inheritable,
+        importcpp: "torch::data::datasets::BatchDataset".}
+        # [Self, Batch, BatchRequest] # TODO: generic inheritable https://github.com/nim-lang/Nim/issues/16653
+      = object
+    ## A BatchDataset type
+    ## Self: is the class type that implements the Dataset API
+    ##   (using the Curious Recurring Template Pattern in underlying C++)
+    ## Batch is by default the type CppVector[T]
+    ##   with T being Example[Data, Target]
+    ## BatchRequest is by default ArrayRef[csize_t]
+
+  Dataset*
+    {.bycopy, pure,
+    importcpp: "torch::data::datasets::Dataset".}
+    [Self, Batch]
+      = object of BatchDataset # [Self, Batch, ArrayRef[csize_t]]
+    ## A Dataset type
+    ## Self: is the class type that implements the Dataset API
+    ##   (using the Curious Recurring Template Pattern in underlying C++)
+    ## Batch is by default the type CppVector[T]
+    ##   with T being Example[Data, Target]
+
+  Mnist*
+    {.bycopy, pure,
+    importcpp: "torch::data::datasets::MNIST".}
+    = object of Dataset[Mnist, CppVector[Example[Tensor, Tensor]]]
+    ## The MNIST dataset
+    ## http://yann.lecun.com/exdb/mnist
+
+  MnistMode* {.size:sizeof(cint),
+      importcpp:"torch::data::datasets::MNIST::Mode".} = enum
+    ## Select the train or test mode of the Mnist data
+    kTrain = 0
+    kTest = 1
+
+func mnist*(rootPath: cstring, mode = kTrain): Mnist {.constructor, importcpp:"MNIST(@)".}
+  ## Loads the MNIST dataset from the `root` path
+  ## The supplied `rootpath` should contain the *content* of the unzipped
+  ## MNIST dataset, available from http://yann.lecun.com/exdb/mnist.
+func get*(dataset: Mnist, index: int): Example[Tensor, Tensor] {.importcpp:"#.get(#)".}
+# func size*(dataset: Mnist): optional(int)
+func is_train*(): bool {.importcpp:"#.is_train()".}
+func images*(dataset: Mnist): lent Tensor {.importcpp: "#.images()".}
+  ## Returns all images stacked into a single tensor
+func targets*(dataset: Mnist): lent Tensor {.importcpp: "#.targets()".}
+
+# #######################################################################
+#
+#                         Dataloader
+#
+# #######################################################################
+
+# #######################################################################
+#
+#                         Samplers
+#
+# #######################################################################
+
+# #######################################################################
+#
+#                         Samplers
+#
+# #######################################################################

flambeau/raw_bindings/neural_nets.nim

@@ -30,6 +30,22 @@ import ./tensors
 # It is suitable for layers with no learning parameters (for example reshaping),
 # or when extra flexibility is required at a small price of ergonomics.
 # The high-level Module API uses Functional internally.
+#
+# Note:
+#   Function exists both in ATen TensorBody.h (namespace at:: or torch::)
+#   and in torch::nn::functional.
+#
+#   We can have
+#     func dropout*(input: Tensor, p = 0.5, training=true): Tensor {.importcpp: "torch::nn::functional::dropout(@)".}
+#     func dropout_mut*(input: var Tensor, p = 0.5, training=true) {.importcpp: "torch::nn::functional::dropout(@, /*inplace=*/ true)".}
+#
+#     OR
+#
+#     func dropout*(input: Tensor, p = 0.5, training=true): Tensor {.importcpp: "torch::dropout(@)".}
+#     func dropout_mut*(input: var Tensor, p = 0.5, training=true) {.importcpp: "torch::dropout_(@)".}
+#
+#   The functions in torch::nn::functional are thin inlined wrapper over TensorBody.h
+#   so we directly use them.
 
 # Linear Layers
 # -------------------------------------------------------------------------
@@ -53,15 +69,23 @@ func linear*(input, weight, bias: Tensor): Tensor {.importcpp: "torch::nn::funct
   ## Bias: (out_features)
   ## Output: (N,∗,out_features)
 
-# Dropout functions
+# Activation functions
 # -------------------------------------------------------------------------
 
-# func dropout*(input: Tensor, p = 0.5, training=true): Tensor {.importcpp: "torch::nn::functional::dropout(@)".}
-# func dropout_mut*(input: var Tensor, p = 0.5, training=true) {.importcpp: "torch::nn::functional::dropout(@, /*inplace=*/ true)".}
+func relu*(input: Tensor): Tensor {.importcpp: "torch::relu(@)".}
+func relu_mut*(input: var Tensor) {.importcpp: "torch::relu_(@)".}
+
+# Dropout functions
+# -------------------------------------------------------------------------
 
 func dropout*(input: Tensor, p = 0.5, training=true): Tensor {.importcpp: "torch::dropout(@)".}
 func dropout_mut*(input: var Tensor, p = 0.5, training=true) {.importcpp: "torch::dropout_(@)".}
 
+# Loss functions
+# -------------------------------------------------------------------------
+
+func log_softmax*(input: Tensor, axis: int64): Tensor {.importcpp: "torch::log_softmax(@)".}
+func log_softmax*(input: Tensor, axis: int64, dtype: ScalarKind): Tensor {.importcpp: "torch::log_softmax(@)".}
 
 # #######################################################################
 #

proof_of_concepts/poc07_datasets.nim

@@ -0,0 +1,13 @@
+import
+  ../flambeau/raw_bindings/[
+    data_api, tensors
+  ]
+
+let mnist = mnist("build/mnist")
+
+echo "Data"
+# mnist.get(0).data.print()
+# echo "\n-----------------------"
+# echo "Target"
+# mnist.get(0).target.print()
+# echo "\n-----------------------"