Implement the reshape layer (#97)

* Interface and constructor for the reshape3d layer

* Use reshape for the constructor function and reshape3d for the internal layer implementation

* Add the submodule for the concrete reshape3d_layer

* Forward and backward passes for the

* Add type guards for reshape layer to forward and backward subroutines

* Test that the resulting shape and values of a reshape layer are correct

* Bump version to 0.8.0 (unreleased)

* Add reshape layer to list of features

* Update CMake build for the reshape layer

* Ignore submodule files

* Enable reading reshape layers from Keras h5

Author: milancurcic

.gitignore

@@ -1,7 +1,8 @@
 *.gz
 *.o
 *.mod
+*.smod
 *.dat
 *.h5
-build
-doc
+/build
+/doc

CMakeLists.txt

@@ -58,6 +58,8 @@ add_library(neural
   src/nf/nf_parallel_submodule.f90
   src/nf/nf_random.f90
   src/nf/nf_random_submodule.f90
+  src/nf/nf_reshape_layer.f90
+  src/nf/nf_reshape_layer_submodule.f90
   src/nf/io/nf_io_binary.f90
   src/nf/io/nf_io_binary_submodule.f90
   src/nf/io/nf_io_hdf5.f90

README.md

@@ -18,7 +18,7 @@ Read the paper [here](https://arxiv.org/abs/1902.06714).
 
 * Dense, fully connected neural layers
 * Convolutional and max-pooling layers (experimental, forward propagation only)
-* Flatten layers (forward and backward pass)
+* Flatten and reshape layers (forward and backward passes)
 * Loading dense and convolutional models from Keras h5 files
 * Stochastic and mini-batch gradient descent for back-propagation
 * Data-based parallelism
@@ -29,10 +29,11 @@ Read the paper [here](https://arxiv.org/abs/1902.06714).
 | Layer type | Constructor name | Supported input layers | Rank of output array | Forward pass | Backward pass |
 |------------|------------------|------------------------|----------------------|--------------|---------------|
 | Input (1-d and 3-d) | `input` | n/a | 1, 3 | n/a | n/a |
-| Dense (fully-connected) | `dense` | `input` (1-d) | 1 | ✅ | ✅ |
-| Convolutional (2-d) | `conv2d` | `input` (3-d), `conv2d`, `maxpool2d` | 3 | ✅ | ❌ |
-| Max-pooling (2-d) | `maxpool2d` | `input` (3-d), `conv2d`, `maxpool2d` | 3 | ✅ | ❌ |
-| Flatten | `flatten` | `input` (3-d), `conv2d`, `maxpool2d` | 1 | ✅ | ✅ |
+| Dense (fully-connected) | `dense` | `input1d` | 1 | ✅ | ✅ |
+| Convolutional (2-d) | `conv2d` | `input3d`, `conv2d`, `maxpool2d` | 3 | ✅ | ❌ |
+| Max-pooling (2-d) | `maxpool2d` | `input3d`, `conv2d`, `maxpool2d` | 3 | ✅ | ❌ |
+| Flatten | `flatten` | `input3d`, `conv2d`, `maxpool2d` | 1 | ✅ | ✅ |
+| Reshape (1-d to 3-d) | `reshape` | `input1d`, `dense`, `flatten` | 3 | ✅ | ✅ |
 
 ## Getting started
 

fpm.toml

@@ -1,5 +1,5 @@
 name = "neural-fortran"
-version = "0.7.0"
+version = "0.8.0"
 license = "MIT"
 author = "Milan Curcic"
 maintainer = "[email protected]"

src/nf.f90

@@ -2,7 +2,8 @@ module nf
   !! User API: everything an application needs to reference directly
   use nf_datasets_mnist, only: label_digits, load_mnist
   use nf_layer, only: layer
-  use nf_layer_constructors, only: conv2d, dense, flatten, input, maxpool2d
+  use nf_layer_constructors, only: &
+    conv2d, dense, flatten, input, maxpool2d, reshape
   use nf_network, only: network
   use nf_optimizers, only: sgd
 end module nf

src/nf/nf_datasets.f90

@@ -12,6 +12,7 @@ module nf_datasets
     download_and_unpack, &
     keras_cnn_mnist_url, &
     keras_dense_mnist_url, &
+    keras_reshape_url, &
     mnist_url
 
   character(*), parameter :: keras_snippets_baseurl = &
@@ -22,6 +23,8 @@ module nf_datasets
     keras_snippets_baseurl // '/8892585/keras_cnn_mnist.tar.gz'
   character(*), parameter :: keras_dense_mnist_url = &
     keras_snippets_baseurl // '/8788739/keras_dense_mnist.tar.gz'
+  character(*), parameter :: keras_reshape_url = &
+    keras_snippets_baseurl // '/9667603/keras_reshape.tar.gz'
   character(*), parameter :: mnist_url = &
     neural_fortran_baseurl // '/8498876/mnist.tar.gz'
 

src/nf/nf_keras.f90

@@ -29,6 +29,9 @@ module nf_keras
     integer, allocatable :: pool_size(:)
     integer, allocatable :: strides(:)
 
+    ! Reshape
+    integer, allocatable :: target_shape(:)
+
   end type keras_layer
 
   interface

src/nf/nf_keras_submodule.f90

@@ -82,6 +82,13 @@ module function get_keras_h5_layers(filename) result(res)
           res(n) % pool_size = reverse(res(n) % pool_size)
           res(n) % strides = reverse(res(n) % strides)
 
+        case('Reshape')
+          ! Only read target shape
+          call json % get(layer_config_json, &
+            'target_shape',  res(n) % target_shape, found)
+          ! Reverse to account for C -> Fortran order
+          res(n) % target_shape = reverse(res(n) % target_shape)
+
         case default
           error stop 'This Keras layer is not supported'
 

src/nf/nf_layer.f90

@@ -24,24 +24,25 @@ module nf_layer
 
   contains
 
-    procedure :: backward
     procedure :: forward
     procedure :: init
     procedure :: print_info
     procedure :: update
 
-    ! Specific output subroutines for different array ranks,
-    ! available via generic `get_output`.
+    ! Specific subroutines for different array ranks
+    procedure, private :: backward_1d
+    procedure, private :: backward_3d
     procedure, private :: get_output_1d
     procedure, private :: get_output_3d
 
+    generic :: backward => backward_1d, backward_3d
     generic :: get_output => get_output_1d, get_output_3d
 
   end type layer
 
-  interface
+  interface backward
 
-    pure module subroutine backward(self, previous, gradient)
+    pure module subroutine backward_1d(self, previous, gradient)
       !! Apply a backward pass on the layer.
       !! This changes the internal state of the layer.
       !! This is normally called internally by the `network % backward`
@@ -52,7 +53,24 @@ pure module subroutine backward(self, previous, gradient)
         !! Previous layer instance
       real, intent(in) :: gradient(:)
         !! Array of gradient values from the next layer
-    end subroutine backward
+    end subroutine backward_1d
+
+    pure module subroutine backward_3d(self, previous, gradient)
+      !! Apply a backward pass on the layer.
+      !! This changes the internal state of the layer.
+      !! This is normally called internally by the `network % backward`
+      !! method.
+      class(layer), intent(in out) :: self
+        !! Layer instance
+      class(layer), intent(in) :: previous
+        !! Previous layer instance
+      real, intent(in) :: gradient(:,:,:)
+        !! Array of gradient values from the next layer
+    end subroutine backward_3d
+
+  end interface backward
+
+  interface
 
     pure module subroutine forward(self, input)
       !! Apply a forward pass on the layer.

src/nf/nf_layer_constructors.f90

@@ -7,7 +7,7 @@ module nf_layer_constructors
   implicit none
 
   private
-  public :: conv2d, dense, flatten, input, maxpool2d
+  public :: conv2d, dense, flatten, input, maxpool2d, reshape
 
   interface input
 
@@ -154,6 +154,17 @@ pure module function maxpool2d(pool_size, stride) result(res)
         !! Resulting layer instance
     end function maxpool2d
 
+    pure module function reshape(output_shape) result(res)
+      !! Rank-1 to rank-any reshape layer constructor.
+      !! Currently implemented is only rank-3 for the output of the reshape.
+      !!
+      !! This layer is for connecting 1-d inputs to conv2d or similar layers.
+      integer, intent(in) :: output_shape(:)
+        !! Shape of the output
+      type(layer) :: res
+        !! Resulting layer instance
+    end function reshape
+
   end interface
 
 end module nf_layer_constructors