Generic reshape (#213)

* Generic reshape

* Update contributors

* Update features list

Author: milancurcic

CMakeLists.txt

@@ -64,10 +64,10 @@ add_library(neural-fortran
   src/nf/nf_parallel.f90
   src/nf/nf_parallel_submodule.f90
   src/nf/nf_random.f90
-  src/nf/nf_reshape_layer.f90
-  src/nf/nf_reshape_layer_submodule.f90
   src/nf/nf_reshape2d_layer.f90
   src/nf/nf_reshape2d_layer_submodule.f90
+  src/nf/nf_reshape3d_layer.f90
+  src/nf/nf_reshape3d_layer_submodule.f90
   src/nf/nf_self_attention_layer.f90
   src/nf/io/nf_io_binary.f90
   src/nf/io/nf_io_binary_submodule.f90

README.md

@@ -15,8 +15,8 @@ Read the paper [here](https://arxiv.org/abs/1902.06714).
 
 ## Features
 
-* Training and inference of dense (fully connected) and convolutional neural
-  networks
+* Training and inference of dense (fully connected), convolutional (1-d and 2-d),
+  and transformer neural networks
 * Stochastic gradient descent optimizers: Classic, momentum, Nesterov momentum,
   RMSProp, Adagrad, Adam, AdamW
 * More than a dozen activation functions and their derivatives
@@ -41,9 +41,8 @@ Read the paper [here](https://arxiv.org/abs/1902.06714).
 | Linear (2-d) | `linear2d` | `input2d`, `layernorm`, `linear2d`, `self_attention` | 2 | ✅ | ✅ |
 | Self-attention | `self_attention` | `input2d`, `layernorm`, `linear2d`, `self_attention` | 2 | ✅ | ✅ |
 | Layer Normalization | `layernorm` | `linear2d`, `self_attention` | 2 | ✅ | ✅ |
-| Flatten | `flatten` | `input2d`, `input3d`, `conv2d`, `maxpool2d`, `reshape` | 1 | ✅ | ✅ |
-| Reshape (1-d to 2-d) | `reshape2d` | `input2d`, `conv1d`, `locally_connected1d`, `maxpool1d` | 2 | ✅ | ✅ |
-| Reshape (1-d to 3-d) | `reshape` | `input1d`, `dense`, `flatten` | 3 | ✅ | ✅ |
+| Flatten | `flatten` | `input2d`, `input3d`, `conv1d`, `conv2d`, `maxpool1d`, `maxpool2d`, `reshape` | 1 | ✅ | ✅ |
+| Reshape (1-d to 2-d or 3-d) | `reshape` | `dense`, `dropout`, `flatten`, `input1d` | 2, 3 | ✅ | ✅ |
 
 ## Getting started
 
@@ -263,11 +262,13 @@ It may be useful to read if you want to contribute a new feature to neural-fortr
 
 Thanks to all open-source contributors to neural-fortran:
 [awvwgk](https://github.com/awvwgk),
+[certik](https://github.com/certik),
 [ggoyman](https://github.com/ggoyman),
 [ivan-pi](https://github.com/ivan-pi),
 [jacobwilliams](https://github.com/jacobwilliams),
 [jvdp1](https://github.com/jvdp1),
 [jvo203](https://github.com/jvo203),
+[mathomp4](https://github.com/mathomp4),
 [milancurcic](https://github.com/milancurcic),
 [OneAdder](https://github.com/OneAdder),
 [pirpyn](https://github.com/pirpyn),

example/cnn_mnist.f90

@@ -20,7 +20,7 @@ program cnn_mnist
 
   net = network([ &
     input(784), &
-    reshape([1,28,28]), &
+    reshape(1, 28, 28), &
     conv2d(filters=8, kernel_size=3, activation=relu()), &
     maxpool2d(pool_size=2), &
     conv2d(filters=16, kernel_size=3, activation=relu()), &

example/cnn_mnist_1d.f90

@@ -1,7 +1,7 @@
 program cnn_mnist_1d
 
     use nf, only: network, sgd, &
-      input, conv1d, maxpool1d, flatten, dense, reshape, reshape2d, locally_connected1d, &
+      input, conv1d, maxpool1d, flatten, dense, reshape, locally_connected1d, &
       load_mnist, label_digits, softmax, relu
 
     implicit none
@@ -20,7 +20,7 @@ program cnn_mnist_1d
 
     net = network([ &
       input(784), &
-      reshape2d([28, 28]), &
+      reshape(28, 28), &
       locally_connected1d(filters=8, kernel_size=3, activation=relu()), &
       maxpool1d(pool_size=2), &
       locally_connected1d(filters=16, kernel_size=3, activation=relu()), &

fpm.toml

@@ -1,5 +1,5 @@
 name = "neural-fortran"
-version = "0.20.0"
+version = "0.21.0"
 license = "MIT"
 author = "Milan Curcic"
 maintainer = "[email protected]"

src/nf.f90

@@ -16,7 +16,6 @@ module nf
     maxpool1d, &
     maxpool2d, &
     reshape, &
-    reshape2d, &
     self_attention
   use nf_loss, only: mse, quadratic
   use nf_metrics, only: corr, maxabs

src/nf/nf_layer_constructors.f90

@@ -20,7 +20,6 @@ module nf_layer_constructors
     maxpool1d, &
     maxpool2d, &
     reshape, &
-    reshape2d, &
     self_attention, &
     embedding, &
     layernorm
@@ -94,6 +93,28 @@ end function input3d
 
   end interface input
 
+
+  interface reshape
+
+    module function reshape2d(dim1, dim2) result(res)
+      !! Rank-1 to rank-2 reshape layer constructor.
+      integer, intent(in) :: dim1, dim2
+        !! Shape of the output
+      type(layer) :: res
+        !! Resulting layer instance
+    end function reshape2d
+
+    module function reshape3d(dim1, dim2, dim3) result(res)
+      !! Rank-1 to rank-3 reshape layer constructor.
+      integer, intent(in) :: dim1, dim2, dim3
+        !! Shape of the output
+      type(layer) :: res
+        !! Resulting layer instance
+    end function reshape3d
+
+  end interface reshape
+
+
   interface
 
     module function dense(layer_size, activation) result(res)
@@ -283,28 +304,6 @@ module function maxpool2d(pool_size, stride) result(res)
         !! Resulting layer instance
     end function maxpool2d
 
-    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
-
-    module function reshape2d(output_shape) result(res)
-      !! Rank-1 to rank-any reshape layer constructor.
-      !! Currently implemented is only rank-2 for the output of the reshape.
-      !!
-      !! This layer is for connecting 1-d inputs to conv1d or similar layers.
-      integer, intent(in) :: output_shape(:)
-        !! Shape of the output
-      type(layer) :: res
-        !! Resulting layer instance
-    end function reshape2d
-
     module function linear2d(out_features) result(res)
       !! Rank-2 (sequence_length, out_features) linear layer constructor.
       !! sequence_length is determined at layer initialization, based on the

src/nf/nf_layer_constructors_submodule.f90

@@ -12,8 +12,8 @@
   use nf_locally_connected1d_layer, only: locally_connected1d_layer
   use nf_maxpool1d_layer, only: maxpool1d_layer
   use nf_maxpool2d_layer, only: maxpool2d_layer
-  use nf_reshape_layer, only: reshape3d_layer
   use nf_reshape2d_layer, only: reshape2d_layer
+  use nf_reshape3d_layer, only: reshape3d_layer
   use nf_linear2d_layer, only: linear2d_layer
   use nf_self_attention_layer, only: self_attention_layer
   use nf_embedding_layer, only: embedding_layer
@@ -229,35 +229,22 @@ module function maxpool2d(pool_size, stride) result(res)
   end function maxpool2d
 
 
-  module function reshape(output_shape) result(res)
-    integer, intent(in) :: output_shape(:)
-    type(layer) :: res
-
-    res % name = 'reshape'
-    res % layer_shape = output_shape
-
-    if (size(output_shape) == 3) then
-      allocate(res % p, source=reshape3d_layer(output_shape))
-    else
-      error stop 'size(output_shape) of the reshape layer must == 3'
-    end if
-
-  end function reshape
-
-  module function reshape2d(output_shape) result(res)
-    integer, intent(in) :: output_shape(:)
+  module function reshape2d(dim1, dim2) result(res)
+    integer, intent(in) :: dim1, dim2
     type(layer) :: res
-
     res % name = 'reshape2d'
-    res % layer_shape = output_shape
+    res % layer_shape = [dim1, dim2]
+    allocate(res % p, source=reshape2d_layer(res % layer_shape))
+  end function reshape2d
 
-    if (size(output_shape) == 2) then
-      allocate(res % p, source=reshape2d_layer(output_shape))
-    else
-      error stop 'size(output_shape) of the reshape layer must == 2'
-    end if
 
-  end function reshape2d
+  module function reshape3d(dim1, dim2, dim3) result(res)
+    integer, intent(in) :: dim1, dim2, dim3
+    type(layer) :: res
+    res % name = 'reshape3d'
+    res % layer_shape = [dim1, dim2, dim3]
+    allocate(res % p, source=reshape3d_layer(res % layer_shape))
+  end function reshape3d
 
 
   module function linear2d(out_features) result(res)

src/nf/nf_layer_submodule.f90

@@ -13,7 +13,7 @@
   use nf_maxpool1d_layer, only: maxpool1d_layer
   use nf_maxpool2d_layer, only: maxpool2d_layer
   use nf_reshape2d_layer, only: reshape2d_layer
-  use nf_reshape_layer, only: reshape3d_layer
+  use nf_reshape3d_layer, only: reshape3d_layer
   use nf_linear2d_layer, only: linear2d_layer
   use nf_self_attention_layer, only: self_attention_layer
   use nf_embedding_layer, only: embedding_layer

src/nf/nf_network_submodule.f90

@@ -12,13 +12,13 @@
   use nf_maxpool1d_layer, only: maxpool1d_layer
   use nf_maxpool2d_layer, only: maxpool2d_layer
   use nf_reshape2d_layer, only: reshape2d_layer
-  use nf_reshape_layer, only: reshape3d_layer
+  use nf_reshape3d_layer, only: reshape3d_layer
   use nf_linear2d_layer, only: linear2d_layer
   use nf_self_attention_layer, only: self_attention_layer
   use nf_embedding_layer, only: embedding_layer
   use nf_layernorm_layer, only: layernorm_layer
   use nf_layer, only: layer
-  use nf_layer_constructors, only: conv1d, conv2d, dense, flatten, input, maxpool1d, maxpool2d, reshape, reshape2d
+  use nf_layer_constructors, only: conv1d, conv2d, dense, flatten, input, maxpool1d, maxpool2d, reshape
   use nf_loss, only: quadratic
   use nf_optimizers, only: optimizer_base_type, sgd
   use nf_parallel, only: tile_indices

src/nf/nf_reshape_layer.f90 renamed to src/nf/nf_reshape3d_layer.f90

@@ -1,4 +1,4 @@
-module nf_reshape_layer
+module nf_reshape3d_layer
 
   !! This module provides the concrete reshape layer type.
   !! It is used internally by the layer type.
@@ -73,4 +73,4 @@ end subroutine init
 
   end interface
 
-end module nf_reshape_layer
+end module nf_reshape3d_layer

src/nf/nf_reshape_layer_submodule.f90 renamed to src/nf/nf_reshape3d_layer_submodule.f90

@@ -1,4 +1,4 @@
-submodule(nf_reshape_layer) nf_reshape_layer_submodule
+submodule(nf_reshape3d_layer) nf_reshape3d_layer_submodule
 
   use nf_base_layer, only: base_layer
 
@@ -48,4 +48,4 @@ module subroutine init(self, input_shape)
 
   end subroutine init
 
-end submodule nf_reshape_layer_submodule
+end submodule nf_reshape3d_layer_submodule

test/test_insert_flatten.f90

@@ -45,7 +45,7 @@ program test_insert_flatten
 
   net = network([ &
     input(4), &
-    reshape([1, 2, 2]), &
+    reshape(1, 2, 2), &
     dense(4) &
   ])
 

test/test_reshape2d_layer.f90

@@ -1,23 +1,21 @@
 program test_reshape2d_layer
 
   use iso_fortran_env, only: stderr => error_unit
-  use nf, only: input, network, reshape2d_layer => reshape2d
-  use nf_datasets, only: download_and_unpack, keras_reshape_url
+  use nf, only: input, network, reshape2d => reshape
 
   implicit none
 
   type(network) :: net
   real, allocatable :: sample_input(:), output(:,:)
   integer, parameter :: output_shape(2) = [4,4]
   integer, parameter :: input_size = product(output_shape)
-  character(*), parameter :: keras_reshape_path = 'keras_reshape.h5'
   logical :: file_exists
   logical :: ok = .true.
 
   ! Create the network
   net = network([ & 
     input(input_size), & 
-    reshape2d_layer(output_shape) & 
+    reshape2d(output_shape(1), output_shape(2)) & 
   ])
 
   if (.not. size(net % layers) == 2) then

test/test_reshape_layer.f90

@@ -1,7 +1,7 @@
 program test_reshape_layer
 
   use iso_fortran_env, only: stderr => error_unit
-  use nf, only: input, network, reshape_layer => reshape
+  use nf, only: input, network, reshape3d => reshape
   use nf_datasets, only: download_and_unpack, keras_reshape_url
 
   implicit none
@@ -17,7 +17,7 @@ program test_reshape_layer
   ! Create the network
   net = network([ &
     input(input_size), &
-    reshape_layer(output_shape) &
+    reshape3d(3, 32, 32) &
   ])
 
   if (.not. size(net % layers) == 2) then