Added valid and fixed padding to conv2d layer (#43)

Added valid and fixed padding to conv2d layer.

Contributors: Adam Hillier

Co-authored-by: Vincent Tjeng <vincent.tjeng@gmail.com>

src/net_components/layers/conv2d.jl

@@ -2,6 +2,15 @@ using JuMP
 using ConditionalJuMP
 
 export Conv2d
+export Padding, SamePadding, ValidPadding
+
+struct SamePadding end
+Base.show(io::IO, p::SamePadding) = print(io, "same")
+struct ValidPadding end
+Base.show(io::IO, p::ValidPadding) = print(io, "valid")
+
+FixedPadding = Union{Int, Tuple{Int, Int}, Tuple{Int, Int, Int, Int}}
+Padding = Union{SamePadding, ValidPadding, FixedPadding}
 
 """
 $(TYPEDEF)
@@ -18,25 +27,30 @@ $(FIELDS)
     filter::Array{T, 4}
     bias::Array{U, 1}
     stride::V
+    padding::Padding
 
-    function Conv2d{T, U, V}(filter::Array{T, 4}, bias::Array{U, 1}, stride::V) where {T<:JuMPReal, U<:JuMPReal, V<:Int64}
+    function Conv2d{T, U, V}(filter::Array{T, 4}, bias::Array{U, 1}, stride::V, padding::Padding) where {T<:JuMPReal, U<:JuMPReal, V<:Int64}
         (filter_height, filter_width, filter_in_channels, filter_out_channels) = size(filter)
         bias_out_channels = length(bias)
         @assert(
             filter_out_channels == bias_out_channels,
             "For this convolution layer, number of output channels in filter, $filter_out_channels, does not match number of output channels in bias, $bias_out_channels."
         )
-        return new(filter, bias, stride)
+        return new(filter, bias, stride, padding)
     end
 
 end
 
+function Conv2d(filter::Array{T, 4}, bias::Array{U, 1}, stride::V, padding::Padding) where {T<:JuMPReal, U<:JuMPReal, V<:Int64}
+    Conv2d{T, U, V}(filter, bias, stride, padding)
+end
+
 function Conv2d(filter::Array{T, 4}, bias::Array{U, 1}, stride::V) where {T<:JuMPReal, U<:JuMPReal, V<:Int64}
-    Conv2d{T, U, V}(filter, bias, stride)
+    Conv2d{T, U, V}(filter, bias, stride, SamePadding())
 end
 
 function Conv2d(filter::Array{T, 4}, bias::Array{U, 1}) where {T<:JuMPReal, U<:JuMPReal}
-    Conv2d(filter, bias, 1)
+    Conv2d(filter, bias, 1, SamePadding())
 end
 
 """
@@ -59,8 +73,9 @@ end
 function Base.show(io::IO, p::Conv2d)
     (filter_height, filter_width, filter_in_channels, filter_out_channels) = size(p.filter)
     stride = p.stride
+    padding = p.padding
     print(io,
-        "Conv2d($filter_in_channels, $filter_out_channels, kernel_size=($(filter_height), $(filter_width)), stride=($(stride), $(stride)), padding=same)"
+        "Conv2d($filter_in_channels, $filter_out_channels, kernel_size=($(filter_height), $(filter_width)), stride=($(stride), $(stride)), padding=$(padding))"
     )
 end
 
@@ -86,15 +101,81 @@ function increment!(s::JuMP.AffExpr, input_val::Real, filter_val::JuMP.Variable)
     push!(s, Float64(input_val), filter_val)
 end
 
+function compute_output_parameters(
+    in_height::Int, in_width::Int,
+    filter_height::Int, filter_width::Int, stride::Int,
+    padding::FixedPadding
+)::Tuple{NTuple{2, Int}, NTuple{2, Int}}
+    (top_padding, bottom_padding, left_padding, right_padding) = compute_padding_values(padding)
+    out_height = round(Int, (in_height + top_padding + bottom_padding - filter_height) / stride, RoundDown) + 1
+    out_width = round(Int, (in_width + left_padding + right_padding - filter_width) / stride, RoundDown) + 1
+
+    output_size = (out_height, out_width)
+    filter_offset = (top_padding, left_padding)
+    return (output_size, filter_offset)
+end
 
+function compute_output_parameters(
+    in_height::Int, in_width::Int,
+    filter_height::Int, filter_width::Int, stride::Int,
+    padding::SamePadding
+)::Tuple{NTuple{2, Int}, NTuple{2, Int}}
+    out_height = round(Int, in_height/stride, RoundUp)
+    out_width = round(Int, in_width/stride, RoundUp)
+    pad_along_height = max((out_height - 1)*stride + filter_height - in_height, 0)
+    pad_along_width = max((out_width - 1)*stride + filter_width - in_width, 0)
+    filter_height_offset = round(Int, pad_along_height/2, RoundDown)
+    filter_width_offset = round(Int, pad_along_width/2, RoundDown)
+
+    output_size = (out_height, out_width)
+    filter_offset = (filter_height_offset, filter_width_offset)
+    return (output_size, filter_offset)
+end
+
+function compute_output_parameters(
+    in_height::Int, in_width::Int,
+    filter_height::Int, filter_width::Int, stride::Int, 
+    padding::ValidPadding
+)::Tuple{NTuple{2, Int}, NTuple{2, Int}}
+    out_height = round(Int, (in_height + 1 - filter_height) / stride, RoundUp)
+    out_width = round(Int, (in_width + 1 - filter_width) / stride, RoundUp)
+    return((out_height, out_width), (0, 0))
+end
+
+function compute_padding_values(
+    padding::Int
+)::NTuple{4, Int}
+    return (padding, padding, padding, padding)
+end
+
+function compute_padding_values(
+    padding::NTuple{2, Int}
+)::NTuple{4, Int}
+    (y_padding, x_padding) = padding
+    return (y_padding, y_padding, x_padding, x_padding)
+end
+
+function compute_padding_values(
+    padding::NTuple{4, Int}
+)::NTuple{4, Int}
+    return padding
+end
 
 """
 $(SIGNATURES)
 
 Computes the result of convolving `input` with the `filter` and `bias` stored in `params`.
 
-Mirrors `tf.nn.conv2d` from the `tensorflow` package, with `strides = [1, 1, 1, 1], 
-padding = 'SAME'`.
+Mirrors `tf.nn.conv2d` from the `tensorflow` package, with 
+`strides = [1, params.stride, params.stride, 1]`.
+
+Supports three types of padding:
+- 'same':  Specify via `SamePadding()`. Padding is added so that the output has the same size as the input.
+- 'valid': Specify via `FixedPadding()`. No padding is added.
+- 'fixed': Specify via:
+  - A single integer, interpreted as padding for both axes
+  - A tuple of two integers, interpreted as (y_padding, x_padding)
+  - A tuple of four integers, interpreted as (top, bottom, left, right)
 
 # Throws
 * AssertionError if `input` and `filter` are not compatible.
@@ -104,34 +185,24 @@ function conv2d(
     params::Conv2d{U, V}) where {T<:JuMPReal, U<:JuMPReal, V<:JuMPReal}
 
     if T<:JuMPLinearType || U<:JuMPLinearType || V<:JuMPLinearType
-        Memento.info(MIPVerify.LOGGER, "Applying $(params) ... ")
+        info(MIPVerify.LOGGER, "Applying $(params) ... ")
     end
     filter = params.filter
     stride = params.stride
+    padding = params.padding
 
     (batch, in_height, in_width, input_in_channels) = size(input)
     (filter_height, filter_width, filter_in_channels, filter_out_channels) = size(filter)
 
     @assert(
-        input_in_channels == filter_in_channels, 
+        input_in_channels == filter_in_channels,
         "Number of channels in input, $input_in_channels, does not match number of channels, $filter_in_channels, that filters operate on."
     )
-
-    out_height = round(Int, in_height/stride, RoundUp)
-    out_width = round(Int, in_width/stride, RoundUp)
-    output_size = (batch, out_height, out_width, filter_out_channels)
 
     # Considered using offset arrays here, but could not get it working.
+    ((out_height, out_width), (filter_height_offset, filter_width_offset)) = compute_output_parameters(in_height, in_width, filter_height, filter_width, stride, padding)
+    output_size = (batch, out_height, out_width, filter_out_channels)
 
-    # Calculating appropriate offsets so that center of kernel is matched with
-    # cell at which correlation is being calculated. Note that tensorflow
-    # chooses a specific convention for a dimension with even size which we
-    # replicate here.
-    pad_along_height = max((out_height - 1)*stride + filter_height - in_height, 0)
-    pad_along_width = max((out_width - 1)*stride + filter_width - in_width, 0)
-    filter_height_offset = round(Int, pad_along_height/2, RoundDown)
-    filter_width_offset = round(Int, pad_along_width/2, RoundDown)
-
     W = Base.promote_op(+, V, Base.promote_op(*, T, U))
     output = Array{W}(undef, output_size)
 

src/utils/import_weights.jl

@@ -3,7 +3,7 @@ export get_matrix_params, get_conv_params, get_example_network_params
 """
 $(SIGNATURES)
 
-Helper function to import the parameters for a layer carrying out matrix multiplication 
+Helper function to import the parameters for a layer carrying out matrix multiplication
     (e.g. fully connected layer / softmax layer) from `param_dict` as a
     [`Linear`](@ref) object.
 
@@ -62,13 +62,15 @@ function get_conv_params(
     expected_size::NTuple{4, Int};
     matrix_name::String = "weight",
     bias_name::String = "bias",
-    expected_stride::Integer = 1
+    expected_stride::Integer = 1,
+    padding::Padding = SamePadding()
     )::Conv2d
 
     params = Conv2d(
         param_dict["$layer_name/$matrix_name"],
         dropdims(param_dict["$layer_name/$bias_name"], dims=1),
-        expected_stride
+        expected_stride,
+        padding
     )
 
     check_size(params, expected_size)