conv1d

Project.toml

@@ -19,6 +19,7 @@ Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MLDatasets = "eb30cadb-4394-5ae3-aed4-317e484a6458"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
+Optim = "429524aa-4258-5aef-a3af-852621145aeb"
 Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
 ProtoBuf = "3349acd9-ac6a-5e09-bcdb-63829b23a429"

REQUIRE

@@ -13,3 +13,4 @@ MacroTools 0.3.6
 AutoHashEquals 0.1.0
 MLDatasets 0.3.0
 SpecialFunctions 0.7.0
+Optim 0.17.0

deps/default_imports.txt

@@ -166,3 +166,6 @@ Rank
 Conv2DBackpropInput
 Svd
 Cross
+FFT
+ComplexAbs
+MatrixSolve

src/TensorFlow.jl

@@ -135,6 +135,7 @@ with_tape
 
 
 using Distributed
+using Optim
 
 # Load these packages here so they are available to the additional
 # process spawned in 'load_python_process. Arslan thinks that will

src/ops/nn.jl

@@ -36,6 +36,25 @@ import .rnn_cell:  zero_state, output_size, state_size
     conv2d(input, filter; padding=padding, strides=strides, kwargs...)
 end
 
+@tf.op function conv1d(input, filter_, strides_::Int64, padding::String; data_format="NHWC", kwargs...)
+    spatial_start_dim = 0
+    if data_format=="NHWC"
+        strides_ = [1,1,strides_,1]
+        spatial_start_dim = 2
+    elseif data_format == "NCHW" || data_format == "NCW"
+        data_format = "NCHW" 
+        spatial_start_dim = 3
+        strides_ = [1,1,1,strides_]
+    else
+        @error "data_format must be NHWC or NCHW or NCW"
+    end
+    input = Ops.expand_dims(input, spatial_start_dim)
+    filter_ = Ops.expand_dims(filter_, 1)
+    result = Ops.conv2d(input, filter_; strides = strides_, padding = padding, data_format=data_format, kwargs...)
+    result = Ops.squeeze(result, squeeze_dims=[spatial_start_dim-1])
+    return result
+end
+
 # Same for max pool
 @tf.op function max_pool(input, ksize, strides, padding; kwargs...)
     max_pool(input; ksize=ksize, strides=strides, padding=padding, kwargs...)

test/nn.jl

@@ -4,6 +4,26 @@ using StatsFuns
 using Random
 import LinearAlgebra
 
+@testset "conv1d" begin
+    let
+        sess = Session(Graph())
+        F = zeros(Float32, 2, 3, 4) #  batch_size = 2, dimension = 3, channle = 4
+        for i = 1:2
+            for j = 1:3
+                for k = 1:4
+                    F[i,j,k] = Float32(i+j+k-3)
+                end
+            end
+        end
+        input = constant(F)
+        filter_ = constant(ones(Float32, 3, 4, 1)) # width = 3, input channel = 4 output channel = 1
+        output = nn.conv1d(input, filter_, 2, "VALID")
+        output_val = run(sess, output)
+        ref_val = reshape(Float32[30.0;42.0], 2, 1, 1)
+        @test ref_val ≈ output_val
+    end
+end
+
 @testset "conv2d_transpose" begin
     let
         sess = Session(Graph())