Fix logger interface for histograms

docs/src/custom_behaviour.md

@@ -5,7 +5,7 @@ value is sent to:
 
   - `::AbstractVector{<:Real}` -> [Histogram backend](https://www.tensorflow.org/guide/tensorboard_histograms) as a vector
   - `::StatsBase.Histogram` -> [Histogram backend](https://www.tensorflow.org/guide/tensorboard_histograms)
-  - `(bin_edges, weights)::Tuple{AbstractVector,AbstractVector}` where `length(bin_edges)==length(weights)+1`, is interpreted as an histogram. (*Will be deprecated.* Please use `TBHistogram(edges, weights)` for this).
+  <!-- - `(bin_edges, weights)::Tuple{AbstractVector,AbstractVector}` where `length(bin_edges)==length(weights)+1`, is interpreted as an histogram. (*Will be deprecated.* Please use `TBHistogram(edges, weights)` for this). -->
   - `::Real` -> Scalar backend
   - `::AbstractArray{<:Colorant}` -> [Image backend](https://www.tensorflow.org/tensorboard/r2/image_summaries)
   - `::Any` -> Text Backend

docs/src/extending_behaviour.md

@@ -12,7 +12,7 @@ At the end of this step, every pair in `objects` will be logged to a specific
 backend, according to the following rules:
 
   - `::AbstractVector{<:Real}` -> [Histogram backend](https://www.tensorflow.org/guide/tensorboard_histograms) as a vector
-  - `::Tuple{AbstractVector,AbstractVector}` [Histogram backend](https://www.tensorflow.org/guide/tensorboard_histograms) as an histogram
+  <!-- - `::Tuple{AbstractVector,AbstractVector}` [Histogram backend](https://www.tensorflow.org/guide/tensorboard_histograms) as an histogram -->
   - `::Real` -> Scalar backend
   - `::AbstractArray{<:Colorant}` -> [Image backend](https://www.tensorflow.org/tensorboard/r2/image_summaries)
   - `::Any` -> Text Backend

docs/src/index.md

@@ -111,7 +111,7 @@ at [Reading back TensorBoard data](@ref)
 We also support logging custom types from a the following third-party libraries:
  - [Plots.jl](https://github.com/JuliaPlots/Plots.jl): the `Plots.Plot` type will be rendered to PNG at the resolution specified by the object and logged as an image
  - [PyPlot.jl](https://github.com/JuliaPy/PyPlot.jl): the `PyPlot.Figure` type will be rendered to PNG at the resolution specified by the object and logged as an image
- - [Gadfly.jl](https://github.com/GiovineItalia/Gadfly.jl) type will be rendered to PNG at the resolution specified by the object and logged as an image. `Cairo` and `Fontconfig` packages must be imported for this functionality to work as it is required by `Gadfly`.
+ - [Gadfly.jl](https://github.com/GiovineItalia/Gadfly.jl): the `Gadfly.Plot` type will be rendered to PNG at the resolution specified by the object and logged as an image. `Cairo` and `Fontconfig` packages must be imported for this functionality to work as it is required by `Gadfly`.
  - [Tracker.jl](https://github.com/FluxML/Tracker.jl): the `TrackedReal` and `TrackedArray` types will be logged as vector data
  - [ValueHistories.jl](https://github.com/JuliaML/ValueHistories.jl): the `MVHistory` type is used to store the deserialized content of .proto files.
 

examples/Histograms.jl

@@ -10,9 +10,10 @@ with_logger(logger) do
         x0 = 0.5+i/30; s0 = 0.5/(i/20);
         edges = collect(-5:0.1:5)
         centers = collect(edges[1:end-1] .+0.05)
-        histvals = [exp(-((c-x0)/s0)^2) for c = centers]
+        histvals = s0 * randn(length(centers)) .+ x0
         data_tuple = (edges, histvals)
-        @info "histogram/loggerinterface" autobin=rand(10).+0.1*i manualbin=data_tuple
+        @info "histogram/loggerinterface" autobin=s0 .* randn(100) .+ x0
+        @info "histogram/loggerinterface" manualbin=data_tuple
     end
 end
 
@@ -21,8 +22,8 @@ for i in 1:100
     x0 = 0.5+i/30; s0 = 0.5/(i/20);
     edges = collect(-5:0.1:5)
     centers = collect(edges[1:end-1] .+0.05)
-    histvals = [exp(-((c-x0)/s0)^2) for c = centers]
+    histvals = s0 * randn(length(centers)) .+ x0
     data_tuple = (edges, histvals)
-    log_histogram(logger, "histogram/explicitinterface/autobin", rand(10).+0.1*i, step = i) #automatic bins
+    log_histogram(logger, "histogram/explicitinterface/autobin", s0 .* randn(100) .+ x0, step = i) #automatic bins
     log_histogram(logger, "histogram/explicitinterface/manualbin", data_tuple, step = i) #manual bins
 end

src/Deserialization/histograms.jl

@@ -1,16 +1,4 @@
 function deserialize_histogram_summary(summary::Summary_Value)
-    # custom deserialization
-    if hasproperty(summary, :metadata)
-        histo = summary.value.value
-        if summary.metadata.plugin_data.plugin_name == TB_PLUGIN_JLARRAY_NAME
-            val = reshape(histo.bucket,
-                          reinterpret(Int,
-                                      summary.metadata.plugin_data.content)...)
-
-            return val
-        end
-    end
-
     # deserialize histogramproto
     hist_proto = summary.value.value
     bin_edges = similar(hist_proto.bucket_limit, length(hist_proto.bucket_limit)+1)

src/Loggers/LogHistograms.jl

@@ -12,7 +12,8 @@ be used to bin the data.
 function log_histogram(logger::TBLogger, name::AbstractString, (bins,weights)::Tuple{AbstractVector, AbstractArray};
                        step=nothing)
     weights = collect(vec(weights))
-    summ = SummaryCollection(histogram_summary(name, collect(bins), weights))
+    hist = fit(Histogram, weights, bins)
+    summ = SummaryCollection(histogram_summary(name, hist))
     write_event(logger.file, make_event(logger, summ, step=step))
 end
 
@@ -24,27 +25,22 @@ Bins the values found in `data` and logs them as an histogram under the tag
 """
 function log_histogram(logger::TBLogger, name::AbstractString, data::AbstractArray;
                        step=nothing)
-    data = vec(data)
-    hvals = fit(Histogram, data)
-    summ = SummaryCollection(histogram_summary(name, collect(hvals.edges[1]), hvals.weights))
+    data = collect(vec(data))
+    hist = fit(Histogram, data)
+    summ = SummaryCollection(histogram_summary(name, hist))
     write_event(logger.file, make_event(logger, summ, step=step))
 end
 
 """
     log_vector(logger, name, data::Vector; step=step(logger))
-
 Logs the vector found in `data` as an histogram under the name `name`.
 """
 function log_vector(logger::TBLogger, name::AbstractString, data::AbstractVector; step=nothing)
-    summ = SummaryCollection(histogram_summary(name, collect(0:length(data)),data))
+    hist = Histogram(collect(0:length(data)), data)
+    summ = SummaryCollection(histogram_summary(name, hist))
     write_event(logger.file, make_event(logger, summ, step=step))
 end
 
-function histogram_summary(name::AbstractString, edges::AbstractVector{<:Number}, hist_vals::AbstractVector{<:Number})
-    @assert length(edges) == length(hist_vals)+1
-    return histogram_summary(name, Histogram(edges, hist_vals))
-end
-
 function histogram_summary(name::AbstractString, hist::Histogram{T,1}) where T
     edges     = first(hist.edges)
     hist_vals = hist.weights
@@ -59,25 +55,3 @@ function histogram_summary(name::AbstractString, hist::Histogram{T,1}) where T
                         hist_vals)
     return Summary_Value(name, name, nothing, OneOf(:histo, hp))
 end
-
-# Writes to an Histogram summary the flattened version of the array.
-# Also stores the shape of the array as a field in a plugin, which allows to
-# reconstruct the original shape when read back into Julia
-function histogram_arr_summary(name::AbstractString, tensor::AbstractArray)
-
-    smpd = SummaryMetadata_PluginData(TB_PLUGIN_JLARRAY_NAME, reinterpret(UInt8, collect(size(tensor))))
-    sm = SummaryMetadata(smpd, name, "", DataClass.DATA_CLASS_TENSOR)
-
-    num = length(tensor)
-    edges = collect(0:num)
-    histsum = sum(tensor)
-    histsumsqr = sum(tensor.^2) 
-    hp = HistogramProto(minimum(edges), maximum(edges),
-                        num,
-                        histsum,
-                        histsumsqr,
-                        edges[2:end],
-                        vec(tensor))
-
-    return Summary_Value(name, name, sm, OneOf(:histo, hp))
-end

src/logger_dispatch.jl

@@ -79,18 +79,7 @@ summary_impl(name, value::Any) = text_summary(name, value)
 preprocess(name, hist::Histogram{<:Any,1}, data) = push!(data, name=>hist)
 summary_impl(name, hist::Histogram) = histogram_summary(name, hist)
 
-# TODO: maybe deprecate? tuple means histogram (only if bins/weights match)
-function preprocess(name,   (bins,weights)::Tuple{AbstractVector,AbstractVector}, data)
-    # if ... this is an histogram
-    if length(bins) == length(weights)+1
-        return preprocess(name, Histogram(bins,weights), data)
-    end
-    preprocess(name*"/1", bins, data)
-    preprocess(name*"/2", weights, data)
-end
-
-preprocess(name, val::AbstractArray{<:Real}, data) = push!(data, name=>val)
-summary_impl(name, val::AbstractArray{<:Real}) = histogram_arr_summary(name, val)
+preprocess(name, val::AbstractArray{<:Real}, data) = return preprocess(name, fit(Histogram, collect(vec(val))), data)
 
 # Split complex numbers into real/complex pairs
 preprocess(name, val::AbstractArray{<:Complex}, data) = push!(data, name*"/re"=>real.(val), name*"/im"=>imag.(val))

test/runtests.jl

@@ -5,6 +5,7 @@ using TestImages
 using ImageCore
 using FileIO
 using LightGraphs
+using StatsBase: fit, Histogram
 
 ENV["DATADEPS_ALWAYS_ACCEPT"] = true
 ENV["GKSwstype"] = "100"
@@ -70,15 +71,15 @@ end
         centers = collect(edges[1:end-1] .+0.05)
         histvals = [exp(-((c-x0)/s0)^2) for c=centers]
         data_tuple = (edges, histvals)
-
-        ss = TensorBoardLogger.histogram_summary("test", edges, histvals)
+        hist = fit(Histogram, histvals, edges)
+        ss = TensorBoardLogger.histogram_summary("test", hist)
         @test isa(ss, TensorBoardLogger.Summary_Value)
         @test ss.tag == "test"
         @test isa(ss.value.value, TensorBoardLogger.HistogramProto)
         @test ss.value.value.min == minimum(edges)
         @test ss.value.value.max == maximum(edges)
         @test all(ss.value.value.bucket_limit .== edges[2:end])
-        @test all(ss.value.value.bucket .== histvals)
+        @test ss.value.value.bucket == hist.weights
 
         log_histogram(logger, "hist/cust", data_tuple, step=step)
         log_histogram(logger, "hist/cust", rand(100), step=step)
@@ -93,7 +94,7 @@ end
         vals = rand(10)
         @test data == preprocess("test1", vals, data)
         @test first(data[1]) == "test1"
-        @test last(data[1])  == vals
+        @test last(data[1])  == fit(Histogram, collect(vec(vals)))
 
         vals = rand(ComplexF32, 10)
         preprocess("test2", vals, data)
@@ -104,7 +105,7 @@ end
 
         vals = rand(10, 10)
         preprocess("test2", vals, data)
-        @test last(data[4]) == vals
+        @test last(data[4]) == fit(Histogram, collect(vec(vals)))
 
     end