Merge remote-tracking branch 'origin/master' into fix_histogram_logge…

…r_interface

.gitignore

@@ -6,4 +6,4 @@ test/test_logs
 docs/Manifest.toml
 
 gen/proto
-gen/protojl
+gen/protojl

Project.toml

@@ -1,7 +1,7 @@
 name = "TensorBoardLogger"
 uuid = "899adc3e-224a-11e9-021f-63837185c80f"
 authors = ["Filippo Vicentini <filippovicentini@gmail.com>"]
-version = "0.1.20"
+version = "0.1.22"
 
 [deps]
 CRC32c = "8bf52ea8-c179-5cab-976a-9e18b702a9bc"
@@ -20,11 +20,14 @@ StatsBase = "0.27, 0.28, 0.29, 0.30, 0.31, 0.32, 0.33, 0.34"
 julia = "1.6"
 
 [extras]
-Minio = "4281f0d9-7ae0-406e-9172-b7277c1efa20"
+Cairo = "159f3aea-2a34-519c-b102-8c37f9878175"
+Fontconfig = "186bb1d3-e1f7-5a2c-a377-96d770f13627"
+Gadfly = "c91e804a-d5a3-530f-b6f0-dfbca275c004"
 ImageMagick = "6218d12a-5da1-5696-b52f-db25d2ecc6d1"
 LightGraphs = "093fc24a-ae57-5d10-9952-331d41423f4d"
 Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"
 MLDatasets = "eb30cadb-4394-5ae3-aed4-317e484a6458"
+Minio = "4281f0d9-7ae0-406e-9172-b7277c1efa20"
 Plots = "91a5bcdd-55d7-5caf-9e0b-520d859cae80"
 PyPlot = "d330b81b-6aea-500a-939a-2ce795aea3ee"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

README.md

@@ -39,8 +39,8 @@ logger in Julia:
 
 You can log to TensorBoard any type. Numeric types will be logged as scalar,
 arrays will be binned into histograms, images and audio will be logged as such,
-and we even support [Plots](https://github.com/JuliaPlots/Plots.jl) and
-[PyPlot](https://github.com/JuliaPlots/Plots.jl) figures!
+and we even support [Plots](https://github.com/JuliaPlots/Plots.jl),
+[PyPlot](https://github.com/JuliaPlots/Plots.jl) and [Gadfly](https://github.com/GiovineItalia/Gadfly.jl) figures!
 
 For details about how types are logged by default, or how to customize this behaviour for your custom types,
 refer to the documentation or the examples folder.
@@ -71,7 +71,7 @@ end
 ```
 
 ## Integration with third party packages
-We also support native logging of the types defined by a few third-party packages, such as `Plots` and `PyPlot` plots.
+We also support native logging of the types defined by a few third-party packages, such as `Plots`, `PyPlot` and `Gadfly` plots.
 If there are other libraries that you think we should include in the list, please open an issue.
 
 ## Roadmap

docs/make.jl

@@ -10,11 +10,13 @@ makedocs(
                 "Backends"     =>  "custom_behaviour.md",
                 "Reading back data" => "deserialization.md",
                 "Extending"    =>  "extending_behaviour.md",
-                "Explicit Interface" => "explicit_interface.md"
+                "Explicit Interface" => "explicit_interface.md",
+                "Hyperparameter logging" => "hyperparameters.md"
             ],
             "Examples"      => Any[
                 "Flux.jl"  => "examples/flux.md"
                 "Optim.jl"  => "examples/optim.md"
+                "Hyperparameter tuning"  => "examples/hyperparameter_tuning.md"
             ]
     ],
     format    = Documenter.HTML(

docs/src/examples/hyperparameter_tuning.md

@@ -0,0 +1,53 @@
+# Hyperparameter tuning
+
+We will start this example by setting up a simple random walk experiment, and seeing the effect of the hyperparameter `bias` on the results.
+
+First, import the packages we will need with:
+```julia
+using TensorBoardLogger, Logging
+using Random
+```
+Next, we will create a function which runs the experiment and logs the results, include the hyperparameters stored in the `config` dictionary.
+```julia
+function run_experiment(id, config)
+    logger = TBLogger("random_walk/run$id", tb_append)
+
+    # Specify all the metrics we want to track in a list
+    metric_names = ["scalar/position"]
+    write_hparams!(logger, config, metric_names)
+
+    epochs = config["epochs"]
+    sigma = config["sigma"]
+    bias = config["bias"]
+    with_logger(logger) do
+        x = 0.0
+        for i in 1:epochs
+            x += sigma * randn() + bias
+            @info "scalar" position = x
+        end
+    end
+    nothing
+end
+```
+Now we can write a script which runs an experiment over a set of parameter values.
+```julia
+id = 0
+for bias in LinRange(-0.1, 0.1, 11)
+    for epochs in [50, 100]
+        config = Dict(
+            "bias"=>bias,
+            "epochs"=>epochs,
+            "sigma"=>0.1
+        )
+        run_experiment(id, config)
+        id += 1
+    end
+end
+```
+
+Below is an example of the dashboard you get when you open Tensorboard with the command:
+```sh
+tensorboard --logdir=random_walk
+```
+
+![tuning plot](tuning.png)

docs/src/hyperparameters.md

@@ -0,0 +1,10 @@
+# Hyperparameter logging
+
+In additition to logging the experiments, you may wish to also visualise the effect of hyperparameters on some plotted metrics. This can be done by logging the hyperparameters via the `write_hparams!` function, which takes a dictionary mapping hyperparameter names to their values (currently limited to `Real`, `Bool` or `String` types), along with the names of any metrics that you want to view the effects of. 
+
+You can see how the HParams dashboard in Tensorboard can be used to tune hyperparameters on the [tensorboard website](https://www.tensorflow.org/tensorboard/hyperparameter_tuning_with_hparams).
+
+## API
+```@docs
+write_hparams!
+```

docs/src/index.md

@@ -111,12 +111,15 @@ 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): the `Gadfly.Plot` 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): 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.
 
 ## Explicit logging
 
-In alternative, you can also log data to TensorBoard through its functional interface,
-by calling the relevant method with a tag string and the data. For information
-on this interface refer to [Explicit interface](@ref)...
+As an alternative, you can also log data to TensorBoard through its functional interface, by calling the relevant method with a tag string and the data. For information on this interface refer to [Explicit interface](@ref).
+
+## Hyperparameter tuning
+
+Many experiments rely on hyperparameters, which can be difficult to tune. Tensorboard allows you to visualise the effect of your hyperparameters on your metrics, giving you an intuition for the correct hyperparameters for your task. For information on this API, see the [Hyperparameter logging](@ref) manual page.
+

examples/Gadfly.jl

@@ -0,0 +1,14 @@
+using TensorBoardLogger #import the TensorBoardLogger package
+using Logging #import Logging package
+using Gadfly, Cairo, Fontconfig
+
+logger = TBLogger("Gadflylogs", tb_append) #create tensorboard logger
+
+################log scalars example: y = x²################
+#using logger interface
+x = rand(100)
+y = rand(100)
+p = plot(x=x, y=y, Geom.point);
+with_logger(logger) do
+    @info "gadfly" plot=p
+end

examples/HParams.jl

@@ -0,0 +1,38 @@
+using TensorBoardLogger #import the TensorBoardLogger package
+using Logging #import Logging package
+using Random # Exports randn
+
+# Run 10 experiments to see a plot
+for j in 1:10
+    logger = TBLogger("random_walks/run$j", tb_append)
+
+    sigma = 0.1
+    epochs = 200
+    bias = (rand()*2 - 1) / 10 # create a random bias
+    use_seed = false
+    # Add in the a dummy loss metric
+    with_logger(logger) do
+        x = 0.0
+        for i in 1:epochs
+            x += sigma * randn() + bias
+            @info "scalar" loss = x
+        end
+    end
+
+    # Hyperparameter is a dictionary of parameter names to their values. This 
+    # supports numerical types, bools and strings. Non-bool numerical types 
+    # are converted to Float64 to be displayed.
+    hparams_config = Dict{String, Any}(
+        "sigma"=>sigma,
+        "epochs"=>epochs,
+        "bias"=>bias,
+        "use_seed"=>use_seed,
+        "method"=>"MC"
+    )
+    # Specify a list of tags that you want to show up in the hyperparameter 
+    # comparison
+    metrics = ["scalar/loss"]
+
+    # Write the hyperparameters and metrics config to the logger.
+    write_hparams!(logger, hparams_config, metrics)
+end

examples/Scalars.jl

@@ -23,3 +23,28 @@ with_logger(logger) do
         @info "scalar/complex" y = z
     end
 end
+
+
+################control step increments with context################
+with_logger(logger) do
+    for epoch in 1:10
+        for i=1:100
+            # increments global_step by default
+            with_TBLogger_hold_step() do
+                # all of these are logged at the same global_step
+                # and the logger global_step is only then increased
+                @info "train1/scalar" val=i
+                @info "train2/scalar" val2=i/2
+                @info "train3/scalar" val3=100-i
+            end
+        end
+        # step increment at end can be disabled for easy train/test sync
+        with_TBLogger_hold_step(;step_at_end=false) do
+            # all of these are logged at the same global_step
+            # and the logger global_step is only then increased
+            @info "test1/scalar" epoch=epoch
+            @info "test2/scalar" epoch2=epoch^2
+            @info "test3/scalar" epoch3=epoch^3
+        end
+    end
+end

gen/Project.toml

@@ -4,5 +4,5 @@ FilePathsBase = "48062228-2e41-5def-b9a4-89aafe57970f"
 Glob = "c27321d9-0574-5035-807b-f59d2c89b15c"
 ProtoBuf = "3349acd9-ac6a-5e09-bcdb-63829b23a429"
 
-[comapt]
+[compat]
 ProtoBuf = "0.9.1"

src/Optional/Gadfly.jl

@@ -0,0 +1,15 @@
+import .Gadfly: Plot, render, draw
+function Base.convert(t::Type{PngImage}, plot::Gadfly.Plot)
+    pb = PipeBuffer();
+    show(pb, MIME("image/png"), render(plot));
+    # draw(Gadfly.PNG(pb), plot); # leaving here for now, does same thing
+    return PngImage(pb)
+end
+
+preprocess(name, plot::Gadfly.Plot, data) = preprocess(name, convert(PngImage, plot), data)
+preprocess(name, plots::AbstractArray{<:Gadfly.Plot}, data) = begin
+    for (i, plot)=enumerate(plots)
+        preprocess(name*"/$i", plot, data)
+    end
+    return data
+end

src/TBLogger.jl

@@ -38,7 +38,8 @@ export tb_append, tb_overwrite, tb_increment
 Creates a TensorBoardLogger in the folder `logdir`. The second (optional)
 argument specifies the behaviour if the `logdir` already exhists: the default
 choice `tb_increment` appends an increasing number 1,2... to `logdir`. Other
-choices are `tb_overwrite`, which overwrites the previous folder, and `tb_append`.
+choices are `tb_overwrite`, which overwrites the previous folder, and `tb_append`,
+which adds to any existing logs.
 
 Optional keyword argument `prefix` can be passed to prepend a path to the file
 name (note, not the log directory). See `create_eventfile()`
@@ -298,3 +299,53 @@ Base.show(io::IO, mime::MIME"text/plain", tbl::TBLogger) = begin
     """
     Base.print(io, str)
 end
+
+"""
+`with_TBLogger_hold_step(f, [step]; step_at_end::Bool=true)`
+Context function to ease control of logging steps and synchronization.
+Amount of step increment can be controlled via `set_step_increment!``.
+
+Example:
+```julia
+with_logger(lg) do
+    for epoch in 1:10
+        for i=1:100
+            # increments global_step by default
+            with_TBLogger_hold_step() do
+                # all of these are logged at the same global_step
+                # and the logger global_step is only then increased
+                @info "train1/scalar" i=i
+                @info "train2/scalar" i2=i^2
+                @info "train3/scalar" i3=i^3
+            end
+        end
+        # step increment at end can be disabled for easy train/test sync
+        with_TBLogger_hold_step(;step_at_end=false) do
+            # all of these are logged at the same global_step
+            # and the logger global_step is only then increased
+            @info "test1/scalar" i=i
+            @info "test2/scalar" i2=i^2
+            @info "test3/scalar" i3=i^3
+        end
+    end
+end
+```
+
+"""
+function with_TBLogger_hold_step(f, step::Int; step_at_end::Bool=true)
+    logger = CoreLogging.current_logger()
+    @assert logger isa TBLogger "with_TBLogger_hold_step: current logger is not a TBLogger, cannot establish current step automatically"
+    curr_step = logger.global_step
+    curr_increment = logger.step_increment
+    set_step!(logger, step)
+    set_step_increment!(logger, 0)
+    f()
+    set_step!(logger, curr_step)
+    set_step_increment!(logger, curr_increment)
+    step_at_end && increment_step!(logger, curr_increment)
+end
+function with_TBLogger_hold_step(f; step_at_end::Bool=true)
+    logger = CoreLogging.current_logger()
+    isa(logger, TBLogger) || error("with_TBLogger_hold_step: current logger is not a TBLogger, cannot establish current step automatically")
+    with_TBLogger_hold_step(f, logger.global_step; step_at_end=step_at_end)
+end

src/TensorBoardLogger.jl

@@ -19,9 +19,9 @@ using  Base.CoreLogging: CoreLogging, AbstractLogger, LogLevel, Info,
     handle_message, shouldlog, min_enabled_level, catch_exceptions, with_logger,
     NullLogger
 
-export TBLogger, reset!, set_step!, increment_step!, set_step_increment!
+export TBLogger, reset!, set_step!, increment_step!, set_step_increment!, with_TBLogger_hold_step
 export log_histogram, log_value, log_vector, log_text, log_image, log_images,
-       log_audio, log_audios, log_graph, log_embeddings, log_custom_scalar
+       log_audio, log_audios, log_graph, log_embeddings, log_custom_scalar, write_hparams!
 export map_summaries, TBReader
 
 export ImageFormat, L, CL, LC, LN, NL, NCL, NLC, CLN, LCN, HW, WH, HWC, WHC,
@@ -62,6 +62,7 @@ include("ImageFormat.jl")
 const TB_PLUGIN_JLARRAY_NAME = "_jl_tbl_array_sz"
 
 include("TBLogger.jl")
+include("hparams.jl")
 include("utils.jl")  # CRC Utils
 include("event.jl")
 include("Loggers/base.jl")
@@ -102,6 +103,16 @@ function __init__()
     @require PyPlot="d330b81b-6aea-500a-939a-2ce795aea3ee" begin
         include("Optional/PyPlot.jl")
     end
+    @require Gadfly="c91e804a-d5a3-530f-b6f0-dfbca275c004" begin
+        @require Fontconfig="186bb1d3-e1f7-5a2c-a377-96d770f13627" begin
+            @require Cairo="159f3aea-2a34-519c-b102-8c37f9878175" begin
+                using .Cairo
+                using .Fontconfig
+                include("Optional/Gadfly.jl")        
+            end
+        end
+    end
+    # @require Gadfly="c91e804a-d5a3-530f-b6f0-dfbca275c004" include("Optional/Gadfly.jl")
     @require Tracker="9f7883ad-71c0-57eb-9f7f-b5c9e6d3789c" begin
         include("Optional/Tracker.jl")
     end

src/event.jl

@@ -29,17 +29,16 @@ format. The format follows the following rule (in bytes)
 #4 N..N+8      UInt32  masked_CRC of #3
 """
 function write_event(out::IO, event::Event)
-    data = PipeBuffer();
-    encode(ProtoEncoder(data), event)
+    event_bytes = serialize_proto(event)
 
     #header
-    header     = collect(reinterpret(UInt8, [data.size]))
+    header     = collect(reinterpret(UInt8, [length(event_bytes)]))
     crc_header = reinterpret(UInt8, UInt32[masked_crc32c(header)])
-    crc_data   = reinterpret(UInt8, UInt32[masked_crc32c(data.data)])
+    crc_data   = reinterpret(UInt8, UInt32[masked_crc32c(event_bytes)])
 
     write(out, header)
     write(out, crc_header)
-    write(out, data.data)
+    write(out, event_bytes)
     write(out, crc_data)
     flush(out)
 end