VisualDL.jl

This package provides a julia wrapper for VisualDL, which is a deep learning visualization tool that can help design deep learning jobs.

Currently, the wrapper is written on top of the Python SDK of VisualDL by PyCall. I have tried to write the wrapper on top of the C++ SDK by leveraging CxxWrap.jl. But unluckily a strange error encountered. Hopefully I'll figured it out later and swap the backend into C++.

Install

• First, install the Python client of VisualDL. Checkout here for a detailed guide.

• Then add this package as a dependent(only tested on Julia v0.7).

  (v0.7) pkg> add VisualDL

Usage

First, initial the logger.

using VisualDL

train_logger = VisualDLLogger("tmp", 1, "train")
test_logger = as_mode(train_logger, "test")

Scalar

for i in 1:100
    with_logger(train_logger) do
        @log_scalar s0=(i,rand()) s1=(i, rand())
    end

    with_logger(test_logger) do
        @log_scalar s0=(i,rand()) s1=(i, rand())
    end
end

Histogram

for i in 1:100
    with_logger(train_logger) do
       @log_histogram h0=(i, randn(100))
    end
end

Text

for i in 1:100
    with_logger(train_logger) do
       @log_text t0=(i, "This is test " * string(i))
    end
end

Image

for i in 1:100
    with_logger(train_logger) do
       @log_image i0=([3,3,3], rand(27) * 255)
    end
end

for i in 1:100
    with_logger(test_logger) do
        @log_image image0=rand(10, 10, 3) * 255
    end
end


# force save and sync
save(train_logger)
save(test_logger)

Finally, run visualDL --logdir ./tmp in current dir. Then launch the visualdl service and watch the above pictures in browser. The default url is http://localhost:8040:

TODO

• More documentation
• Add LogReader and tests
• Precompile
• Travis
• Make Release
• Move out the start_sampling and finish_sampling from @log_image and @log_audio