Merge branch 'master' into stable

Author: vchuravy

.travis.yml

@@ -6,7 +6,7 @@ os:
   - osx
 julia:
   - 0.5
-  - nightly
+#  - nightly 0.6 supports depends on #170
 
 # dependent apt packages
 addons:

deps/build.jl

@@ -31,8 +31,7 @@ if !libmxnet_detected
     base_url = "https://github.com/dmlc/mxnet/releases/download/20160531/20160531_win10_x64_cpu.7z"
     if libmxnet_curr_ver == "master"
       # download_cmd uses powershell 2, but we need powershell 3 to do this
-      ps_wget(url, file) = run(`powershell -NoProfile -Command "wget \"$url\" -o \"$file\""`)
-      ps_wget("https://api.github.com/repos/yajiedesign/mxnet/releases/latest", "mxnet.json")
+	  run(`powershell -NoProfile -Command Invoke-WebRequest -Uri "https://api.github.com/repos/yajiedesign/mxnet/releases/latest" -OutFile "mxnet.json"`)
       curr_win = JSON.parsefile("mxnet.json")["tag_name"]
       info("Can't use MXNet master on Windows, using latest binaries from $curr_win.")
     end

examples/char-lstm/seq-data.jl

@@ -5,7 +5,7 @@ using MXNet
 function build_vocabulary(corpus_fn::AbstractString, vocab_fn::AbstractString; max_vocab=10000)
   if isfile(vocab_fn)
     info("Vocabulary already exists, reusing $vocab_fn...")
-    vocab = Dict{Char,Int}([w => i for (i,w) in enumerate(readall(vocab_fn))])
+    vocab = Dict{Char,Int}([w => i for (i,w) in enumerate(readstring(vocab_fn))])
   else
     # count symbol frequency
     dict = Dict{Char,Int}()

examples/char-lstm/train.jl

@@ -14,7 +14,7 @@ lstm = LSTM(LSTM_N_LAYER, SEQ_LENGTH, DIM_HIDDEN, DIM_EMBED,
 
 #--data
 # load data
-text_all  = readall(INPUT_FILE)
+text_all  = readstring(INPUT_FILE)
 len_train = round(Int, length(text_all)*DATA_TR_RATIO)
 text_tr   = text_all[1:len_train]
 text_val  = text_all[len_train+1:end]

examples/mnist/mlp-test.jl

@@ -5,6 +5,8 @@ module MNISTTest
 using MXNet
 using Base.Test
 
+include("mnist-data.jl")
+
 function get_mnist_mlp()
   mlp = @mx.chain mx.Variable(:data)             =>
     mx.FullyConnected(name=:fc1, num_hidden=128) =>
@@ -17,7 +19,6 @@ function get_mnist_mlp()
 end
 
 function get_mnist_data(batch_size=100)
-  include("mnist-data.jl")
   return get_mnist_providers(batch_size)
 end
 
@@ -40,7 +41,7 @@ function mnist_fit_and_predict(optimizer, initializer, n_epoch)
   end
   mlp_load = mx.load("$cp_prefix-symbol.json", mx.SymbolicNode)
   @test mx.to_json(mlp_load) == mx.to_json(mlp)
-  mlp_load = mx.from_json(readall("$cp_prefix-symbol.json"), mx.SymbolicNode)
+  mlp_load = mx.from_json(readstring("$cp_prefix-symbol.json"), mx.SymbolicNode)
   @test mx.to_json(mlp_load) == mx.to_json(mlp)
 
   #--------------------------------------------------------------------------------

src/MXNet.jl

@@ -11,6 +11,10 @@ using Compat
 import Compat.String
 import Compat.view
 
+if VERSION >= v"0.6.0-dev.1024"
+  import Base.Iterators: filter
+end
+
 using Formatting
 
 # Functions from base that we can safely extend and that are defined by libmxnet.

src/model.jl

@@ -389,7 +389,7 @@ function fit(self :: FeedForward, optimizer :: AbstractOptimizer, data :: Abstra
     end
   end
 
-  train_execs = Array(Executor, num_dev)
+  train_execs = Array{Executor}(num_dev)
   for i = 1:num_dev
     data_shapes = Dict(map((x) -> x[1] => tuple(x[2][1:end-1]...,length(slices[i])), provide_data(data)))
     label_shapes = Dict(map((x) -> x[1] => tuple(x[2][1:end-1]...,length(slices[i])), provide_label(data)))

src/optimizers/sgd.jl

@@ -49,18 +49,21 @@ function create_state(self :: SGD, index :: Int, weight :: NDArray)
   end
 end
 
-function update(self :: SGD, index :: Int, weight :: NDArray, grad :: NDArray, state :: Union{Void, NDArray})
+function update(self :: SGD, index :: Int, weight :: NDArray, grad :: NDArray, state :: Void)
   lr = get_learning_rate(self.opts.lr_scheduler, self.state)
   grad = normalized_gradient(self.opts, self.state, weight, grad)
+  
+  @inplace weight += -lr * grad
+end
 
-  if isa(state, Void)
-    # vanilla SGD, without momentum
-    @inplace weight += -lr * grad
-  else
-    mom = state :: NDArray
-    coef = get_momentum(self.opts.momentum_scheduler, self.state)
-    @inplace mom    .*= coef
-    @inplace mom    .+= -lr * grad
-    @inplace weight .+= mom
-  end
+# update with momentum
+function update(self :: SGD, index :: Int, weight :: NDArray, grad :: NDArray, state :: NDArray)
+  lr = get_learning_rate(self.opts.lr_scheduler, self.state)
+  grad = normalized_gradient(self.opts, self.state, weight, grad)
+
+  mom = state :: NDArray
+  coef = get_momentum(self.opts.momentum_scheduler, self.state)
+  @inplace mom    .*= coef
+  @inplace mom    .+= -lr * grad
+  @inplace weight .+= mom
 end

src/symbolic-node.jl

@@ -627,8 +627,8 @@ function _define_atomic_symbol_creator(name :: String)
 
     $(if key_narg != ""
       quote
-        if !in(Symbol($key_narg), param_keys)
-          push!(param_keys, Symbol($key_narg))
+        if !in($key_narg, param_keys)
+          push!(param_keys, $key_narg)
           push!(param_vals, string(length(args)))
         end
       end

test/unittest/kvstore.jl

@@ -62,7 +62,7 @@ function test_aggregator()
 
   for vv in vals
     for v in vv
-      @test maximum(abs(copy(v)) - 2num_devs) == 0
+      @test maximum(abs.(copy(v)) - 2 * num_devs) == 0
     end
   end
 end

test/unittest/symbolic-node.jl

@@ -130,6 +130,14 @@ function test_dot()
   @test reldiff(ret, 2*ones(100, 200)) < 1e-6
 end
 
+function test_misc()
+  info("SymbolicNode::Miscellaneous")
+  # Test for #189
+  a = mx.Variable("a")
+  b = mx.Variable("b")
+  symb = mx.ElementWiseSum(a,b)
+end
+
 ################################################################################
 # Run tests
 ################################################################################
@@ -143,6 +151,7 @@ end
   test_attrs()
   test_functions()
   test_dot()
+  test_misc()
 end
 
 end