Merge pull request #105 from rfourquet/rf/MT-seed

seed explicitly MersenneTwister(0)

deprecated/src7/deprecated/model.jl

@@ -18,7 +18,7 @@ gnorm(m::Model,g=0)=(for p in params(m); g += vnorm(p.dif); end; g)           #
 item2xy(item)=(isa(item, Tuple) ? (item[1:end-1],item[end]) : item==nothing ? (nothing,nothing) : ((),item))
 
 # So gradient checking does not mess up random seed:
-const gradcheck_rng = MersenneTwister()
+const gradcheck_rng = MersenneTwister(0)
 
 
 # train(m, d; seq=false, a...)=(!seq ? train1(m,d;a...) : train2(m,d;a...))
@@ -48,7 +48,7 @@ const gradcheck_rng = MersenneTwister()
 # end
 
 # NO: make the model interface more functional:
-# back and loss rely on hidden state info.  
+# back and loss rely on hidden state info.
 # forw has to allocate.
 # purely functional models are impossible.
 # forw needs to compute intermediate values.
@@ -71,7 +71,7 @@ const gradcheck_rng = MersenneTwister()
 # end
 
 # Use test with percloss instead:
-# 
+#
 # function accuracy(m::Model, d) # TODO: this only works if y is a single item
 #     numcorr = numinst = 0
 #     z = nothing
@@ -108,4 +108,3 @@ const gradcheck_rng = MersenneTwister()
 #     end
 #     return sumloss/numloss
 # end
-

deprecated/src7/gradcheck.jl

@@ -1,8 +1,8 @@
 function gradcheck(m, grad, loss; gcheck=10, _eps=cbrt(eps(eltype(m))), delta=_eps, atol=_eps, rtol=_eps, o...)
     # 6e-6 for Float64, 5e-3 for Float32 works best
     # rnum = 42 # time_ns() #DBG
-    # isdefined(data,:rng) && (data_rng_save = data.rng; data.rng=MersenneTwister(); srand(data.rng,rnum))
-    # isdefined(m,:rng) && (m_rng_save = m.rng; m.rng=MersenneTwister(); srand(m.rng,rnum))
+    # isdefined(data,:rng) && (data_rng_save = data.rng; data.rng=MersenneTwister(0); srand(data.rng,rnum))
+    # isdefined(m,:rng) && (m_rng_save = m.rng; m.rng=MersenneTwister(0); srand(m.rng,rnum))
     # l = zeros(2)
     # train(m, data, loss; gcheck=true, losscnt=fill!(l,0), o...)
     # loss0 = l[1]

examples/deprecated/adding.jl

@@ -65,8 +65,7 @@ end
 
 function gradloss(f, data, loss; grad=false, seed=42)
     data_rng = data.rng
-    data.rng = MersenneTwister()
-    srand(data.rng, seed)
+    data.rng = MersenneTwister(seed)
     reset!(f)
     myforw = grad ? sforw : forw
     loss1 = 0
@@ -93,7 +92,7 @@ import Base: start, next, done
 
 type Data; len; batchsize; epochsize; batch; sum; cnt; rng;
     Data(len, batchsize, epochsize; rng=Base.GLOBAL_RNG) =
-    new(len, batchsize, epochsize, zeros(Float32,2,batchsize), 
+    new(len, batchsize, epochsize, zeros(Float32,2,batchsize),
         zeros(Float32,1,batchsize), zeros(Int,1,batchsize), rng)
 end
 
@@ -109,7 +108,7 @@ function next(a::Data, s)
     fill!(sub(a.batch,2,:),0)
     togo = a.len - t
     for b=1:a.batchsize
-        if (a.cnt[b]==0 ? rand(a.rng) <= 2/togo : 
+        if (a.cnt[b]==0 ? rand(a.rng) <= 2/togo :
             a.cnt[b]==1 ? rand(a.rng) <= 1/togo : false)
             a.batch[2,b] = 1
             a.cnt[b] += 1
@@ -407,7 +406,7 @@ end #module
 #     return(xx, yy)
 # end
 # net0 = (args["type"] == "irnn" ? Net(irnn(nh),quadlosslayer(ny)) :
-#         args["type"] == "lstm" ? Net(lstm(nh),quadlosslayer(ny)) : 
+#         args["type"] == "lstm" ? Net(lstm(nh),quadlosslayer(ny)) :
 #         error("Unknown network type "*args["type"]))
 # setparam!(net; lr=args["lr"], gc=args["gc"])  # do a global gclip instead of per parameter
         # "--test"
@@ -529,7 +528,7 @@ end #module
 
 # OLD GENERATOR FOR COMPARISON:
 
-# type Adding1; len; batchsize; epochsize; b; x; y; 
+# type Adding1; len; batchsize; epochsize; b; x; y;
 #     Adding1(len, batchsize, epochsize; o...)=new(len,batchsize,epochsize,Adding0(len,batchsize,epochsize; o...))
 # end
 
@@ -548,7 +547,7 @@ end #module
 # end
 
 # type Adding0; len; batchsize; epochsize; rng;
-#     Adding0(len, batchsize, epochsize; rng=MersenneTwister())=new(len, batchsize, epochsize, rng)
+#     Adding0(len, batchsize, epochsize; rng=MersenneTwister(0))=new(len, batchsize, epochsize, rng)
 # end
 
 # start(a::Adding0)=0
@@ -574,7 +573,7 @@ end #module
 
 
     # p1 = (opts["nettype"] == "irnn" ? Net(irnn; out=opts["hidden"], winit=Gaussian(0,opts["winit"])) :
-    #       opts["nettype"] == "lstm" ? Net(lstm; out=opts["hidden"], fbias=opts["fbias"]) : 
+    #       opts["nettype"] == "lstm" ? Net(lstm; out=opts["hidden"], fbias=opts["fbias"]) :
     #       error("Unknown network type "*opts["nettype"]))
     # p2 = Net(wb; out=1, winit=Gaussian(0,opts["winit"]))
 
@@ -585,4 +584,3 @@ end #module
 # @knet function p2(x; o...)
 #     y = wb(x; o..., out=1)
 # end
-

examples/deprecated/mnistpixels.jl

@@ -86,8 +86,7 @@ end
 
 function gradloss(f, data, loss; grad=false, seed=42)
     data_rng = data.rng
-    data.rng = MersenneTwister()
-    srand(data.rng, seed)
+    data.rng = MersenneTwister(seed)
     reset!(f)
     myforw = grad ? sforw : forw
     loss1 = 0
@@ -118,7 +117,7 @@ import Base: start, next, done
 # the last pixel should be served as x(1,batch), y(10,batch)
 
 type Pixels; x; y; rng; datasize; epochsize; batchsize; bootstrap; shuffle; xbatch; ybatch; images;
-    function Pixels(x, y; rng=MersenneTwister(), epoch=ccount(x), batch=16, bootstrap=false, shuffle=false)
+    function Pixels(x, y; rng=MersenneTwister(0), epoch=ccount(x), batch=16, bootstrap=false, shuffle=false)
         nx = ccount(x)
         nx == ccount(y) || error("Item count mismatch")
         shuf = (shuffle ? shuffle!(rng,[1:nx;]) : nothing)
@@ -176,7 +175,7 @@ function parse_commandline(args)
         "--lrate"
         help = "learning rate"
         arg_type = Float64
-        default = 0.005          # paper says 1e-8? 
+        default = 0.005          # paper says 1e-8?
         "--gclip"
         help = "gradient clip"
         arg_type = Float64
@@ -277,6 +276,6 @@ end # module
 
 # S2C no longer accepts Net, it expects kfun:
     # p1 = (nettype == "irnn" ? Net(irnn; out=hidden, winit=Gaussian(0,winit)) :
-    #       nettype == "lstm" ? Net(lstm; out=hidden, fbias=fbias) : 
+    #       nettype == "lstm" ? Net(lstm; out=hidden, fbias=fbias) :
     #       error("Unknown network type "*nettype))
     # p2 = Net(wbf; out=10, winit=Gaussian(0,winit), f=soft)