Merge pull request #491 from malmaud/eager_mode

Eager mode

Project.toml

@@ -5,8 +5,10 @@ version = "0.12.0"
 
 [deps]
 AutoHashEquals = "15f4f7f2-30c1-5605-9d31-71845cf9641f"
+CRC32c = "8bf52ea8-c179-5cab-976a-9e18b702a9bc"
 Compat = "34da2185-b29b-5c13-b0c7-acf172513d20"
 Conda = "8f4d0f93-b110-5947-807f-2305c1781a2d"
+Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"
 Distributed = "8ba89e20-285c-5b6f-9357-94700520ee1b"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
 FileIO = "5789e2e9-d7fb-5bc7-8068-2c6fae9b9549"

deps/build.jl

@@ -1,8 +1,8 @@
 using PyCall
 using Conda
 
-const cur_version = "1.12.0"
-const cur_py_version = "1.12.0"
+const cur_version = "1.13.1"
+const cur_py_version = "1.12"  # Temporarily downgrade Python version until 1.13.1 is released on Conda
 
 
 ############################

examples/diffeq.jl

@@ -0,0 +1,8 @@
+using DifferentialEquations
+
+f(u,p,t)=1.01 .* u
+
+u0=constant(0.5)
+tspan=(0.0,1.0)
+prob=ODEProblem(f, u0, tspan)
+s=solve(prob)

examples/keras.jl

@@ -0,0 +1,13 @@
+using TensorFlow
+tf=TensorFlow
+tf.enable_eager_execution()
+m = tf.Sequential()
+
+tf.add(m, tf.Dense(3,10))
+tf.add(m, tf.Relu())
+tf.add(m, tf.Dense(10, 3))
+
+x=constant(randn(5,3))
+y=3x+5
+tf.compile(m, optimizer=tf.SGD(lr=1e-4), loss=tf.mse)
+tf.fit(m, x, y, n_epochs=200)

src/TensorFlow.jl

@@ -124,7 +124,14 @@ Ops,
 slice,
 import_op,
 @tfimport,
-tf_versioninfo
+tf_versioninfo,
+copy_to_device,
+enable_eager_execution,
+EagerTensor,
+summary,
+create_tape,
+set_tape,
+with_tape
 
 
 using Distributed
@@ -141,8 +148,13 @@ function deallocator(data, len, arg)
 
 end
 
+include("context.jl")
+
 function __init__()
     c_deallocator[] = @cfunction(deallocator, Cvoid, (Ptr{Cvoid}, Csize_t, Ptr{Cvoid}))
+    for context in default_context()
+        push!(global_context, context)
+    end
 end
 
 function load_python_process(;force_reload=false)
@@ -198,6 +210,7 @@ include("meta.jl")
 include("constants.jl")
 include("tensorflow_protos.jl")
 include("core.jl")
+include("eager.jl")
 include("run.jl")
 include("version.jl")
 include("ops.jl")
@@ -211,5 +224,7 @@ include("summary.jl")
 include("deprecated.jl")
 include("show.jl")
 include("generate_ops.jl")
+include("tape.jl")
+include("keras.jl")
 
 end

src/context.jl

@@ -0,0 +1,44 @@
+abstract type Context
+end
+
+struct ContextStack
+    contexts::Vector{Context}
+end
+
+ContextStack() = ContextStack(Context[])
+
+function Base.push!(stack::ContextStack, context::Context)
+    push!(stack.contexts, context)
+end
+
+function Base.pop!(stack::ContextStack)
+    pop!(stack.contexts)
+end
+
+function default_context()
+    return [ExecutionMode(eager=false)]
+end
+
+function context_value(context_type)
+    return global_context[context_type]
+end
+
+function Base.getindex(c::ContextStack, context_type)
+    value = nothing
+    for context in c.contexts
+        if isa(context, context_type)
+            value = context
+        end
+    end
+    return value
+end
+
+function with_context(block, ctx)
+    push!(global_context, ctx)
+    res = block()
+    # This assumes the block doesn't adjust the context. We should pop explicitly the pushed context.
+    pop!(global_context)
+    return res
+end
+
+const global_context = ContextStack()

src/core.jl

@@ -507,20 +507,20 @@ end
 mutable struct DeviceList
     ptr::Ptr{Cvoid}
     count::Int
+end
 
-    function DeviceList(s::Session)
-        status = Status()
-        ptr = @tfcall(:TF_SessionListDevices, Ptr{Cvoid},
-            (Ptr{Cvoid}, Ptr{Cvoid}), s, status)
-        check_status(status)
-        count = @tfcall(:TF_DeviceListCount, Cint, (Ptr{Cvoid},),
-            ptr)
-        this = new(ptr, count)
-        finalizer(this) do self
-            close(self)
-        end
-        this
+function DeviceList(s::Session)
+    status = Status()
+    ptr = @tfcall(:TF_SessionListDevices, Ptr{Cvoid},
+        (Ptr{Cvoid}, Ptr{Cvoid}), s, status)
+    check_status(status)
+    count = @tfcall(:TF_DeviceListCount, Cint, (Ptr{Cvoid},),
+        ptr)
+    this = DeviceList(ptr, count)
+    finalizer(this) do self
+        close(self)
     end
+    this
 end
 
 struct DeviceInfo
@@ -663,6 +663,8 @@ RawTensor(data::AbstractArray) = RawTensor(collect(data))
 
 RawTensor(t::RawTensor) = t
 
+Base.unsafe_convert(::Type{Ptr{Cvoid}}, t::RawTensor) = t.ptr
+
 function varint_encode(b::IO, n::Integer)
     while n ≥ 2^7
         write(b, UInt8(0b10000000 | (n & 0b1111111)))
@@ -803,7 +805,7 @@ function Base.sizeof(t::RawTensor)
     @tfcall(:TF_TensorByteSize, Csize_t, (Ptr{Cvoid},), t.ptr) |> Int
 end
 
-function set_device(node_desc, device::String)
+function set_device(node_desc, device)
     @tfcall(:TF_SetDevice, Cvoid,
         (Ptr{Cvoid}, Cstring),
         node_desc.ptr, device)
@@ -1168,7 +1170,10 @@ function load_proto(value::tensorflow.AttrValue)
         load_proto(value.list)
     elseif has_field(value, :_type)
         type_ = value._type
-        proto_type_map[type_]
+        get(proto_type_map, type_) do
+            @warn "Unrecognized type. Defaulting to Float32." type_
+            Float32
+        end
     end
 end
 
@@ -1218,10 +1223,6 @@ Represents the output of an operation in the computation graph
     value_index::Int
 end
 
-get_graph(t::AbstractTensor) = Tensor(t).op.graph
-
-node_name(t::AbstractTensor) = (node_name(Tensor(t).op), Tensor(t).value_index)
-
 function Tensor(op::Operation, value_index::Int)
     base_tensor = Tensor{Any}(op, value_index)
     Tensor{get_output_type(base_tensor)}(op, value_index)
@@ -1242,6 +1243,10 @@ Base.convert(::Type{Tensor{Any}}, value::Tensor{R}) where {R} = value
 
 Base.convert(::Type{Tensor{T}}, value) where {T} =  convert(Tensor{T}, constant(value))
 
+get_graph(t::AbstractTensor) = Tensor(t).op.graph
+
+node_name(t::AbstractTensor) = (node_name(Tensor(t).op), Tensor(t).value_index)
+
 function operation_output_type(port::Port)
     @tfcall(:TF_OperationOutputType, TF_DataType, (Port,), port)
 end