Merge pull request #25 from pluskid/nd-tensor

Cutsomizable "channel" dimension for ND-tensor

src/Mocha.jl

@@ -15,6 +15,7 @@ include("exception.jl")
 include("utils/blas.jl")
 include("utils/math.jl")
 include("utils/io.jl")
+include("utils/tensor.jl")
 
 if Config.use_native_extension
   include("utils/im2col-native.jl")

src/cuda/layers/accuracy.jl

@@ -1,6 +1,7 @@
-function setup_etc(backend::GPUBackend, layer::AccuracyLayer, inputs)
-  width, height, channels, num = get_whcn(inputs[1])
-  etc = make_blob(backend, eltype(inputs[1]), (width,height,1,num))
+function setup_etc(backend::GPUBackend, layer::AccuracyLayer, op_dim::Int, inputs)
+  dims = [size(inputs[1])...]
+  dims[op_dim] = 1
+  etc = make_blob(backend, eltype(inputs[1]), dims...)
   return etc
 end
 function shutdown(backend::GPUBackend, state::AccuracyLayerState)
@@ -11,8 +12,7 @@ function forward(backend::GPUBackend, state::AccuracyLayerState, inputs::Vector{
   pred = inputs[1]
   label = inputs[2]
 
-  width, height, channels, num = get_whcn(pred)
-  spatial_dim = width*height
+  spatial_dim, pred_dim, num = split_dims(pred, state.op_dim)
   data_type = eltype(pred)
 
   x_block = int(ceil(float64(num)/CUDA.THREADS_PER_BLOCK_X));
@@ -26,7 +26,7 @@ function forward(backend::GPUBackend, state::AccuracyLayerState, inputs::Vector{
     error("Unsupported data type $data_type")
   end
   CUDA.launch(kernel, (x_block,y_block),(CUDA.THREADS_PER_BLOCK_X,CUDA.THREADS_PER_BLOCK_Y),
-      (pred.ptr.p, label.ptr.p, state.etc.ptr.p, num, channels, spatial_dim));
+      (pred.ptr.p, label.ptr.p, state.etc.ptr.p, num, pred_dim, spatial_dim));
 
   N = num * spatial_dim
   accuracy = CuBLAS.dot(backend.cublas_ctx, data_type, N, state.etc.ptr, 1, state.etc.ptr, 1)

src/cuda/layers/argmax.jl

@@ -3,8 +3,7 @@ function forward(backend::GPUBackend, state::ArgmaxLayerState, inputs::Vector{Bl
     input = inputs[i]
     output = state.blobs[i]
 
-    width, height, channels, num = get_whcn(input)
-    spatial_dim = width*height
+    spatial_dim, channels, num = split_dims(input, state.dims[i])
     data_type = eltype(input)
 
     x_block = int(ceil(float64(num)/CUDA.THREADS_PER_BLOCK_X));

src/cuda/layers/channel-pooling.jl

@@ -1,16 +1,14 @@
-function setup_etc(backend::GPUBackend, layer::ChannelPoolingLayer, inputs, pooled_chann)
+function setup_etc(backend::GPUBackend, layer::ChannelPoolingLayer, inputs, blobs)
   if isa(layer.pooling, Pooling.Max)
     masks = Array(CuPtr, length(inputs))
     for i = 1:length(inputs)
-      masks[i] = CUDA.cualloc(Csize_t, get_width(inputs[i]) * get_height(inputs[i]) *
-          pooled_chann[i] * get_num(inputs[i]))
+      masks[i] = CUDA.cualloc(Csize_t, length(blobs[i]))
     end
     etc = masks
   elseif isa(layer.pooling, Pooling.Mean)
     integrals = Array(CuPtr, length(inputs))
     for i = 1:length(inputs)
-      integrals[i] = CUDA.cualloc(eltype(inputs[i]), get_width(inputs[i]) * get_height(inputs[i]) *
-          get_chann(inputs[i]))
+      integrals[i] = CUDA.cualloc(eltype(inputs[i]), prod(size(inputs[i])[1:end-1]))
     end
     etc = integrals
   else
@@ -39,9 +37,9 @@ function forward(backend::GPUBackend, pool::StdPoolingFunction,
     output = state.blobs[i]
 
     if isa(pool, Pooling.Max)
-      cuda_max_channel_pooling_forward(backend, input, output, state.etc[i], state.layer)
+      cuda_max_channel_pooling_forward(backend, input, output, state.etc[i], state.layer, state.op_dims[i])
     elseif isa(pool, Pooling.Mean)
-      cuda_mean_channel_pooling_forward(backend, input, output, state.etc[i], state.layer)
+      cuda_mean_channel_pooling_forward(backend, input, output, state.etc[i], state.layer, state.op_dims[i])
     else
       error("Pooling for $pool not implemented yet")
     end
@@ -59,9 +57,9 @@ function backward(backend::GPUBackend, pool::StdPoolingFunction, state::ChannelP
     diff = diffs[i]
     if !isa(diff, NullBlob)
       if isa(pool, Pooling.Max)
-        cuda_max_channel_pooling_backward(backend, diff, state.blobs_diff[i], state.etc[i], state.layer)
+        cuda_max_channel_pooling_backward(backend, diff, state.blobs_diff[i], state.etc[i], state.layer, state.op_dims[i])
       elseif isa(pool, Pooling.Mean)
-        cuda_mean_channel_pooling_backward(backend, diff, state.blobs_diff[i], state.layer)
+        cuda_mean_channel_pooling_backward(backend, diff, state.blobs_diff[i], state.layer, state.op_dims[i])
       else
         error("Pooling for $pool not implemented yet")
       end
@@ -72,15 +70,14 @@ function backward(backend::GPUBackend, pool::StdPoolingFunction, state::ChannelP
 end
 
 function cuda_mean_channel_pooling_forward{T}(backend::GPUBackend, input::CuTensorBlob{T},
-    output::CuTensorBlob{T}, integral::CuPtr, layer)
+    output::CuTensorBlob{T}, integral::CuPtr, layer, op_dim)
 
-  width, height, channels, num = size(input)
-  pooled_chann = size(output, 3)
+  spatial_dim_T, channels, num = split_dims(input, op_dim)
+  pooled_chann = size(output, op_dim)
   one = convert(T, 1)
   neg_one = convert(T, -1)
   scale = convert(T, 1.0/layer.kernel)
 
-  spatial_dim_T = width*height
   spatial_dim = spatial_dim_T * sizeof(T)
   fea_dim = spatial_dim * channels
   output_fea_dim = spatial_dim * pooled_chann
@@ -118,15 +115,14 @@ function cuda_mean_channel_pooling_forward{T}(backend::GPUBackend, input::CuTens
 end
 
 function cuda_mean_channel_pooling_backward{T}(backend::GPUBackend, input::CuTensorBlob{T},
-    output::CuTensorBlob{T}, layer)
+    output::CuTensorBlob{T}, layer, op_dim)
 
-  width, height, channels, num = size(input)
-  pooled_chann = size(output, 3)
+  spatial_dim_T, channels, num = split_dims(input, op_dim)
+  pooled_chann = size(output, op_dim)
   scale = 1/convert(T, layer.kernel)
 
   fill!(input, 0)
 
-  spatial_dim_T = width*height
   spatial_dim = spatial_dim_T * sizeof(T)
   fea_dim = spatial_dim * channels
   output_fea_dim = spatial_dim * pooled_chann
@@ -159,11 +155,10 @@ function cuda_geometry_max_chann_pool(sp_dim::Int, num::Int)
 
 end
 function cuda_max_channel_pooling_forward{T}(backend::GPUBackend, input::CuTensorBlob{T},
-    output::CuTensorBlob{T}, mask::CuPtr, layer)
+    output::CuTensorBlob{T}, mask::CuPtr, layer, op_dim)
 
-  width, height, channels, num = size(input)
-  sp_dim = width*height
-  pooled_chann = get_chann(output)
+  sp_dim, channels, num = split_dims(input, op_dim)
+  pooled_chann = size(output, op_dim)
 
   cuda_dim = cuda_geometry_max_chann_pool(sp_dim, num);
   if T == Float32
@@ -179,11 +174,10 @@ function cuda_max_channel_pooling_forward{T}(backend::GPUBackend, input::CuTenso
 end
 
 function cuda_max_channel_pooling_backward{T}(backend::GPUBackend, input::CuTensorBlob{T},
-    output::CuTensorBlob{T}, mask::CuPtr, layer)
+    output::CuTensorBlob{T}, mask::CuPtr, layer, op_dim)
 
-  width, height, channels, num = size(input)
-  sp_dim = width*height
-  pooled_chann = get_chann(output)
+  sp_dim, channels, num = split_dims(input, op_dim)
+  pooled_chann = size(output, op_dim)
 
   cuda_dim = cuda_geometry_max_chann_pool(sp_dim, num);
   if T == Float32

src/cuda/layers/multinomial-logistic-loss.jl

@@ -3,9 +3,7 @@ function forward(backend::GPUBackend, state::MultinomialLogisticLossLayerState,
   label     = inputs[2]
   data_type = eltype(pred)
 
-  width, height, channels, num = get_whcn(pred)
-
-  spatial_dim = height*width
+  spatial_dim, channels, num = split_dims(pred, state.op_dim)
   prob_dim = channels
 
   x_block = int(ceil(float64(num)/CUDA.THREADS_PER_BLOCK_X))

src/cuda/layers/softmax-loss.jl

@@ -4,9 +4,7 @@ function backward(backend::GPUBackend, state::SoftmaxLossLayerState, inputs::Vec
     copy!(diff, state.softmax.blobs[1])
 
     data_type = eltype(diff)
-    height, width, channels, num = get_whcn(diff)
-
-    spatial_dim = height*width
+    spatial_dim, channels, num = split_dims(diff, state.logistic.op_dim)
     prob_dim = channels
 
     x_block = int(ceil(float64(num)/CUDA.THREADS_PER_BLOCK_X))

src/cuda/layers/softmax.jl

@@ -3,12 +3,13 @@ type CuDNNSoftmaxState
   outputs_desc :: Vector{CuDNN.Tensor4dDescriptor}
 end
 
-function setup_etc(backend::GPUBackend, layer::SoftmaxLayer, data_type, inputs)
+function setup_etc(backend::GPUBackend, layer::SoftmaxLayer, dims::Vector{Int}, data_type, inputs)
   inputs_desc = Array(CuDNN.Tensor4dDescriptor, length(inputs))
   outputs_desc = Array(CuDNN.Tensor4dDescriptor, length(inputs))
   for i = 1:length(inputs)
-    inputs_desc[i] = CuDNN.create_tensor4d_descriptor(data_type, get_whcn(inputs[i]))
-    outputs_desc[i] = CuDNN.create_tensor4d_descriptor(data_type, get_whcn(inputs[i]))
+    dim_sp, dim_prob, dim_num = split_dims(inputs[i], dims[i])
+    inputs_desc[i] = CuDNN.create_tensor4d_descriptor(data_type, (1,dim_sp,dim_prob,dim_num))
+    outputs_desc[i] = CuDNN.create_tensor4d_descriptor(data_type, (1,dim_sp,dim_prob,dim_num))
   end
   etc = CuDNNSoftmaxState(inputs_desc, outputs_desc)
   return etc

src/layers/accuracy.jl

@@ -1,6 +1,7 @@
 @defstruct AccuracyLayer Layer (
   name :: String = "accuracy",
   report_error :: Bool = false,
+  (dim :: Int = -2, dim != 0),
   (bottoms :: Vector{Symbol} = Symbol[], length(bottoms) == 2),
 )
 @characterize_layer(AccuracyLayer,
@@ -11,18 +12,24 @@
 type AccuracyLayerState <: LayerState
   layer :: AccuracyLayer
 
+  op_dim   :: Int
   accuracy :: Float64
   n_accum  :: Int
   etc      :: Any
 end
 
-function setup_etc(backend::CPUBackend, layer::AccuracyLayer, inputs)
+function setup_etc(backend::CPUBackend, layer::AccuracyLayer, op_dim::Int, inputs)
   nothing
 end
 
 function setup(backend::Backend, layer::AccuracyLayer, inputs::Vector{Blob}, diffs::Vector{Blob})
-  etc = setup_etc(backend, layer, inputs)
-  return AccuracyLayerState(layer, 0.0, 0, etc)
+  total_dim = ndims(inputs[1])
+  dim = layer.dim < 0 ? layer.dim + total_dim + 1 : layer.dim
+  @assert 1 <= dim <= total_dim
+  @assert dim != total_dim
+
+  etc = setup_etc(backend, layer, dim, inputs)
+  return AccuracyLayerState(layer, dim, 0.0, 0, etc)
 end
 function shutdown(backend::CPUBackend, state::AccuracyLayerState)
 end
@@ -48,20 +55,18 @@ function forward(backend::CPUBackend, state::AccuracyLayerState, inputs::Vector{
   pred = inputs[1].data
   label = inputs[2].data
 
-  width, height, channels, num = get_whcn(pred)
-  canonical_pred  = reshape(pred, (width,height,channels,num))
-  canonical_label = reshape(label, (width,height,1,num))
+  dim_pre, dim_prob, dim_post = split_dims(pred, state.op_dim)
 
   accuracy = 0.0
-  for w = 1:width
-    for h = 1:height
-      for n = 1:num
-        if int(canonical_label[w,h,1,n])+1 == indmax(canonical_pred[w,h,:,n])
-          accuracy += 1.0
-        end
+  for i = 0:dim_pre-1
+    for j = 0:dim_post-1
+      idx = Int[i + dim_pre*(k + dim_prob*j) for k=0:dim_prob-1] + 1
+      @inbounds if int(label[i + dim_pre*j + 1])+1 == indmax(pred[idx])
+        accuracy += 1.0
       end
     end
   end
+
   state.accuracy = float64(state.accuracy * state.n_accum + accuracy) / (state.n_accum + length(label))
   state.n_accum += length(label)
 end

src/layers/argmax.jl

@@ -1,45 +1,52 @@
 @defstruct ArgmaxLayer Layer (
   name :: String = "argmax",
+  (dim :: Int = -2, dim != 0),
   (tops :: Vector{Symbol} = Symbol[], length(tops) > 0),
   (bottoms :: Vector{Symbol} = Symbol[], length(bottoms) == length(tops)),
 )
 
 type ArgmaxLayerState <: LayerState
   layer :: ArgmaxLayer
   blobs :: Vector{Blob}
+
+  dims  :: Vector{Int}
 end
 
 function setup(backend::Backend, layer::ArgmaxLayer, inputs::Vector{Blob}, diffs::Vector{Blob})
-  blobs = map(inputs) do input
-    width, height, channels, num = get_whcn(input)
-    data_type = eltype(input)
-
-    blob = make_blob(backend, data_type, width, height, 1, num)
-    blob
+  dims = Array(Int, length(inputs))
+  blobs = Array(Blob, length(inputs))
+  for i = 1:length(inputs)
+    total_dim = ndims(inputs[i])
+    dim = layer.dim < 0 ? layer.dim + total_dim + 1 : layer.dim
+    @assert 1 <= dim <= total_dim
+    @assert dim != total_dim
+    dims[i] = dim
+    shape = [size(inputs[i])...]
+    shape[dim] = 1
+    blobs[i] = make_blob(backend, eltype(inputs[i]), shape...)
   end
 
-  return ArgmaxLayerState(layer, blobs)
+  return ArgmaxLayerState(layer, blobs, dims)
 end
 
 function forward(backend::CPUBackend, state::ArgmaxLayerState, inputs::Vector{Blob})
   for i = 1:length(inputs)
     input = inputs[i].data
     output = state.blobs[i].data
-    width, height, channels, num = get_whcn(input)
-    canonical_input = reshape(input, (width,height,channels,num))
-    for n = 1:num
-      for w = 1:width
-        for h = 1:height
-          maxc = 1; maxval = canonical_input[w,h,maxc,n]
-          for c = 2:channels
-            @inbounds val = canonical_input[w,h,c,n]
-            if val > maxval
-              maxval = val
-              maxc = c
-            end
+    pre_dim, mid_dim, post_dim = split_dims(input, state.dims[i])
+    for x = 0:pre_dim-1
+      for z = 0:post_dim-1
+        idx = Int[x + pre_dim*(y + mid_dim*z) for y=0:mid_dim-1] + 1
+        maxc = 1
+        @inbounds maxval = input[idx[1]]
+        for y = 2:length(idx)
+          @inbounds val = input[idx[y]]
+          if val > maxval
+            maxval = val
+            maxc = y
           end
-          @inbounds output[w,h,1,n] = maxc-1
         end
+        @inbounds output[x + pre_dim*z + 1] = maxc-1
       end
     end
   end