Add support for resnet_14l in tools/export.lua; Add resnet_14l.prototxt

appendix/caffe_prototxt/resnet_14l.prototxt

@@ -0,0 +1,524 @@
+name: "resnet_14l"
+layer {
+  name: "input"
+  type: "Input"
+  top: "input"
+  input_param { shape: { dim: 1 dim: 3 dim: 156 dim: 156 } }
+}
+layer {
+  name: "Convolution1"
+  type: "Convolution"
+  bottom: "input"
+  top: "Convolution1"
+  convolution_param {
+    num_output: 32
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU1"
+  type: "ReLU"
+  bottom: "Convolution1"
+  top: "Convolution1"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Convolution2"
+  type: "Convolution"
+  bottom: "Convolution1"
+  top: "Convolution2"
+  convolution_param {
+    num_output: 64
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU2"
+  type: "ReLU"
+  bottom: "Convolution2"
+  top: "Convolution2"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Convolution3"
+  type: "Convolution"
+  bottom: "Convolution2"
+  top: "Convolution3"
+  convolution_param {
+    num_output: 64
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU3"
+  type: "ReLU"
+  bottom: "Convolution3"
+  top: "Convolution3"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Convolution4"
+  type: "Convolution"
+  bottom: "Convolution1"
+  top: "Convolution4"
+  convolution_param {
+    num_output: 64
+    bias_term: true
+    pad: 0
+    kernel_size: 1
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "Crop1"
+  type: "Crop"
+  bottom: "Convolution4"
+  bottom: "Convolution3"
+  top: "Crop1"
+  crop_param {
+    axis: 2
+    offset: 2
+    offset: 2
+  }
+}
+layer {
+  name: "Eltwise1"
+  type: "Eltwise"
+  bottom: "Convolution3"
+  bottom: "Crop1"
+  top: "Eltwise1"
+  eltwise_param {
+    operation: SUM
+  }
+}
+layer {
+  name: "Convolution5"
+  type: "Convolution"
+  bottom: "Eltwise1"
+  top: "Convolution5"
+  convolution_param {
+    num_output: 64
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU4"
+  type: "ReLU"
+  bottom: "Convolution5"
+  top: "Convolution5"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Convolution6"
+  type: "Convolution"
+  bottom: "Convolution5"
+  top: "Convolution6"
+  convolution_param {
+    num_output: 64
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU5"
+  type: "ReLU"
+  bottom: "Convolution6"
+  top: "Convolution6"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Crop2"
+  type: "Crop"
+  bottom: "Eltwise1"
+  bottom: "Convolution6"
+  top: "Crop2"
+  crop_param {
+    axis: 2
+    offset: 2
+    offset: 2
+  }
+}
+layer {
+  name: "Eltwise2"
+  type: "Eltwise"
+  bottom: "Convolution6"
+  bottom: "Crop2"
+  top: "Eltwise2"
+  eltwise_param {
+    operation: SUM
+  }
+}
+layer {
+  name: "Convolution7"
+  type: "Convolution"
+  bottom: "Eltwise2"
+  top: "Convolution7"
+  convolution_param {
+    num_output: 128
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU6"
+  type: "ReLU"
+  bottom: "Convolution7"
+  top: "Convolution7"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Convolution8"
+  type: "Convolution"
+  bottom: "Convolution7"
+  top: "Convolution8"
+  convolution_param {
+    num_output: 128
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU7"
+  type: "ReLU"
+  bottom: "Convolution8"
+  top: "Convolution8"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Convolution9"
+  type: "Convolution"
+  bottom: "Eltwise2"
+  top: "Convolution9"
+  convolution_param {
+    num_output: 128
+    bias_term: true
+    pad: 0
+    kernel_size: 1
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "Crop3"
+  type: "Crop"
+  bottom: "Convolution9"
+  bottom: "Convolution8"
+  top: "Crop3"
+  crop_param {
+    axis: 2
+    offset: 2
+    offset: 2
+  }
+}
+layer {
+  name: "Eltwise3"
+  type: "Eltwise"
+  bottom: "Convolution8"
+  bottom: "Crop3"
+  top: "Eltwise3"
+  eltwise_param {
+    operation: SUM
+  }
+}
+layer {
+  name: "Convolution10"
+  type: "Convolution"
+  bottom: "Eltwise3"
+  top: "Convolution10"
+  convolution_param {
+    num_output: 128
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU8"
+  type: "ReLU"
+  bottom: "Convolution10"
+  top: "Convolution10"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Convolution11"
+  type: "Convolution"
+  bottom: "Convolution10"
+  top: "Convolution11"
+  convolution_param {
+    num_output: 128
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU9"
+  type: "ReLU"
+  bottom: "Convolution11"
+  top: "Convolution11"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Crop4"
+  type: "Crop"
+  bottom: "Eltwise3"
+  bottom: "Convolution11"
+  top: "Crop4"
+  crop_param {
+    axis: 2
+    offset: 2
+    offset: 2
+  }
+}
+layer {
+  name: "Eltwise4"
+  type: "Eltwise"
+  bottom: "Convolution11"
+  bottom: "Crop4"
+  top: "Eltwise4"
+  eltwise_param {
+    operation: SUM
+  }
+}
+layer {
+  name: "Convolution12"
+  type: "Convolution"
+  bottom: "Eltwise4"
+  top: "Convolution12"
+  convolution_param {
+    num_output: 256
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU10"
+  type: "ReLU"
+  bottom: "Convolution12"
+  top: "Convolution12"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Convolution13"
+  type: "Convolution"
+  bottom: "Convolution12"
+  top: "Convolution13"
+  convolution_param {
+    num_output: 256
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU11"
+  type: "ReLU"
+  bottom: "Convolution13"
+  top: "Convolution13"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Convolution14"
+  type: "Convolution"
+  bottom: "Eltwise4"
+  top: "Convolution14"
+  convolution_param {
+    num_output: 256
+    bias_term: true
+    pad: 0
+    kernel_size: 1
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "Crop5"
+  type: "Crop"
+  bottom: "Convolution14"
+  bottom: "Convolution13"
+  top: "Crop5"
+  crop_param {
+    axis: 2
+    offset: 2
+    offset: 2
+  }
+}
+layer {
+  name: "Eltwise5"
+  type: "Eltwise"
+  bottom: "Convolution13"
+  bottom: "Crop5"
+  top: "Eltwise5"
+  eltwise_param {
+    operation: SUM
+  }
+}
+layer {
+  name: "Convolution15"
+  type: "Convolution"
+  bottom: "Eltwise5"
+  top: "Convolution15"
+  convolution_param {
+    num_output: 256
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU12"
+  type: "ReLU"
+  bottom: "Convolution15"
+  top: "Convolution15"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Convolution16"
+  type: "Convolution"
+  bottom: "Convolution15"
+  top: "Convolution16"
+  convolution_param {
+    num_output: 256
+    bias_term: true
+    pad: 0
+    kernel_size: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
+layer {
+  name: "ReLU13"
+  type: "ReLU"
+  bottom: "Convolution16"
+  top: "Convolution16"
+  relu_param {
+    negative_slope: 0.1
+  }
+}
+layer {
+  name: "Crop6"
+  type: "Crop"
+  bottom: "Eltwise5"
+  bottom: "Convolution16"
+  top: "Crop6"
+  crop_param {
+    axis: 2
+    offset: 2
+    offset: 2
+  }
+}
+layer {
+  name: "Eltwise6"
+  type: "Eltwise"
+  bottom: "Convolution16"
+  bottom: "Crop6"
+  top: "Eltwise6"
+  eltwise_param {
+    operation: SUM
+  }
+}
+layer {
+  name: "Deconvolution1"
+  type: "Deconvolution"
+  bottom: "Eltwise6"
+  top: "Deconvolution1"
+  convolution_param {
+    num_output: 3
+    pad: 3
+    kernel_size: 4
+    stride: 2
+  }
+}

tools/export_model.lua

@@ -22,7 +22,6 @@ local function includes(s, a)
    end
    return false
 end
-
 local function get_bias(mod)
    if mod.bias then
       return mod.bias:float()
@@ -31,20 +30,18 @@ local function get_bias(mod)
       return torch.FloatTensor(mod.nOutputPlane):zero()
    end
 end
-local function export(model, output)
+local function export_weight(jmodules, seq)
    local targets = {"nn.SpatialConvolutionMM",
 		    "cudnn.SpatialConvolution",
 		    "nn.SpatialFullConvolution",
 		    "cudnn.SpatialFullConvolution"
    }
-   local jmodules = {}
-   local model_config = meta_data(model)
-   local first_layer = true
-
-   for k = 1, #model.modules do
-      local mod = model.modules[k]
+   for k = 1, #seq.modules do
+      local mod = seq.modules[k]
       local name = torch.typename(mod)
-      if includes(name, targets) then
+      if name == "nn.Sequential" or name == "nn.ConcatTable" then
+	 export_weight(jmodules, mod)
+      elseif includes(name, targets) then
 	 local weight = mod.weight:float()
 	 if name:match("FullConvolution") then
 	    weight = torch.totable(weight:reshape(mod.nInputPlane, mod.nOutputPlane, mod.kH, mod.kW))
@@ -71,6 +68,14 @@ local function export(model, output)
 	 table.insert(jmodules, jmod)
       end
    end
+end
+local function export(model, output)
+   local jmodules = {}
+   local model_config = meta_data(model)
+   local first_layer = true
+
+   export_weight(jmodules, model)
+
    local fp = io.open(output, "w")
    if not fp then
       error("IO Error: " .. output)