Merge pull request xxradon#1 from xxradon/master

update

Caffe/ReadMe.md

@@ -1,4 +1,4 @@
-# The Caffe in nn_tools Provides some convenient API
+# The Caffe in PytorchToCaffe Provides some convenient API
 If there are some problem in parse your prototxt or caffemodel, Please replace
 the caffe.proto with your own version and compile it with command
                    `protoc --python_out ./ caffe.proto`

Caffe/caffe.proto

@@ -565,6 +565,7 @@ message LayerParameter {
   optional RecurrentParameter recurrent_param = 146;
   optional ReductionParameter reduction_param = 136;
   optional ReLUParameter relu_param = 123;
+  optional ReLU6Parameter relu6_param = 208;
   optional ReshapeParameter reshape_param = 133;
   optional ROIPoolingParameter roi_pooling_param = 8266711; //roi pooling
   optional ScaleParameter scale_param = 142;
@@ -600,7 +601,7 @@ message LayerParameter {
   optional CropParameter crop_param = 167;
   optional DetectionEvaluateParameter detection_evaluate_param = 168;
   optional DetectionOutputParameter detection_output_param = 169;
-  //optional NormalizeParameter normalize_param = 170;
+  optional NormalizeParameter norm_param = 170;
   optional MultiBoxLossParameter multibox_loss_param = 171;
   optional PermuteParameter permute_param = 172;
   optional VideoDataParameter video_data_param = 173;
@@ -632,7 +633,7 @@ message LayerParameter {
   optional UpsampleParameter  upsample_param = 100003;
   optional MatMulParameter matmul_param = 100005;
   optional PassThroughParameter pass_through_param = 100004;
-  optional NormalizeParameter norm_param = 100001;
+  //optional NormalizeParameter norm_param = 100001;
 }
 
 //*********************add by wdd******************
@@ -658,13 +659,13 @@ message PassThroughParameter {
   optional uint32 block_width = 3 [default = 0];
 }
 
-message NormalizeParameter{
-optional bool across_spatial = 1 [default = true];
-optional FillerParameter scale_filler = 2;
-optional bool channel_shared = 3 [default = true];
-optional float eps = 4 [default = 1e-10];
-optional float sqrt_a = 5 [default = 1];
-}
+//message NormalizeParameter{
+//optional bool across_spatial = 1 [default = true];
+//optional FillerParameter scale_filler = 2;
+//optional bool channel_shared = 3 [default = true];
+//optional float eps = 4 [default = 1e-10];
+//optional float sqrt_a = 5 [default = 1];
+//}
 
 
 
@@ -1592,19 +1593,15 @@ message MultiBoxLossParameter {
 }
 
 // Message that stores parameters used by NormalizeLayer
-//message NormalizeParameter {
-//  //optional bool across_spatial = 1 [default = true];
-//  // Initial value of scale. Default is 1.0 for all
-//  //optional FillerParameter scale_filler = 2;
-//  // Whether or not scale parameters are shared across channels.
-//  //optional bool channel_shared = 3 [default = true];
-//  // Epsilon for not dividing by zero while normalizing variance
-//  //optional float eps = 4 [default = 1e-10];
-//  //**************************************************
-//  optional string normalize_type = 1 [default = "L2"];
-//  optional bool fix_gradient = 2 [default = false];
-//  optional bool bp_norm = 3 [default = false];
-//}
+message NormalizeParameter {
+  optional bool across_spatial = 1 [default = true];
+  // Initial value of scale. Default is 1.0 for all
+  optional FillerParameter scale_filler = 2;
+  // Whether or not scale parameters are shared across channels.
+  optional bool channel_shared = 3 [default = true];
+  // Epsilon for not dividing by zero while normalizing variance
+  optional float eps = 4 [default = 1e-10];
+}
 
 message PermuteParameter {
   // The new orders of the axes of data. Notice it should be with
@@ -1792,6 +1789,16 @@ message ReLUParameter {
   optional Engine engine = 2 [default = DEFAULT];
 }
 
+// Message that stores parameters used by ReLU6Layer
+message ReLU6Parameter {
+  enum Engine {
+    DEFAULT = 0;
+    CAFFE = 1;
+    CUDNN = 2;
+  }
+  optional Engine engine = 2 [default = DEFAULT];
+}
+
 message ReshapeParameter {
   // Specify the output dimensions. If some of the dimensions are set to 0,
   // the corresponding dimension from the bottom layer is used (unchanged).

Caffe/caffe_net.py

@@ -7,6 +7,7 @@
 class _Net(object):
     def __init__(self):
         self.net=pb.NetParameter()
+        self.needChange = {}
 
     def layer_index(self,layer_name):
         # find a layer's index by name. if the layer was found, return the layer position in the net, else return -1.
@@ -39,6 +40,18 @@ def remove_layer_by_name(self,layer_name):
                 return
         raise(AttributeError, "cannot found layer %s" % str(layer_name))
 
+
+
+    def remove_layer_by_type(self,type_name):
+        for i,layer in enumerate(self.net.layer):
+            if layer.type == type_name:
+                # self.change_layer_bottom(layer.top,layer.bottom)
+                self.needChange[layer.top[0]]=layer.bottom[0]
+                del self.net.layer[i]
+        return
+
+
+
     def get_layer_by_name(self, layer_name):
         # get the layer by layer_name
         for layer in self.net.layer:
@@ -52,7 +65,10 @@ def save_prototxt(self,path):
         for layer in prototxt.layer:
             del layer.blobs[:]
         with open(path,'w') as f:
-            f.write(text_format.MessageToString(prototxt))
+            string = text_format.MessageToString(prototxt)
+            for origin_name in self.needChange.keys():
+                string = string.replace(origin_name,self.needChange[origin_name])
+            f.write(string)
 
     def layer(self,layer_name):
         return self.get_layer_by_name(layer_name)

Caffe/layer_param.py

@@ -71,15 +71,57 @@ def conv_param(self, num_output, kernel_size, stride=(1), pad=(0,),
             conv_param.dilation.extend(pair_reduce(dilation))
         if groups:
             conv_param.group=groups
+            if groups != 1:
+                conv_param.engine = 1
         self.param.convolution_param.CopyFrom(conv_param)
 
-    def pool_param(self,type='MAX',kernel_size=2,stride=2,pad=None):
+    def norm_param(self, eps):
+        """
+        add a conv_param layer if you spec the layer type "Convolution"
+        Args:
+            num_output: a int
+            kernel_size: int list
+            stride: a int list
+            weight_filler_type: the weight filer type
+            bias_filler_type: the bias filler type
+        Returns:
+        """
+        l2norm_param = pb.NormalizeParameter()
+        l2norm_param.across_spatial = False
+        l2norm_param.channel_shared = False
+        l2norm_param.eps = eps
+        self.param.norm_param.CopyFrom(l2norm_param)
+
+
+    def permute_param(self, order1, order2, order3, order4):
+        """
+        add a conv_param layer if you spec the layer type "Convolution"
+        Args:
+            num_output: a int
+            kernel_size: int list
+            stride: a int list
+            weight_filler_type: the weight filer type
+            bias_filler_type: the bias filler type
+        Returns:
+        """
+        permute_param = pb.PermuteParameter()
+        permute_param.order.extend([order1, order2, order3, order4])
+
+        self.param.permute_param.CopyFrom(permute_param)
+
+
+    def pool_param(self,type='MAX',kernel_size=2,stride=2,pad=None, ceil_mode = True):
         pool_param=pb.PoolingParameter()
         pool_param.pool=pool_param.PoolMethod.Value(type)
         pool_param.kernel_size=pair_process(kernel_size)
         pool_param.stride=pair_process(stride)
+        pool_param.ceil_mode=ceil_mode
         if pad:
-            pool_param.pad=pad
+            if isinstance(pad,tuple):
+                pool_param.pad_h = pad[0]
+                pool_param.pad_w = pad[1]
+            else:
+                pool_param.pad=pad
         self.param.pooling_param.CopyFrom(pool_param)
 
     def batch_norm_param(self,use_global_stats=0,moving_average_fraction=None,eps=None):

README.md

@@ -9,9 +9,11 @@
 
 1. Converting a pytorch model to caffe model.
 2. Some convenient tools of manipulate caffemodel and prototxt quickly(like get or set weights of layers).
-3. Support pytorch version >= 0.2.(Have tested on 0.3,0.3.1, 0.4, 0.4.1 ,1.0)
+3. Support pytorch version >= 0.2.(Have tested on 0.3,0.3.1, 0.4, 0.4.1 ,1.0, 1.2)
 4. Analysing a model, get the operations number(ops) in every layers.
 
+Noting: pytorch version 1.1 is not supported now
+
 ### requirements
 
 - Python2.7 or Python3.x
@@ -53,7 +55,7 @@ For example `python pytorch_analyser.py example/resnet_pytorch_analysis_example.
 
 ## Pytorch to Caffe
 
-The new version of pytorch_to_caffe supporting the newest version(from 0.2.0 to 0.4.1) of pytorch.
+The new version of pytorch_to_caffe supporting the newest version(from 0.2.0 to 1.2.0) of pytorch.
 NOTICE: The transfer output will be somewhat different with the original model, caused by implementation difference.
 
 - Supporting layers types: 
@@ -74,29 +76,46 @@ dropout -> Dropout,
  batch_norm -> BatchNorm,Scale, 
  instance_norm -> BatchNorm,Scale,
  _interpolate  ->  Upsample
+ _hardtanh -> ReLU6
+ _permute -> Permute
+ _l2Norm -> Normalize
+
 
-- Supporting operations: torch.split, torch.max, torch.cat
+- Supporting operations: torch.split, torch.max, torch.cat ,torch.sigmoid, torch.div
 - Supporting tensor Variable operations: var.view, + (add), += (iadd), -(sub), -=(isub)
- \* (mul) *= (imul)
+ \* (mul) *= (imul), torch.Tensor.contiguous(_contiguous), torch.Tensor.pow(_pow), 
+ \* torch.Tensor.sum(_sum), torch.Tensor.sqrt(_sqrt), torch.Tensor.unsqueeze(_unsqueeze)
+ \* torch.Tensor.expand_as(_expand_as),
 
 Need to be added for caffe in the future:
-- Normalize,DepthwiseConv
+- DepthwiseConv
 
 The supported above can transfer many kinds of nets, 
 such as AlexNet(tested), VGG(tested), ResNet(fixed the bug in origin repo which mainly caused by ReLu layer function.), Inception_V3(tested).
 
 The supported layers concluded the most popular layers and operations.
  The other layer types will be added soon, you can ask me to add them in issues.
 
-Example: please see file `example/alexnet_pytorch_to_caffe.py`. Just Run `python3 example/alexnet_pytorch_to_caffe.py`
+Example: please see file `example/alexnet_pytorch_to_caffe.py`. Just Run `python3 example/alexnet_pytorch_to_caffe.py`.
 
-# Deploy verify
+Attention:
+the main difference from convert model is the BN layer,you should pay more attention to the BN parameters like  momentum=0.1, eps=1e-5.
+
+# Deploy verify(Very Important)
 After Converter,we should use verify_deploy.py to verify the output of pytorch model and the convertted caffe model.
 If you want to verify the outputs of caffe and pytorch,you should make caffe and pytorch install in the same environment,anaconda is recommended.
 using following script,we can install caffe-gpu(master branch). 
 ```angular2html
 conda install caffe-gpu pytorch cudatoolkit=9.0 -c pytorch 
 
+```
+other way,we can use docker,and in https://github.com/ufoym/deepo,for cuda9
+```
+docker pull ufoym/deepo:all-py36-cu90
+```
+for cuda10
+```
+docker pull ufoym/deepo:all-py36-cu100
 ```
 
 please see file `example/verify_deploy.py`,it can verify the output of pytorch model and the convertted caffe model in the same input.