diff --git a/metal/paddle-mobile-demo/paddle-mobile-demo/ViewController.swift b/metal/paddle-mobile-demo/paddle-mobile-demo/ViewController.swift
index e2351d14e317ada506d1f15c59efda22c984cab8..6e832a612107162adce250c3172602d4d5553547 100644
--- a/metal/paddle-mobile-demo/paddle-mobile-demo/ViewController.swift
+++ b/metal/paddle-mobile-demo/paddle-mobile-demo/ViewController.swift
@@ -79,7 +79,7 @@ class ViewController: UIViewController {
       return
     }
     do {
-      let max = 10
+      let max = 1
       let startDate = Date.init()
       for i in 0..<max {
         try net.predict(inTexture: inTexture) { [weak self] (result) in
diff --git a/metal/paddle-mobile/paddle-mobile/Executor.swift b/metal/paddle-mobile/paddle-mobile/Executor.swift
index 47c91de096ebaae1899a5ee6a0d4b85f5b1f99da..3296d106569048c6de673c9abfa4e5e1fc0ec79a 100644
--- a/metal/paddle-mobile/paddle-mobile/Executor.swift
+++ b/metal/paddle-mobile/paddle-mobile/Executor.swift
@@ -111,6 +111,11 @@ public class Executor<P: PrecisionType> {
     
     buffer.addCompletedHandler { (commandbuffer) in
     
+//      let inputArr = resInput.floatArray(res: { (p:P) -> P in
+//        return p
+//      })
+//      print(inputArr.strideArray())
+//
 //      let inputArr = resInput.floatArray(res: { (p:P) -> P in
 //        return p
 //      })
@@ -124,12 +129,12 @@ public class Executor<P: PrecisionType> {
       //            print(stridableInput)
       
       //            let _: Flo? = input.logDesc(header: "input: ", stridable: true)
-//      for i in 0..<self.ops.count {
-//        let op = self.ops[i]
-//        print(" 第 \(i) 个 op: ")
-//        op.delogOutput()
-//      }
-//      
+      for i in 0..<self.ops.count {
+        let op = self.ops[i]
+        print(" 第 \(i) 个 op: ")
+        op.delogOutput()
+      }
+//
 //      return
       
       let afterDate = Date.init()
diff --git a/metal/paddle-mobile/paddle-mobile/Operators/BoxcoderOp.swift b/metal/paddle-mobile/paddle-mobile/Operators/BoxcoderOp.swift
index eaa596e071b7628339be185a7e3599a370763041..e7d118afe83a200a464d191c8d391df128fa3936 100644
--- a/metal/paddle-mobile/paddle-mobile/Operators/BoxcoderOp.swift
+++ b/metal/paddle-mobile/paddle-mobile/Operators/BoxcoderOp.swift
@@ -76,10 +76,9 @@ class BoxcoderOp<P: PrecisionType>: Operator<BoxcoderKernel<P>, BoxcoderParam<P>
     print(targetBoxArray.strideArray())
     
     let originDim = para.output.originDim
-    
     let outputArray = para.output.metalTexture.realNHWC(dim: (n: originDim[0], h: originDim[1], w: originDim[2], c: originDim[3]))
+    print(" output ")
     print(outputArray.strideArray())
-    
   }
   
 }
diff --git a/metal/paddle-mobile/paddle-mobile/Operators/ConcatOp.swift b/metal/paddle-mobile/paddle-mobile/Operators/ConcatOp.swift
index 117fd8c39fadec55b88cc1aeaf3b91e5a0dd966d..84e58a0f3d537ed96d24744fc3803d0885375e14 100644
--- a/metal/paddle-mobile/paddle-mobile/Operators/ConcatOp.swift
+++ b/metal/paddle-mobile/paddle-mobile/Operators/ConcatOp.swift
@@ -65,10 +65,16 @@ class ConcatOp<P: PrecisionType>: Operator<ConcatKernel<P>, ConcatParam<P>>, Run
   
   func delogOutput() {
     print(" \(type) output: ")
+    
     let originDim = para.output.originDim
-    let outputArray = para.output.metalTexture.realNHWC(dim: (n: originDim[0], h: originDim[1], w: originDim[2], c: originDim[3]))
-    print(outputArray.strideArray())
-    print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+    if para.output.transpose == [0, 1, 2, 3] {
+      let outputArray = para.output.metalTexture.realNHWC(dim: (n: originDim[0], h: originDim[1], w: originDim[2], c: originDim[3]))
+      print(outputArray.strideArray())
+    } else if para.output.transpose == [0, 2, 3, 1] {
+          print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+    } else {
+      fatalError()
+    }
   }
 }
 
diff --git a/metal/paddle-mobile/paddle-mobile/Operators/ConvAddBatchNormReluOp.swift b/metal/paddle-mobile/paddle-mobile/Operators/ConvAddBatchNormReluOp.swift
index 6f67014444e5ef82fe4cdc30f99bc371fef2d417..8bb9349be7bd054a5e00005131b9baf1afa370d5 100644
--- a/metal/paddle-mobile/paddle-mobile/Operators/ConvAddBatchNormReluOp.swift
+++ b/metal/paddle-mobile/paddle-mobile/Operators/ConvAddBatchNormReluOp.swift
@@ -125,13 +125,13 @@ class ConvAddBatchNormReluOp<P: PrecisionType>: Operator<ConvAddBatchNormReluKer
     //        let _: P? = para.newBiase?.logDesc(header: "new biase: ", stridable: false)
     //        let _: P? = para.newScale?.logDesc(header: "new scale: ", stridable: false)
     
-    let output = para.output.metalTexture.floatArray { (p: P) -> P in
-      return p
-    }
-    //
-    writeToLibrary(fileName: "output_112x112x32_2", array: output)
-    print(" write done")
-    
-    //        let _: P? = para.output.metalTexture.logDesc(header: "conv add batchnorm relu output: ", stridable: false)
+//    let output = para.output.metalTexture.floatArray { (p: P) -> P in
+//      return p
+//    }
+//    //
+//    writeToLibrary(fileName: "output_112x112x32_2", array: output)
+//    print(" write done")
+//    
+//    //        let _: P? = para.output.metalTexture.logDesc(header: "conv add batchnorm relu output: ", stridable: false)
   }
 }
diff --git a/metal/paddle-mobile/paddle-mobile/Operators/FeedOp.swift b/metal/paddle-mobile/paddle-mobile/Operators/FeedOp.swift
index 59e5443a0f8c372dc366e2e1c65422dffbf42252..93560582166c53b6de308e53b1cda431151ad741 100644
--- a/metal/paddle-mobile/paddle-mobile/Operators/FeedOp.swift
+++ b/metal/paddle-mobile/paddle-mobile/Operators/FeedOp.swift
@@ -60,9 +60,8 @@ class FeedOp<P: PrecisionType>: Operator<Texture2DTo2DArrayKernel<P>, FeedParam<
   }
   
   func delogOutput() {
-    //        para.input.mtlTexture.logDesc()
-    //        let _: P? = para.input.mtlTexture.logDesc(header: "feed input: ", stridable: true)
-    //        let _: P? = para.output.metalTexture.logDesc(header: "feed output: ", stridable: false)
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.originDim[0], c: para.output.originDim[1], h: para.output.originDim[2], w: para.output.originDim[3])).strideArray())
   }
 }
 
diff --git a/metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvKernel.metal b/metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvKernel.metal
index a17366cbe0e67b4cd38cafe4e09909f537d269bf..3254346c8da2b63420c6dcaf540e3c28f663b890 100644
--- a/metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvKernel.metal
+++ b/metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvKernel.metal
@@ -769,3 +769,4 @@ kernel void depthwise_conv_batch_norm_relu_3x3(texture2d_array<float, access::sa
   outTexture.write(output, gid.xy, gid.z);
 }
 
+
diff --git a/metal/paddle-mobile/paddle-mobile/Operators/TransposeOp.swift b/metal/paddle-mobile/paddle-mobile/Operators/TransposeOp.swift
index 3f05a3db3e0de7abaa208ba8b0700688fe349d4e..8281ba543365410f49eaad1f4d992f54af30531d 100644
--- a/metal/paddle-mobile/paddle-mobile/Operators/TransposeOp.swift
+++ b/metal/paddle-mobile/paddle-mobile/Operators/TransposeOp.swift
@@ -48,9 +48,15 @@ class TransposeOp<P: PrecisionType>: Operator<TransposeKernel<P>, TransposeParam
   
   func delogOutput() {
     print(" \(type) output: ")
-    let originDim = para.output.tensorDim
-    let outputArray = para.output.metalTexture.realNHWC(dim: (n: originDim[0], h: originDim[1], w: originDim[2], c: originDim[3]))
-    print(outputArray.strideArray())
+    let originDim = para.output.originDim
+    if para.output.transpose == [0, 1, 2, 3] {
+      let outputArray = para.output.metalTexture.realNHWC(dim: (n: originDim[0], h: originDim[1], w: originDim[2], c: originDim[3]))
+      print(outputArray.strideArray())
+    } else if para.output.transpose == [0, 2, 3, 1] {
+      print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+    } else {
+      print(" not implement")
+    }
   }
 }