Fix the padding logic in dilated conv pass.
*The previous logic of padding will run into issue in certain cases since it ignores the fact that `base_padding` (padding - crop) need to be explicitly checked to see if the original padding is 'SAME' or 'VALID'. *Updated several TODO items to clarify the possible future improvements to this pass. PiperOrigin-RevId: 304730637 Change-Id: Ie65f750a3d690856824bc3cde907a9ba254b28d7
This commit is contained in:
Haoliang Zhang
TensorFlower Gardener
parent
67570827b3
commit
3d1ea6a550
@ -1,37 +1,53 @@
|
||||
// RUN: tf-opt %s -tfl-identify-dilated-conv | FileCheck %s --dump-input-on-failure
|
||||
|
||||
func @testDilatedConv(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
func @testDilatedConv(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%cst_0 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.Conv2D"(%0, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %cst_0) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
return %2 : tensor<1x128x128x8xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedConv
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>)
|
||||
// CHECK-NEXT: [[RESULT:%.*]] = "tf.Conv2D"([[INPUT]], [[FILTER]]) {dilations = [1, 2, 2, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x128x128x3xf32>, tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32>
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128x8xf32>
|
||||
}
|
||||
|
||||
func @testDilatedConvWithNonConstantPadAndCrops(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.Conv2D"(%0, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %arg1) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
return %2 : tensor<1x128x128x8xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedConv
|
||||
// CHECK-LABEL: testDilatedConvWithNonConstantPadAndCrops
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[PADDING:%.*]]: tensor<2x2xi32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>)
|
||||
// CHECK-NEXT: [[RESULT:%.*]] = "tf.Conv2D"([[INPUT]], [[FILTER]]) {dilations = [1, 2, 2, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x128x128x3xf32>, tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32>
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128x8xf32>
|
||||
}
|
||||
|
||||
func @testDilatedConvWithNonZeroSTBPadding(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
func @testDilatedConvWithNonZeroBasePadding(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%cst_0 = constant dense<2> : tensor<2x2xi32>
|
||||
%cst_1 = constant dense<1> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.Conv2D"(%0, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %arg1) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%1 = "tf.Conv2D"(%0, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %cst_1) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
return %2 : tensor<1x128x128x8xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedConvWithNonZeroSTBPadding
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[PADDING:%.*]]: tensor<2x2xi32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>)
|
||||
// CHECK-LABEL: testDilatedConvWithNonZeroBasePadding
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>)
|
||||
// CHECK-NEXT: [[RESULT:%.*]] = "tf.Conv2D"([[INPUT]], [[FILTER]]) {dilations = [1, 2, 2, 1], padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<1x128x128x3xf32>, tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32>
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128x8xf32>
|
||||
}
|
||||
|
||||
func @testDilatedConvWithNonTrivialDilations(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
func @testDilatedConvWithNonTrivialDilations(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.Conv2D"(%0, %arg2) {padding = "VALID", dilations = [1, 2, 2, 1], strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %arg1) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%cst_0 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.Conv2D"(%0, %arg1) {padding = "VALID", dilations = [1, 2, 2, 1], strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %cst_0) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
return %2 : tensor<1x128x128x8xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedConvWithNonTrivialDilations
|
||||
@ -41,25 +57,27 @@ func @testDilatedConvWithNonTrivialDilations(%arg0: tensor<1x128x128x3xf32>, %ar
|
||||
// CHECK-NEXT: return [[RESULT]]
|
||||
}
|
||||
|
||||
func @testDilatedDepthWiseConv(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
func @testDilatedDepthWiseConv(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.DepthwiseConv2dNative"(%0, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %arg1) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%cst_0 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.DepthwiseConv2dNative"(%0, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %cst_0) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
return %2 : tensor<1x128x128x8xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedDepthWiseConv
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[PADDING:%.*]]: tensor<2x2xi32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>)
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>)
|
||||
// CHECK-NEXT: [[RESULT:%.*]] = "tf.DepthwiseConv2dNative"([[INPUT]], [[FILTER]]) {dilations = [1, 2, 2, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x128x128x3xf32>, tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32>
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128x8xf32>
|
||||
}
|
||||
|
||||
func @testDilatedConvWithPad(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x3x8xf32>, %arg3: tensor<8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%cst_0 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.Conv2D"(%0, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.Pad"(%1, %arg1) : (tensor<4x64x64x8xf32>, tensor<2x2xi32>) -> tensor<4x64x64x8xf32>
|
||||
%3 = "tf.BatchToSpaceND"(%2, %cst, %arg1) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%3 = "tf.BatchToSpaceND"(%2, %cst, %cst_0) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%4 = "tf.BiasAdd"(%3, %arg3) : (tensor<1x128x128x8xf32>, tensor<8xf32>) -> tensor<1x128x128x8xf32>
|
||||
return %4 : tensor<1x128x128x8xf32>
|
||||
|
||||
@ -72,10 +90,11 @@ func @testDilatedConvWithPad(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<2x2xi
|
||||
|
||||
func @testDilatedDepthWiseConvWithPad(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x3x8xf32>, %arg3: tensor<8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%cst_0 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.DepthwiseConv2dNative"(%0, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.Pad"(%1, %arg1) : (tensor<4x64x64x8xf32>, tensor<2x2xi32>) -> tensor<4x64x64x8xf32>
|
||||
%3 = "tf.BatchToSpaceND"(%2, %cst, %arg1) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%3 = "tf.BatchToSpaceND"(%2, %cst, %cst_0) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%4 = "tf.BiasAdd"(%3, %arg3) : (tensor<1x128x128x8xf32>, tensor<8xf32>) -> tensor<1x128x128x8xf32>
|
||||
return %4 : tensor<1x128x128x8xf32>
|
||||
|
||||
@ -86,49 +105,52 @@ func @testDilatedDepthWiseConvWithPad(%arg0: tensor<1x128x128x3xf32>, %arg1: ten
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128x8xf32>
|
||||
}
|
||||
|
||||
func @testDilatedConvWithBiasAdd(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x3x8xf32>, %arg3: tensor<8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
func @testDilatedConvWithBiasAdd(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<5x5x3x8xf32>, %arg2: tensor<8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.Conv2D"(%0, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %arg1) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%3 = "tf.BiasAdd"(%2, %arg3) : (tensor<1x128x128x8xf32>, tensor<8xf32>) -> tensor<1x128x128x8xf32>
|
||||
%cst_0 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.Conv2D"(%0, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %cst_0) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%3 = "tf.BiasAdd"(%2, %arg2) : (tensor<1x128x128x8xf32>, tensor<8xf32>) -> tensor<1x128x128x8xf32>
|
||||
return %3 : tensor<1x128x128x8xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedConvWithBiasAdd
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[PADDING:%.*]]: tensor<2x2xi32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>, [[BIAS:%.*]]: tensor<8xf32>)
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>, [[BIAS:%.*]]: tensor<8xf32>)
|
||||
// CHECK-NEXT: [[CONV:%.*]] = "tf.Conv2D"([[INPUT]], [[FILTER]]) {dilations = [1, 2, 2, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x128x128x3xf32>, tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32>
|
||||
// CHECK-NEXT: [[RESULT:%.*]] = "tf.BiasAdd"([[CONV]], [[BIAS]]) : (tensor<1x128x128x8xf32>, tensor<8xf32>) -> tensor<1x128x128x8xf32>
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128x8xf32>
|
||||
}
|
||||
|
||||
func @testDilatedDepthWiseConvWithBiasAdd(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x3x8xf32>, %arg3: tensor<8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
func @testDilatedDepthWiseConvWithBiasAdd(%arg0: tensor<1x128x128x3xf32>, %arg1: tensor<5x5x3x8xf32>, %arg2: tensor<8xf32>) -> tensor<1x128x128x8xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.DepthwiseConv2dNative"(%0, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %arg1) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%3 = "tf.BiasAdd"(%2, %arg3) : (tensor<1x128x128x8xf32>, tensor<8xf32>) -> tensor<1x128x128x8xf32>
|
||||
%cst_0 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>
|
||||
%1 = "tf.DepthwiseConv2dNative"(%0, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x3xf32>, tensor<5x5x3x8xf32>) -> tensor<4x64x64x8xf32>
|
||||
%2 = "tf.BatchToSpaceND"(%1, %cst, %cst_0) : (tensor<4x64x64x8xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x8xf32>
|
||||
%3 = "tf.BiasAdd"(%2, %arg2) : (tensor<1x128x128x8xf32>, tensor<8xf32>) -> tensor<1x128x128x8xf32>
|
||||
return %3 : tensor<1x128x128x8xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedDepthWiseConvWithBiasAdd
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[PADDING:%.*]]: tensor<2x2xi32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>, [[BIAS:%.*]]: tensor<8xf32>)
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128x3xf32>, [[FILTER:%.*]]: tensor<5x5x3x8xf32>, [[BIAS:%.*]]: tensor<8xf32>)
|
||||
// CHECK-NEXT: [[CONV:%.*]] = "tf.DepthwiseConv2dNative"([[INPUT]], [[FILTER]]) {dilations = [1, 2, 2, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x128x128x3xf32>, tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32>
|
||||
// CHECK-NEXT: [[RESULT:%.*]] = "tf.BiasAdd"([[CONV]], [[BIAS]]) : (tensor<1x128x128x8xf32>, tensor<8xf32>) -> tensor<1x128x128x8xf32>
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128x8xf32>
|
||||
}
|
||||
|
||||
func @testDilatedConvWithExpandSqueeze1(%arg0: tensor<1x128x128xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x1x1xf32>, %arg3: tensor<128xf32>) -> tensor<1x128x128xf32> {
|
||||
func @testDilatedConvWithExpandSqueeze1(%arg0: tensor<1x128x128xf32>, %arg1: tensor<5x5x1x1xf32>, %arg2: tensor<128xf32>) -> tensor<1x128x128xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%cst_0 = "tf.Const"() { value = dense<3> : tensor<i32> } : () -> tensor<i32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68xf32>
|
||||
%cst_1 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68xf32>
|
||||
%1 = "tf.ExpandDims"(%0, %cst_0) : (tensor<4x68x68xf32>, tensor<i32>) -> tensor<4x68x68x1xf32>
|
||||
%2 = "tf.Conv2D"(%1, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x64x64x1xf32>
|
||||
%2 = "tf.Conv2D"(%1, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x64x64x1xf32>
|
||||
%3 = "tf.Squeeze"(%2) {squeeze_dims = [3]} : (tensor<4x64x64x1xf32>) -> tensor<4x64x64xf32>
|
||||
%4 = "tf.BatchToSpaceND"(%3, %cst, %arg1) : (tensor<4x64x64xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
%5 = "tf.BiasAdd"(%4, %arg3) : (tensor<1x128x128xf32>, tensor<128xf32>) -> tensor<1x128x128xf32>
|
||||
%4 = "tf.BatchToSpaceND"(%3, %cst, %cst_1) : (tensor<4x64x64xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
%5 = "tf.BiasAdd"(%4, %arg2) : (tensor<1x128x128xf32>, tensor<128xf32>) -> tensor<1x128x128xf32>
|
||||
return %5 : tensor<1x128x128xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedConvWithExpandSqueeze1
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128xf32>, [[PADDING:%.*]]: tensor<2x2xi32>, [[FILTER:%.*]]: tensor<5x5x1x1xf32>, [[BIAS:%.*]]: tensor<128xf32>)
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128xf32>, [[FILTER:%.*]]: tensor<5x5x1x1xf32>, [[BIAS:%.*]]: tensor<128xf32>)
|
||||
// CHECK-NEXT: [[AXIS:%.*]] = "tf.Const"() {value = dense<3> : tensor<i32>} : () -> tensor<i32>
|
||||
// CHECK-NEXT: [[EXPAND:%.*]] = "tf.ExpandDims"([[INPUT]], [[AXIS]]) : (tensor<1x128x128xf32>, tensor<i32>) -> tensor<1x128x128x1xf32>
|
||||
// CHECK-NEXT: [[CONV:%.*]] = "tf.Conv2D"([[EXPAND]], [[FILTER]]) {dilations = [1, 2, 2, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x128x128x1xf32>, tensor<5x5x1x1xf32>) -> tensor<1x128x128x1xf32>
|
||||
@ -137,19 +159,20 @@ func @testDilatedConvWithExpandSqueeze1(%arg0: tensor<1x128x128xf32>, %arg1: ten
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128xf32>
|
||||
}
|
||||
|
||||
func @testDilatedDepthWiseConvWithExpandSqueeze1(%arg0: tensor<1x128x128xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x1x1xf32>, %arg3: tensor<128xf32>) -> tensor<1x128x128xf32> {
|
||||
func @testDilatedDepthWiseConvWithExpandSqueeze1(%arg0: tensor<1x128x128xf32>, %arg1: tensor<5x5x1x1xf32>, %arg2: tensor<128xf32>) -> tensor<1x128x128xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%cst_0 = "tf.Const"() { value = dense<3> : tensor<i32> } : () -> tensor<i32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68xf32>
|
||||
%cst_1 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68xf32>
|
||||
%1 = "tf.ExpandDims"(%0, %cst_0) : (tensor<4x68x68xf32>, tensor<i32>) -> tensor<4x68x68x1xf32>
|
||||
%2 = "tf.DepthwiseConv2dNative"(%1, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x64x64x1xf32>
|
||||
%2 = "tf.DepthwiseConv2dNative"(%1, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x64x64x1xf32>
|
||||
%3 = "tf.Squeeze"(%2) {squeeze_dims = [3]} : (tensor<4x64x64x1xf32>) -> tensor<4x64x64xf32>
|
||||
%4 = "tf.BatchToSpaceND"(%3, %cst, %arg1) : (tensor<4x64x64xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
%5 = "tf.BiasAdd"(%4, %arg3) : (tensor<1x128x128xf32>, tensor<128xf32>) -> tensor<1x128x128xf32>
|
||||
%4 = "tf.BatchToSpaceND"(%3, %cst, %cst_1) : (tensor<4x64x64xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
%5 = "tf.BiasAdd"(%4, %arg2) : (tensor<1x128x128xf32>, tensor<128xf32>) -> tensor<1x128x128xf32>
|
||||
return %5 : tensor<1x128x128xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedDepthWiseConvWithExpandSqueeze1
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128xf32>, [[PADDING:%.*]]: tensor<2x2xi32>, [[FILTER:%.*]]: tensor<5x5x1x1xf32>, [[BIAS:%.*]]: tensor<128xf32>)
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128xf32>, [[FILTER:%.*]]: tensor<5x5x1x1xf32>, [[BIAS:%.*]]: tensor<128xf32>)
|
||||
// CHECK-NEXT: [[AXIS:%.*]] = "tf.Const"() {value = dense<3> : tensor<i32>} : () -> tensor<i32>
|
||||
// CHECK-NEXT: [[EXPAND:%.*]] = "tf.ExpandDims"([[INPUT]], [[AXIS]]) : (tensor<1x128x128xf32>, tensor<i32>) -> tensor<1x128x128x1xf32>
|
||||
// CHECK-NEXT: [[CONV:%.*]] = "tf.DepthwiseConv2dNative"([[EXPAND]], [[FILTER]]) {dilations = [1, 2, 2, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x128x128x1xf32>, tensor<5x5x1x1xf32>) -> tensor<1x128x128x1xf32>
|
||||
@ -158,19 +181,20 @@ func @testDilatedDepthWiseConvWithExpandSqueeze1(%arg0: tensor<1x128x128xf32>, %
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128xf32>
|
||||
}
|
||||
|
||||
func @testDilatedConvWithExpandSqueeze2(%arg0: tensor<1x128x128xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x1x1xf32>, %arg3: tensor<?xf32>) -> tensor<1x128x128xf32> {
|
||||
func @testDilatedConvWithExpandSqueeze2(%arg0: tensor<1x128x128xf32>, %arg1: tensor<5x5x1x1xf32>, %arg2: tensor<?xf32>) -> tensor<1x128x128xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%cst_0 = "tf.Const"() { value = dense<3> : tensor<i32> } : () -> tensor<i32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x?x?xf32>
|
||||
%cst_1 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x?x?xf32>
|
||||
%1 = "tf.ExpandDims"(%0, %cst_0) : (tensor<4x?x?xf32>, tensor<i32>) -> tensor<4x?x?x1xf32>
|
||||
%2 = "tf.Conv2D"(%1, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x?x?x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x?x?x1xf32>
|
||||
%2 = "tf.Conv2D"(%1, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x?x?x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x?x?x1xf32>
|
||||
%3 = "tf.Squeeze"(%2) {squeeze_dims = [3]} : (tensor<4x?x?x1xf32>) -> tensor<4x?x?xf32>
|
||||
%4 = "tf.BiasAdd"(%3, %arg3) : (tensor<4x?x?xf32>, tensor<?xf32>) -> tensor<4x?x?xf32>
|
||||
%5 = "tf.BatchToSpaceND"(%4, %cst, %arg1) : (tensor<4x?x?xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
%4 = "tf.BiasAdd"(%3, %arg2) : (tensor<4x?x?xf32>, tensor<?xf32>) -> tensor<4x?x?xf32>
|
||||
%5 = "tf.BatchToSpaceND"(%4, %cst, %cst_1) : (tensor<4x?x?xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
return %5 : tensor<1x128x128xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedConvWithExpandSqueeze2
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128xf32>, [[PADDING:%.*]]: tensor<2x2xi32>, [[FILTER:%.*]]: tensor<5x5x1x1xf32>, [[BIAS:%.*]]: tensor<?xf32>)
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128xf32>, [[FILTER:%.*]]: tensor<5x5x1x1xf32>, [[BIAS:%.*]]: tensor<?xf32>)
|
||||
// CHECK-NEXT: [[AXIS:%.*]] = "tf.Const"() {value = dense<3> : tensor<i32>} : () -> tensor<i32>
|
||||
// CHECK-NEXT: [[EXPAND:%.*]] = "tf.ExpandDims"([[INPUT]], [[AXIS]]) : (tensor<1x128x128xf32>, tensor<i32>) -> tensor<1x128x128x1xf32>
|
||||
// CHECK-NEXT: [[CONV:%.*]] = "tf.Conv2D"([[EXPAND]], [[FILTER]]) {dilations = [1, 2, 2, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x128x128x1xf32>, tensor<5x5x1x1xf32>) -> tensor<1x128x128x1xf32>
|
||||
@ -179,19 +203,20 @@ func @testDilatedConvWithExpandSqueeze2(%arg0: tensor<1x128x128xf32>, %arg1: ten
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128xf32>
|
||||
}
|
||||
|
||||
func @testDilatedDepthWiseConvWithExpandSqueeze2(%arg0: tensor<1x128x128xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x1x1xf32>, %arg3: tensor<?xf32>) -> tensor<1x128x128xf32> {
|
||||
func @testDilatedDepthWiseConvWithExpandSqueeze2(%arg0: tensor<1x128x128xf32>, %arg1: tensor<5x5x1x1xf32>, %arg2: tensor<?xf32>) -> tensor<1x128x128xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%cst_0 = "tf.Const"() { value = dense<3> : tensor<i32> } : () -> tensor<i32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x?x?xf32>
|
||||
%cst_1 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x?x?xf32>
|
||||
%1 = "tf.ExpandDims"(%0, %cst_0) : (tensor<4x?x?xf32>, tensor<i32>) -> tensor<4x?x?x1xf32>
|
||||
%2 = "tf.DepthwiseConv2dNative"(%1, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x?x?x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x?x?x1xf32>
|
||||
%2 = "tf.DepthwiseConv2dNative"(%1, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x?x?x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x?x?x1xf32>
|
||||
%3 = "tf.Squeeze"(%2) {squeeze_dims = [3]} : (tensor<4x?x?x1xf32>) -> tensor<4x?x?xf32>
|
||||
%4 = "tf.BiasAdd"(%3, %arg3) : (tensor<4x?x?xf32>, tensor<?xf32>) -> tensor<4x?x?xf32>
|
||||
%5 = "tf.BatchToSpaceND"(%4, %cst, %arg1) : (tensor<4x?x?xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
%4 = "tf.BiasAdd"(%3, %arg2) : (tensor<4x?x?xf32>, tensor<?xf32>) -> tensor<4x?x?xf32>
|
||||
%5 = "tf.BatchToSpaceND"(%4, %cst, %cst_1) : (tensor<4x?x?xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
return %5 : tensor<1x128x128xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedDepthWiseConvWithExpandSqueeze2
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128xf32>, [[PADDING:%.*]]: tensor<2x2xi32>, [[FILTER:%.*]]: tensor<5x5x1x1xf32>, [[BIAS:%.*]]: tensor<?xf32>)
|
||||
// CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128xf32>, [[FILTER:%.*]]: tensor<5x5x1x1xf32>, [[BIAS:%.*]]: tensor<?xf32>)
|
||||
// CHECK-NEXT: [[AXIS:%.*]] = "tf.Const"() {value = dense<3> : tensor<i32>} : () -> tensor<i32>
|
||||
// CHECK-NEXT: [[EXPAND:%.*]] = "tf.ExpandDims"([[INPUT]], [[AXIS]]) : (tensor<1x128x128xf32>, tensor<i32>) -> tensor<1x128x128x1xf32>
|
||||
// CHECK-NEXT: [[CONV:%.*]] = "tf.DepthwiseConv2dNative"([[EXPAND]], [[FILTER]]) {dilations = [1, 2, 2, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x128x128x1xf32>, tensor<5x5x1x1xf32>) -> tensor<1x128x128x1xf32>
|
||||
@ -203,12 +228,13 @@ func @testDilatedDepthWiseConvWithExpandSqueeze2(%arg0: tensor<1x128x128xf32>, %
|
||||
func @testDilatedConvWithExpandSqueeze3(%arg0: tensor<1x128x128xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x1x1xf32>, %arg3: tensor<128xf32>) -> tensor<1x128x128xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%cst_0 = "tf.Const"() { value = dense<3> : tensor<i32> } : () -> tensor<i32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68xf32>
|
||||
%cst_1 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68xf32>
|
||||
%1 = "tf.ExpandDims"(%0, %cst_0) : (tensor<4x68x68xf32>, tensor<i32>) -> tensor<4x68x68x1xf32>
|
||||
%2 = "tf.Conv2D"(%1, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x64x64x1xf32>
|
||||
%3 = "tf.Squeeze"(%2) {squeeze_dims = [3]} : (tensor<4x64x64x1xf32>) -> tensor<4x64x64xf32>
|
||||
%4 = "tf.Pad"(%3, %arg1) : (tensor<4x64x64xf32>, tensor<2x2xi32>) -> tensor<4x64x64xf32>
|
||||
%5 = "tf.BatchToSpaceND"(%4, %cst, %arg1) : (tensor<4x64x64xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
%5 = "tf.BatchToSpaceND"(%4, %cst, %cst_1) : (tensor<4x64x64xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
%6 = "tf.BiasAdd"(%5, %arg3) : (tensor<1x128x128xf32>, tensor<128xf32>) -> tensor<1x128x128xf32>
|
||||
return %6 : tensor<1x128x128xf32>
|
||||
|
||||
@ -225,12 +251,13 @@ func @testDilatedConvWithExpandSqueeze3(%arg0: tensor<1x128x128xf32>, %arg1: ten
|
||||
func @testDilatedDepthWiseConvWithExpandSqueeze3(%arg0: tensor<1x128x128xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x1x1xf32>, %arg3: tensor<128xf32>) -> tensor<1x128x128xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%cst_0 = "tf.Const"() { value = dense<3> : tensor<i32> } : () -> tensor<i32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68xf32>
|
||||
%cst_1 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68xf32>
|
||||
%1 = "tf.ExpandDims"(%0, %cst_0) : (tensor<4x68x68xf32>, tensor<i32>) -> tensor<4x68x68x1xf32>
|
||||
%2 = "tf.DepthwiseConv2dNative"(%1, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x64x64x1xf32>
|
||||
%3 = "tf.Squeeze"(%2) {squeeze_dims = [3]} : (tensor<4x64x64x1xf32>) -> tensor<4x64x64xf32>
|
||||
%4 = "tf.Pad"(%3, %arg1) : (tensor<4x64x64xf32>, tensor<2x2xi32>) -> tensor<4x64x64xf32>
|
||||
%5 = "tf.BatchToSpaceND"(%4, %cst, %arg1) : (tensor<4x64x64xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
%5 = "tf.BatchToSpaceND"(%4, %cst, %cst_1) : (tensor<4x64x64xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128xf32>
|
||||
%6 = "tf.BiasAdd"(%5, %arg3) : (tensor<1x128x128xf32>, tensor<128xf32>) -> tensor<1x128x128xf32>
|
||||
return %6 : tensor<1x128x128xf32>
|
||||
|
||||
@ -244,14 +271,15 @@ func @testDilatedDepthWiseConvWithExpandSqueeze3(%arg0: tensor<1x128x128xf32>, %
|
||||
// CHECK-NEXT: return [[RESULT]] : tensor<1x128x128xf32>
|
||||
}
|
||||
|
||||
func @testDilatedConvWithDifferentExpandSqueezeAxis(%arg0: tensor<1x128x128xf32>, %arg1: tensor<2x2xi32>, %arg2: tensor<5x5x1x1xf32>, %arg3: tensor<128xf32>) -> tensor<1x128x128x1xf32> {
|
||||
func @testDilatedConvWithDifferentExpandSqueezeAxis(%arg0: tensor<1x128x128xf32>, %arg1: tensor<5x5x1x1xf32>) -> tensor<1x128x128x1xf32> {
|
||||
%cst = constant dense<[2, 2]> : tensor<2xi32>
|
||||
%cst_0 = "tf.Const"() { value = dense<3> : tensor<i32> } : () -> tensor<i32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %arg1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68xf32>
|
||||
%cst_1 = constant dense<2> : tensor<2x2xi32>
|
||||
%0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_1) : (tensor<1x128x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68xf32>
|
||||
%1 = "tf.ExpandDims"(%0, %cst_0) : (tensor<4x68x68xf32>, tensor<i32>) -> tensor<4x68x68x1xf32>
|
||||
%2 = "tf.Conv2D"(%1, %arg2) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x64x64x1xf32>
|
||||
%2 = "tf.Conv2D"(%1, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<4x68x68x1xf32>, tensor<5x5x1x1xf32>) -> tensor<4x64x64x1xf32>
|
||||
%3 = "tf.Squeeze"(%2) {squeeze_dims = [2]} : (tensor<4x64x64x1xf32>) -> tensor<4x64x64x1xf32>
|
||||
%4 = "tf.BatchToSpaceND"(%3, %cst, %arg1) : (tensor<4x64x64x1xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x1xf32>
|
||||
%4 = "tf.BatchToSpaceND"(%3, %cst, %cst_1) : (tensor<4x64x64x1xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<1x128x128x1xf32>
|
||||
return %4 : tensor<1x128x128x1xf32>
|
||||
|
||||
// CHECK-LABEL: testDilatedConvWithDifferentExpandSqueezeAxis
|
||||
|
||||
@ -152,7 +152,6 @@ LogicalResult ConvertTFDilatedConvOp<Conv2dOpTy>::matchAndRewrite(
|
||||
}
|
||||
|
||||
// BatchToSpaceND + BiasAdd.
|
||||
// TODO(b/149936532): Check the `crops` input, currently ignored.
|
||||
TF::BatchToSpaceNDOp bts_op;
|
||||
TF::BiasAddOp biasadd_op;
|
||||
bool final_op_is_bts = true;
|
||||
@ -179,16 +178,50 @@ LogicalResult ConvertTFDilatedConvOp<Conv2dOpTy>::matchAndRewrite(
|
||||
if (!dilations_attr.hasValue()) return failure();
|
||||
op.setAttr("dilations", dilations_attr.getValue());
|
||||
|
||||
// Padding is set to 'SAME' when `stb_op` has non-zero paddings.
|
||||
// TODO(b/149936532): This assumption only holds when the input width & height
|
||||
// is multiple of dilation width & height. We should fix it in order to
|
||||
// support other use cases.
|
||||
// TODO(b/149936532): Check that the input width & height are multiples of
|
||||
// dilation rate.
|
||||
// TF python library will rewrite dilated conv to
|
||||
// "SpaceToBatch->Conv->BatchToSpace" pattern, and the Conv in the middle
|
||||
// always has 'VALID' padding. The padding tensor in `SpaceToBatch` has two
|
||||
// parts of contributions, one is to reduce padding of CONV from 'SAME' to
|
||||
// 'VALID', and another is to make input shape multiples of dilation rate. The
|
||||
// first part of padding, which is also called `base_padding` will be used
|
||||
// here to determine if the original padding format is 'SAME' or 'VALID'.
|
||||
// According to the following formula we will compute the `base_padding` if
|
||||
// it's a constant. Basically, `paddings` tensor in `SpaceToBatch` and `crops`
|
||||
// tensor in `BatchToSpace` must satisfy the following:
|
||||
// paddings[i, 0] = base_paddings[i, 0].
|
||||
// 0 <= paddings[i, 1] - base_paddings[i, 1] < block_shape[i]
|
||||
// (input_shape[i] + paddings[i, 0] + paddings[i, 1]) % block_shape[i] == 0.
|
||||
// crops[i, 0] = 0.
|
||||
// crops[i, 1] = paddings[i, 1] - base_paddings[i, 1].
|
||||
|
||||
// If `paddings` - `crops` != 0, this means that `base_paddings` != 0, which
|
||||
// tells us the original padding is 'SAME' (with one caveat presented below).
|
||||
// Here we need to reset the padding back to `SAME` if `base_padding`
|
||||
// != 0.
|
||||
// TODO(b/149936532): We might not simply rely on `paddings - crops != 0` to
|
||||
// determine the original padding format. For example, users can build
|
||||
// arbitrary valid examples of `STB->Conv->BTS` which doesn't represent a
|
||||
// dilated conv, hence we shouldn't pattern match here. Instead, we need to
|
||||
// check values of `paddings` and `crops` to make sure it really stands for
|
||||
// a dilated conv.
|
||||
auto stb_paddings = stb_op.paddings();
|
||||
ElementsAttr stb_paddings_attr;
|
||||
if (matchPattern(stb_paddings, m_Constant(&stb_paddings_attr))) {
|
||||
if (llvm::any_of(stb_paddings_attr.getValues<IntegerAttr>(),
|
||||
[](IntegerAttr attr) { return attr.getInt() != 0; })) {
|
||||
op.setAttr("padding", rewriter.getStringAttr("SAME"));
|
||||
auto bts_crops = bts_op.crops();
|
||||
ElementsAttr stb_paddings_attr, bts_crops_attr;
|
||||
if (matchPattern(stb_paddings, m_Constant(&stb_paddings_attr)) &&
|
||||
matchPattern(bts_crops, m_Constant(&bts_crops_attr))) {
|
||||
if (stb_paddings_attr.getNumElements() != bts_crops_attr.getNumElements())
|
||||
return failure();
|
||||
// padding - crop.
|
||||
auto paddings = stb_paddings_attr.getValues<IntegerAttr>();
|
||||
auto crops = bts_crops_attr.getValues<IntegerAttr>();
|
||||
for (auto it1 = paddings.begin(), it2 = crops.begin();
|
||||
it1 != paddings.end() && it2 != crops.end(); it1++, it2++) {
|
||||
if ((*it1).getInt() != (*it2).getInt()) {
|
||||
op.setAttr("padding", rewriter.getStringAttr("SAME"));
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
Loading…
Reference in New Issue
Block a user