- Sort Score
- Result 10 results
- Languages All
Results 1 - 10 of 80 for Convolution (0.33 sec)
-
tensorflow/compiler/mlir/tensorflow/g3doc/space_to_depth.md
speedup and reduce memory usage in the first convolution. The first convolution in many image models, including ResNet or ResNet-like, is a (kernel=7, stride=2) 2D convolution. The input of the convolution is images, which usually has RGB channels. The input of this first convolution is of shape [batch\_size, height, width, 3] and the kernel size is [kernel\_size, kernel\_size, 3, out\_channel]. Space to depth is to transform this first
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Sat Oct 24 02:51:43 UTC 2020 - 8.3K bytes - Viewed (0) -
tensorflow/compiler/mlir/quantization/stablehlo/tests/passes/nchw_convolution_to_nhwc.mlir
%2 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, 0, 1, f]x[o, i, 0, 1]->[b, f, 0, 1], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x4x4x8xf32>, tensor<8x8x3x3xf32>) -> tensor<1x8x4x4xf32> return %2 : tensor<1x8x4x4xf32> } // CHECK-NOT: stablehlo.transpose // CHECK: %[[CONV:.+]] = stablehlo.convolution
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Mon Mar 25 23:00:47 UTC 2024 - 5.5K bytes - Viewed (0) -
tensorflow/compiler/mlir/quantization/stablehlo/tests/pipelines/process_nchw_tensor.mlir
// CHECK: return %[[TRANSPOSE_1]] // ----- // Tests that a `add(convolution(%activation, %weight), %bias)` pattern with the // activation tensor of NCHW format and non-constant bias is converted to NHWC // convolution, but without the deferred transpose for `stablehlo.add`. // Transpose ops are inserted to the activation and output of // `stablehlo.convolution`. The weight constants is transposed.
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Thu Apr 18 20:32:46 UTC 2024 - 12.6K bytes - Viewed (0) -
tensorflow/compiler/mlir/quantization/stablehlo/tests/bridge/optimize.mlir
) -> tensor<?x2x2x1xi32> { // CHECK-DAG: %[[conv:.*]] = mhlo.convolution // CHECK-DAG: %[[combined:.*]] = chlo.broadcast_add %[[zp_offset:.*]], %[[bias:.*]] // CHECK-DAG: %[[result:.*]] = chlo.broadcast_add %[[conv]], %[[combined]] // CHECK: return %[[result]] %0 = mhlo.convolution(%lhs, %rhs) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f],
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Sat Feb 24 02:26:47 UTC 2024 - 10.7K bytes - Viewed (0) -
tensorflow/compiler/mlir/tensorflow/transforms/tpu_space_to_depth_pass.cc
// Iterate through block argument and its convolution users. Space to depth // transform will be applied only if all the below conditions are satisfied: // 1. All the users of the block argument will lead to convolutions; // 2. block_size of for the space to depth transform for these convolutions // are the same; // 3. block_size of for the space to depth transform for these convolutions // are larger than 1.
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Thu Apr 25 16:01:03 UTC 2024 - 29.3K bytes - Viewed (0) -
tensorflow/compiler/mlir/lite/stablehlo/tests/fuse_mhlo_convolution.mlir
// RUN: odml-to-stablehlo-opt %s -fuse-mhlo-convolution-pass -cse | FileCheck %s // CHECK-LABEL: @fuseMulAndConv2D // CHECK-SAME: %[[INPUT:[^:[:space:]]+]] func.func @fuseMulAndConv2D(%input: tensor<1x256x256x3xf32>) -> (tensor<1x256x256x2xf32>) { // CHECK-DAG: %[[FILTER:.+]] = mhlo.constant dense<{{\[\[\[\[}}1.000000e+00, 2.000000e+00], [3.000000e+00, 4.000000e+00], [5.000000e+00, 6.000000e+00]]]]> : tensor<1x1x3x2xf32>
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Sat Apr 06 15:32:52 UTC 2024 - 4.4K bytes - Viewed (0) -
tensorflow/compiler/mlir/quantization/stablehlo/tests/passes/lift_quantizable_spots_as_functions.mlir
%3 = stablehlo.constant dense<6.000000e+00> : tensor<f32> %4 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<?x28x28x1xf32>, tensor<3x3x1x16xf32>) -> tensor<?x28x28x16xf32>
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Fri May 10 04:07:09 UTC 2024 - 49.8K bytes - Viewed (0) -
tensorflow/compiler/mlir/lite/stablehlo/tests/legalize-tfl-stablehlo-conv.mlir
module { func.func @main(%arg0: tensor<8x8x1x207xf32>, %arg1: tensor<3x3x16x207xf32>) -> tensor<16x8x8x1xf32> {
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Wed Jan 24 06:08:43 UTC 2024 - 1.6K bytes - Viewed (0) -
tensorflow/compiler/mlir/lite/stablehlo/transforms/uniform_quantized_stablehlo_to_tfl_pass.cc
rewriter.replaceAllUsesExcept(rhs, dq.getOutput(), dq); } }; // Splits hybrid quantized `stablehlo.convolution` into `tfl.dequantize` and // float `stablehlo.convolution` op. Weight tensor is transposed to match the // filter tensor format for TFLite convolution. // Legalization of float `stablehlo.convolution` op relies on existing passes // for conversion of StableHLO -> MHLO -> TF -> TFL.
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Mon Apr 22 09:00:19 UTC 2024 - 99.8K bytes - Viewed (0) -
tensorflow/compiler/mlir/quantization/stablehlo/tests/passes/quantize/quantize_weight_only.mlir
Registered: Sun Jun 16 05:45:23 UTC 2024 - Last Modified: Tue May 14 17:10:32 UTC 2024 - 4.8K bytes - Viewed (0)