diff --git a/csrc/backend_ops/onnxruntime/roi_align/roi_align.cpp b/csrc/backend_ops/onnxruntime/roi_align/roi_align.cpp deleted file mode 100644 index 78cd13c922..0000000000 --- a/csrc/backend_ops/onnxruntime/roi_align/roi_align.cpp +++ /dev/null @@ -1,255 +0,0 @@ -// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved. -// modified from -// https://github.com/facebookresearch/maskrcnn-benchmark/blob/main/maskrcnn_benchmark/csrc/cpu/ROIAlign_cpu.cpp -#include "roi_align.h" - -#include "ort_utils.h" - -namespace mmdeploy { -// implementation taken from Caffe2 -struct PreCalc { - int pos1; - int pos2; - int pos3; - int pos4; - float w1; - float w2; - float w3; - float w4; -}; - -void pre_calc_for_bilinear_interpolate(const int height, const int width, const int pooled_height, - const int pooled_width, const int iy_upper, - const int ix_upper, float roi_start_h, float roi_start_w, - float bin_size_h, float bin_size_w, int roi_bin_grid_h, - int roi_bin_grid_w, std::vector &pre_calc) { - int pre_calc_index = 0; - for (int ph = 0; ph < pooled_height; ph++) { - for (int pw = 0; pw < pooled_width; pw++) { - for (int iy = 0; iy < iy_upper; iy++) { - const float yy = roi_start_h + ph * bin_size_h + - static_cast(iy + .5f) * bin_size_h / - static_cast(roi_bin_grid_h); // e.g., 0.5, 1.5 - for (int ix = 0; ix < ix_upper; ix++) { - const float xx = - roi_start_w + pw * bin_size_w + - static_cast(ix + .5f) * bin_size_w / static_cast(roi_bin_grid_w); - - float x = xx; - float y = yy; - // deal with: inverse elements are out of feature map boundary - if (y < -1.0 || y > height || x < -1.0 || x > width) { - // empty - PreCalc pc; - pc.pos1 = 0; - pc.pos2 = 0; - pc.pos3 = 0; - pc.pos4 = 0; - pc.w1 = 0; - pc.w2 = 0; - pc.w3 = 0; - pc.w4 = 0; - pre_calc[pre_calc_index] = pc; - pre_calc_index += 1; - continue; - } - - if (y <= 0) { - y = 0; - } - if (x <= 0) { - x = 0; - } - - int y_low = (int)y; - int x_low = (int)x; - int y_high; - int x_high; - - if (y_low >= height - 1) { - y_high = y_low = height - 1; - y = (float)y_low; - } else { - y_high = y_low + 1; - } - - if (x_low >= width - 1) { - x_high = x_low = width - 1; - x = (float)x_low; - } else { - x_high = x_low + 1; - } - - float ly = y - y_low; - float lx = x - x_low; - float hy = 1. - ly, hx = 1. - lx; - float w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx; - - // save weights and indices - PreCalc pc; - pc.pos1 = y_low * width + x_low; - pc.pos2 = y_low * width + x_high; - pc.pos3 = y_high * width + x_low; - pc.pos4 = y_high * width + x_high; - pc.w1 = w1; - pc.w2 = w2; - pc.w3 = w3; - pc.w4 = w4; - pre_calc[pre_calc_index] = pc; - - pre_calc_index += 1; - } - } - } - } -} - -void ROIAlignForwardCPU(const int nthreads, const float *input, const float *rois, float *output, - float *argmax_y, float *argmax_x, const int pooled_height, - const int pooled_width, const float spatial_scale, const int sampling_ratio, - const int pool_mode, // 0 - max pool, 1 - avg pool - const bool aligned, const int channels, const int height, const int width) { - int n_rois = nthreads / channels / pooled_width / pooled_height; - // (n, c, ph, pw) is an element in the pooled output - // can be parallelized using omp - // #pragma omp parallel for num_threads(32) - for (int n = 0; n < n_rois; n++) { - int index_n = n * channels * pooled_width * pooled_height; - - const float *offset_rois = rois + n * 5; - int roi_batch_ind = offset_rois[0]; - - // Do not use rounding; this implementation detail is critical - float offset = aligned ? (float)0.5 : (float)0.0; - float roi_start_w = offset_rois[1] * spatial_scale - offset; - float roi_start_h = offset_rois[2] * spatial_scale - offset; - float roi_end_w = offset_rois[3] * spatial_scale - offset; - float roi_end_h = offset_rois[4] * spatial_scale - offset; - - float roi_width = roi_end_w - roi_start_w; - float roi_height = roi_end_h - roi_start_h; - if (aligned) { - /*AT_ASSERTM(roi_width >= 0 && roi_height >= 0, - "ROIs in ROIAlign cannot have non-negative size!");*/ - assert(roi_width >= 0 && roi_height >= 0); - } else { // for backward-compatibility only - roi_width = std::max(roi_width, (float)1.); - roi_height = std::max(roi_height, (float)1.); - } - float bin_size_h = static_cast(roi_height) / static_cast(pooled_height); - float bin_size_w = static_cast(roi_width) / static_cast(pooled_width); - - // We use roi_bin_grid to sample the grid and mimic integral - int roi_bin_grid_h = - (sampling_ratio > 0) ? sampling_ratio : ceil(roi_height / pooled_height); // e.g., = 2 - int roi_bin_grid_w = (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width); - - // When the grid is empty, output zeros == 0/1, instead of NaN. - const float count = std::max(roi_bin_grid_h * roi_bin_grid_w, 1); // e.g. = 4 - - // we want to precalculate indices and weights shared by all channels, - // this is the key point of optimization - std::vector pre_calc(roi_bin_grid_h * roi_bin_grid_w * pooled_width * pooled_height); - pre_calc_for_bilinear_interpolate(height, width, pooled_height, pooled_width, roi_bin_grid_h, - roi_bin_grid_w, roi_start_h, roi_start_w, bin_size_h, - bin_size_w, roi_bin_grid_h, roi_bin_grid_w, pre_calc); - - for (int c = 0; c < channels; c++) { - int index_n_c = index_n + c * pooled_width * pooled_height; - const float *offset_input = input + (roi_batch_ind * channels + c) * height * width; - int pre_calc_index = 0; - - for (int ph = 0; ph < pooled_height; ph++) { - for (int pw = 0; pw < pooled_width; pw++) { - int index = index_n_c + ph * pooled_width + pw; - - float output_val = 0.; - float maxval = -10000; - float maxidx_y = -1.f, maxidx_x = -1.f; - for (int iy = 0; iy < roi_bin_grid_h; iy++) { - const float y = - roi_start_h + ph * bin_size_h + - static_cast(iy + .5f) * bin_size_h / static_cast(roi_bin_grid_h); - for (int ix = 0; ix < roi_bin_grid_w; ix++) { - const float x = - roi_start_w + pw * bin_size_w + - static_cast(ix + .5f) * bin_size_w / static_cast(roi_bin_grid_w); - PreCalc pc = pre_calc[pre_calc_index]; - float val = pc.w1 * offset_input[pc.pos1] + pc.w2 * offset_input[pc.pos2] + - pc.w3 * offset_input[pc.pos3] + pc.w4 * offset_input[pc.pos4]; - if (val > maxval) { - maxval = val; - maxidx_y = y; - maxidx_x = x; - } - output_val += val; - pre_calc_index += 1; - } - } - if (pool_mode == 0) { - // We do max pooling inside a bin - output[index] = maxval; - argmax_y[index] = maxidx_y; - argmax_x[index] = maxidx_x; - } else if (pool_mode == 1) { - // We do average (integral) pooling inside a bin - output[index] = output_val / count; - } // if - } // for pw - } // for ph - } // for c - } // for n -} - -void MMCVRoiAlignKernel::Compute(OrtKernelContext *context) { - // Setup inputs - const OrtValue *input_X = ort_.KernelContext_GetInput(context, 0); - const float *X_data = reinterpret_cast(ort_.GetTensorData(input_X)); - const OrtValue *input_rois = ort_.KernelContext_GetInput(context, 1); - const float *rois = - reinterpret_cast(ort_.GetTensorData(input_rois)); - - // Setup output - OrtTensorDimensions out_dimensions(ort_, input_X); - OrtTensorDimensions roi_dimensions(ort_, input_rois); - - int batch_size = out_dimensions.data()[0]; - int input_channels = out_dimensions.data()[1]; - int input_height = out_dimensions.data()[2]; - int input_width = out_dimensions.data()[3]; - - out_dimensions.data()[0] = roi_dimensions.data()[0]; - out_dimensions.data()[2] = aligned_height_; - out_dimensions.data()[3] = aligned_width_; - - OrtValue *output = - ort_.KernelContext_GetOutput(context, 0, out_dimensions.data(), out_dimensions.size()); - float *out = ort_.GetTensorMutableData(output); - OrtTensorTypeAndShapeInfo *output_info = ort_.GetTensorTypeAndShape(output); - ort_.ReleaseTensorTypeAndShapeInfo(output_info); - - // TODO: forward here - int output_size = out_dimensions.data()[0]; - for (auto i = 1; i < out_dimensions.size(); ++i) { - output_size *= out_dimensions.data()[i]; - } - - int poolMod = 1; - if (pool_mode_ == "max") poolMod = 0; - - float *argmax_x = nullptr, *argmax_y = nullptr; - if (poolMod == 0) { - argmax_y = new float[output_size]; - argmax_x = new float[output_size]; - } - - ROIAlignForwardCPU(output_size, X_data, rois, out, argmax_y, argmax_x, aligned_height_, - aligned_width_, spatial_scale_, sampling_ratio_, poolMod, aligned_, - input_channels, input_height, input_width); - - if (argmax_x) delete argmax_x; - if (argmax_y) delete argmax_y; -} - -REGISTER_ONNXRUNTIME_OPS(mmdeploy, MMCVRoiAlignCustomOp); -} // namespace mmdeploy diff --git a/csrc/backend_ops/onnxruntime/roi_align/roi_align.h b/csrc/backend_ops/onnxruntime/roi_align/roi_align.h deleted file mode 100644 index 0c7afa67da..0000000000 --- a/csrc/backend_ops/onnxruntime/roi_align/roi_align.h +++ /dev/null @@ -1,59 +0,0 @@ -// Copyright (c) OpenMMLab. All rights reserved. -#ifndef ONNXRUNTIME_ROI_ALIGN_H -#define ONNXRUNTIME_ROI_ALIGN_H - -#include -#include - -#include -#include -#include -#include - -namespace mmdeploy { -struct MMCVRoiAlignKernel { - public: - MMCVRoiAlignKernel(Ort::CustomOpApi ort, const OrtKernelInfo* info) : ort_(ort) { - aligned_ = ort_.KernelInfoGetAttribute(info, "aligned"); - aligned_height_ = ort_.KernelInfoGetAttribute(info, "output_height"); - aligned_width_ = ort_.KernelInfoGetAttribute(info, "output_width"); - pool_mode_ = ort_.KernelInfoGetAttribute(info, "mode"); - sampling_ratio_ = ort_.KernelInfoGetAttribute(info, "sampling_ratio"); - spatial_scale_ = ort_.KernelInfoGetAttribute(info, "spatial_scale"); - } - - void Compute(OrtKernelContext* context); - - private: - Ort::CustomOpApi ort_; - - int aligned_height_; - int aligned_width_; - float spatial_scale_; - int sampling_ratio_; - std::string pool_mode_; - int aligned_; -}; - -struct MMCVRoiAlignCustomOp : Ort::CustomOpBase { - void* CreateKernel(Ort::CustomOpApi api, const OrtKernelInfo* info) const { - return new MMCVRoiAlignKernel(api, info); - } - const char* GetName() const { return "MMCVRoiAlign"; } - - size_t GetInputTypeCount() const { return 2; } - ONNXTensorElementDataType GetInputType(size_t) const { - return ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT; - } - - size_t GetOutputTypeCount() const { return 1; } - ONNXTensorElementDataType GetOutputType(size_t) const { - return ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT; - } - - // force cpu - const char* GetExecutionProviderType() const { return "CPUExecutionProvider"; } -}; -} // namespace mmdeploy - -#endif // ONNXRUNTIME_ROI_ALIGN_H diff --git a/docs/en/backends/onnxruntime.md b/docs/en/backends/onnxruntime.md index a47569d802..181e655094 100644 --- a/docs/en/backends/onnxruntime.md +++ b/docs/en/backends/onnxruntime.md @@ -57,7 +57,6 @@ make -j$(nproc) | Operator | CPU | GPU | MMDeploy Releases | | :--------------------------------------------------------------------------- | :---: | :---: | :---------------- | -| [RoIAlign](../ops/onnxruntime.md#roialign) | Y | N | master | | [grid_sampler](../ops/onnxruntime.md#grid_sampler) | Y | N | master | | [MMCVModulatedDeformConv2d](../ops/onnxruntime.md#mmcvmodulateddeformconv2d) | Y | N | master | diff --git a/docs/en/ops/onnxruntime.md b/docs/en/ops/onnxruntime.md index 2e4d741e0d..51791ebc9f 100644 --- a/docs/en/ops/onnxruntime.md +++ b/docs/en/ops/onnxruntime.md @@ -3,64 +3,21 @@ - [ONNX Runtime Ops](#onnx-runtime-ops) - - [RoIAlign](#roialign) + - [grid_sampler](#grid_sampler) - [Description](#description) - [Parameters](#parameters) - [Inputs](#inputs) - [Outputs](#outputs) - [Type Constraints](#type-constraints) - - [grid_sampler](#grid_sampler) + - [MMCVModulatedDeformConv2d](#mmcvmodulateddeformconv2d) - [Description](#description-1) - [Parameters](#parameters-1) - [Inputs](#inputs-1) - [Outputs](#outputs-1) - [Type Constraints](#type-constraints-1) - - [MMCVModulatedDeformConv2d](#mmcvmodulateddeformconv2d) - - [Description](#description-2) - - [Parameters](#parameters-2) - - [Inputs](#inputs-2) - - [Outputs](#outputs-2) - - [Type Constraints](#type-constraints-2) -### RoIAlign - -#### Description - -Perform RoIAlign on output feature, used in bbox_head of most two-stage detectors. - -#### Parameters - -| Type | Parameter | Description | -| ------- | ---------------- | ------------------------------------------------------------------------------------------------------------- | -| `int` | `output_height` | height of output roi | -| `int` | `output_width` | width of output roi | -| `float` | `spatial_scale` | used to scale the input boxes | -| `int` | `sampling_ratio` | number of input samples to take for each output sample. `0` means to take samples densely for current models. | -| `str` | `mode` | pooling mode in each bin. `avg` or `max` | -| `int` | `aligned` | If `aligned=0`, use the legacy implementation in MMDetection. Else, align the results more perfectly. | - -#### Inputs - -
-
input: T
-
Input feature map; 4D tensor of shape (N, C, H, W), where N is the batch size, C is the numbers of channels, H and W are the height and width of the data.
-
rois: T
-
RoIs (Regions of Interest) to pool over; 2-D tensor of shape (num_rois, 5) given as [[batch_index, x1, y1, x2, y2], ...]. The RoIs' coordinates are the coordinate system of input.
-
- -#### Outputs - -
-
feat: T
-
RoI pooled output, 4-D tensor of shape (num_rois, C, output_height, output_width). The r-th batch element feat[r-1] is a pooled feature map corresponding to the r-th RoI RoIs[r-1].
-
- -#### Type Constraints - -- T:tensor(float32) - ### grid_sampler #### Description diff --git a/mmdeploy/codebase/mmdet/models/roi_heads/single_level_roi_extractor.py b/mmdeploy/codebase/mmdet/models/roi_heads/single_level_roi_extractor.py index de0ea22dad..7c262c1dca 100644 --- a/mmdeploy/codebase/mmdet/models/roi_heads/single_level_roi_extractor.py +++ b/mmdeploy/codebase/mmdet/models/roi_heads/single_level_roi_extractor.py @@ -137,14 +137,13 @@ def single_roi_extractor__forward(ctx, device=target_lvls.device) target_lvls = torch.cat((_tmp, _tmp, target_lvls)) for i in range(num_levels): - # use the roi align in torhcvision to accelerate the inference - # roi_align in MMCV is same as torchvision when pool mode is 'avg' - if backend == Backend.TORCHSCRIPT or self.roi_layers[ - i].pool_mode == 'avg': - self.roi_layers[i].use_torchvision = True mask = target_lvls == i inds = mask.nonzero(as_tuple=False).squeeze(1) - roi_feats_t = self.roi_layers[i](feats[i], rois[inds]) + rois_t = rois[inds] + # use the roi align in torhcvision + if backend == Backend.TORCHSCRIPT: + self.roi_layers[i].use_torchvision = True + roi_feats_t = self.roi_layers[i](feats[i], rois_t) roi_feats[inds] = roi_feats_t # slice to recover original size roi_feats = roi_feats[num_levels * 2:] diff --git a/mmdeploy/mmcv/ops/roi_align.py b/mmdeploy/mmcv/ops/roi_align.py index c6da740fe5..33cd7342d5 100644 --- a/mmdeploy/mmcv/ops/roi_align.py +++ b/mmdeploy/mmcv/ops/roi_align.py @@ -1,10 +1,11 @@ # Copyright (c) OpenMMLab. All rights reserved. from typing import List +import torch from torch import Tensor from mmdeploy.core import SYMBOLIC_REWRITER -from mmdeploy.utils import Backend, get_backend +from mmdeploy.utils import Backend, get_backend, get_ir_config # Here using mmcv.ops.roi_align.__self__ to find @@ -17,7 +18,10 @@ def roi_align_default(ctx, g, input: Tensor, rois: Tensor, sampling_ratio: int, pool_mode: str, aligned: bool): """Rewrite symbolic function for default backend. - Replace onnx::RoiAlign with mmdeploy::MMCVRoiAlign. + Replace onnx::RoiAlign with mmcv::MMCVRoiAlign for PPLNN. For ONNXRuntime, + align operation get done outside the inference engine for opset versions + lower than 16. By default, onnx::RoiAlign get replaced to + mmdeploy::MMCVRoiAlign. Args: ctx (ContextCaller): The context with additional information. @@ -40,6 +44,58 @@ def roi_align_default(ctx, g, input: Tensor, rois: Tensor, backend = get_backend(ctx.cfg) if backend == Backend.PPLNN: domain = 'mmcv' + elif backend == Backend.ONNXRUNTIME: + from torch.onnx.symbolic_opset9 import _cast_Long + from torch.onnx.symbolic_opset11 import add, select, squeeze + batch_indices = _cast_Long( + g, + squeeze( + g, + select( + g, rois, 1, + g.op( + 'Constant', + value_t=torch.tensor([0], dtype=torch.long))), 1), + False) + rois = select( + g, rois, 1, + g.op( + 'Constant', + value_t=torch.tensor([1, 2, 3, 4], dtype=torch.long))) + ir_cfg = get_ir_config(ctx.cfg) + opset_version = ir_cfg.get('opset_version', 11) + if opset_version < 16: + # preprocess rois to make compatible with opset 16- + # as for opset 16+, `aligned` get implemented inside onnxruntime. + if aligned is True: + rois = add( + g, rois, + g.op( + 'Constant', + value_t=torch.tensor([-0.5 / spatial_scale], + dtype=torch.float))) + return g.op( + 'RoiAlign', + input, + rois, + batch_indices, + output_height_i=output_size[0], + output_width_i=output_size[1], + spatial_scale_f=spatial_scale, + sampling_ratio_i=sampling_ratio, + mode_s=pool_mode) + else: + return g.op( + 'RoiAlign', + input, + rois, + batch_indices, + output_height_i=output_size[0], + output_width_i=output_size[1], + spatial_scale_f=spatial_scale, + sampling_ratio_i=sampling_ratio, + mode_s=pool_mode, + aligned_i=aligned) else: domain = 'mmdeploy' return g.op( diff --git a/tests/test_codebase/test_mmdet/test_mmdet_models.py b/tests/test_codebase/test_mmdet/test_mmdet_models.py index 706cc05be6..b3a1a94173 100644 --- a/tests/test_codebase/test_mmdet/test_mmdet_models.py +++ b/tests/test_codebase/test_mmdet/test_mmdet_models.py @@ -546,7 +546,6 @@ def test_single_roi_extractor(backend_type: Backend): wrapped_model=single_roi_extractor, model_inputs=model_inputs, deploy_cfg=deploy_cfg) - if isinstance(backend_outputs, dict): backend_outputs = backend_outputs.values() for model_output, backend_output in zip(model_outputs[0], backend_outputs): diff --git a/tests/test_ops/test_ops.py b/tests/test_ops/test_ops.py index 7a5784b0ac..72ff750d1e 100644 --- a/tests/test_ops/test_ops.py +++ b/tests/test_ops/test_ops.py @@ -16,7 +16,7 @@ TEST_NCNN = TestNCNNExporter() -@pytest.mark.parametrize('backend', [TEST_ONNXRT, TEST_TENSORRT]) +@pytest.mark.parametrize('backend', [TEST_TENSORRT]) @pytest.mark.parametrize('pool_h,pool_w,spatial_scale,sampling_ratio', [(2, 2, 1.0, 2), (4, 4, 2.0, 4)]) def test_roi_align(backend,