Support onnxruntime fp16 (open-mmlab#2269)

* support ort-fp16 * update configs * update * update reg ci * fix mmrotate mmdet3d ort fp16 * fix dead links
zhjunqin · Jul 21, 2023 · 64e96e3 · 64e96e3
1 parent 553f9b8
commit 64e96e3
Show file tree

Hide file tree

Showing 29 changed files with 191 additions and 16 deletions.
diff --git a/.github/workflows/regression-test.yml b/.github/workflows/regression-test.yml
@@ -111,7 +111,7 @@ jobs:
           apt update && apt install unzip
           python -V
           python -m pip install --upgrade pip
-          python -m pip install openmim numpy pycuda xlsxwriter packaging prettytable
+          python -m pip install openmim numpy pycuda xlsxwriter packaging prettytable onnxconverter-common
           python -m pip install opencv-python==4.5.4.60 opencv-python-headless==4.5.4.60 opencv-contrib-python==4.5.4.60
           python .github/scripts/prepare_reg_test.py --torch-version ${{ matrix.torch_version }} --codebases ${{ matrix.codebase}}
           python -m pip install -r requirements.txt
@@ -221,7 +221,7 @@ jobs:
           conda activate $env:TEMP_ENV
           python -V
           python -m pip install --upgrade pip
-          python -m pip install openmim numpy pycuda xlsxwriter packaging prettytable
+          python -m pip install openmim numpy pycuda xlsxwriter packaging prettytable onnxconverter-common
           python -m pip install opencv-python==4.5.4.60 opencv-python-headless==4.5.4.60 opencv-contrib-python==4.5.4.60
           python .github/scripts/prepare_reg_test.py --torch-version ${{ matrix.torch_version }} --codebases ${{ matrix.codebase}}
           python -m pip install -r requirements.txt

diff --git a/configs/_base_/backends/onnxruntime-fp16.py b/configs/_base_/backends/onnxruntime-fp16.py
@@ -0,0 +1,10 @@
+backend_config = dict(
+    type='onnxruntime',
+    precision='fp16',
+    common_config=dict(
+        min_positive_val=1e-7,
+        max_finite_val=1e4,
+        keep_io_types=False,
+        disable_shape_infer=False,
+        op_block_list=None,
+        node_block_list=None))
diff --git a/configs/mmaction/video-recognition/video-recognition_onnxruntime-fp16_static.py b/configs/mmaction/video-recognition/video-recognition_onnxruntime-fp16_static.py
@@ -0,0 +1,6 @@
+_base_ = [
+    './video-recognition_static.py',
+    '../../_base_/backends/onnxruntime-fp16.py'
+]
+
+onnx_config = dict(input_shape=None)
diff --git a/configs/mmagic/super-resolution/super-resolution_onnxruntime-fp16_dynamic.py b/configs/mmagic/super-resolution/super-resolution_onnxruntime-fp16_dynamic.py
@@ -0,0 +1,4 @@
+_base_ = [
+    './super-resolution_dynamic.py',
+    '../../_base_/backends/onnxruntime-fp16.py'
+]
diff --git a/configs/mmdet/detection/detection_onnxruntime-fp16_dynamic.py b/configs/mmdet/detection/detection_onnxruntime-fp16_dynamic.py
@@ -0,0 +1,3 @@
+_base_ = [
+    '../_base_/base_dynamic.py', '../../_base_/backends/onnxruntime-fp16.py'
+]
diff --git a/configs/mmdet/instance-seg/instance-seg_onnxruntime-fp16_dynamic.py b/configs/mmdet/instance-seg/instance-seg_onnxruntime-fp16_dynamic.py
@@ -0,0 +1,4 @@
+_base_ = [
+    '../_base_/base_instance-seg_dynamic.py',
+    '../../_base_/backends/onnxruntime-fp16.py'
+]
diff --git a/configs/mmdet3d/voxel-detection/voxel-detection_onnxruntime-fp16_dynamic.py b/configs/mmdet3d/voxel-detection/voxel-detection_onnxruntime-fp16_dynamic.py
@@ -0,0 +1,3 @@
+_base_ = [
+    './voxel-detection_dynamic.py', '../../_base_/backends/onnxruntime-fp16.py'
+]
diff --git a/configs/mmocr/text-detection/text-detection_onnxruntime-fp16_dynamic.py b/configs/mmocr/text-detection/text-detection_onnxruntime-fp16_dynamic.py
@@ -0,0 +1,3 @@
+_base_ = [
+    './text-detection_dynamic.py', '../../_base_/backends/onnxruntime-fp16.py'
+]
diff --git a/configs/mmocr/text-recognition/text-recognition_onnxruntime-fp16_dynamic.py b/configs/mmocr/text-recognition/text-recognition_onnxruntime-fp16_dynamic.py
@@ -0,0 +1,4 @@
+_base_ = [
+    './text-recognition_dynamic.py',
+    '../../_base_/backends/onnxruntime-fp16.py'
+]
diff --git a/configs/mmpose/pose-detection_onnxruntime-fp16_static.py b/configs/mmpose/pose-detection_onnxruntime-fp16_static.py
@@ -0,0 +1,3 @@
+_base_ = [
+    './pose-detection_static.py', '../_base_/backends/onnxruntime-fp16.py'
+]
diff --git a/configs/mmpose/pose-detection_simcc_onnxruntime-fp16_dynamic.py b/configs/mmpose/pose-detection_simcc_onnxruntime-fp16_dynamic.py
@@ -0,0 +1,18 @@
+_base_ = [
+    './pose-detection_static.py', '../_base_/backends/onnxruntime-fp16.py'
+]
+
+onnx_config = dict(
+    input_shape=[192, 256],
+    output_names=['simcc_x', 'simcc_y'],
+    dynamic_axes={
+        'input': {
+            0: 'batch',
+        },
+        'simcc_x': {
+            0: 'batch'
+        },
+        'simcc_y': {
+            0: 'batch'
+        }
+    })
diff --git a/configs/mmpretrain/classification_onnxruntime-fp16_dynamic.py b/configs/mmpretrain/classification_onnxruntime-fp16_dynamic.py
@@ -0,0 +1,3 @@
+_base_ = [
+    './classification_dynamic.py', '../_base_/backends/onnxruntime-fp16.py'
+]
diff --git a/configs/mmrotate/rotated-detection_onnxruntime-fp16_dynamic.py b/configs/mmrotate/rotated-detection_onnxruntime-fp16_dynamic.py
@@ -0,0 +1,25 @@
+_base_ = [
+    './rotated-detection_static.py', '../_base_/backends/onnxruntime-fp16.py'
+]
+
+onnx_config = dict(
+    output_names=['dets', 'labels'],
+    input_shape=[1024, 1024],
+    dynamic_axes={
+        'input': {
+            0: 'batch',
+            2: 'height',
+            3: 'width'
+        },
+        'dets': {
+            0: 'batch',
+            1: 'num_dets',
+        },
+        'labels': {
+            0: 'batch',
+            1: 'num_dets',
+        },
+    })
+
+backend_config = dict(
+    common_config=dict(op_block_list=['NMSRotated', 'Resize']))
diff --git a/configs/mmseg/segmentation_onnxruntime-fp16_dynamic.py b/configs/mmseg/segmentation_onnxruntime-fp16_dynamic.py
@@ -0,0 +1,3 @@
+_base_ = [
+    './segmentation_dynamic.py', '../_base_/backends/onnxruntime-fp16.py'
+]
diff --git a/docs/en/05-supported-backends/onnxruntime.md b/docs/en/05-supported-backends/onnxruntime.md
@@ -22,6 +22,14 @@ pip install onnxruntime==1.8.1 # if you want to use cpu version
 pip install onnxruntime-gpu==1.8.1 # if you want to use gpu version
 ```
 
+### Install float16 conversion tool (optional)
+
+If you want to use float16 precision, install the tool by running the following script:
+
+```bash
+pip install onnx onnxconverter-common
+```
+
 ## Build custom ops
 
 ### Download ONNXRuntime Library

diff --git a/docs/en/05-supported-backends/openvino.md b/docs/en/05-supported-backends/openvino.md
@@ -83,7 +83,7 @@ Notes:
 
 - Custom operations from OpenVINO use the domain `org.openvinotoolkit`.
 - For faster work in OpenVINO in the Faster-RCNN, Mask-RCNN, Cascade-RCNN, Cascade-Mask-RCNN models
-  the RoiAlign operation is replaced with the [ExperimentalDetectronROIFeatureExtractor](https://docs.openvinotoolkit.org/latest/openvino_docs_ops_detection_ExperimentalDetectronROIFeatureExtractor_6.html) operation in the ONNX graph.
+  the RoiAlign operation is replaced with the [ExperimentalDetectronROIFeatureExtractor](https://docs.openvino.ai/2022.3/openvino_docs_ops_detection_ExperimentalDetectronROIFeatureExtractor_6.html) operation in the ONNX graph.
 - Models "VFNet" and "Faster R-CNN + DCN" use the custom "DeformableConv2D" operation.
 
 ## Deployment config

diff --git a/docs/zh_cn/05-supported-backends/onnxruntime.md b/docs/zh_cn/05-supported-backends/onnxruntime.md
@@ -22,6 +22,14 @@ pip install onnxruntime==1.8.1 # if you want to use cpu version
 pip install onnxruntime-gpu==1.8.1 # if you want to use gpu version
 ```
 
+### Install float16 conversion tool (optional)
+
+If you want to use float16 precision, install the tool by running the following script:
+
+```bash
+pip install onnx onnxconverter-common
+```
+
 ## Build custom ops
 
 ### Download ONNXRuntime Library

diff --git a/docs/zh_cn/05-supported-backends/openvino.md b/docs/zh_cn/05-supported-backends/openvino.md
@@ -83,7 +83,7 @@ Notes:
 
 - Custom operations from OpenVINO use the domain `org.openvinotoolkit`.
 - For faster work in OpenVINO in the Faster-RCNN, Mask-RCNN, Cascade-RCNN, Cascade-Mask-RCNN models
-  the RoiAlign operation is replaced with the [ExperimentalDetectronROIFeatureExtractor](https://docs.openvinotoolkit.org/latest/openvino_docs_ops_detection_ExperimentalDetectronROIFeatureExtractor_6.html) operation in the ONNX graph.
+  the RoiAlign operation is replaced with the [ExperimentalDetectronROIFeatureExtractor](https://docs.openvino.ai/2022.3/openvino_docs_ops_detection_ExperimentalDetectronROIFeatureExtractor_6.html) operation in the ONNX graph.
 - Models "VFNet" and "Faster R-CNN + DCN" use the custom "DeformableConv2D" operation.
 
 ## Deployment config

diff --git a/mmdeploy/backend/onnxruntime/backend_manager.py b/mmdeploy/backend/onnxruntime/backend_manager.py
@@ -3,6 +3,7 @@
 import os.path as osp
 from typing import Any, Callable, Optional, Sequence
 
+from mmdeploy.utils import get_backend_config, get_common_config
 from ..base import BACKEND_MANAGERS, BaseBackendManager
 
 
@@ -125,6 +126,7 @@ def check_env(cls, log_callback: Callable = lambda _: _) -> str:
     def to_backend(cls,
                    ir_files: Sequence[str],
                    work_dir: str,
+                   deploy_cfg: Any,
                    log_level: int = logging.INFO,
                    device: str = 'cpu',
                    **kwargs) -> Sequence[str]:
@@ -134,9 +136,21 @@ def to_backend(cls,
             ir_files (Sequence[str]): The intermediate representation files.
             work_dir (str): The work directory, backend files and logs should
                 be saved in this directory.
+            deploy_cfg (Any): The deploy config.
             log_level (int, optional): The log level. Defaults to logging.INFO.
             device (str, optional): The device type. Defaults to 'cpu'.
         Returns:
             Sequence[str]: Backend files.
         """
+        backend_cfg = get_backend_config(deploy_cfg)
+
+        precision = backend_cfg.get('precision', 'fp32')
+        if precision == 'fp16':
+            import onnx
+            from onnxconverter_common import float16
+
+            common_cfg = get_common_config(deploy_cfg)
+            model = onnx.load(ir_files[0])
+            model_fp16 = float16.convert_float_to_float16(model, **common_cfg)
+            onnx.save(model_fp16, ir_files[0])
         return ir_files
diff --git a/mmdeploy/backend/onnxruntime/wrapper.py b/mmdeploy/backend/onnxruntime/wrapper.py
@@ -2,6 +2,7 @@
 import os.path as osp
 from typing import Dict, Optional, Sequence
 
+import numpy as np
 import onnxruntime as ort
 import torch
 
@@ -58,6 +59,7 @@ def __init__(self,
         if output_names is None:
             output_names = [_.name for _ in sess.get_outputs()]
         self.sess = sess
+        self._input_metas = {_.name: _ for _ in sess.get_inputs()}
         self.io_binding = sess.io_binding()
         self.device_id = device_id
         self.device_type = 'cpu' if device == 'cpu' else 'cuda'
@@ -75,6 +77,9 @@ def forward(self, inputs: Dict[str,
         """
         for name, input_tensor in inputs.items():
             # set io binding for inputs/outputs
+            input_type = self._input_metas[name].type
+            if 'float16' in input_type:
+                input_tensor = input_tensor.to(torch.float16)
             input_tensor = input_tensor.contiguous()
             if self.device_type == 'cpu':
                 input_tensor = input_tensor.cpu()
@@ -98,6 +103,8 @@ def forward(self, inputs: Dict[str,
         output_list = self.io_binding.copy_outputs_to_cpu()
         outputs = {}
         for output_name, numpy_tensor in zip(self._output_names, output_list):
+            if numpy_tensor.dtype == np.float16:
+                numpy_tensor = numpy_tensor.astype(np.float32)
             outputs[output_name] = torch.from_numpy(numpy_tensor)
 
         return outputs

diff --git a/tests/regression/mmaction.yml b/tests/regression/mmaction.yml
@@ -28,6 +28,10 @@ onnxruntime:
     convert_image: *convert_image
     deploy_config: configs/mmaction/video-recognition/video-recognition_onnxruntime_static.py
     backend_test: *default_backend_test
+  pipeline_ort_static_fp16: &pipeline_ort_static_fp16
+    convert_image: *convert_image
+    deploy_config: configs/mmaction/video-recognition/video-recognition_onnxruntime-fp16_static.py
+    backend_test: *default_backend_test
 
 torchscript:
   pipeline_torchscript_fp32: &pipeline_torchscript_fp32
@@ -51,7 +55,7 @@ models:
     model_configs:
       - configs/recognition/tsn/tsn_imagenet-pretrained-r50_8xb32-1x1x3-100e_kinetics400-rgb.py
     pipelines:
-      - *pipeline_ort_static_fp32
+      - *pipeline_ort_static_fp16
       - *pipeline_trt_2d_static_fp32
       - *pipeline_torchscript_fp32
 

diff --git a/tests/regression/mmagic.yml b/tests/regression/mmagic.yml
@@ -46,6 +46,11 @@ onnxruntime:
     convert_image: *convert_image
     deploy_config: configs/mmagic/super-resolution/super-resolution_onnxruntime_dynamic.py
 
+  pipeline_ort_dynamic_fp16: &pipeline_ort_dynamic_fp16
+    convert_image: *convert_image
+    deploy_config: configs/mmagic/super-resolution/super-resolution_onnxruntime-fp16_dynamic.py
+
+
 tensorrt:
   pipeline_trt_static_fp32: &pipeline_trt_static_fp32
     convert_image: *convert_image
@@ -114,7 +119,7 @@ models:
       - configs/srcnn/srcnn_x4k915_1xb16-1000k_div2k.py
     pipelines:
       - *pipeline_ts_fp32
-      - *pipeline_ort_dynamic_fp32
+      - *pipeline_ort_dynamic_fp16
 #      - *pipeline_trt_dynamic_fp32
       - *pipeline_trt_dynamic_fp16
 #      - *pipeline_trt_dynamic_int8

diff --git a/tests/regression/mmdet.yml b/tests/regression/mmdet.yml
@@ -38,6 +38,11 @@ onnxruntime:
     backend_test: False
     deploy_config: configs/mmdet/detection/detection_onnxruntime_dynamic.py
 
+  pipeline_ort_dynamic_fp16: &pipeline_ort_dynamic_fp16
+    convert_image: *convert_image
+    backend_test: False
+    deploy_config: configs/mmdet/detection/detection_onnxruntime-fp16_dynamic.py
+
   pipeline_seg_ort_static_fp32: &pipeline_seg_ort_static_fp32
     convert_image: *convert_image
     backend_test: False
@@ -48,6 +53,11 @@ onnxruntime:
     backend_test: False
     deploy_config: configs/mmdet/instance-seg/instance-seg_onnxruntime_dynamic.py
 
+  pipeline_seg_ort_dynamic_fp16: &pipeline_seg_ort_dynamic_fp16
+    convert_image: *convert_image
+    backend_test: False
+    deploy_config: configs/mmdet/instance-seg/instance-seg_onnxruntime-fp16_dynamic.py
+
 tensorrt:
   pipeline_trt_static_fp32: &pipeline_trt_static_fp32
     convert_image: *convert_image
@@ -203,7 +213,7 @@ models:
       - configs/retinanet/retinanet_r50_fpn_1x_coco.py
     pipelines:
       - *pipeline_ts_fp32
-      - *pipeline_ort_dynamic_fp32
+      - *pipeline_ort_dynamic_fp16
       - *pipeline_trt_dynamic_fp32
       - *pipeline_ncnn_static_fp32
       - *pipeline_pplnn_dynamic_fp32
@@ -323,7 +333,7 @@ models:
       - configs/mask_rcnn/mask-rcnn_r50_fpn_1x_coco.py
     pipelines:
       - *pipeline_seg_ts_fp32
-      - *pipeline_seg_ort_dynamic_fp32
+      - *pipeline_seg_ort_dynamic_fp16
       - *pipeline_seg_trt_dynamic_fp32
       - *pipeline_seg_openvino_dynamic_fp32
 

diff --git a/tests/regression/mmdet3d.yml b/tests/regression/mmdet3d.yml
@@ -42,6 +42,11 @@ onnxruntime:
     backend_test: False
     deploy_config: configs/mmdet3d/voxel-detection/voxel-detection_onnxruntime_dynamic.py
 
+  pipeline_ort_dynamic_kitti_fp16: &pipeline_ort_dynamic_kitti_fp16
+    convert_image: *convert_image
+    backend_test: False
+    deploy_config: configs/mmdet3d/voxel-detection/voxel-detection_onnxruntime-fp16_dynamic.py
+
   pipeline_ort_dynamic_nus_fp32: &pipeline_ort_dynamic_nus_fp32
     convert_image: *convert_image_nus
     backend_test: False
@@ -86,7 +91,7 @@ models:
       - configs/pointpillars/pointpillars_hv_secfpn_8xb6-160e_kitti-3d-3class.py
       - configs/pointpillars/pointpillars_hv_secfpn_8xb6-160e_kitti-3d-car.py
     pipelines:
-      - *pipeline_ort_dynamic_kitti_fp32
+      - *pipeline_ort_dynamic_kitti_fp16
       - *pipeline_openvino_dynamic_kitti_fp32
       - *pipeline_trt_dynamic_kitti_fp32
   - name: PointPillars

diff --git a/tests/regression/mmocr.yml b/tests/regression/mmocr.yml
@@ -43,6 +43,10 @@ onnxruntime:
     convert_image: *convert_image_det
     deploy_config: configs/mmocr/text-detection/text-detection_onnxruntime_dynamic.py
 
+  pipeline_ort_detection_dynamic_fp16: &pipeline_ort_detection_dynamic_fp16
+    convert_image: *convert_image_det
+    deploy_config: configs/mmocr/text-detection/text-detection_onnxruntime-fp16_dynamic.py
+
   pipeline_ort_detection_mrcnn_dynamic_fp32: &pipeline_ort_detection_mrcnn_dynamic_fp32
     convert_image: *convert_image_det
     deploy_config: configs/mmocr/text-detection/text-detection_mrcnn_onnxruntime_dynamic.py
@@ -56,6 +60,11 @@ onnxruntime:
     convert_image: *convert_image_rec
     deploy_config: configs/mmocr/text-recognition/text-recognition_onnxruntime_dynamic.py
 
+  pipeline_ort_recognition_dynamic_fp16: &pipeline_ort_recognition_dynamic_fp16
+    convert_image: *convert_image_rec
+    deploy_config: configs/mmocr/text-recognition/text-recognition_onnxruntime-fp16_dynamic.py
+
+
 tensorrt:
   # ======= detection =======
   pipeline_trt_detection_static_fp32: &pipeline_trt_detection_static_fp32
@@ -239,7 +248,7 @@ models:
       - configs/textdet/dbnet/dbnet_resnet18_fpnc_1200e_icdar2015.py
     pipelines:
       - *pipeline_ts_detection_fp32
-      - *pipeline_ort_detection_dynamic_fp32
+      - *pipeline_ort_detection_dynamic_fp16
       - *pipeline_trt_detection_dynamic_fp16
       - *pipeline_ncnn_detection_static_fp32
       - *pipeline_pplnn_detection_dynamic_fp32
@@ -303,7 +312,7 @@ models:
       - configs/textrecog/crnn/crnn_mini-vgg_5e_mj.py
     pipelines:
       - *pipeline_ts_recognition_fp32
-      - *pipeline_ort_recognition_dynamic_fp32
+      - *pipeline_ort_recognition_dynamic_fp16
       - *pipeline_trt_recognition_dynamic_fp16_H32_C1
       - *pipeline_ncnn_recognition_static_fp32
       - *pipeline_pplnn_recognition_dynamic_fp32