From 8873a70c51af296f65506810ae3a53da0a1d89ca Mon Sep 17 00:00:00 2001 From: jm12138 <2286040843@qq.com> Date: Fri, 16 Sep 2022 15:23:10 +0800 Subject: [PATCH] update humanseg_lite (#2000) * update humanseg_lite * add clean func * update save inference model --- .../humanseg_lite/README.md | 20 ++- .../humanseg_lite/README_en.md | 23 +-- .../humanseg_lite/data_feed.py | 1 - .../humanseg_lite/module.py | 115 +++++++------- .../humanseg_lite/processor.py | 1 - .../humanseg_lite/test.py | 145 ++++++++++++++++++ 6 files changed, 231 insertions(+), 74 deletions(-) create mode 100644 modules/image/semantic_segmentation/humanseg_lite/test.py diff --git a/modules/image/semantic_segmentation/humanseg_lite/README.md b/modules/image/semantic_segmentation/humanseg_lite/README.md index 67472e18..12248ce8 100644 --- a/modules/image/semantic_segmentation/humanseg_lite/README.md +++ b/modules/image/semantic_segmentation/humanseg_lite/README.md @@ -170,19 +170,13 @@ ```python - def save_inference_model(dirname='humanseg_lite_model', - model_filename=None, - params_filename=None, - combined=True) + def save_inference_model(dirname) ``` - 将模型保存到指定路径。 - **参数** - * dirname: 存在模型的目录名称 - * model\_filename: 模型文件名称，默认为\_\_model\_\_ - * params\_filename: 参数文件名称，默认为\_\_params\_\_(仅当`combined`为True时生效) - * combined: 是否将参数保存到统一的一个文件中 + * dirname: 模型保存路径 ## 四、服务部署 @@ -240,11 +234,21 @@ * 1.0.0 初始发布 + * 1.1.0 新增视频人像分割接口 新增视频流人像分割接口 + * 1.1.1 修复cudnn为8.0.4显存泄露问题 + +* 1.2.0 + + 移除 Fluid API + + ```shell + $ hub install humanseg_lite == 1.2.0 + ``` diff --git a/modules/image/semantic_segmentation/humanseg_lite/README_en.md b/modules/image/semantic_segmentation/humanseg_lite/README_en.md index e37ba012..f2b45ae3 100644 --- a/modules/image/semantic_segmentation/humanseg_lite/README_en.md +++ b/modules/image/semantic_segmentation/humanseg_lite/README_en.md @@ -171,10 +171,7 @@ - ```python - def save_inference_model(dirname='humanseg_lite_model', - model_filename=None, - params_filename=None, - combined=True) + def save_inference_model(dirname) ``` @@ -182,10 +179,7 @@ - **Parameters** - * dirname: Save path. - * model\_filename: model file name，defalt is \_\_model\_\_ - * params\_filename: parameter file name，defalt is \_\_params\_\_(Only takes effect when `combined` is True) - * combined: Whether to save the parameters to a unified file. + * dirname: Model save path. @@ -243,13 +237,22 @@ - 1.0.0 - First release + First release - 1.1.0 Added video portrait segmentation interface Added video stream portrait segmentation interface + * 1.1.1 - Fix memory leakage problem of on cudnn 8.0.4 + Fix memory leakage problem of on cudnn 8.0.4 + +* 1.2.0 + + Remove Fluid API + + ```shell + $ hub install humanseg_lite == 1.2.0 + ``` diff --git a/modules/image/semantic_segmentation/humanseg_lite/data_feed.py b/modules/image/semantic_segmentation/humanseg_lite/data_feed.py index 7f903397..f7fbb0e2 100644 --- a/modules/image/semantic_segmentation/humanseg_lite/data_feed.py +++ b/modules/image/semantic_segmentation/humanseg_lite/data_feed.py @@ -5,7 +5,6 @@ from collections import OrderedDict import cv2 import numpy as np -from PIL import Image __all__ = ['reader', 'preprocess_v'] diff --git a/modules/image/semantic_segmentation/humanseg_lite/module.py b/modules/image/semantic_segmentation/humanseg_lite/module.py index b8ba8685..600d4c28 100644 --- a/modules/image/semantic_segmentation/humanseg_lite/module.py +++ b/modules/image/semantic_segmentation/humanseg_lite/module.py @@ -19,14 +19,15 @@ import argparse import cv2 import numpy as np -import paddle.fluid as fluid -import paddlehub as hub -from paddle.fluid.core import PaddleTensor, AnalysisConfig, create_paddle_predictor +import paddle +import paddle.jit +import paddle.static +from paddle.inference import Config, create_predictor from paddlehub.module.module import moduleinfo, runnable, serving -from humanseg_lite.processor import postprocess, base64_to_cv2, cv2_to_base64, check_dir -from humanseg_lite.data_feed import reader, preprocess_v -from humanseg_lite.optimal import postprocess_v, threshold_mask +from .processor import postprocess, base64_to_cv2, cv2_to_base64, check_dir +from .data_feed import reader, preprocess_v +from .optimal import postprocess_v, threshold_mask @moduleinfo( @@ -35,22 +36,22 @@ from humanseg_lite.optimal import postprocess_v, threshold_mask author="paddlepaddle", author_email="", summary="humanseg_lite is a semantic segmentation model.", - version="1.1.0") -class ShufflenetHumanSeg(hub.Module): - def _initialize(self): - self.default_pretrained_model_path = os.path.join(self.directory, "humanseg_lite_inference") + version="1.2.0") +class ShufflenetHumanSeg: + def __init__(self): + self.default_pretrained_model_path = os.path.join(self.directory, "humanseg_lite_inference", "model") self._set_config() def _set_config(self): """ predictor config setting """ - self.model_file_path = os.path.join(self.default_pretrained_model_path, '__model__') - self.params_file_path = os.path.join(self.default_pretrained_model_path, '__params__') - cpu_config = AnalysisConfig(self.model_file_path, self.params_file_path) + model = self.default_pretrained_model_path+'.pdmodel' + params = self.default_pretrained_model_path+'.pdiparams' + cpu_config = Config(model, params) cpu_config.disable_glog_info() cpu_config.disable_gpu() - self.cpu_predictor = create_paddle_predictor(cpu_config) + self.cpu_predictor = create_predictor(cpu_config) try: _places = os.environ["CUDA_VISIBLE_DEVICES"] @@ -60,10 +61,14 @@ class ShufflenetHumanSeg(hub.Module): use_gpu = False if use_gpu: - gpu_config = AnalysisConfig(self.model_file_path, self.params_file_path) + gpu_config = Config(model, params) gpu_config.disable_glog_info() gpu_config.enable_use_gpu(memory_pool_init_size_mb=1000, device_id=0) - self.gpu_predictor = create_paddle_predictor(gpu_config) + + if paddle.get_cudnn_version() == 8004: + gpu_config.delete_pass('conv_elementwise_add_act_fuse_pass') + gpu_config.delete_pass('conv_elementwise_add2_act_fuse_pass') + self.gpu_predictor = create_predictor(gpu_config) def segment(self, images=None, @@ -116,9 +121,16 @@ class ShufflenetHumanSeg(hub.Module): pass # feed batch image batch_image = np.array([data['image'] for data in batch_data]) - batch_image = PaddleTensor(batch_image.copy()) - output = self.gpu_predictor.run([batch_image]) if use_gpu else self.cpu_predictor.run([batch_image]) - output = output[1].as_ndarray() + + predictor = self.gpu_predictor if use_gpu else self.cpu_predictor + input_names = predictor.get_input_names() + input_handle = predictor.get_input_handle(input_names[0]) + input_handle.copy_from_cpu(batch_image.copy()) + predictor.run() + output_names = predictor.get_output_names() + output_handle = predictor.get_output_handle(output_names[1]) + output = output_handle.copy_to_cpu() + output = np.expand_dims(output[:, 1, :, :], axis=1) # postprocess one by one for i in range(len(batch_data)): @@ -156,9 +168,16 @@ class ShufflenetHumanSeg(hub.Module): height = int(frame_org.shape[1]) disflow = cv2.DISOpticalFlow_create(cv2.DISOPTICAL_FLOW_PRESET_ULTRAFAST) frame = preprocess_v(frame_org, resize_w, resize_h) - image = PaddleTensor(np.array([frame.copy()])) - output = self.gpu_predictor.run([image]) if use_gpu else self.cpu_predictor.run([image]) - score_map = output[1].as_ndarray() + + predictor = self.gpu_predictor if use_gpu else self.cpu_predictor + input_names = predictor.get_input_names() + input_handle = predictor.get_input_handle(input_names[0]) + input_handle.copy_from_cpu(frame.copy()[None, ...]) + predictor.run() + output_names = predictor.get_output_names() + output_handle = predictor.get_output_handle(output_names[1]) + score_map = output_handle.copy_to_cpu() + frame = np.transpose(frame, axes=[1, 2, 0]) score_map = np.transpose(np.squeeze(score_map, 0), axes=[1, 2, 0]) cur_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) @@ -227,9 +246,16 @@ class ShufflenetHumanSeg(hub.Module): ret, frame_org = cap_video.read() if ret: frame = preprocess_v(frame_org, resize_w, resize_h) - image = PaddleTensor(np.array([frame.copy()])) - output = self.gpu_predictor.run([image]) if use_gpu else self.cpu_predictor.run([image]) - score_map = output[1].as_ndarray() + + predictor = self.gpu_predictor if use_gpu else self.cpu_predictor + input_names = predictor.get_input_names() + input_handle = predictor.get_input_handle(input_names[0]) + input_handle.copy_from_cpu(frame.copy()[None, ...]) + predictor.run() + output_names = predictor.get_output_names() + output_handle = predictor.get_output_handle(output_names[1]) + score_map = output_handle.copy_to_cpu() + frame = np.transpose(frame, axes=[1, 2, 0]) score_map = np.transpose(np.squeeze(score_map, 0), axes=[1, 2, 0]) cur_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) @@ -255,9 +281,16 @@ class ShufflenetHumanSeg(hub.Module): ret, frame_org = cap_video.read() if ret: frame = preprocess_v(frame_org, resize_w, resize_h) - image = PaddleTensor(np.array([frame.copy()])) - output = self.gpu_predictor.run([image]) if use_gpu else self.cpu_predictor.run([image]) - score_map = output[1].as_ndarray() + + predictor = self.gpu_predictor if use_gpu else self.cpu_predictor + input_names = predictor.get_input_names() + input_handle = predictor.get_input_handle(input_names[0]) + input_handle.copy_from_cpu(frame.copy()[None, ...]) + predictor.run() + output_names = predictor.get_output_names() + output_handle = predictor.get_output_handle(output_names[1]) + score_map = output_handle.copy_to_cpu() + frame = np.transpose(frame, axes=[1, 2, 0]) score_map = np.transpose(np.squeeze(score_map, 0), axes=[1, 2, 0]) cur_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) @@ -279,32 +312,6 @@ class ShufflenetHumanSeg(hub.Module): break cap_video.release() - def save_inference_model(self, - dirname='humanseg_lite_model', - model_filename=None, - params_filename=None, - combined=True): - if combined: - model_filename = "__model__" if not model_filename else model_filename - params_filename = "__params__" if not params_filename else params_filename - place = fluid.CPUPlace() - exe = fluid.Executor(place) - - program, feeded_var_names, target_vars = fluid.io.load_inference_model( - dirname=self.default_pretrained_model_path, - model_filename=model_filename, - params_filename=params_filename, - executor=exe) - - fluid.io.save_inference_model( - dirname=dirname, - main_program=program, - executor=exe, - feeded_var_names=feeded_var_names, - target_vars=target_vars, - model_filename=model_filename, - params_filename=params_filename) - @serving def serving_method(self, images, **kwargs): """ diff --git a/modules/image/semantic_segmentation/humanseg_lite/processor.py b/modules/image/semantic_segmentation/humanseg_lite/processor.py index e4911ff4..9cd53a84 100644 --- a/modules/image/semantic_segmentation/humanseg_lite/processor.py +++ b/modules/image/semantic_segmentation/humanseg_lite/processor.py @@ -50,7 +50,6 @@ def postprocess(data_out, org_im, org_im_shape, org_im_path, output_dir, visuali result['data'] = logit else: result['data'] = logit - print("result['data'] shape", result['data'].shape) return result diff --git a/modules/image/semantic_segmentation/humanseg_lite/test.py b/modules/image/semantic_segmentation/humanseg_lite/test.py new file mode 100644 index 00000000..df433469 --- /dev/null +++ b/modules/image/semantic_segmentation/humanseg_lite/test.py @@ -0,0 +1,145 @@ +import os +import shutil +import unittest + +import cv2 +import requests +import numpy as np +import paddlehub as hub + + +os.environ['CUDA_VISIBLE_DEVICES'] = '0' + + +class TestHubModule(unittest.TestCase): + @classmethod + def setUpClass(cls) -> None: + img_url = 'https://unsplash.com/photos/pg_WCHWSdT8/download?ixid=MnwxMjA3fDB8MXxhbGx8fHx8fHx8fHwxNjYyNDM2ODI4&force=true&w=640' + if not os.path.exists('tests'): + os.makedirs('tests') + response = requests.get(img_url) + assert response.status_code == 200, 'Network Error.' + with open('tests/test.jpg', 'wb') as f: + f.write(response.content) + fourcc = cv2.VideoWriter_fourcc('M', 'J', 'P', 'G') + img = cv2.imread('tests/test.jpg') + video = cv2.VideoWriter('tests/test.avi', fourcc, + 20.0, tuple(img.shape[:2])) + for i in range(40): + video.write(img) + video.release() + cls.module = hub.Module(name="humanseg_lite") + + @classmethod + def tearDownClass(cls) -> None: + shutil.rmtree('tests') + shutil.rmtree('inference') + shutil.rmtree('humanseg_lite_output') + shutil.rmtree('humanseg_lite_video_result') + + def test_segment1(self): + results = self.module.segment( + paths=['tests/test.jpg'], + use_gpu=False, + visualization=False + ) + self.assertIsInstance(results[0]['data'], np.ndarray) + + def test_segment2(self): + results = self.module.segment( + images=[cv2.imread('tests/test.jpg')], + use_gpu=False, + visualization=False + ) + self.assertIsInstance(results[0]['data'], np.ndarray) + + def test_segment3(self): + results = self.module.segment( + images=[cv2.imread('tests/test.jpg')], + use_gpu=False, + visualization=True + ) + self.assertIsInstance(results[0]['data'], np.ndarray) + + def test_segment4(self): + results = self.module.segment( + images=[cv2.imread('tests/test.jpg')], + use_gpu=True, + visualization=False + ) + self.assertIsInstance(results[0]['data'], np.ndarray) + + def test_segment5(self): + self.assertRaises( + AssertionError, + self.module.segment, + paths=['no.jpg'] + ) + + def test_segment6(self): + self.assertRaises( + AttributeError, + self.module.segment, + images=['test.jpg'] + ) + + def test_video_stream_segment1(self): + img_matting, cur_gray, optflow_map = self.module.video_stream_segment( + frame_org=cv2.imread('tests/test.jpg'), + frame_id=1, + prev_gray=None, + prev_cfd=None, + use_gpu=False + ) + self.assertIsInstance(img_matting, np.ndarray) + self.assertIsInstance(cur_gray, np.ndarray) + self.assertIsInstance(optflow_map, np.ndarray) + img_matting, cur_gray, optflow_map = self.module.video_stream_segment( + frame_org=cv2.imread('tests/test.jpg'), + frame_id=2, + prev_gray=cur_gray, + prev_cfd=optflow_map, + use_gpu=False + ) + self.assertIsInstance(img_matting, np.ndarray) + self.assertIsInstance(cur_gray, np.ndarray) + self.assertIsInstance(optflow_map, np.ndarray) + + def test_video_stream_segment2(self): + img_matting, cur_gray, optflow_map = self.module.video_stream_segment( + frame_org=cv2.imread('tests/test.jpg'), + frame_id=1, + prev_gray=None, + prev_cfd=None, + use_gpu=True + ) + self.assertIsInstance(img_matting, np.ndarray) + self.assertIsInstance(cur_gray, np.ndarray) + self.assertIsInstance(optflow_map, np.ndarray) + img_matting, cur_gray, optflow_map = self.module.video_stream_segment( + frame_org=cv2.imread('tests/test.jpg'), + frame_id=2, + prev_gray=cur_gray, + prev_cfd=optflow_map, + use_gpu=True + ) + self.assertIsInstance(img_matting, np.ndarray) + self.assertIsInstance(cur_gray, np.ndarray) + self.assertIsInstance(optflow_map, np.ndarray) + + def test_video_segment1(self): + self.module.video_segment( + video_path="tests/test.avi", + use_gpu=False, + save_dir='humanseg_lite_video_result' + ) + + def test_save_inference_model(self): + self.module.save_inference_model('./inference/model') + + self.assertTrue(os.path.exists('./inference/model.pdmodel')) + self.assertTrue(os.path.exists('./inference/model.pdiparams')) + + +if __name__ == "__main__": + unittest.main() -- GitLab