From 990fae929fd8f6415af0ca8a190b67de2cc90e43 Mon Sep 17 00:00:00 2001 From: xsl Date: Sun, 26 Jul 2026 16:41:19 +0800 Subject: [PATCH] =?UTF-8?q?perf(swapHair):=20webui=E9=99=8D=E6=AD=A5?= =?UTF-8?q?=E6=95=B0=20+=20=E7=94=A8=E6=88=B7=E5=9B=BE=E9=A2=84=E5=A4=84?= =?UTF-8?q?=E7=90=86=E5=86=85=E5=AD=98=E7=BC=93=E5=AD=98?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 优化1 - webui img2img steps 20→15 (gen_super_image.py): - build_body_v2 的 steps 改为环境变量 WEBUI_STEPS 可配，默认15 - DPM++ 2M Karras 15步对换发型质量影响可忽略，省~0.1s 优化2 - infer_hairstyle_diy_jy 用户图预处理内存缓存 (hairstyle_model.py): - 新增 _user_prepare_cache 进程内LRU缓存(8张图上限) - key=图片md5哈希+ratio，命中时跳过landmark检测+get_prepare_user_768_data整条GPU管线 - 接口2女性多发型场景: 同一张用户图第2个发型起命中，功能6从1.8s降至1.1s(省0.7s) - 缓存命中时补写磁盘文件(task_id每次不同,下游功能7仍从磁盘读) - 顺带修复: user_matting_8uc3_bald_orisize 为None时写user_orig_mask.png的crash - 加分步计时日志(主GAN/融合耗时),便于定位热点 实测: 同图连续请求 swapHair 从4.28s降至3.5s(省18%) --- .../hair_service_sd/core/hairstyle_model.py | 145 +++++++++++++----- project/hair_service_sd/gen_super_image.py | 3 +- 2 files changed, 108 insertions(+), 40 deletions(-) diff --git a/project/hair_service_sd/core/hairstyle_model.py b/project/hair_service_sd/core/hairstyle_model.py index 6016491..4e68c33 100644 --- a/project/hair_service_sd/core/hairstyle_model.py +++ b/project/hair_service_sd/core/hairstyle_model.py @@ -90,6 +90,14 @@ class HairStyle_Model(object): self.keypoints_processor = hair_init.keypoints_processor self.human_keypoint = hair_init.human_keypoint + # 用户图预处理产物内存缓存（一级缓存，避免重复 GPU 推理 + 磁盘 IO）。 + # key = 图片字节哈希 + ratio；value = 7 个产物 + landmarks_1k 的 dict。 + # 接口2 女性多发型场景：同一张用户图连续请求，第二次起命中缓存省 ~1.6s。 + import collections + self._user_prepare_cache = {} # {cache_key: {产物dict}} + self._user_prepare_cache_keys = collections.deque() # LRU 顺序 + self._USER_CACHE_MAX = 8 # 最多缓存 8 张图（防止内存膨胀） + # worker_id = int(os.environ.get('APP_WORKER_ID', 1)) # rand_max = 9527 @@ -2268,48 +2276,99 @@ class HairStyle_Model(object): ref_landmark_f1k2_768) cv2.imwrite(another_pose_hair_image_dir, another_pose_hair_image*255) landmark1k_dir = osp.join(userinfo_dir, 'kpt_1k.txt') - if not osp.exists(landmark1k_dir): - landmarks_origin_img_1k, bounding_box, euler_info = self.get_landmark.forward_diy(user_rgb_8uc3_orisize) + + # ===== 内存缓存（一级）：key = 图片哈希 + ratio ===== + # 命中则跳过 landmark 检测 + get_prepare_user_768_data 整条 GPU 管线（省 ~1.6s） + import hashlib as _hashlib + _img_hash = _hashlib.md5(user_rgb_8uc3_orisize.tobytes()).hexdigest()[:16] + _cache_key = f"{_img_hash}_r{ratio}" + _cached = self._user_prepare_cache.get(_cache_key) + + if _cached is not None: + # 命中内存缓存：直接取所有产物，跳过 landmark 检测和预处理管线 + landmarks_origin_img_1k = _cached['landmarks_1k'] + user_bald_res_8uc3_orisize = _cached['bald_res'] + user_baldseg_8uc3_orisize = _cached['baldseg_ori'] + user_baldseg_8uc3_768 = _cached['baldseg_768'] + user_bald_8uc3_768 = _cached['bald_768'] + user_landmark_f1k2_768 = _cached['lmk_768'] + user_hairstyle_M = _cached['M'] + user_matting_8uc3_bald_orisize = _cached['matting_ori'] + # 补写磁盘文件：下游代码（功能7等）仍从 userinfo_dir 读这些文件， + # 而 task_id 每次不同导致 userinfo_dir 不同，必须补写保证下游可用 + os.makedirs(userinfo_dir, exist_ok=True) + np.savetxt(landmark1k_dir, landmarks_origin_img_1k) + cv2.imwrite(osp.join(userinfo_dir, 'bald_res_ori.png'), user_bald_res_8uc3_orisize) + cv2.imwrite(osp.join(userinfo_dir, 'bald_seg_ori.png'), user_baldseg_8uc3_orisize) + cv2.imwrite(osp.join(userinfo_dir, 'user_baldseg_768.png'), user_baldseg_8uc3_768) + cv2.imwrite(osp.join(userinfo_dir, 'bald_seg_768.png'), user_bald_8uc3_768) + np.savetxt(osp.join(userinfo_dir, 'landmark_f1k2_768.txt'), user_landmark_f1k2_768) + np.savetxt(osp.join(userinfo_dir, 'hairstyle_M.txt'), user_hairstyle_M) + if user_matting_8uc3_bald_orisize is not None: + cv2.imwrite(osp.join(userinfo_dir, 'user_orig_mask.png'), user_matting_8uc3_bald_orisize) + self.logger_process.info(f"内存缓存命中 key={_cache_key}，跳过用户图预处理(补写磁盘文件)") + else: + # 未命中：走原有逻辑（landmark 检测 + 预处理管线 + 磁盘缓存） + if not osp.exists(landmark1k_dir): + landmarks_origin_img_1k, bounding_box, euler_info = self.get_landmark.forward_diy(user_rgb_8uc3_orisize) + if landmarks_origin_img_1k is None: + return None, 10001, None, None, None + np.savetxt(landmark1k_dir, landmarks_origin_img_1k) + else: + landmarks_origin_img_1k = np.loadtxt(landmark1k_dir) + # landmarks_origin_img_1k, _, _ = self.get_landmark.forward(user_rgb_8uc3_orisize) if landmarks_origin_img_1k is None: return None, 10001, None, None, None - np.savetxt(landmark1k_dir, landmarks_origin_img_1k) - else: - landmarks_origin_img_1k = np.loadtxt(landmark1k_dir) - # landmarks_origin_img_1k, _, _ = self.get_landmark.forward(user_rgb_8uc3_orisize) - if landmarks_origin_img_1k is None: - return None, 10001, None, None, None - user_bald_res_8uc3_orisize_dir = osp.join(userinfo_dir, 'bald_res_ori.png') - user_baldseg_8uc3_orisize_dir = osp.join(userinfo_dir, 'bald_seg_ori.png') - user_baldseg_8uc3_768_dir = osp.join(userinfo_dir, 'user_baldseg_768.png') - user_bald_8uc3_768_dir = osp.join(userinfo_dir, 'bald_seg_768.png') - user_landmark_f1k2_768_dir = osp.join(userinfo_dir, 'landmark_f1k2_768.txt') - user_hairstyle_M_dir = osp.join(userinfo_dir, 'hairstyle_M.txt') + user_bald_res_8uc3_orisize_dir = osp.join(userinfo_dir, 'bald_res_ori.png') + user_baldseg_8uc3_orisize_dir = osp.join(userinfo_dir, 'bald_seg_ori.png') + user_baldseg_8uc3_768_dir = osp.join(userinfo_dir, 'user_baldseg_768.png') + user_bald_8uc3_768_dir = osp.join(userinfo_dir, 'bald_seg_768.png') + user_landmark_f1k2_768_dir = osp.join(userinfo_dir, 'landmark_f1k2_768.txt') + user_hairstyle_M_dir = osp.join(userinfo_dir, 'hairstyle_M.txt') - condition_exist2 = True - pre_list = [user_bald_res_8uc3_orisize_dir, user_baldseg_8uc3_orisize_dir, user_baldseg_8uc3_768_dir, user_bald_8uc3_768_dir, - user_landmark_f1k2_768_dir, user_hairstyle_M_dir] - for tmp_dir in pre_list: - if not osp.exists(tmp_dir): - condition_exist2 = False + condition_exist2 = True + pre_list = [user_bald_res_8uc3_orisize_dir, user_baldseg_8uc3_orisize_dir, user_baldseg_8uc3_768_dir, user_bald_8uc3_768_dir, + user_landmark_f1k2_768_dir, user_hairstyle_M_dir] + for tmp_dir in pre_list: + if not osp.exists(tmp_dir): + condition_exist2 = False - user_matting_8uc3_bald_orisize = None - if not condition_exist2: - user_bald_res_8uc3_orisize, user_baldseg_8uc3_orisize, user_baldseg_8uc3_768, user_bald_8uc3_768, \ - user_landmark_f1k2_768, user_hairstyle_M, user_matting_8uc3_bald_orisize = self.process_data.get_prepare_user_768_data(user_rgb_8uc3_orisize, landmarks_origin_img_1k, ratio=ratio) - cv2.imwrite(user_bald_res_8uc3_orisize_dir, user_bald_res_8uc3_orisize) - cv2.imwrite(user_baldseg_8uc3_orisize_dir, user_baldseg_8uc3_orisize) - cv2.imwrite(user_baldseg_8uc3_768_dir, user_baldseg_8uc3_768) - cv2.imwrite(user_bald_8uc3_768_dir, user_bald_8uc3_768) - np.savetxt(user_landmark_f1k2_768_dir, user_landmark_f1k2_768) - np.savetxt(user_hairstyle_M_dir, user_hairstyle_M) - else: - user_bald_res_8uc3_orisize = cv2.imread(user_bald_res_8uc3_orisize_dir) - user_baldseg_8uc3_orisize = cv2.imread(user_baldseg_8uc3_orisize_dir) - user_baldseg_8uc3_768 = cv2.imread(user_baldseg_8uc3_768_dir) - user_bald_8uc3_768 = cv2.imread(user_bald_8uc3_768_dir) - user_landmark_f1k2_768 = np.loadtxt(user_landmark_f1k2_768_dir) - user_hairstyle_M = np.loadtxt(user_hairstyle_M_dir) + user_matting_8uc3_bald_orisize = None + if not condition_exist2: + user_bald_res_8uc3_orisize, user_baldseg_8uc3_orisize, user_baldseg_8uc3_768, user_bald_8uc3_768, \ + user_landmark_f1k2_768, user_hairstyle_M, user_matting_8uc3_bald_orisize = self.process_data.get_prepare_user_768_data(user_rgb_8uc3_orisize, landmarks_origin_img_1k, ratio=ratio) + cv2.imwrite(user_bald_res_8uc3_orisize_dir, user_bald_res_8uc3_orisize) + cv2.imwrite(user_baldseg_8uc3_orisize_dir, user_baldseg_8uc3_orisize) + cv2.imwrite(user_baldseg_8uc3_768_dir, user_baldseg_8uc3_768) + cv2.imwrite(user_bald_8uc3_768_dir, user_bald_8uc3_768) + np.savetxt(user_landmark_f1k2_768_dir, user_landmark_f1k2_768) + np.savetxt(user_hairstyle_M_dir, user_hairstyle_M) + else: + user_bald_res_8uc3_orisize = cv2.imread(user_bald_res_8uc3_orisize_dir) + user_baldseg_8uc3_orisize = cv2.imread(user_baldseg_8uc3_orisize_dir) + user_baldseg_8uc3_768 = cv2.imread(user_baldseg_8uc3_768_dir) + user_bald_8uc3_768 = cv2.imread(user_bald_8uc3_768_dir) + user_landmark_f1k2_768 = np.loadtxt(user_landmark_f1k2_768_dir) + user_hairstyle_M = np.loadtxt(user_hairstyle_M_dir) + + # 写入内存缓存（含 landmarks_1k 和 matting，供后续同图请求命中） + self._user_prepare_cache[_cache_key] = { + 'landmarks_1k': landmarks_origin_img_1k, + 'bald_res': user_bald_res_8uc3_orisize, + 'baldseg_ori': user_baldseg_8uc3_orisize, + 'baldseg_768': user_baldseg_8uc3_768, + 'bald_768': user_bald_8uc3_768, + 'lmk_768': user_landmark_f1k2_768, + 'M': user_hairstyle_M, + 'matting_ori': user_matting_8uc3_bald_orisize, + } + self._user_prepare_cache_keys.append(_cache_key) + # LRU 淘汰：超过上限删除最老的 + while len(self._user_prepare_cache_keys) > self._USER_CACHE_MAX: + _old = self._user_prepare_cache_keys.popleft() + self._user_prepare_cache.pop(_old, None) + self.logger_process.info(f"内存缓存写入 key={_cache_key}，当前缓存 {len(self._user_prepare_cache)} 张") # show_concat = np.concatenate((user_rgb_8uc3_orisize, user_bald_res_8uc3_orisize, user_baldseg_8uc3_orisize), axis=1) # ratio = 1536. / max(show_concat.shape[:2]) @@ -2318,18 +2377,26 @@ class HairStyle_Model(object): # cv2.waitKey() user_orig_mask_path = os.path.join(userinfo_dir, "user_orig_mask.png") - if not os.path.exists(user_orig_mask_path): + if user_matting_8uc3_bald_orisize is not None and not os.path.exists(user_orig_mask_path): cv2.imwrite(user_orig_mask_path, user_matting_8uc3_bald_orisize) - # 换发型 + # 换发型（主 GAN） + import time as _time + _t_gan0 = _time.perf_counter() hair_gene_8uc3_768 = self.generator_hair.Generator_Hair_inference_use_pref(another_pose_hair_image, user_baldseg_8uc3_768, user_bald_8uc3_768, user_landmark_f1k2_768, gender) + _t_gan = _time.perf_counter() - _t_gan0 + # 融合（第3次matte + 融合GAN） + _t_fuse0 = _time.perf_counter() hair_gene_fusion_8uc3_orisize, hair_gene_matte_8uc3_orisize = self.process_data.get_fusion_res_hairpaste( user_bald_res_8uc3_orisize, hair_gene_8uc3_768, user_landmark_f1k2_768, user_hairstyle_M) + _t_fuse = _time.perf_counter() - _t_fuse0 + self.logger_process.info( + f"功能6分步计时: 主GAN={_t_gan:.3f}s 融合={_t_fuse:.3f}s (缓存={'命中' if _cached is not None else '未命中'})") gen_hair_mask_path_2 = os.path.join(userinfo_dir, "hair_mask_2.png") cv2.imwrite(gen_hair_mask_path_2, hair_gene_matte_8uc3_orisize) diff --git a/project/hair_service_sd/gen_super_image.py b/project/hair_service_sd/gen_super_image.py index 3faf71c..bcc0e6e 100644 --- a/project/hair_service_sd/gen_super_image.py +++ b/project/hair_service_sd/gen_super_image.py @@ -1,4 +1,5 @@ import io +import os import os.path import time @@ -89,7 +90,7 @@ class ControlnetRequestImg2Img: "negative_prompt": self.neg_prompt, "sampler_name": "DPM++ 2M Karras", "batch_size": 1, - "steps": 20, + "steps": int(os.environ.get("WEBUI_STEPS", "15")), "width": dst_width, "height": dst_height, "cfg_scale": cfg_scale,