版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
DataFunCon#
2023多模态预训练模型在OPPO端云场景的落地实践陈宸-OPPO研究院-高级算法工程师03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11Contents目录端侧图文检索技术研究图文生成&理解模型的应用优化文图生成模型的端侧轻量化03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11端侧图文检索技术研究03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11query1:
和女朋友去迪士尼 query2:
山顶婚纱照一句话搜索的意义:用户体验:真正解决用户想什么就能搜什么的痛点,“智慧搜图,搜你所想”。依托于大模型预训练技术,不再依赖于标签的迭代和扩展大模型轻量端侧化的技术意义:成本节约:将云侧大模型才能体验的效果搬向到端侧,大幅节约计算资源;隐私保护:直接在端侧处理用户的私人照片,无需上传到云端,保护用户隐私;https://b /s端侧图文检索技术研究——解决了什么问题?端侧检索demo实现端侧智慧搜索的关键因素:其一,“人话”解读能力。智慧搜图不仅能单独搜词,也能放一起搜,实现真正的口语化表达搜索,所想即所得,如“去年在动物园拍的老虎”等。因此需要类似多模态大模型
CLIP(OpenAI)的图文理解能力。其二,高效搜索速度。相比动辄翻上十几分钟半个小时的相册,现在无论从桌面下拉智慧搜索、打开相册、或是用语音助手,都只需要一句话就能搜到想要的图片,系统级地提升了找信息的效率。因此如何实现大模型在端侧的轻量化部署有重大的意义。大模型轻量化端侧技术落地的难点:压缩多模态大模型并确保精度。这并非简单用剪枝或量化等方法,直接压缩几倍模型大小就能搞定。毕竟对于端侧而言,算力有限的情况下,能部署的模型大小是往往只能达到大模型的几十分之一。与算法模型升级相对应的,需要在端侧实现一个性能鲁棒的向量检索引擎,保证大模型下端的工程性能03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11端侧图文检索技术研究——算法优化CLIP双塔模型ALBEF单流模型单双流多教师蒸馏架构损失函数检索引擎的计算分位两部分:离线部分:扫描相册所有图片,通过图片编码器将所有图片转成向量;并经过fp16量化存储成Nx200的fp矩阵在线部分:每次输入query,通过文本编码器将query转成向量;并经过fp16量化降低计算内存;最后通过矩阵相乘计算query向量跟所有图片的相似分数,并通过排序输出一个有序列表。Lei,Youbo,etal."MCAD:Multi-teacherCross-modalAlignmentDistillationforefficient
image-textretrieval."arXivpreprintarXiv:2310.19654
(2023).03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11端侧图文检索技术研究——学术集效果各种蒸馏方法的效果对比Modelnameimagemodeltextmodelfusion
modelimage
encodingtimeretrieval
timeparameternumbertestsetplatformCLIPVIT-L/1412-layertransformerdot
product11.0ms32.5ms427.62Mfilckr5KV100
GPUALBEFVIT-B/166-layertransformer6-layertransformer7.6ms265ms(k=16)1945ms
(k=128)3865ms
(k=256)419.12Mfilckr5KV100
GPU自研小模型mobileVitV2-1.54-layerTinyBertdoc
product3.8
ms14.1
ms25.9
Mfilckr5KV100
GPU自研小模型mobileVitV2-1.54-layerTinyBertdoc
product17.3
ms14.6
ms25.9
Mfilckr5KMTK
DX3大小模型的性能对比03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11端侧图文检索技术研究——真实场景效果数据量:11个用户真实相册共2万+图片,手写5400+query数据分布:测试集R@1R@5R@10MRmAP010.47280.6710.74950.63110.6080020.49560.7580.82510.69290.5306030.40190.56650.61080.52640.4889040.45320.68470.73890.62560.6048050.58430.7530.79520.71080.6428060.53230.68550.750.65590.5890070.350.52940.60880.49610.4771080.64170.80830.84170.76390.5943090.59650.68420.71930.66670.5622100.51210.70590.76470.66090.5441110.56540.74180.7810.69610.6336平均0.48480.67680.73600.63250.584003872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11端侧图文检索技术研究——细粒度优化Doveh,Sivan,etal."Teachingstructuredvision&languageconceptstovision&languagemodels."ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognition.2023.细粒度属性词替换+hard
negative
sampling+
LwF抗遗忘算法03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11文图生成&理解态模型的应用优化03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——中文文生图大模型继续预训练如何做高质量低成本的继续预训练如何对齐中文的语言文化如何提升生成图像的细节质量Parameterefficient
adapterOrthogonal
FinetuningQiu,Zeju,etal."Controllingtext-to-imagediffusionbyorthogonalfinetuning."Thirty-seventhConferenceonNeuralInformationProcessingSystems.
2023.03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11中文语境迁移效果图文生成&理解模型的应用优化——中文文生图大模型继续预训练收敛速度03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11Finetune模型鸳鸯双栖蝶双飞,满园春色惹人醉LoRAControlnetSSD1.3B小模型一只超级可爱的兔子穿着僧侣服装,肖像照,皮克斯动画SDXLinpainting青花瓷版的恐龙在长椅上江南,夏日湖畔的一个村庄图文生成&理解模型的应用优化——中文文生图大模型继续预训练一个漂亮的亚洲女孩,电影灯光 西湖,塔和瀑布,日出3D电影,4k,高度细致,男人坐在马桶上读报带着墨镜的猫咪手里拿着剑,在恶魔城堡里,仙剑奇侠风格LatentCM03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——通用优化应用壁纸生成春节档热度top1春节档热度top3文生图模型+超分辨率生成2k高清壁纸03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——通用优化应用锁屏杂志生成文生图模型+微调LLAVA+LLM
生成图文并茂的杂志Liu,Haotian,etal."Visualinstructiontuning."arXivpreprintarXiv:2304.08485
(2023).03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——通用优化应用Zhang,Pan,etal."Internlm-xcomposer:Avision-languagelargemodelforadvanced
text-imagecomprehensionandcomposition."arXivpreprintarXiv:2309.15112
(2023).Internlm-xcomposer训练框架03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——垂域优化-人像垂域AI模型画人的几个问题:1.
人脸人手等身体部位的崩坏。2.
过于精致标准,渲染过度光滑,在质感上失真。3. 细粒度属性和文本描述的不对齐。03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——垂域优化-人像垂域构建细粒度的人像属性数据03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——垂域优化-人像垂域U-Net中模块与图像中特征的对应关系,可用于指导LoRA微调的参数选择厚嘴唇薄嘴唇03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——垂域优化-人像垂域小鼻子大鼻子03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——垂域优化-人像垂域细眉毛粗眉毛03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——垂域优化-人像垂域垂域微调经验:大量数据粗调,增加模型对新概念的泛化能力少量高质量数据精调,提升模型的图片生成质量人脸修复逻辑:穿着华丽盔甲的玄幻战士与巨龙激战,雷霆与火焰交织在一起。(随机6张,无cherry-pick)03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——垂域优化-古风人像效果古道边一骑红尘客正巍然马背,身披白色斗篷,踏寂静落阿叶(随机6张,无cherry-pick)树丛中,翩翩少女,红衣绿裙,手提花伞,踏泥寻径,仿佛踏入了一幅画卷(随机6张,无cherry-pick)03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——垂域优化应用广告营销工具(内测版)03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——文字渲染-问题定义如何在文生图模型中渲染出正确的文字?Ma,Jian,etal."GlyphDraw:LearningtoDrawChineseCharactersinImageSynthesisModelsCoherently."arXivpreprintarXiv:2303.17870
(2023).03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——文字渲染-算法GlyphDraw训练框架GlyphDraw推理框架数据集图文对数量文字数量中文数据集792k3.3M
字英文数据集1.9M2.3M
wordsMa,Jian,etal."GlyphDraw:LearningtoDrawChineseCharactersinImageSynthesisModelsCoherently."arXivpreprintarXiv:2303.17870
(2023).GlyphDraw数据集构建03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——文字渲染-客观效果Ma,Jian,etal."GlyphDraw:LearningtoDrawChineseCharactersinImageSynthesisModelsCoherently."arXivpreprintarXiv:2303.17870
(2023).03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——文字渲染-主观效果Ma,Jian,etal."GlyphDraw:LearningtoDrawChineseCharactersinImageSynthesisModelsCoherently."arXivpreprintarXiv:2303.17870
(2023).03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成-问题定义Ma,Jian,etal."Subject-diffusion:Opendomainpersonalizedtext-to-imagegenerationwithouttest-timefine-tuning."arXivpreprintarXiv:2307.11410
(2023).如何使用一张参考图像快速生成新图片并平衡保真度和泛化性?03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成-数据集SDD数据集统计数据SDD数据集词云Ma,Jian,etal."Subject-diffusion:Opendomainpersonalizedtext-to-imagegenerationwithouttest-timefine-tuning."arXivpreprintarXiv:2307.11410
(2023).03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成-算法Ma,Jian,etal."Subject-diffusion:Opendomainpersonalizedtext-to-imagegenerationwithouttest-timefine-tuning."arXivpreprintarXiv:2307.11410
(2023).03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成-效果单实体生成与各种方法的对比Ma,Jian,etal."Subject-diffusion:Opendomainpersonalizedtext-to-imagegenerationwithouttest-timefine-tuning."arXivpreprintarXiv:2307.11410
(2023).双实体生成与各种方法的对比03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成-效果Ma,Jian,etal."Subject-diffusion:Opendomainpersonalizedtext-to-imagegenerationwithouttest-timefine-tuning."arXivpreprintarXiv:2307.11410
(2023).03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成-效果Ma,Jian,etal."Subject-diffusion:Opendomainpersonalizedtext-to-imagegenerationwithouttest-timefine-tuning."arXivpreprintarXiv:2307.11410
(2023).03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成-效果Ma,Jian,etal."Subject-diffusion:Opendomainpersonalizedtext-to-imagegenerationwithouttest-timefine-tuning."arXivpreprintarXiv:2307.11410
(2023).03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成应用外观多角度生成品牌调性干预产品外观描述生成根据参照图生成效果图描述生成品牌调性/风格干预广告营销工具产品外观设计(从0-1设计) 产品效果图生成(工作室拍摄的效果图)产品营销素材生成(海报/banner)营销文案&图片生成素材布局生成布局描述生成参照物干预设计草图生图Ayellow
hatAgirlwearingthehatandfacing
forest选择生成【海报】根据参考素材生成根据品牌VI,历史产品调性生成产品设计03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成-应用商品设计个性化图片生成海报设计03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成-应用Subject-diffusion的个性化生成03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11图文生成&理解模型的应用优化——个性化生成-应用Stable-diffusion的outpainting03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11文图生成模型的端侧轻量化03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11文图生成模型的端侧轻量化——技术路线-模型结构优化Unet结构示意图删除某个模块之后的效果和参数量波动分析模型采样时间(DPMsolver++25步)运行内存UNet参数量SD
1.51.34s4105M859.52MSD
base-2m0.9s3458M579.38MSD
small-2m0.83s3287M482.35MSD
tiny-2m0.76s2979M323.38MSD
small0.88s3477M579.38MSD
tiny0.75s3043M323.38M不同剪枝模型在V100上测试结果03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11文图生成模型的端侧轻量化——技术路线-模型结构优化采用SDXL蒸馏SD
small模型03872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11-29803872412023-11文图生成模型的端侧轻量化——技术路线-采样加速Salimans,Tim,andJonathanHo."Progressivedistil
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 二零二五年度2025版木材行业标准制定合作合同2篇
- 福建省泉州市南安市2024-2025学年八年级上学期期末英语试题(无答案)
- 创新创业-职业核心能力课件
- 丝印精加工在微型电子设备制造领域的应用考核试卷
- 二零二五年度墓地陵园土地租赁与使用权转让合同4篇
- 母婴行业2025年度母婴用品环保认证服务合同2篇
- 二零二五版钢材货物流动银行托管运输合同3篇
- 二零二五年度木制品生产与销售承包合同3篇
- 2025年公司内部竞业保密协议
- 2025年太阳能光伏电站智能监控工程施工合同
- 2024年高纯氮化铝粉体项目可行性分析报告
- 安检人员培训
- 山东省潍坊市2024-2025学年高三上学期1月期末 英语试题
- 危险性较大分部分项工程及施工现场易发生重大事故的部位、环节的预防监控措施
- 《榜样9》观后感心得体会四
- 2023事业单位笔试《公共基础知识》备考题库(含答案)
- 化学-广东省广州市2024-2025学年高一上学期期末检测卷(一)试题和答案
- 2025四川中烟招聘高频重点提升(共500题)附带答案详解
- EHS工程师招聘笔试题与参考答案(某大型央企)2024年
- 营销策划 -丽亭酒店品牌年度传播规划方案
- 2025年中国蛋糕行业市场规模及发展前景研究报告(智研咨询发布)
评论
0/150
提交评论