版權(quán)說(shuō)明:本文檔由用戶提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請(qǐng)進(jìn)行舉報(bào)或認(rèn)領(lǐng)
文檔簡(jiǎn)介
MACHINELEARNINGINPYTHON
(PART4):
DIFFUSIONMODELSINPYTORCH
LUKESHENEMAN
GENERATIVEARTIFICIALINTELLIGENCE
TexttoImage(StableDiffusion)
TexttoVideo
GenerativeAI:Learnalatentrepresentationofthedistributionofourcomplextrainingdataandthensamplefromit
TrainingData
DeepLearning
Diffusion,etc.
Transformers,etc.
DIFFUSIONMODELS
CONDITIONINGIMAGEGENERATION
Providenaturallanguagetextpromptsto
guidereversediffusionprocess
Text-to-ImageDiffusionModelsareboth:
ImageGenerationModels
LanguageModels
StableDiffusionArchitecture
OVERVIEW
RecapfromParts1-3
MachineLearningBasics
NeuralNetworks
Tensors
ConvolutionalNeuralNetworks(CNNs)
GPUsandCUDA
PyTorch
WhyusePyTorch?
ImplementingaDiffusionModelinPython
TrainandTestourDiffusionModel
REVIEWOFBASICS
Machinelearningisadata-drivenmethodforcreatingmodelsforprediction,optimization,classification,generation,andmore
Pythonandscikit-learn
MNIST
ArtificialNeuralNetworks(ANNs)
MNIST
NEURALNETWORKBASICS
WeightsandBiases
FULLY-CONNECTEDNEURALNETWORKS
Imagesaretensors!
FEATUREHIERARCHIES
Weneedimagefilterstohelpusextractfeatures
EXAMPLE: SOBELFILTER
Sobelkernels=
CONVOLUTIONALNEURALNETWORK
GPUvs.CPU
“Moore’sLawforCPUsisDead”
WHYGPUSEXACTLY?
CNNsareallaboutmatrixandvectoroperations(multiplication,addition)
GPUscanperformparallelmultiplicationandadditionstepspereachclockcycle.
FrameworksmakeGPUsEasy
DIFFUSIONMODELS
FORWARDDIFFUSION
Definehowmanytimestepswillbeused(commontousehundredsormore)
EstablishanoiseschedulewhichdescribestherateatwhichGaussiannoiseisadded
Linear
Cosine
T=0 T=1 T=2 T=3 …
T=n
Iused100timesteps. LargermodelslikeStableDiffusionusethousandsofsmallersteps.Iusedacosinenoiseschedule.
TIMESTEPENCODING
30
30
+ =
RGBImage Integertimestep 4-ChannelRGB+Timestep
Iencodetimestepasanotherbandintheimageinpixel-space
U-NetArchitecture
4
U-NetDenoiser
3
RGB+T RGB
TrainingourNeuralNetwork
PossiblelossfunctionsforourU-Net
loss1=MSE(predt,original)loss2=MSE(predt,noisyt-1)loss3=predt-noisyt
OtherHyperparameters:
Epochs=100
Timesteps=100BatchSize=1250Optimizer=AdamLearningRate=0.001
CoreTrainingLoop
schedule=cosine_schedule(TIMESTEPS)foreachEpoch:
foreachBatchb:
foreachTimestept:
img=add_gaussian_noise(img,schedule(t))predicted=UNet(img)
loss=loss_function(img,predicted)backward_propagationandoptimization
CELEBFACESATTRIBUTES(CELEBA)DATASET
202,599numberoffaceimagesofvariouscelebrities
10,177uniqueidentities,butnamesofidentitiesarenotgiven
40binaryattributeannotationsperimage
5landmarklocations
Images”inthewild”orCropped/Aligned
SOMEPRELIMINARYOUTPUT
Ohno!
UseseparateAImodelforupsampling
64
64
512
SRResNet
512
/twtygqyy/pytorch-SRResNet
MyModel
Mightnotbeterrific,but…
Itwastrainedononly5000imagesforafewhoursonasingleRTX4090GPU
StableDiffusionwastrainedon600millioncaptionedimages
Took256NVIDIAA100GPUsonAmazonWebServicesatotalof150,000GPU-hoursAtacostof$600,000
StableDiffusion
ConditioningreverseDiffusiononTextprompts
PRE-PROCESSINGCELEBADATASET
Readfirst5000annotationsintoPANDASdataframe(easy!)
Foreachimage,gettheheadingnamesforpositiveattributes
Convertheadingnamesintoatextprompt:
e.g.“Photoofperson<attribute_x>,<attribute_y>,<attribute_z>,…”
e.g.“Photoofpersonbushyeyebrows,beard,mouthslightlyopen,wearinghat.”
Cropthelargestsquarefromtheimage,thenresizeto64x64x3numpyarray
UseOpenAICLIPmodeltofindtheimageembeddingsandtextembeddingsforeveryimage/promptpair.
Createa5000elementPythonlistof4-tuples:
(filename,64x64xRGBimage_array,image_embedding,prompt_embedding)
Picklelisttoafilewecanquickyloadintomemorywhenwetrainourmodel!
OPENAICLIPMODEL (CONTRASTIVELANGUAGE–IMAGEPRE-TRAINING)
Opensource/weightsmulti-modalAImodeltrainedonimage,captionpairs
Sharedembeddingspace!
Usetransformermodel(GPT-2)tocreatetokenembeddingsfromtext
Usevisiontransformer(VIT)tocreatetokenembeddingsfromimages
CLIPExamples/research/clip
USINGCLIPISTRIVIAL
/openai/CLIP
Zero-shotclassifications!ConditioningGenerativeAI(DALL-E)
GeneratingcaptionsforimagesorvideoImagesimilaritysearch
ContentModerationObjectTracking
CLIPusesvectorswith512dimensionsGPT3(Davinci)uses12888dimensions
Vectorembeddingscapturethedeepersemanticcontextofawordortextchunk…orimage…oranything.
Thesemanticsofanobjectaredefinedbyitsmulti-dimensionalandmulti-scaleco-occurrenceandrelationshipswithotherobjectsinthetrainingdata
Semanticvectorembeddingsarelearnedfromvastamountsofdata.
400,000,000(image,text)pairs
CLIPwastrainedon256largeGPUsfor2weeks.
SizeofEmbeddingVector
Onewaytotrainamulti-modalembeddinglayer
“AcuteWelshCorgidog.”
Learningyoursemanticembeddingsf
溫馨提示
- 1. 本站所有資源如無(wú)特殊說(shuō)明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請(qǐng)下載最新的WinRAR軟件解壓。
- 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請(qǐng)聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
- 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁(yè)內(nèi)容里面會(huì)有圖紙預(yù)覽,若沒(méi)有圖紙預(yù)覽就沒(méi)有圖紙。
- 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
- 5. 人人文庫(kù)網(wǎng)僅提供信息存儲(chǔ)空間,僅對(duì)用戶上傳內(nèi)容的表現(xiàn)方式做保護(hù)處理,對(duì)用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對(duì)任何下載內(nèi)容負(fù)責(zé)。
- 6. 下載文件中如有侵權(quán)或不適當(dāng)內(nèi)容,請(qǐng)與我們聯(lián)系,我們立即糾正。
- 7. 本站不保證下載資源的準(zhǔn)確性、安全性和完整性, 同時(shí)也不承擔(dān)用戶因使用這些下載資源對(duì)自己和他人造成任何形式的傷害或損失。
最新文檔
- 二零二五年度生態(tài)循環(huán)農(nóng)業(yè)技術(shù)支持合同范本4篇
- 二零二五年度影視作品宣傳推廣合同樣本3篇
- 69252號(hào)離婚合同-2024年版版B版
- 2025年度電梯維保人員培訓(xùn)與資質(zhì)認(rèn)證合同4篇
- 2025年度充電樁充電服務(wù)收費(fèi)標(biāo)準(zhǔn)制定合同3篇
- 2025年度高速打印機(jī)采購(gòu)及專業(yè)打印解決方案合同4篇
- 2025年度派駐研發(fā)團(tuán)隊(duì)技術(shù)合作協(xié)議范本4篇
- 2025年城市道路照明改造工程合同4篇
- 2025版教育培訓(xùn)機(jī)構(gòu)課程合作合同4篇
- 2025年度全國(guó)范圍內(nèi)促銷員勞動(dòng)合同標(biāo)準(zhǔn)范本2篇
- 2024公路瀝青路面結(jié)構(gòu)內(nèi)部狀況三維探地雷達(dá)快速檢測(cè)規(guī)程
- 2024年高考真題-地理(河北卷) 含答案
- 中國(guó)高血壓防治指南(2024年修訂版)解讀課件
- 2024風(fēng)力發(fā)電葉片維保作業(yè)技術(shù)規(guī)范
- 《思想道德與法治》課程教學(xué)大綱
- 2024光儲(chǔ)充一體化系統(tǒng)解決方案
- 處理后事授權(quán)委托書
- 封條(標(biāo)準(zhǔn)A4打印封條)
- 中醫(yī)治療“濕疹”醫(yī)案72例
- 2023年大學(xué)生《思想道德與法治》考試題庫(kù)附答案(712題)
- 清代文學(xué)緒論
評(píng)論
0/150
提交評(píng)論