up

Author: owenliang

README.md

@@ -2,4 +2,214 @@
 
 手写数字数据集的diffusion图像生成
 
-TODO：引入cross attention实现生成数字引导
+## 训练效果
+
+```
+epoch:0 loss=0.09367454051971436
+epoch:1 loss=0.08872738480567932
+epoch:2 loss=0.07516557723283768
+epoch:3 loss=0.06611064076423645
+epoch:4 loss=0.0668148323893547
+epoch:5 loss=0.07564781606197357
+epoch:6 loss=0.06627427041530609
+epoch:7 loss=0.06185731664299965
+epoch:8 loss=0.06242090463638306
+epoch:9 loss=0.06153980270028114
+epoch:10 loss=0.05845785513520241
+epoch:11 loss=0.05370307341217995
+epoch:12 loss=0.05589013174176216
+epoch:13 loss=0.057146105915308
+epoch:14 loss=0.05727981775999069
+epoch:15 loss=0.049735818058252335
+epoch:16 loss=0.05287974327802658
+epoch:17 loss=0.048421893268823624
+epoch:18 loss=0.04815121367573738
+epoch:19 loss=0.054195549339056015
+epoch:20 loss=0.04814799875020981
+epoch:21 loss=0.04964037984609604
+epoch:22 loss=0.04809208959341049
+epoch:23 loss=0.04784173145890236
+epoch:24 loss=0.044943470507860184
+epoch:25 loss=0.04372747614979744
+epoch:26 loss=0.047794587910175323
+epoch:27 loss=0.045284222811460495
+epoch:28 loss=0.04716828092932701
+epoch:29 loss=0.04249279201030731
+epoch:30 loss=0.0484805665910244
+epoch:31 loss=0.04517273232340813
+epoch:32 loss=0.04286431148648262
+epoch:33 loss=0.0415775366127491
+epoch:34 loss=0.044873639941215515
+epoch:35 loss=0.04109857603907585
+epoch:36 loss=0.04397314414381981
+epoch:37 loss=0.045988839119672775
+epoch:38 loss=0.0408693365752697
+epoch:39 loss=0.040605854243040085
+epoch:40 loss=0.04187183454632759
+epoch:41 loss=0.04328584298491478
+epoch:42 loss=0.04206966981291771
+epoch:43 loss=0.042049579322338104
+epoch:44 loss=0.03962571918964386
+epoch:45 loss=0.04218325391411781
+epoch:46 loss=0.045767635107040405
+epoch:47 loss=0.04328799620270729
+epoch:48 loss=0.04084816947579384
+epoch:49 loss=0.04190921038389206
+epoch:50 loss=0.042815081775188446
+epoch:51 loss=0.0385555773973465
+epoch:52 loss=0.04020746052265167
+epoch:53 loss=0.03904496133327484
+epoch:54 loss=0.03913990780711174
+epoch:55 loss=0.03791428729891777
+epoch:56 loss=0.038833633065223694
+epoch:57 loss=0.040748436003923416
+epoch:58 loss=0.038732051849365234
+epoch:59 loss=0.04364493116736412
+epoch:60 loss=0.03879616782069206
+epoch:61 loss=0.04526577889919281
+epoch:62 loss=0.037386633455753326
+epoch:63 loss=0.037982646375894547
+epoch:64 loss=0.03987875208258629
+epoch:65 loss=0.038053322583436966
+epoch:66 loss=0.03936714679002762
+epoch:67 loss=0.038860950618982315
+epoch:68 loss=0.03833260387182236
+epoch:69 loss=0.0384192168712616
+epoch:70 loss=0.03793536126613617
+epoch:71 loss=0.038762498646974564
+epoch:72 loss=0.040717121213674545
+epoch:73 loss=0.03743593394756317
+epoch:74 loss=0.03939107432961464
+epoch:75 loss=0.03890211507678032
+epoch:76 loss=0.03750889375805855
+epoch:77 loss=0.038216933608055115
+epoch:78 loss=0.03754616901278496
+epoch:79 loss=0.03699401021003723
+epoch:80 loss=0.03747900202870369
+epoch:81 loss=0.03875776007771492
+epoch:82 loss=0.036669716238975525
+epoch:83 loss=0.0362057164311409
+epoch:84 loss=0.037177979946136475
+epoch:85 loss=0.0386287197470665
+epoch:86 loss=0.03745108097791672
+epoch:87 loss=0.035838667303323746
+epoch:88 loss=0.03874886408448219
+epoch:89 loss=0.03735683485865593
+epoch:90 loss=0.03665066137909889
+epoch:91 loss=0.03970884904265404
+epoch:92 loss=0.037125006318092346
+epoch:93 loss=0.03512052074074745
+epoch:94 loss=0.03688664734363556
+epoch:95 loss=0.03524479269981384
+epoch:96 loss=0.03926272317767143
+epoch:97 loss=0.036016158759593964
+epoch:98 loss=0.03794339671730995
+epoch:99 loss=0.037227191030979156
+epoch:100 loss=0.03673290088772774
+epoch:101 loss=0.03483697026968002
+epoch:102 loss=0.038553137332201004
+epoch:103 loss=0.037967342883348465
+epoch:104 loss=0.033836644142866135
+epoch:105 loss=0.03644904866814613
+epoch:106 loss=0.03620164841413498
+epoch:107 loss=0.037461355328559875
+epoch:108 loss=0.037306372076272964
+epoch:109 loss=0.0383845716714859
+epoch:110 loss=0.03473557159304619
+epoch:111 loss=0.0360155887901783
+epoch:112 loss=0.037127867341041565
+epoch:113 loss=0.03884747251868248
+epoch:114 loss=0.03486650809645653
+epoch:115 loss=0.034917399287223816
+epoch:116 loss=0.03442040830850601
+epoch:117 loss=0.037376388907432556
+epoch:118 loss=0.03415631130337715
+epoch:119 loss=0.034467242658138275
+epoch:120 loss=0.03787562623620033
+epoch:121 loss=0.03824310377240181
+epoch:122 loss=0.03426727280020714
+epoch:123 loss=0.03559662029147148
+epoch:124 loss=0.034789443016052246
+epoch:125 loss=0.036057088524103165
+epoch:126 loss=0.033968184143304825
+epoch:127 loss=0.034011874347925186
+epoch:128 loss=0.03501100093126297
+epoch:129 loss=0.034487612545490265
+epoch:130 loss=0.03293808922171593
+epoch:131 loss=0.03524317964911461
+epoch:132 loss=0.03606180101633072
+epoch:133 loss=0.03861376270651817
+epoch:134 loss=0.03357643261551857
+epoch:135 loss=0.035601940006017685
+epoch:136 loss=0.03599358722567558
+epoch:137 loss=0.03669115528464317
+epoch:138 loss=0.03771153464913368
+epoch:139 loss=0.03636263683438301
+epoch:140 loss=0.037684496492147446
+epoch:141 loss=0.03500496596097946
+epoch:142 loss=0.035157471895217896
+epoch:143 loss=0.03328583762049675
+epoch:144 loss=0.033992815762758255
+epoch:145 loss=0.034569159150123596
+epoch:146 loss=0.032364796847105026
+epoch:147 loss=0.033583834767341614
+epoch:148 loss=0.0324256494641304
+epoch:149 loss=0.03645392879843712
+epoch:150 loss=0.03714875131845474
+epoch:151 loss=0.035763125866651535
+epoch:152 loss=0.034584891051054
+epoch:153 loss=0.03313685953617096
+epoch:154 loss=0.03428370878100395
+epoch:155 loss=0.03502508997917175
+epoch:156 loss=0.03611355274915695
+epoch:157 loss=0.03500016778707504
+epoch:158 loss=0.0349695198237896
+epoch:159 loss=0.03749793395400047
+epoch:160 loss=0.03525347262620926
+epoch:161 loss=0.03660665079951286
+epoch:162 loss=0.03322933241724968
+epoch:163 loss=0.033140361309051514
+epoch:164 loss=0.03343517705798149
+epoch:165 loss=0.03476928547024727
+epoch:166 loss=0.03314348682761192
+epoch:167 loss=0.0333263985812664
+epoch:168 loss=0.03442295640707016
+epoch:169 loss=0.03410166874527931
+epoch:170 loss=0.03464585915207863
+epoch:171 loss=0.03307883068919182
+epoch:172 loss=0.03434787318110466
+epoch:173 loss=0.033934805542230606
+epoch:174 loss=0.03369530290365219
+epoch:175 loss=0.03294972702860832
+epoch:176 loss=0.035837724804878235
+epoch:177 loss=0.035677727311849594
+epoch:178 loss=0.03211052343249321
+epoch:179 loss=0.03525029495358467
+epoch:180 loss=0.033026374876499176
+epoch:181 loss=0.03467435762286186
+epoch:182 loss=0.033783476799726486
+epoch:183 loss=0.03337103873491287
+epoch:184 loss=0.03317328169941902
+epoch:185 loss=0.03287879377603531
+epoch:186 loss=0.03362761437892914
+epoch:187 loss=0.0347357839345932
+epoch:188 loss=0.0342991016805172
+epoch:189 loss=0.03391407057642937
+epoch:190 loss=0.03683505207300186
+epoch:191 loss=0.03795667365193367
+epoch:192 loss=0.03495040908455849
+epoch:193 loss=0.034872621297836304
+epoch:194 loss=0.0343380831182003
+epoch:195 loss=0.033570654690265656
+epoch:196 loss=0.03375042602419853
+epoch:197 loss=0.03282016143202782
+epoch:198 loss=0.03280553221702576
+epoch:199 loss=0.03433068096637726
+```
+
+![模型效果](inference.png)
+
+
+## TODO
+
+引入cross attention，支持用数字输入引导相应图片生成