2023-01-01ãã1å¹´éã®è¨äºä¸è¦§
ç¶ºç¾ æã«æ§ãããããã¼ã¯ããè½ã¡æ°å³ã«ãªã£ã¦æ²è¦³ããã®ã¯ç´ 人ã ç¡äºä¹å馬ã CV/BVã«ãã£ã¨åãåã£ã¦ããã¼ã³ã¬ã¼ãã管çããªãããã£ã¨ç¾å®ãåãéãã¦ãã®ã大人ã
ã¨ã³ã¿ã¡ã¯éç¨ãæ¬è³ªãéã³çµãã£ãææå ã«ããã¢ãã¯ãã¾ãéè¦ã§ãªãã è»äºã¯çµæãæ¬è³ªãè»äºä½æ¦çµäºæã«ãããå½ã®ç¶æ (çµæ«ç¶æ ã»ã¨ã³ãã¹ãã¼ã) ãçæ³ã«è¿ã¥ããæ¦ãã
ããï¼åææ¡ä»¶ï¼å ¬çï¼ ããï¼èªç¶ã«å°ãããæ¦å¿µã¸ã®å½åï¼å®ç¾©ï¼ ãç·å½¢çµåã âãspan(S)ãç·å½¢çµåã§å¼µããã空é âãçæç³»ã空éå´ããè¦ããèªå·±ãå¼µãå ã®éå âãç·å½¢ç¬ç«ãç·å½¢çµåã§è¡¨ããªãå âãåºåºã常ã«å®æ¿åº¦ã§åå¨ãããç·å½¢ç¬ç«ãªçæç³»
disentanglement ã¯æ å ±ç§å¦ã«ãããéè¦æ¦å¿µè¶³ãããã ãããèªã¨ãã¦é·ããé·ãããã è¯ãç¥èª/å¥ç§°ãèãããã ã¢ã¤ã㢠untanglement entanglement ã®å¦å®è¾ã ããé·ããã¨ããããã§å¯¾å¿ãã untanglement ãç´ ç´ã«ä½¿ãã untangled latents ã¨ãç· â¦
ææå°éæ©æ§ã¨ãªã¼ã«ãã¤ã³ãã¯æµå¯¾ãã¦ããã¨æãããã ç®çæè ã触ããªå°å°ããããªææå°éæ©æ§ vs ãã³ã¼ã©ã«ãéç©ãã¦ãªãªã¼ã¹ããããªãªã¼ã«ãã¤ã³ã è¡åæ¹é ãã³ã¼ã©ã«åããªããªææå°éæ©æ§ vs ããã¹ãã¥ã©ã¼ãã©ã³ãè¨å¶ãã¦ãããã¢ã¼ãâ¦
ãªã¼ã«ãã¤ã³ããæ´¾é£ãã¦ããã¹ãã«ã¹æ©ä½ãã´ã¼ã¹ããã ãæå±ä¸ææ©ä½ãã¨ãã¦ä½ä¸ã«å¤æ°åºç¾ããããã·ã§ã³ãã¶ã¤ã¬ã å¶å¾¡ä¿®æ£ãã«ã¦ãªã¼ã«ãã¤ã³ãã®ç®¡çä¸ã¨æããããã åãªãæµã¨ãã¦ã§ã¯ãªããæ å ±ãã°ãæ±ããæå¾ æ®éª¸ãªã©ã®å½¢ã§ãç»å ´ã ãªã¼ã«â¦
èªä¿¡ã¯å ¨ã¦ã«éããã èªä¿¡ã®æºã¨ãªãååã®ã¢ã¤ãã ãæ¬ ããã¦ã¯ãããªãã ä»äººã®è¨è ãè¯å®ãã¦ãããä»äººãã¨ãã®äººã®è¨èã æ ¹æ ãç¡ãã¦ãä»äººãè¨ã£ã¦ãããã°ããæããã æ ¹æ ã®ãªãèªä¿¡ãä¸ãã¦ãããä»äººãèªåã®æ ¹ã£ãã«ããã¨ããã èªåã®å®â¦
NOLA / Nonzero Overlap Add: overlapped window ãå ¨åã§éã¼ã ï¼æ å ±ãè½ã¡ãªãï¼(scipy) COLA / Constant OverLap Add: overlapped window ãä¸å¤®é¨ã§ãã©ãã (scipy)
FP16㸠autocast ãããCUDA Ops: Convç³» Convç³» conv1d conv2d conv3d ConvTç³» conv_transpose1d conv_transpose2d conv_transpose3d RNNç³» RNNCell LSTMCell GRUCell Linearç³» linear matmul __matmul__ chain_matmul baddbmm addbmm addmm addmv addr bmmâ¦
ãã£ã¹ãã¨ãã¼ã¿è»¢éãæ ãé¢æ°ã device ã§è»¢éå ããã¤ã¹ãdtype ã§å¤æåãæå®ãæ¢åTensorã渡ãã¦ãã®deviceã¨dtypeãåç §ããããã¨ãå¯è½ã ãªãã·ã§ã³ã¯3ã¤: non_blocking: éåæã®ãã¼ã¿è»¢é copy: ãã£ã¹ã/転éãå®ç¾æ¸ã¿ãå¦ãã«é¢ãããæ°è¦â¦
PyTorch Lightning: Class device unit internal SimpleProfiler L's action PyTorchProfiler PyTorchâs Autograd Profiler AdvancedProfiler XLAProfiler TPU (PassThroughProfiler) any (default) Just pass-through, do nothing SimpleProfilerã¨DataLoadâ¦
åºå (output) ã§ãªãææ (outcome) ã«æã姿å¢ã ãããåããªããåºåãææã«ç¹ãã£ã¦ãããã¨ãå¿ é æ¡ä»¶ã ããã確ãããã®ããè©ä¾¡ãã®æ®µéã ã ããè©ä¾¡æ¹æ³ã¨ã¿ã¤ãã³ã°ãäºåã«ãã£ããã¨æ§æ³ãã¦ãããã¨ã¯å¿ é ã ããããã¦ããªãã¨ãããã¨ã¯æâ¦
ã½ããã¦ã§ã¢ç£æ¥ã¯ãä¸å ·åæãã§åºè·ãå¾ã«ä¿®æ£ãã許ãããæåãããã ãªãããã許容å¯è½ã«ãªã£ã¦ããã®ãï¼ ãå¾ã«ä¿®æ£ããç¾å®çã«å¯è½ãã©ãããå¿ è¦æ¡ä»¶ã ä¾ãã°è»ã ã£ãããããã¯ãªã³ã¼ã«ã«å¤ããã ä¿®æ£è¨è¨ãå®äºããå¾ãè»ã®ååã¨æ¹ä¿®ã¨åâ¦
Pythonã¯è¨èªä»æ§ã¨ãã¦ããã©ã«ãå¼æ°ãåå©ç¨ããããã㯠mutable ãªãã¸ã§ã¯ãã¨ã®çµã¿åããã§å®¹æã«ãã°ãå¼ãèµ·ããã è¨èªä»æ§ é¢æ°å®ç¾©æã«ããã©ã«ãå¼æ°ãè©ä¾¡ããã¦ç¢ºå®ãï¼=è©ä¾¡ã¯ä¸åº¦ããï¼ããããé¢æ°å¼ã³åºãæã«æ¸¡ãããã The default vaâ¦
OmegaConf ã¯æ§é åã³ã³ãã£ã°ã®å ¥ãåï¼Nesting structured configsï¼ããµãã¼ããã¦ããï¼docs - Nesting structured configsï¼ã ãã®éãdefault_factory ã®å©ç¨ãæ¨å¥¨ããã¦ããã @dataclass class Group: name: str = MISSING admin: User = field(deâ¦
æéãããã«ä½ãåºããç©é â çç£æ§ æéãããã«ç£ãã 価å¤ã®é ï¼ çç£æ§ ã´ããé«éã§ä½ãç¶ãã¦ãçç£æ§ãç¡ãã æ©æã®ç¨¼åçã¯çç£æ§ã¨æ£æ¯ä¾ããªãã
確çã¨å¯åº¦é¢æ° åæã¨ãã¦ãç¹ç¢ºçã¨ç¢ºçå¯åº¦é¢æ°ã¯ä¸è¬ã«ä¸è´ããªãã é£ç¶ç¢ºçå¤æ°ã«ã¤ãã¦ã常㫠probability(x=x_i) == 0 ã ã PDF(x=x_i) ã¯éã¼ããåãããã é¢æ£ç¢ºçå¤æ°ã«ã¤ãã¦ã¯ãprobability(x=x_i) == PMF(x=x_i) ã常ã«æç«ããã ãªã®ã§ç¢ºâ¦
ããã®å£°ãå¦ä½ã«ã¢ãã«åãããããã®ã¢ãã«ãä½ãåç¾åºæ¥ã¦ä½ãåãæ¨ã¦ã¦ããã®ãã ãã®éã®ããªãåºæ¬çãªè¦³ç¹ã¨ãã¦ãããã®å£°ã¯monophonicã¨è¦åãããããããã polyphonicã«ãèããããä¾ ç°ãªãæå¾ã2ã¤èãããä¾ã¨ãã¦ãã®åç»ãæããããâ¦
èªåã®ããã°ã§ã¾ã¨ã¾ã£ããwikipediaã«ç§»ç®¡ãã¦ããã é³é¿ç¹å¾´é: wikipedia/é³å£°åæ#é³é¿ç¹å¾´é
OmegaConfã®ä¸»è¦æ©è½ã«mergeãããã ãã®mergeæ©è½ãListã«é¢ãã¦ãçã¯éã£ã¦ãããçã®ããæ¯ãèãããããã Merge == Reassign OmegaConfã®Listã«å¯¾ããmerge㯠reassign ã¨ãã¦ã®ã¿æ©è½ããã from omegaconf import OmegaConf conf1 = OmegaConf.creâ¦
AWS GPU EC2 Fargate ECS1 Batch2 SageMaker NVIDIA A100 p4d/p4de 3 - p4d p4d p4d/p4de A10G g5 4 - g5 g5 g5 V100 p3/p3dn 5 - p3/p3dn p3/p3dn p3/p3dn T4G g5g 6 - - - - T4 g4dn 7 - g4dn g4dn g4dn M60 g3s/g3 8 - g3s/g3 g3s/g3 - K80 p2 9 - p2 p2 â¦
å¤æ°ã®è¨å®ãå¯è½ãã¤å¤ã¦ã¼ã¶ã¼ãªã·ã¹ãã ã§ãããã©ã«ãè¨å®ã»ãããããã±ã¼ã¸ç®¡ç+baseè¨å®ã§ç®¡çããã ç¶æ³ å¤æ°ã®è¨å®é ç®ãåå¨ ãã¹ããã©ã¯ãã£ã¹ã®ææ¡èªä½ãé«ã³ã¹ã 大ã¾ãã«ã¯ã¿ããªå ±éããããç°å¢ãããã¸ã§ã¯ãäºæ ã§å¾®å¦ã«ç°ãªã å¤ã¦ã¼â¦
両æ¹ã¨ãã¦ã¼ã¶ã¼å©ç¨ãè¶ éè¦ãã¦ããMVPã¯æ´ã«ãå¸å ´ã®æèããããã MVPã¯ã客ã«è²·ãããããviableã®ææ¨ã¨ãããã¢ã¸ã£ã¤ã«å®£è¨ã¯ãã¦ã¼ã¶ã¼ã®ä½¿ç¨æããé©å¿ã®ææ¨ã«ãã¦ãã MVPã¯ä¾¡å¤ææ¡ (ããªã¥ã¼ãããã¸ã·ã§ã³) ã®ã¿ã®æ¤è¨¼ãé©ç¨ç¯å²ãªã®ã§ããâ¦
ãµã¤ãå¿å vs ã³ã³ãã³ãå¿å ã¨ããæè¦ãããã ãµã¤ãå¿å ããã®å ´æã¯é¢ç½ããã®ããããããããã¨ããã¤ã³ã¿ã¼ãããä¸ã®å ´æ=ãµã¤ãã 1ã¤1ã¤ã®ã³ã³ãã³ãã«ä¾¡å¤ãããã®ã¯åæã¨ãã¦ããããªãä»å¾ãè¯ããã®ãæã«å ¥ããããã¨ããä¿¡é ¼ã»ãã©ã³ãâ¦
è³é調é èè³ã«å¼·ãã æè³ã¨éã£ã¦èè³ã¯ä½åãæ±ããããã¡ã äºæ¥éã·ãã¸ã¼ãç¡ããã¨ã¯ããããä¸ã¤ã®ã«ã´ã«å ¨ã¦ã®åµãå ¥ãã¦ããªããã¨ããã¡ãªããã«ãããªãã è³æ¬å¸å ´ãããçºéãã¦ããã¨æè³ã«ããè³é調éãå½ããåã«ãªãããã®ã¡ãªããã¯ç¸â¦
ç®æ¨ã¨è¨ç»ã¯å¾å±é¢ä¿ã«ããã®ãããããã¯ç¸äºä½ç¨ãããã®ãã æéã®è½å ä¸äººã®ããã«ã¯éçãããã è¿ãæéã¹ã±ã¼ã«ã§ã¯ããããéæä¸å¯è½ãªå ´åãããããé ã足ããå®ç§ãªè¨ç»ãç«æ¡ã§ããªãå ´åãããã ä¸å®å ¨ãªè¨ç»ã¨ç¾å®çãªè¨ç» ç®æ¨ã確å®ãâ¦
ãããã¸ã¡ã³ãã¯ã対象ã®å·è¡ãå¶ç¶ããä¸æãå®è¡ãããã¨ããåªåã§ããã æãåããã°çç£æ´»åã¯ã§ãããæ°´æµã«æè¨ã®ãã¼ããæãè¾¼ãã°ãã¾ã«æè¨ãçµã¿ä¸ãããã§ãã©ã³ãã ã«æãåããã¦ãéããå¶ç¶ã§ãã製åã¯å®æããªãã ã ããåªããçç£è â¦
éä¸æ¦ç¥ãåãå ´åã1çªã·ã³ãã«ãªæ¦è¡ã¯1ãµãã»ã°ã¡ã³ãæ¦è¡ã ä¸è¯æç â 麻å©è±è ãã¿ãããªé¸ã³æ¹ã å¥ã®æ¦è¡ã¨ãã¦ããã«ããµãã»ã°ã¡ã³ãæ¦è¡ãããã ç°ãªãè¦æ±ãæã¤è¤æ°ãµãã»ã°ã¡ã³ãã«å¯¾ãã¦è¤æ°ã®å´é¢ããã£ã製åãæä¾ãããã¨ã§ãåä¸è£½åâ¦
åæ£/ãããã¯ã¼ã¯ã³ã³ãã¥ã¼ãã£ã³ã°ã®ææ³ã¯ç¾ããã åå°éä¼æ¥ã«ãã£ã¦æä¾ãããåªããæ©è½ã»ãµã¼ãã¹ããããã¯ã¼ã¯çµç±ã§ç¹ãåããã1ã¤ã®ã·ã¹ãã ã¨ãã¦æä¾ããã ã§ã¯ãç¾å®ã®åé¡ã解ãéã«ãå¤é¨ã¸ä»»ããããä»»ããã«å¤ãããé¨åã¯ã©ããªã®ãâ¦
HTMLã«é¢ããå¦æ³ãé£ããã çæã¨é ä¿¡ã»è¡¨ç¤ºã®åé¢ HTMLã¯ãã¼ã¯ã¢ããè¨èªã¨ãã¦çºæããããããããã¹ãå½¢å¼ã§ããã ã¾ãHTMLã«ã¯ã¿ã°åã®å¶ç´ããããæ¨æºã¿ã°ã«ã¯idãæ¯ããã¨æãã°æ¯ããã ããã¦DOMãããããããã«ãHTMLã«ã¯æ¨æ§é å¶ç´ããããâ¦